From 440e28b08fa0f503c229f5787be4f775ad20484c Mon Sep 17 00:00:00 2001 From: carlushuang Date: Tue, 26 Nov 2024 11:14:56 +0800 Subject: [PATCH] [CK_TILE] fused-moe first version (#1634) * moe pipeline * update code * compile OK * update * update cpu reference * update pipeline_gemm0 * compiler ok * update pipeline * rename to ex pipeline * block-asm * update * update * update first gemm ok * compute correct * update file structure * update README * update * update * update code * update API * return unsupport case * add comment * update readme * update * uncomment * update * fix build err --------- Co-authored-by: valarLip <340077269@qq.com> --- .../alternative_impl/matrix_core_swizzle.cpp | 4 +- .../matrix_core_swizzle_kernel.hpp | 12 +- example/ck_tile/06_permute/permute.cpp | 2 +- .../13_moe_sorting/moe_sorting_api.hpp | 2 +- example/ck_tile/15_fused_moe/CMakeLists.txt | 19 + example/ck_tile/15_fused_moe/README.md | 69 ++ example/ck_tile/15_fused_moe/fused_moe.hpp | 52 ++ .../ck_tile/15_fused_moe/fused_moegemm.hpp | 84 ++ .../ck_tile/15_fused_moe/fused_moesorting.hpp | 20 + .../15_fused_moe/instances/fused_moe_api.cpp | 80 ++ .../instances/fused_moegemm_api.cpp | 33 + .../instances/fused_moegemm_api_internal.hpp | 60 ++ .../instances/fused_moegemm_api_traits.hpp | 53 ++ .../instances/fused_moegemm_bf16_m32.cpp | 14 + .../instances/fused_moegemm_fp16_m32.cpp | 14 + .../instances/fused_moesorting_api.cpp | 73 ++ example/ck_tile/15_fused_moe/main.cpp | 603 +++++++++++++ example/ck_tile/15_fused_moe/misc/moe-0.png | Bin 0 -> 76830 bytes example/ck_tile/15_fused_moe/misc/moe-1.png | Bin 0 -> 92535 bytes example/ck_tile/15_fused_moe/misc/moe-2.png | Bin 0 -> 126766 bytes example/ck_tile/15_fused_moe/misc/moe-3.png | Bin 0 -> 18655 bytes example/ck_tile/CMakeLists.txt | 2 + include/ck_tile/core.hpp | 2 + .../core/arch/amd_buffer_addressing.hpp | 103 +++ include/ck_tile/core/arch/arch.hpp | 18 + include/ck_tile/core/arch/utility.hpp | 24 + include/ck_tile/core/tensor/buffer_view.hpp | 86 +- include/ck_tile/core/tensor/load_tile.hpp | 54 +- .../core/tensor/static_distributed_tensor.hpp | 26 + include/ck_tile/core/tensor/tensor_view.hpp | 42 + include/ck_tile/core/tensor/tile_window.hpp | 74 +- .../core/tensor/tile_window_linear.hpp | 159 +++- .../ck_tile/core/tensor/tile_window_utils.hpp | 54 ++ include/ck_tile/core/tensor/update_tile.hpp | 56 +- .../ck_tile/core/utility/static_counter.hpp | 116 +++ include/ck_tile/host.hpp | 2 + include/ck_tile/host/device_memory.hpp | 35 + include/ck_tile/host/fill.hpp | 113 ++- include/ck_tile/host/host_tensor.hpp | 121 ++- include/ck_tile/host/joinable_thread.hpp | 27 + .../host/reference/reference_fused_moe.hpp | 196 +++++ .../host/reference/reference_permute.hpp | 23 +- .../unary_element_wise_operation.hpp | 99 +++ include/ck_tile/ops/flatmm.hpp | 10 + .../flatmm_32x512x128_1x4x1_16x16x32.hpp | 615 +++++++++++++ .../flatmm_sn_32x128x512_1x4x1_16x16x32.hpp | 562 ++++++++++++ .../ops/flatmm/block/flatmm_uk_config.hpp | 10 + include/ck_tile/ops/flatmm/block/uk/README.md | 1 + ...m_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc | 613 +++++++++++++ ...atmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc | 516 +++++++++++ .../block_fmha_pipeline_qr_ks_vs_async.hpp | 19 +- include/ck_tile/ops/fused_moe.hpp | 8 + .../fused_moe/kernel/fused_moegemm_kernel.hpp | 421 +++++++++ .../fused_moe/kernel/fused_moegemm_shape.hpp | 125 +++ .../kernel/fused_moegemm_tile_partitioner.hpp | 33 + .../fused_moegemm_pipeline_flatmm_ex.hpp | 651 ++++++++++++++ .../fused_moegemm_pipeline_flatmm_policy.hpp | 831 ++++++++++++++++++ .../fused_moegemm_pipeline_flatmm_uk.hpp | 354 ++++++++ .../fused_moegemm_pipeline_problem.hpp | 46 + .../pipeline/fused_moegemm_traits.hpp | 48 + include/ck_tile/ops/gemm/warp/warp_gemm.hpp | 130 +-- .../gemm/warp/warp_gemm_attribute_mfma.hpp | 170 +++- .../warp/warp_gemm_attribute_mfma_impl.hpp | 457 +++++++--- .../ops/gemm/warp/warp_gemm_dispatcher.hpp | 58 +- .../ck_tile/ops/gemm/warp/warp_gemm_impl.hpp | 61 +- include/ck_tile/ops/moe_sorting.hpp | 11 - 66 files changed, 8067 insertions(+), 309 deletions(-) create mode 100644 example/ck_tile/15_fused_moe/CMakeLists.txt create mode 100644 example/ck_tile/15_fused_moe/README.md create mode 100644 example/ck_tile/15_fused_moe/fused_moe.hpp create mode 100644 example/ck_tile/15_fused_moe/fused_moegemm.hpp create mode 100644 example/ck_tile/15_fused_moe/fused_moesorting.hpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moegemm_api.cpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moegemm_api_traits.hpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moegemm_bf16_m32.cpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moegemm_fp16_m32.cpp create mode 100644 example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp create mode 100644 example/ck_tile/15_fused_moe/main.cpp create mode 100644 example/ck_tile/15_fused_moe/misc/moe-0.png create mode 100644 example/ck_tile/15_fused_moe/misc/moe-1.png create mode 100644 example/ck_tile/15_fused_moe/misc/moe-2.png create mode 100644 example/ck_tile/15_fused_moe/misc/moe-3.png create mode 100644 include/ck_tile/core/tensor/tile_window_utils.hpp create mode 100644 include/ck_tile/core/utility/static_counter.hpp create mode 100644 include/ck_tile/host/joinable_thread.hpp create mode 100644 include/ck_tile/host/reference/reference_fused_moe.hpp create mode 100644 include/ck_tile/ops/flatmm.hpp create mode 100644 include/ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp create mode 100644 include/ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp create mode 100644 include/ck_tile/ops/flatmm/block/flatmm_uk_config.hpp create mode 100644 include/ck_tile/ops/flatmm/block/uk/README.md create mode 100644 include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc create mode 100644 include/ck_tile/ops/flatmm/block/uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc create mode 100644 include/ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp create mode 100644 include/ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp create mode 100644 include/ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp create mode 100644 include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp create mode 100644 include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp create mode 100644 include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp create mode 100644 include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp create mode 100644 include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp delete mode 100644 include/ck_tile/ops/moe_sorting.hpp diff --git a/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp b/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp index 93c662a288..e5ded0ef3b 100644 --- a/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp +++ b/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle.cpp @@ -40,7 +40,7 @@ float matrix_core_swizzle(matrix_core_swizzle_traits t, else if(t.permute.compare("0,1,3,4,2,5") == 0) { constexpr matrix_core_permute_style pstyle = - matrix_core_permute_style::permute_b_nr_kr_kw_nw_kv; + matrix_core_permute_style::b_nr_kr_kw_nw_kv; using Kernel = matrix_core_swizzle_kernel; @@ -83,7 +83,7 @@ float matrix_core_swizzle(matrix_core_swizzle_traits t, else if(t.permute.compare("0,1,3,4,2,5") == 0) { constexpr matrix_core_permute_style pstyle = - matrix_core_permute_style::permute_b_nr_kr_kw_nw_kv; + matrix_core_permute_style::b_nr_kr_kw_nw_kv; using Kernel = matrix_core_swizzle_kernel; diff --git a/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp b/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp index 60ac103ec3..28f4c452bc 100644 --- a/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp +++ b/example/ck_tile/06_permute/alternative_impl/matrix_core_swizzle_kernel.hpp @@ -42,8 +42,8 @@ enum class matrix_core_permute_style { permute_b_n0_k0_n1_k1_n2_k2 = 0, // 0,1,4,2,5,3,6 permute_b_n0_n1_k0_k1_n2_k2 = 1, // 0,1,2,4,5,3,6 - permute_b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5 - permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv, + b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5 + b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv, }; // assume this is B matrix, originally we have batch*n*k @@ -203,7 +203,7 @@ struct matrix_core_swizzle_kernel else { // clang-format off - // permute_b_nr_kr_kw_nw_kv or permute_b_nr_kr_waveflatten + // b_nr_kr_kw_nw_kv or b_nr_kr_waveflatten constexpr index_t Kv = Alignment; constexpr index_t Nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane; constexpr index_t Kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane; @@ -332,7 +332,7 @@ struct matrix_core_swizzle_kernel make_tuple(sequence<0>{}, sequence<1>{})); return tmp_1; #else - // permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv, + // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv, constexpr index_t kv = Alignment; constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane; constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane; @@ -376,13 +376,13 @@ struct matrix_core_swizzle_kernel else { #if MERGE_2D_013425 - // permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv + // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv return make_tile_window(dst_view, make_tuple(number{}, number{}), {i_n * NPerBlock, i_k * KPerBlock}, get_dst_dist()); #else - // permute_b_nr_kr_waveflatten = permute_b_nr_kr_kw_nw_kv + // b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv constexpr index_t kv = Alignment; constexpr index_t nw = WarpGemm::WarpGemmAttribute::Impl::kAMLane; constexpr index_t kw = WarpGemm::WarpGemmAttribute::Impl::kABKLane; diff --git a/example/ck_tile/06_permute/permute.cpp b/example/ck_tile/06_permute/permute.cpp index af95b64e69..477ae370b9 100644 --- a/example/ck_tile/06_permute/permute.cpp +++ b/example/ck_tile/06_permute/permute.cpp @@ -264,7 +264,7 @@ bool run(const ck_tile::ArgParser& arg_parser) { if(arg_parser.get_str("perm") == std::string("0,1,3,4,2,5")) { - // permute_b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5 + // b_nr_kr_kw_nw_kv = 2, // 0,1,3,4,2,5 matrix_core_swizzle_traits t; t.data_type = data_type; t.permute = arg_parser.get_str("perm"); diff --git a/example/ck_tile/13_moe_sorting/moe_sorting_api.hpp b/example/ck_tile/13_moe_sorting/moe_sorting_api.hpp index 91b54932ce..0cb393f7de 100644 --- a/example/ck_tile/13_moe_sorting/moe_sorting_api.hpp +++ b/example/ck_tile/13_moe_sorting/moe_sorting_api.hpp @@ -5,7 +5,7 @@ #include #include "ck_tile/core.hpp" #include "ck_tile/host.hpp" -#include "ck_tile/ops/moe_sorting.hpp" +#include "ck_tile/ops/fused_moe.hpp" struct moe_sorting_trait { diff --git a/example/ck_tile/15_fused_moe/CMakeLists.txt b/example/ck_tile/15_fused_moe/CMakeLists.txt new file mode 100644 index 0000000000..a716eef19e --- /dev/null +++ b/example/ck_tile/15_fused_moe/CMakeLists.txt @@ -0,0 +1,19 @@ +set(TILE_EXAPMLE_FUSED_MOE "tile_example_fused_moe") +# not using add_example_executable() to add this target, since we don't want this to have +# to be included in "make all/install/check" +message("adding ${TILE_EXAPMLE_FUSED_MOE}") +file(GLOB INSTANCE_SRCS instances/*.cpp) +add_executable(${TILE_EXAPMLE_FUSED_MOE} EXCLUDE_FROM_ALL main.cpp) +target_include_directories(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${CMAKE_CURRENT_LIST_DIR}) +target_sources(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${INSTANCE_SRCS}) + +set(TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS) + +# NOTE: we turn off undefined-func-template to let source compile without explicit declare function specializations +list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -Wno-undefined-func-template -Wno-float-equal) +list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_BUFFER_LOAD_AGPR=1) # TODO: enable load to a +list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -DCK_TILE_FLOAT_TO_BFLOAT16_DEFAULT=4) # rta +# list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -mllvm -greedy-reverse-local-assignment=1) +# list(APPEND TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS -v --save-temps -Wno-gnu-line-marker) + +target_compile_options(${TILE_EXAPMLE_FUSED_MOE} PRIVATE ${TILE_EXAPMLE_FUSED_MOE_COMPILE_OPTIONS}) diff --git a/example/ck_tile/15_fused_moe/README.md b/example/ck_tile/15_fused_moe/README.md new file mode 100644 index 0000000000..dd566c1667 --- /dev/null +++ b/example/ck_tile/15_fused_moe/README.md @@ -0,0 +1,69 @@ +# fused-moe +Implementing the fused-moe block operator using ck-tile. This is a scatter/gather-group-gemm based solution, similiar to that of [vllm moe](https://github.com/vllm-project/vllm/blob/main/benchmarks/kernels/benchmark_moe.py), but we introduce more kernel fusion to boost performance +![](misc/moe-0.png) + +The benifit of this fused-moe: +* 1.5~2x perf boost compared with current vllm solution +* zero workspace to reduce memory footprint +* much less kernel instance, easy to maintain + +# Implementation and feature support +## moe-sorting +this is a common pre-process step before the actual moe-gemm. The purpose is to transform the moe loop over from token-by-token to expert-by-expert, make sure very workgroup is working for a single expert (B matrix). Besides, we extend this op to do the zeroing of the output buffer(to be used for reduce buffer with atomic) + +## moe-gemm +`moe-gemm` is a group-gemm based back-to-back gemm, where the row-id of input token comes from another buffer. Naive understanding of fused-moe is from token-by-token view as below picture: +![](misc/moe-1.png) +After `moe-sorting`, we can view this algorithm as expert-by-expert, as below: +![](misc/moe-2.png) + +## optimization +summary of the key design of this fused-moe operator: +* fuse 2 group-gemm + activation + `topk-weight` multiply into single kernel, using atomic for 2nd gemm accumualation +* fuse buffer-zeroing in `moe-sorgin`, user no longer need call extra torch.zero() for the out buffer +* fused scatter-gather for row index(same as vllm) +* pre-shuffle B matric(weight) to maximize memory throughput. input(activation) keep original layout `[batch, hidden]`. +* extrem optimized pipeline using block-inline-asm(we call it `micro-kernel` or `uk`), while not breaking the *composable* design of ck + +## +``` +// [indexing implementation-1] +// using M_a as constexpr block_size to partition all tokens into different slices +// each slice map to one expert, and one expert can have multiple slices +// e.g. num_experts = 6, topk=3, M_a = 4, input_tokens = 5 +// before sort, topk_ids is : [[0, 3, 5], [2, 3, 5], [1, 3, 5], [1, 2, 3], [1, 3, 5]] +// tok-0 tok-1 tok-2 tok-3 tok-4 +// topk_weight is : [[a, b, c], [d, e, f], [g, h, i], [j, k, l], [m, n, o]] (some float number) +// +// token_id_per_expert is : [[0], [2, 3, 4], [1, 3], [0, 1, 2, 3, 4], [], [0, 1, 2, 5]] +// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5 +// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]] +// +// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1) +// * this could be larger than actual, since actual tokens are on GPU +// +// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5] +// |- exp-0 -|- exp-1 -|- exp-2 -|- exp-3 -|- exp-4 -|- exp-5 -| +// sorted_weight_ptr : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *, c, f, i, o] +// +// * length is max_num_tokens_padded, actual size is num_tokens_post_padded_ptr +// +// sorted_expert_ids_ptr : [0, 1, 2, 3, 3, 4, 5] +// * length is (max_num_tokens_padded + block_size - 1) / block_size +// +// num_tokens_post_padded_ptr : [28] +// num_sorted_tiles_ptr : [7] +// +// * different from vLLM +// 1) token_id stored in sorted_token_ids_ptr is actual token_id, not token_id*top_K expanded id +// 2）need sorted_weight_ptr +// 3) use num_sorted_tiles_ptr, already divided by M_a +// +// * below used for indexing +// 1) sorted_token_ids_ptr [max_num_tokens_padded] +// 2) sorted_weight_ptr +// 3) sorted_expert_ids_ptr +// 4）num_tokens_post_padded_ptr/num_sorted_tiles_ptr (select one) +// +// max_num_tokens_padded: opk_ids.numel() + num_experts * (block_size - 1) +``` \ No newline at end of file diff --git a/example/ck_tile/15_fused_moe/fused_moe.hpp b/example/ck_tile/15_fused_moe/fused_moe.hpp new file mode 100644 index 0000000000..6bd7688d8a --- /dev/null +++ b/example/ck_tile/15_fused_moe/fused_moe.hpp @@ -0,0 +1,52 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "fused_moesorting.hpp" +#include "fused_moegemm.hpp" + +struct fused_moe_args +{ + const void* a_ptr; // [m, k], input token + const void* a_scale_ptr; // [m, 1], token scale + const void* g_ptr; // [e, n, k]/[e, 2*n, k], pre-shuffle([e, nr, kr, w]) + const void* d_ptr; // [e, n, k], pre-shuffle([e, nr, kr, w]) + const void* g_scale_ptr; // [e, 1, n], gate(up) scale + const void* d_scale_ptr; // [e, 1, k], down scale + const void* y_smooth_scale_ptr; // [e, 1, n], smooth-quant-scale for 2nd gemm input + void* o_ptr; // [m, k], output token (no need to do zeroing) + + const void* topk_ids_ptr; // [tokens, topk] + const void* topk_weight_ptr; // [tokens, topk] + void* sorted_token_ids_ptr; // [max_num_tokens_padded] + void* sorted_weight_ptr; // [max_num_tokens_padded] + void* sorted_expert_ids_ptr; // [(max_num_tokens_padded + block_size - 1) / block_size] + void* num_sorted_tiles_ptr; // [1] + + ck_tile::index_t block_m; // block_m, used to devide the input + ck_tile::index_t hidden_size; // k + ck_tile::index_t intermediate_size; // n / TP, for Gate. if Gate+Up, Down need divide by 2 + ck_tile::index_t num_tokens; // input number of tokens for current iteration + ck_tile::index_t num_experts; // number of groups + ck_tile::index_t topk; // need this? + + ck_tile::index_t stride_token; // for input/output, stride for each row, should >= hidden_size +}; + +// This is the public API, will be generated by script +struct fused_moe_traits +{ + std::string prec_i; // input precision + std::string prec_w; // weight precision + std::string prec_o; // output precision + std::string prec_st; // token scale data type + std::string prec_sw; // weight scale data type + std::string prec_sq; // smooth quant scale + std::string prec_kw; // topk-weight data type + int block_m; + int gate_only; + int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant +}; + +float fused_moe(fused_moe_traits, fused_moe_args, const ck_tile::stream_config&); diff --git a/example/ck_tile/15_fused_moe/fused_moegemm.hpp b/example/ck_tile/15_fused_moe/fused_moegemm.hpp new file mode 100644 index 0000000000..b8e51475ad --- /dev/null +++ b/example/ck_tile/15_fused_moe/fused_moegemm.hpp @@ -0,0 +1,84 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/host/kernel_launch.hpp" +#include "ck_tile/ops/fused_moe.hpp" +#include + +// this is only a convenient structure for creating an example +// this is not part of the host API +template +struct FusedMoeGemmTypeConfig; + +template +struct FusedMoeGemmTypeConfig +{ + using ADataType = ck_tile::bf16_t; + using GDataType = ck_tile::bf16_t; + using DDataType = ck_tile::bf16_t; + using AccDataType = float; + using ODataType = ck_tile::bf16_t; + using AScaleDataType = ck_tile::remove_cvref_t; + using GScaleDataType = ck_tile::remove_cvref_t; + using DScaleDataType = ck_tile::remove_cvref_t; + using YSmoothScaleDataType = ck_tile::remove_cvref_t; + using TopkWeightDataType = ck_tile::remove_cvref_t; + using IndexDataType = ck_tile::index_t; +}; + +template +struct FusedMoeGemmTypeConfig +{ + using ADataType = ck_tile::fp16_t; + using GDataType = ck_tile::fp16_t; + using DDataType = ck_tile::fp16_t; + using AccDataType = float; + using ODataType = ck_tile::fp16_t; + using AScaleDataType = ck_tile::remove_cvref_t; + using GScaleDataType = ck_tile::remove_cvref_t; + using DScaleDataType = ck_tile::remove_cvref_t; + using YSmoothScaleDataType = ck_tile::remove_cvref_t; + using TopkWeightDataType = ck_tile::remove_cvref_t; + using IndexDataType = ck_tile::index_t; +}; + +template +struct FusedMoeGemmTypeConfig +{ + using ADataType = ck_tile::int8_t; + using GDataType = ck_tile::int8_t; + using DDataType = ck_tile::int8_t; + using AccDataType = int32_t; + using ODataType = ck_tile::bf16_t; + using AScaleDataType = ck_tile::remove_cvref_t; + using GScaleDataType = ck_tile::remove_cvref_t; + using DScaleDataType = ck_tile::remove_cvref_t; + using YSmoothScaleDataType = ck_tile::remove_cvref_t; + using TopkWeightDataType = ck_tile::remove_cvref_t; + using IndexDataType = ck_tile::index_t; +}; + +// runtime args +struct fused_moegemm_args : public ck_tile::FusedMoeGemmHostArgs +{ +}; + +// This is the public API, will be generated by script +struct fused_moegemm_traits +{ + std::string prec_i; // input precision + std::string prec_w; // weight precision + std::string prec_o; // output precision + std::string prec_st; // token scale data type + std::string prec_sw; // weight scale data type + std::string prec_sq; // smooth quant scale + std::string prec_kw; // topk-weight data type + int block_m; + int gate_only; + int fused_quant; // 0:no-sweep, 1:smooth-dynamic-quant, 2:dynamic-quant +}; + +float fused_moegemm(fused_moegemm_traits, fused_moegemm_args, const ck_tile::stream_config&); diff --git a/example/ck_tile/15_fused_moe/fused_moesorting.hpp b/example/ck_tile/15_fused_moe/fused_moesorting.hpp new file mode 100644 index 0000000000..57dace9b41 --- /dev/null +++ b/example/ck_tile/15_fused_moe/fused_moesorting.hpp @@ -0,0 +1,20 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once +#include +#include "ck_tile/core.hpp" +#include "ck_tile/host.hpp" +#include "ck_tile/ops/fused_moe.hpp" + +struct fused_moesorting_trait +{ + std::string index_type; + std::string weight_type; // currently always float +}; + +struct fused_moesorting_args : public ck_tile::MoeSortingHostArgs +{ +}; + +float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s); diff --git a/example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp b/example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp new file mode 100644 index 0000000000..bfc0ce4096 --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moe_api.cpp @@ -0,0 +1,80 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "fused_moe.hpp" + +float fused_moe(fused_moe_traits t, fused_moe_args a, const ck_tile::stream_config& s) +{ + auto s_sub = ck_tile::stream_config{s.stream_id_, false, s.log_level_, 0, 1}; + + auto o_data_bytes = [&]() { + if(t.prec_o == "fp32") + return 4; + else if(t.prec_o == "fp16" || t.prec_o == "bf16") + return 2; + else if(t.prec_o == "int8" || t.prec_o == "fp8") + return 1; + return 1; + }(); + + auto t0 = fused_moesorting_trait{"int32", "fp32"}; + auto a0 = fused_moesorting_args{ + a.topk_ids_ptr, // const void* p_topk_ids; + a.topk_weight_ptr, // const void* p_weights; + a.sorted_token_ids_ptr, // void* p_sorted_token_ids; + a.sorted_weight_ptr, // void* p_sorted_weights; + a.sorted_expert_ids_ptr, // void* p_sorted_expert_ids; + a.num_sorted_tiles_ptr, // void* p_total_tokens_post_pad; + a.o_ptr, // void* p_moe_buf; + a.num_tokens, // index_t tokens; + a.block_m, // index_t unit_size; + a.num_experts, // index_t num_experts; + a.topk, // index_t topk; + a.num_tokens * a.stride_token * o_data_bytes // index_t moe_buf_bytes; + }; + + auto t1 = fused_moegemm_traits{t.prec_i, + t.prec_w, + t.prec_o, + t.prec_st, + t.prec_sw, + t.prec_sq, + t.prec_kw, + t.block_m, + t.gate_only, + t.fused_quant}; + auto a1 = fused_moegemm_args{ + a.a_ptr, // const void* a_ptr; + a.a_scale_ptr, // const void* a_scale_ptr; + a.g_ptr, // const void* g_ptr; + a.d_ptr, // const void* d_ptr; + a.g_scale_ptr, // const void* g_scale_ptr; + a.d_scale_ptr, // const void* d_scale_ptr; + a.y_smooth_scale_ptr, // const void* y_smooth_scale_ptr; + a.o_ptr, // void* o_ptr; + a.sorted_token_ids_ptr, // const void* sorted_token_ids_ptr; + a.sorted_weight_ptr, // const void* sorted_weight_ptr; + a.sorted_expert_ids_ptr, // const void* sorted_expert_ids_ptr; + a.num_sorted_tiles_ptr, // const void* num_sorted_tiles_ptr; + a.hidden_size, // index_t hidden_size; + a.intermediate_size, // index_t intermediate_size; + a.num_tokens, // index_t num_tokens; + a.num_experts, // index_t num_experts; + a.topk, // index_t topk; + a.stride_token // index_t stride_token; + }; + + float r0 = -1; + float r1 = -1; + + float r = ck_tile::launch_kernel( + s, + [=, &r0](const ck_tile::stream_config&) { r0 = fused_moesorting(t0, a0, s_sub); }, + [=, &r1](const ck_tile::stream_config&) { r1 = fused_moegemm(t1, a1, s_sub); }); + + // keep unsupported case return negative + if(r0 < 0 || r1 < 0) + return -1; + + return r; +} diff --git a/example/ck_tile/15_fused_moe/instances/fused_moegemm_api.cpp b/example/ck_tile/15_fused_moe/instances/fused_moegemm_api.cpp new file mode 100644 index 0000000000..c1a4c495c3 --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moegemm_api.cpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include "fused_moegemm.hpp" +#include "fused_moegemm_api_traits.hpp" + +// Note: this internal API only declare, not define here, otherwise will block `make -j` +template +float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a); + +template +using S = ck_tile::sequence; + +float fused_moegemm(fused_moegemm_traits t, fused_moegemm_args a, const ck_tile::stream_config& s) +{ + // clang-format off + float r = -1; + if(t.prec_i == "bf16" && t.prec_w == "bf16" && t.prec_o == "bf16" && t.prec_st == "fp32" && + t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1) + { + using t_ = fmoe_, S<1, 4, 1>, S<16, 16, 32>, 1, 0>; + r = fused_moegemm_(s, a); + } + else if(t.prec_i == "fp16" && t.prec_w == "fp16" && t.prec_o == "fp16" && t.prec_st == "fp32" && + t.prec_sw == "fp32" && t.prec_sq == "fp32" && t.prec_kw == "fp32" && t.block_m == 32 && t.gate_only == 1) + { + using t_ = fmoe_, S<1, 4, 1>, S<16, 16, 32>, 1, 0>; + r = fused_moegemm_(s, a); + } + // clang-format on + return r; +} diff --git a/example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp b/example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp new file mode 100644 index 0000000000..5872179ef7 --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moegemm_api_internal.hpp @@ -0,0 +1,60 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "fused_moegemm_api_traits.hpp" +#include "ck_tile/ops/fused_moe.hpp" +#include + +template +using S = ck_tile::sequence; + +// do not the define of this tepmlate function inside the _api.cpp, otherwise will block make -j +template +float fused_moegemm_(const ck_tile::stream_config& s, fused_moegemm_args a) +{ + using f_traits = ck_tile::FusedMoeGemmTraits; + using f_shape = ck_tile::FusedMoeGemmShape; + using f_problem = + ck_tile::FusedMoeGemmPipelineProblem; + + // using f_pipeline = ck_tile::FusedMoeGemmPipeline_FlatmmEx; + using f_pipeline = ck_tile::FusedMoeGemmPipeline_FlatmmUk; + using f_partitioner = ck_tile::FusedMoeGemmTilePartitioner_Linear; + using f_kernel = ck_tile::FusedMoeGemmKernel; + + const dim3 grids = f_kernel::GridSize(a); + constexpr dim3 blocks = f_kernel::BlockSize(); + constexpr ck_tile::index_t kBlockPerCu = 1; + + static int printed = 0; + + auto kargs = f_kernel::MakeKargs(a); + if(s.log_level_ > 0 && printed == 0) + { + std::cout << ", " << f_kernel::GetName() << std::flush; + printed = 1; + } + + return ck_tile::launch_kernel( + s, ck_tile::make_kernel(f_kernel{}, grids, blocks, 0, kargs)); +} diff --git a/example/ck_tile/15_fused_moe/instances/fused_moegemm_api_traits.hpp b/example/ck_tile/15_fused_moe/instances/fused_moegemm_api_traits.hpp new file mode 100644 index 0000000000..cc476685de --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moegemm_api_traits.hpp @@ -0,0 +1,53 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include + +// this is used to pattern-match internl kernel implementation, not to instantiate kernel +template + typename WarpPerBlock_, + typename WarpTile_, // seq<*,*,*>, used to select mfma + ck_tile::index_t GateOnly_ = 0, + ck_tile::index_t FusedQuant_ = 0> +struct fmoe_ // traits, ugly name, only used for internal +{ + using TypeConfig = FusedMoeGemmTypeConfig; + + using ADataType = ck_tile::remove_cvref_t; + using GDataType = ck_tile::remove_cvref_t; + using DDataType = ck_tile::remove_cvref_t; + using AccDataType = ck_tile::remove_cvref_t; + using ODataType = ck_tile::remove_cvref_t; + using AScaleDataType = ck_tile::remove_cvref_t; + using GScaleDataType = ck_tile::remove_cvref_t; + using DScaleDataType = ck_tile::remove_cvref_t; + using YSmoothScaleDataType = ck_tile::remove_cvref_t; + using TopkWeightDataType = ck_tile::remove_cvref_t; + using IndexDataType = ck_tile::remove_cvref_t; + + static constexpr ck_tile::index_t BT_ = BlockTIle_::at(ck_tile::number<0>{}); // block token + static constexpr ck_tile::index_t BI_ = + BlockTIle_::at(ck_tile::number<1>{}); // block intermediate + static constexpr ck_tile::index_t BH_ = BlockTIle_::at(ck_tile::number<2>{}); // block hidden + static constexpr ck_tile::index_t BD_ = BlockTIle_::at(ck_tile::number<3>{}); // block down + + using BlockTile_0 = ck_tile::sequence; + using WarpPerBlock_0 = ck_tile::remove_cvref_t; + using WarpTile_0 = ck_tile::remove_cvref_t; + + using BlockTile_1 = ck_tile::sequence; + using WarpPerBlock_1 = ck_tile::remove_cvref_t; + using WarpTile_1 = ck_tile::remove_cvref_t; + + static constexpr ck_tile::index_t GateOnly = GateOnly_; + static constexpr ck_tile::index_t FusedQuant = FusedQuant_; +}; diff --git a/example/ck_tile/15_fused_moe/instances/fused_moegemm_bf16_m32.cpp b/example/ck_tile/15_fused_moe/instances/fused_moegemm_bf16_m32.cpp new file mode 100644 index 0000000000..93f9c77869 --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moegemm_bf16_m32.cpp @@ -0,0 +1,14 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include "fused_moegemm.hpp" +#include "fused_moegemm_api_traits.hpp" +#include "fused_moegemm_api_internal.hpp" + +// clang-format off +template float fused_moegemm_< + fmoe_, S<1, 4, 1>, S<16, 16, 32>, 1, 0> +>(const ck_tile::stream_config& s, fused_moegemm_args a); + +// clang-format on diff --git a/example/ck_tile/15_fused_moe/instances/fused_moegemm_fp16_m32.cpp b/example/ck_tile/15_fused_moe/instances/fused_moegemm_fp16_m32.cpp new file mode 100644 index 0000000000..b8a823e8ed --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moegemm_fp16_m32.cpp @@ -0,0 +1,14 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include "fused_moegemm.hpp" +#include "fused_moegemm_api_traits.hpp" +#include "fused_moegemm_api_internal.hpp" + +// clang-format off +template float fused_moegemm_< + fmoe_, S<1, 4, 1>, S<16, 16, 32>, 1, 0> +>(const ck_tile::stream_config& s, fused_moegemm_args a); + +// clang-format on diff --git a/example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp b/example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp new file mode 100644 index 0000000000..75aaf86b74 --- /dev/null +++ b/example/ck_tile/15_fused_moe/instances/fused_moesorting_api.cpp @@ -0,0 +1,73 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "fused_moesorting.hpp" + +#define MOE_SORTING_DISPATCH(unroll_num_) \ + constexpr ck_tile::index_t unroll_num = unroll_num_; \ + using ms_problem = ck_tile::MoeSortingProblem; \ + using kernel = ck_tile::MoeSortingKernel; \ + auto kargs = kernel::MakeKargs(a); \ + const dim3 grids = kernel::GridSize(a); \ + const dim3 blocks = kernel::BlockSize(a); \ + const auto lds_bytes = kernel::GetSmemSize(a); \ + float ave_time = ck_tile::launch_kernel( \ + s, ck_tile::make_kernel(kernel{}, grids, blocks, lds_bytes, kargs)); \ + return ave_time; + +float fused_moesorting(fused_moesorting_trait t, fused_moesorting_args a, ck_tile::stream_config s) +{ + if(t.weight_type == "fp32" && t.index_type == "int32") + { + if(a.num_experts > 127) + { + printf("lds size exceed, only support experts <127 \n"); + return -1; + } + if(a.moe_buf_bytes % 16) + { + printf("buf set size %d unaligned, must be multiple of 16\n", a.moe_buf_bytes); + return -1; + } + using index_t = ck_tile::index_t; + using ms_weight_type = float; + index_t smem_io_unroll_num = ck_tile::integer_divide_ceil(a.tokens * a.topk, 64); + switch(smem_io_unroll_num) + { + case(1): { + MOE_SORTING_DISPATCH(1); + } + case(2): { + MOE_SORTING_DISPATCH(2); + } + case(3): { + MOE_SORTING_DISPATCH(3); + } + case(5): { + MOE_SORTING_DISPATCH(5); + } + case(6): { + MOE_SORTING_DISPATCH(6); + } + case(7): { + MOE_SORTING_DISPATCH(7); + } + case(8): { + MOE_SORTING_DISPATCH(8); + } + case(9): { + MOE_SORTING_DISPATCH(9); + } + case(10): { + MOE_SORTING_DISPATCH(10); + } + case(11): { + MOE_SORTING_DISPATCH(11); + } + default: { + MOE_SORTING_DISPATCH(4); + } + } + } + return -1; +} diff --git a/example/ck_tile/15_fused_moe/main.cpp b/example/ck_tile/15_fused_moe/main.cpp new file mode 100644 index 0000000000..2f44f903e9 --- /dev/null +++ b/example/ck_tile/15_fused_moe/main.cpp @@ -0,0 +1,603 @@ +#include +#include +#include +#include +#include + +#include "ck_tile/host.hpp" +#include "fused_moe.hpp" + +// different threshold for different dtype +template +auto get_elimit() +{ + double rtol = 1e-2; + double atol = 1e-2; + return ck_tile::make_tuple(rtol, atol); +} + +template <> +auto get_elimit() +{ + double rtol = 1e-2; + double atol = 1e-2; + return ck_tile::make_tuple(rtol, atol); +} + +// mfma_type, 0:32x32, 1:16x16 +// TODO: padding? +template +auto shuffle_moe_weight(const ck_tile::HostTensor& t, std::string mfma_dtype, int mfma_type = 0) +{ + assert(t.get_lengths().size() == 3); + int b_ = t.get_lengths()[0]; + int n_ = t.get_lengths()[1]; + int k_ = t.get_lengths()[2]; + if((mfma_dtype == "bf16" || mfma_dtype == "fp16") && mfma_type == 0) + { + ck_tile::HostTensor t_view({b_, n_ / 32, 32, k_ / 16, 2, 8}); + std::copy(t.begin(), t.end(), t_view.begin()); + return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5}); + } + else if((mfma_dtype == "bf16" || mfma_dtype == "fp16") && mfma_type == 1) + { + ck_tile::HostTensor t_view({b_, n_ / 16, 16, k_ / 32, 4, 8}); + std::copy(t.begin(), t.end(), t_view.begin()); + return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5}); + } + else if((mfma_dtype == "int8" || mfma_dtype == "fp8") && mfma_type == 0) + { + ck_tile::HostTensor t_view({b_, n_ / 32, 32, k_ / 32, 2, 16}); + std::copy(t.begin(), t.end(), t_view.begin()); + return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5}); + } + else if((mfma_dtype == "int8" || mfma_dtype == "fp8") && mfma_type == 1) + { + ck_tile::HostTensor t_view({b_, n_ / 16, 16, k_ / 64, 4, 16}); + std::copy(t.begin(), t.end(), t_view.begin()); + return ck_tile::reference_permute(t_view, {0, 1, 3, 4, 2, 5}); + } + return t; +} + +template +void topid_unique_gen( + std::vector& host_tensor, int tokens, int topk, int num_expert, int seed) +{ + size_t total_size = topk * tokens; + std::srand(seed); + std::set unique_set; + IndexType current_v; + for(size_t i = 0; i < total_size; i++) + { + if(i % topk == 0) + { + unique_set.clear(); + } + current_v = std::rand() % num_expert; + while(unique_set.find(current_v) != unique_set.end()) + { + current_v = std::rand() % num_expert; + } + unique_set.insert(current_v); + host_tensor[i] = current_v; + } +} + +auto create_args(int argc, char* argv[]) +{ + ck_tile::ArgParser arg_parser; + arg_parser.insert("t", "128", "num input tokens") + .insert("e", "32", "num of experts") + .insert("k", "5", "topk") + .insert("h", "8192", "hidden_size of this model") + .insert("i", "8192", "intermediate_size between 2 gemms of FFN") + .insert("stride", "-1", "stride per row, if -1 then equal to hidden_size") + .insert("bm", "32", "blocking factor for sorted tokens") + .insert("tp", "8", "tensor parallel size") + .insert("v", "1", "cpu validation or not") + .insert("kname", "1", "print kernel name or not") + .insert("prec_i", "bf16", "input precision") + .insert("prec_w", "bf16", "weight precision") + .insert("prec_o", "bf16", "output precision") + .insert("prec_st", "auto", "token scale data type. auto will set to fp32") + .insert("prec_sw", "auto", "weight scale data type. auto will set to fp32") + .insert("prec_sq", "auto", "(dynamic) smooth quant data type. auto will set to fp32") + .insert("prec_kw", "auto", "topk-weight data type. auto will set to fp32") + .insert("fquant", "0", "fused-quant, 0:no, 1:smooth-dynamic-quant, 2:dynamic-quant") + .insert( + "gate_only", "1", "w0(gate/up) style, 0:gate+up will double interm size, 1:only gate") + .insert("api", "0", "benchmark api set: 0:fused-moe(moe-gemm+moe-sorting), 1:moe-gemm") + .insert("balance", + "0", + "if set to 1, will try balance the expert in topk-ids(convenient for testing)") + .insert("init", + "2", + "init method. 0:random stepped float(fast). 1: random uniform, 2:rand normalized" + "normalized(slow)") + .insert("seed", "11939", "seed used to do random") + .insert("warmup", "5", "cold iter") + .insert("repeat", "20", "hot iter"); + + bool result = arg_parser.parse(argc, argv); + return std::make_tuple(result, arg_parser); +} + +// I:input-type, W:weight-type, O:output-type, ST:toke-scale-tpye, SW:weight-scale-type, +// SQ:smooth-quant-type, KW:topk-weight-type +template +bool run(const ck_tile::ArgParser& arg_parser) +{ + ck_tile::index_t tokens = arg_parser.get_int("t"); + ck_tile::index_t experts = arg_parser.get_int("e"); + ck_tile::index_t topk = arg_parser.get_int("k"); + ck_tile::index_t hidden_size = arg_parser.get_int("h"); + ck_tile::index_t intermediate_size = arg_parser.get_int("i"); + ck_tile::index_t stride = arg_parser.get_int("stride"); + ck_tile::index_t block_m = arg_parser.get_int("bm"); + if(stride < 0) + stride = hidden_size; + std::string prec_i = arg_parser.get_str("prec_i"); + std::string prec_w = arg_parser.get_str("prec_w"); + std::string prec_o = arg_parser.get_str("prec_o"); + std::string prec_st = arg_parser.get_str("prec_st"); + std::string prec_sw = arg_parser.get_str("prec_sw"); + std::string prec_sq = arg_parser.get_str("prec_sq"); + std::string prec_kw = arg_parser.get_str("prec_kw"); + prec_st = (prec_st == "auto") ? "fp32" : prec_st; + prec_sw = (prec_sw == "auto") ? "fp32" : prec_sw; + prec_sq = (prec_sq == "auto") ? "fp32" : prec_sq; + prec_kw = (prec_kw == "auto") ? "fp32" : prec_kw; + int kname = arg_parser.get_int("kname"); + int do_validation = arg_parser.get_int("v"); + int warmup = arg_parser.get_int("warmup"); + int repeat = arg_parser.get_int("repeat"); + int fused_quant = arg_parser.get_int("fquant"); + int gate_only = arg_parser.get_int("gate_only"); + int api = arg_parser.get_int("api"); + int balance = arg_parser.get_int("balance"); + int tp = arg_parser.get_int("tp"); + int init = arg_parser.get_int("init"); + uint32_t seed = arg_parser.get_uint32("seed"); + + // w0 (Gate+Up or Gate only, N size) + ck_tile::index_t shared_intermediate_size_0 = intermediate_size * (gate_only ? 1 : 2) / tp; + // w1 (Down, N size) + ck_tile::index_t shared_intermediate_size_1 = intermediate_size / tp; + + auto prec_str = [&]() { + auto base_str = prec_i; + if(prec_i != prec_w) + base_str += "x" + prec_w; + if(prec_i != prec_o) + base_str += "=" + prec_o; + if(fused_quant != 0) + { + base_str += std::string("(") + prec_st + "|" + prec_sw + "|" + prec_sq + ")"; + } + return base_str; + }(); + auto api_str = [&]() { + if(api == 0) + return std::string("fmoe"); + else if(api == 1) + return std::string("moeg"); + else if(api == 2) + return std::string("moes"); + return std::string(""); + }(); + + auto stride_str = [&]() { + if(stride == hidden_size) + return std::string(""); + else + return std::string(", st:") + std::to_string(stride); + }(); + + std::cout << "[" << api_str << "|" << prec_str << "]" + << " t:" << tokens << ", e:" << experts << ", k:" << topk << stride_str + << ", hidden:" << hidden_size << ", interm:" << intermediate_size << ", tp:" << tp + << ", shrd_interm:" << shared_intermediate_size_0 << "|" << shared_intermediate_size_1 + << ", go:" << gate_only << ", q:" << fused_quant << std::flush; + + using TypeConfig = FusedMoeGemmTypeConfig; + using ADataType = typename TypeConfig::ADataType; + using GDataType = typename TypeConfig::GDataType; + using DDataType = typename TypeConfig::DDataType; + using AccDataType = typename TypeConfig::AccDataType; + using ODataType = typename TypeConfig::ODataType; + using AScaleDataType = typename TypeConfig::AScaleDataType; + using GScaleDataType = typename TypeConfig::GScaleDataType; + using DScaleDataType = typename TypeConfig::DScaleDataType; + using YSmoothScaleDataType = typename TypeConfig::YSmoothScaleDataType; + using TopkWeightDataType = typename TypeConfig::TopkWeightDataType; + using IndexDataType = typename TypeConfig::IndexDataType; + + // host verify + ck_tile::HostTensor a_host({tokens, hidden_size}, {stride, 1}); + ck_tile::HostTensor g_host({experts, shared_intermediate_size_0, hidden_size}); + ck_tile::HostTensor d_host({experts, hidden_size, shared_intermediate_size_1}); + ck_tile::HostTensor o_host({tokens, hidden_size}, {stride, 1}); + ck_tile::HostTensor sa_host({tokens}); + ck_tile::HostTensor sg_host({shared_intermediate_size_0}); + ck_tile::HostTensor sd_host({shared_intermediate_size_1}); + ck_tile::HostTensor sy_host({shared_intermediate_size_1}); // smooth-quant + ck_tile::HostTensor topk_ids_host({tokens, topk}); // to be sort + ck_tile::HostTensor topk_weight_host({tokens, topk}); // to be sort + + int max_num_tokens_padded = topk * tokens + experts * block_m - topk; + ck_tile::HostTensor sorted_token_ids_host({max_num_tokens_padded}); + ck_tile::HostTensor sorted_weight_host({max_num_tokens_padded}); + ck_tile::HostTensor sorted_expert_ids_host( + {(max_num_tokens_padded + block_m - 1) / block_m}); + ck_tile::HostTensor num_sorted_tiles_host({1}); + + if(init == 0) + { + ck_tile::FillStepRange{-.5f, .5f, 0.01f}(a_host); + ck_tile::FillStepRange{-.5f, .5f, 0.01f}(g_host); + ck_tile::FillStepRange{.5f, -.5f, -0.01f}(d_host); + ck_tile::FillStepRange{0.f, 1.f, 0.01f}(sa_host); + ck_tile::FillStepRange{0.f, 1.f, 0.01f}(sg_host); + ck_tile::FillStepRange{0.f, 1.f, 0.01f}(sd_host); + ck_tile::FillStepRange{0.f, 1.f, 0.01f}(sy_host); + ck_tile::FillStepRange{-.5f, .5f, 0.01f}(topk_weight_host); + } + else if(init == 1) + { + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(a_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(g_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(d_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(sa_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(sg_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(sd_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}(sy_host); + ck_tile::FillUniformDistribution{-.5f, .5f, seed, true}( + topk_weight_host); + } + else if(init == 2) + { + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(a_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(g_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(d_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(sa_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(sg_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(sd_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(sy_host); + ck_tile::FillNormalDistribution{0.f, 1.f, seed, true}(topk_weight_host); + } + + // permute weight + ck_tile::HostTensor g_perm_host = shuffle_moe_weight(g_host, prec_w, 1); + ck_tile::HostTensor d_perm_host = shuffle_moe_weight(d_host, prec_w, 1); + + // do moe sorting + if(balance) + { + int e_cnt = 0; + for(int i = 0; i < static_cast(topk_ids_host.mData.size()); i++) + { + topk_ids_host.mData[i] = e_cnt; + e_cnt++; + if(e_cnt >= experts) + e_cnt = 0; + } + } + else + { + topid_unique_gen(topk_ids_host.mData, tokens, topk, experts, 11913); + } + +// leave it here for future debug purpose +#if 0 + a_host.loadtxt("../../ater/input_torch.txt"); + + topk_ids_host.loadtxt("../../ater/topk_ids_torch.txt", "int"); + // topk_ids_host.savetxt("topk_ids_2.txt"); + topk_weight_host.loadtxt("../../ater/topk_weights_torch.txt", "float"); + std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl; + + g_host.loadtxt("../../ater/w1_torch.txt", "float"); + std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl; + d_host.loadtxt("../../ater/w2_torch.txt", "float"); + std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl; + + ck_tile::HostTensor g_perm_host = shuffle_moe_weight(g_host, prec_w, 1); + std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl; + ck_tile::HostTensor d_perm_host = shuffle_moe_weight(d_host, prec_w, 1); + std::cout << "------- @@@ " << __LINE__ << std::flush << std::endl; +#endif + +#if 0 + std::cout << "sorted_token_ids_host:" << sorted_token_ids_host << std::endl; + std::cout << "num_sorted_tiles_host:" << num_sorted_tiles_host << std::endl; + std::cout << "sorted_expert_ids_host:" << sorted_expert_ids_host << std::endl; + std::cout << "topk_weight_host:" << topk_weight_host << std::endl; + std::cout << "sorted_weight_host:" << sorted_weight_host << std::endl; +#endif + auto cal_tflops = [&](auto ms) { + double flop_gemm_0 = + 2 * static_cast(tokens) * topk * shared_intermediate_size_0 * hidden_size; + double flop_gemm_1 = + 2 * static_cast(tokens) * topk * shared_intermediate_size_1 * hidden_size; + return (flop_gemm_0 + flop_gemm_1) / (static_cast(ms) * 1e-3) / 1e12; + }; + + // TODO: this method we use expert-by-expert view, just for reference + auto cal_tbps = [&](auto ms) { + double token_bytes = + static_cast(tokens) * topk / experts * hidden_size * sizeof(ADataType); + double w0_bytes = static_cast(shared_intermediate_size_0) * experts * hidden_size * + sizeof(GDataType); + double w1_bytes = static_cast(shared_intermediate_size_1) * experts * hidden_size * + sizeof(DDataType); + double o_bytes = + static_cast(tokens) * topk / experts * hidden_size * sizeof(ODataType); + double topk_weights_bytes = static_cast(tokens) * topk * sizeof(TopkWeightDataType); + // ignore index, they are too small + + return (token_bytes + w0_bytes + w1_bytes + o_bytes + topk_weights_bytes) / + (static_cast(ms) * 1e-3) / 1e12; + }; + + if(api == 0) + { + ck_tile::DeviceMem a_buf(a_host); + ck_tile::DeviceMem g_perm_buf(g_perm_host); + ck_tile::DeviceMem d_perm_buf(d_perm_host); + ck_tile::DeviceMem sa_buf(sa_host); + ck_tile::DeviceMem sg_buf(sg_host); + ck_tile::DeviceMem sd_buf(sd_host); + ck_tile::DeviceMem sy_buf(sy_host); + ck_tile::DeviceMem o_buf(o_host.get_element_space_size_in_bytes()); + + ck_tile::DeviceMem topk_ids_buf(topk_ids_host); + ck_tile::DeviceMem topk_weight_buf(topk_weight_host); + + ck_tile::DeviceMem sorted_token_ids_buf( + sorted_token_ids_host.get_element_space_size_in_bytes()); + ck_tile::DeviceMem sorted_weight_buf(sorted_weight_host.get_element_space_size_in_bytes()); + ck_tile::DeviceMem sorted_expert_ids_buf( + sorted_expert_ids_host.get_element_space_size_in_bytes()); + ck_tile::DeviceMem num_sorted_tiles_buf( + num_sorted_tiles_host.get_element_space_size_in_bytes()); + + fused_moe_traits traits{prec_i, + prec_w, + prec_o, + prec_st, + prec_sw, + prec_sq, + prec_kw, + block_m, + gate_only, + fused_quant}; + + fused_moe_args args{a_buf.GetDeviceBuffer(), + fused_quant != 0 ? sa_buf.GetDeviceBuffer() : nullptr, + g_perm_buf.GetDeviceBuffer(), + d_perm_buf.GetDeviceBuffer(), + fused_quant != 0 ? sg_buf.GetDeviceBuffer() : nullptr, + fused_quant != 0 ? sd_buf.GetDeviceBuffer() : nullptr, + fused_quant == 1 ? sy_buf.GetDeviceBuffer() : nullptr, + o_buf.GetDeviceBuffer(), + topk_ids_buf.GetDeviceBuffer(), + topk_weight_buf.GetDeviceBuffer(), + sorted_token_ids_buf.GetDeviceBuffer(), + sorted_weight_buf.GetDeviceBuffer(), + sorted_expert_ids_buf.GetDeviceBuffer(), + num_sorted_tiles_buf.GetDeviceBuffer(), + block_m, + hidden_size, + shared_intermediate_size_0, + tokens, + experts, + topk, + stride}; + float ave_time = fused_moe( + traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat}); + + if(ave_time < 0) + { + std::cout << " not supported!" << std::endl << std::flush; + return false; + } + + // float gb_per_sec = num_byte / 1.E6 / ave_time; + std::cout << ", " << ave_time * 1.E3 << " us, " << cal_tflops(ave_time) << " tflops, " + << cal_tbps(ave_time) << " TB/s" << std::flush; + bool pass = true; + + if(do_validation) + { + ck_tile::reference_moe_sorting( + topk_ids_host, + topk_weight_host, + sorted_token_ids_host, + sorted_weight_host, + sorted_expert_ids_host, + num_sorted_tiles_host.mData[0], + experts, + block_m); + + ck_tile::reference_fused_moe( + a_host, + g_host, + d_host, + sa_host, + sg_host, + sd_host, + sy_host, + o_host, + sorted_token_ids_host, + sorted_weight_host, + sorted_expert_ids_host, + num_sorted_tiles_host, + topk_ids_host, + block_m, + tokens, + experts, + hidden_size, + shared_intermediate_size_0, + topk, + gate_only); + + auto o_dev = o_buf.ToHost(); + // o_dev.savetxt("gpu-out.txt", "float"); + auto [rtol, atol] = get_elimit(); + pass &= ck_tile::check_err( + o_dev, o_host, std::string("OUT Error: Incorrect results!"), rtol, atol); + std::cout << ", valid:" << (pass ? "y" : "n") << std::flush; + } + std::cout << std::flush << std::endl; + return pass; + } + else if(api == 1) + { + ck_tile::reference_moe_sorting( + topk_ids_host, + topk_weight_host, + sorted_token_ids_host, + sorted_weight_host, + sorted_expert_ids_host, + num_sorted_tiles_host.mData[0], + experts, + block_m); + + // done, preparing GPU buffer + ck_tile::DeviceMem a_buf(a_host); + ck_tile::DeviceMem g_perm_buf(g_perm_host); + ck_tile::DeviceMem d_perm_buf(d_perm_host); + ck_tile::DeviceMem sa_buf(sa_host); + ck_tile::DeviceMem sg_buf(sg_host); + ck_tile::DeviceMem sd_buf(sd_host); + ck_tile::DeviceMem sy_buf(sy_host); + ck_tile::DeviceMem o_buf(o_host); + + // manually clear output buffer for atomic + o_buf.SetZero(); + // + + ck_tile::DeviceMem sorted_token_ids_buf(sorted_token_ids_host); + ck_tile::DeviceMem sorted_weight_buf(sorted_weight_host); + ck_tile::DeviceMem sorted_expert_ids_buf(sorted_expert_ids_host); + ck_tile::DeviceMem num_sorted_tiles_buf(num_sorted_tiles_host); + + fused_moegemm_traits traits{prec_i, + prec_w, + prec_o, + prec_st, + prec_sw, + prec_sq, + prec_kw, + block_m, + gate_only, + fused_quant}; + + fused_moegemm_args args{a_buf.GetDeviceBuffer(), + fused_quant != 0 ? sa_buf.GetDeviceBuffer() : nullptr, + g_perm_buf.GetDeviceBuffer(), + d_perm_buf.GetDeviceBuffer(), + fused_quant != 0 ? sg_buf.GetDeviceBuffer() : nullptr, + fused_quant != 0 ? sd_buf.GetDeviceBuffer() : nullptr, + fused_quant == 1 ? sy_buf.GetDeviceBuffer() : nullptr, + o_buf.GetDeviceBuffer(), + sorted_token_ids_buf.GetDeviceBuffer(), + sorted_weight_buf.GetDeviceBuffer(), + sorted_expert_ids_buf.GetDeviceBuffer(), + num_sorted_tiles_buf.GetDeviceBuffer(), + hidden_size, + shared_intermediate_size_0, + tokens, + experts, + topk, + stride}; + + float ave_time = fused_moegemm( + traits, args, ck_tile::stream_config{nullptr, true, kname ? 1 : 0, warmup, repeat}); + + if(ave_time < 0) + { + std::cout << " not supported!" << std::endl << std::flush; + return false; + } + + // float gb_per_sec = num_byte / 1.E6 / ave_time; + std::cout << ", " << ave_time * 1.E3 << " us, " << cal_tflops(ave_time) << " tflops, " + << cal_tbps(ave_time) << " TB/s" << std::flush; + bool pass = true; + + if(do_validation) + { + ck_tile::reference_fused_moe( + a_host, + g_host, + d_host, + sa_host, + sg_host, + sd_host, + sy_host, + o_host, + sorted_token_ids_host, + sorted_weight_host, + sorted_expert_ids_host, + num_sorted_tiles_host, + topk_ids_host, + block_m, + tokens, + experts, + hidden_size, + shared_intermediate_size_0, + topk, + gate_only); + + auto o_dev = o_buf.ToHost(); + // o_dev.savetxt("gpu-out.txt", "float"); + auto [rtol, atol] = get_elimit(); + pass &= ck_tile::check_err( + o_dev, o_host, std::string("OUT Error: Incorrect results!"), rtol, atol); + std::cout << ", valid:" << (pass ? "y" : "n") << std::flush; + } + std::cout << std::flush << std::endl; + + return pass; + } + return false; +} + +int main(int argc, char* argv[]) +{ + auto [result, arg_parser] = create_args(argc, argv); + if(!result) + return -1; + + std::string prec_i = arg_parser.get_str("prec_i"); + std::string prec_w = arg_parser.get_str("prec_w"); + std::string prec_o = arg_parser.get_str("prec_o"); + std::string prec_st = arg_parser.get_str("prec_st"); + std::string prec_sw = arg_parser.get_str("prec_sw"); + std::string prec_sq = arg_parser.get_str("prec_sq"); + std::string prec_kw = arg_parser.get_str("prec_kw"); + prec_st = (prec_st == "auto") ? "fp32" : prec_st; + prec_sw = (prec_sw == "auto") ? "fp32" : prec_sw; + prec_sq = (prec_sq == "auto") ? "fp32" : prec_sq; + prec_kw = (prec_kw == "auto") ? "fp32" : prec_kw; + + // no dynamic quant case + if(prec_i == "bf16" && prec_w == "bf16" && prec_o == "bf16" && prec_kw == "fp32") + { + return run( + arg_parser) + ? 0 + : -2; + } + else if(prec_i == "fp16" && prec_w == "fp16" && prec_o == "fp16" && prec_kw == "fp32") + { + return run( + arg_parser) + ? 0 + : -2; + } + + return -3; +} diff --git a/example/ck_tile/15_fused_moe/misc/moe-0.png b/example/ck_tile/15_fused_moe/misc/moe-0.png new file mode 100644 index 0000000000000000000000000000000000000000..aed1964f2802c4e7f65d7080f338309c8c2287a6 GIT binary patch literal 76830 zcmdSA^;=t8)HPZ>c+lbyytuZwJ4K7TLveS4yB065#ibN0?ouoicXxM(FXx>1yZ8PH z@BM-2A$fMP*Pd&vF~=NpCQ?~Z8XbiQ<=wk?=&~{rs_)*xmO=mA0kF__u9#x1-@SW# zla&xv^ISa6M#{pKTpVci+1!1{gbGLu-liw$0?@7Nd&Bj__e_2b~xSX|8}W3GCaJ{=5;lm&A-v1l8B6q{QdiPyT$tM&Yc&t zkyPfZ-Lc?7)*4ku(xxRrv!R6aeG_BjjX6XA*!pbm!!l;w#+bkQzCvqHXPs{cCAo0~ z?3)9`4wWyqvmDTs|FWD|?eOE5YjD{c-*uXw2tmf#h2F{iSZmsIsK!XMc7AzT4|i_8 zJ6N~g3ajFp3VrH_LT~4L15e{$7GowUF^9D_bJc79Z=@%uiSjJ6pFhXAIa{Y6A0LOJ zzVERC1q%lE@kwqIL`2=D(XnY5u_*UxS-uf?p0(dqcfLMK#S(Rs-|SCrS%HJ<+V^sN z47?9$e1-0oT`qc~Cdhf6x24JbLNI}J7a;hWxu?_Cv)8-Mx0PDs-qV%Vv8@Qn_2miv zv$jn8mJd`^RL|iK?Kk)xi~wku>$))RRwxma|>Pd9}5* z=~mj#4O2V#Cr2oaO#l|^t@2|qw%M&@(;69E-J zLfH5tJ6a+z==x}m6$tQs*hxz>@cvqvs6yItzZpHk+DStG(QYwhUto5)OcppBL&&+C z>Cl1Y-_zc9F+luN5tkRSp!8**vdMl$2SqOLM)w(c!?4b5_@elQ-2eF^MbluRTA$yu zpz!^W|Laxm+v8kk-`;Q%-J17N)lN3A%kJo}aXxw9xcLP4<9Vks?z!jNC5KI=h++b; z$R~`Peu4YLiW;#EG4NCF+w)$_QMLHbE%<3vp^&TqDsyEzw6 z+>5Si{9kO`SVHY7#Y)vHGz4F+r$ELM1_rAu9&51#7;pbPRzvQXj>SZdpr)P^!ejP? zz}4v5+pEVFPuoW(!a+l?os=c})-i!Nt&O@77AQtkwgzI2O97POXoQp*=L50CSpA#f zICP2vC-qa-bT6e6^$}3$Q;3MBV;^l+qOAGdt*pJ?ehN8`us8i%S1?$rQn(6!vTIt3 z-3P{cbOpfRp{XbG$>27SpTdpPY!bY)_vSK%bBj8(>KwBOZf>-mLFtU;u~1f4X0w7& zxVrY`ZY*|!Ummfjq#_29=;xv6_~rL>yW6zi|GM_}vUW8g%qsDcW(Y?9=5KX4qiIM= zPL645CPm;DPBab1?U$v?uJ?a?z1SLrg8F5HE-ylv^yj#fopR1?nkEMHxP)PC_D+|5 znGk;SJ<^1pzt;VNX!o;K^r!CdMOop7&T&}z=IumzAnmH#oW61%s&wPA&eS`86 z{pc4ZLGDepFYuq)7}Ve^Q8mJIw>CEdb?og=k;r{(eIJ|w44W~}&(DV72lMeP-sTPh zI_d9fIx+Xd(1(bkKM5nYmywf0=@h~a_SHuEdX591SK9H7N_taix(B+f7Kb$GVOCa5$tKqWRXxM!L+Ei0#X4R7jk7`ZN+Rg(DkcUSLyj09^zwW)@ofVy@9`B}GM9(6CxzMWiV4pVVL0 zlgFX%75-I$1>CoIoW*@qy&wgtu}NO}*2^XPd(mxm2*2J_F#!hxu zf>f21l)||-ADep&UoX?3UxF9NRH@&33&|H}184dF`B0MO*n^ysidQZbLrXb3_yFbO zHQy`rXFU7qLg{#U)^@HwT^-SPNJ#-V1@9*7l;g3GoRvwx{P-N06hhqny;l8x0?Oh@ zjZsqY@&R<^& zcWI~`HzS;eiir$|XCcmAtooeg>*v8c!WN-a6W>H*mKt)A&?|a{q7mbOF_{r=mw%c{ zXBn>Q+2YpFGd{~D;~l?Q`O0ud{=6-`R#7q;;%nnaO@p6^u$DrtA5RVx?Rq&`Z6}5` zB9p-C70U5B{hbhGv_wLL+pSEPPMhNlLm4v%X(caKJkTR6 zZWN0eNBl_pt}M9|l&XJZM3`o1 zAs1$W^9GQm^VzxRquvw9sA&z%|JSyiN`OO*4_7ECK0!C{;cbijP2}n)A56WC1OaQ? zx!C>n{QDrSk|{aSsrI{7A1s1UOBlqk(alE-ZwL#-G!Q8~3}m>mi1|U%8uUCH983IZ zpoYwQKR-$y%_Lr#;4vyuQyot>NFigb;6wf5r{B~+X@BDoHb#==hPiUR(o`@jr%rj7Y^Zz=h;l=K6gs z;?0XK)VWAXNb27v?xURL)nX&VrFeAMAdM%Yh9QF5yn!ZbU{UAcM%aJ@IaE9|@%n$lqQ z&Q@CA8!BY;&FFNnU5v1_qBX>oi)}|3Zg3uwq2qPnot9ZI#w#KoBci~VN}8;PK>+>8 zP?;;#2KR`=1G9lzjb;Ow1p7w60yp)RTmEu;)@qv7LWRzLHAt@^mD&uMxje|R0ah7< zo<$9Q87BlD3z6t!)J|DR4)%sHQAxua3S+o(%i`a;pcO4Uj(>X zB2f7|RTP9!tscA{YM`V-jsW3hB#v^Pe6AiFp`lWd&Jf_ zk@g^;jSdSYrz8v24D>9-{>ETTY;s*4Bqm<1f=bVqj2QBTNmT6f`K+!3Xk5f4Kv-gM z_!-$5#R1tov^x)C#Cy)bs8QMVV3kNuxV*@B88r}18w04Rtp(CigqCVw8|XCI#{!XP zLuLcZyRqtFub))POZlIHq{Q@}>cZ5zM1WIheC59jZ-iQAYYls>m`zj978bPOt zq0TR+Fr+702AD^LbhCB~h>c7CSf_X7E`4_y?JSZ4vZ41)#)BoL#{3JcgO!BezkjzD zeNC}Yio1;}6%-6&CQ6f(TW?los?}9PISuOTb?dcVAHL|_ANGeTNq1|KwfCJs94Ec(fy-%R5TnjYSV!y;Yja$ zS0YX~Rv5Asg5^Qd0D0rI0`9295UB!qWjV+F*|=8RYt(bVIqim@2aXy{RtD~Mkp{W3 zP<*Y>L$FQosa&@g+H4ME^TTfDWEaHyLL4`mM~iBaHLmxmFwRkwhZUecq(o`mSxfIL zk`(-$KBz+4jL;*h1&cNyU~k4yhA~ympf9Ivegc*KD3o|MfqC%jzbRB$+pn=7 zMRwidRG$!b^@MpOu0twigIq-qF3zC3Rctq!qGZ3i+uj&7I~)vfbTU*sf)yaS*M61q%OhM1QmF= zM4!hG*(pLOSpwZa=|H4OIPup%GCb5T>V6*nClwQlu;|?HFC~gQctqPsYnyB5u06#RxZ(<+5_+#+ z6#@04+#C2nf_$T~D=er&yjQI^@llF zFp6j)qW`kTfO(LH1!jnkV0}s-G~PSM1M19oH9Kvsw^n95JTt!kDdq+G>U6!sSW0-*gw-F&-fGnrEM<8}+68pMzCgcFBpfQBCiz;fbRw2Qdy#Q^5^ zRBfs5n`wGW4}qLdoPX!NIEGrtMsb7B5AGR?{yLF(gyV(PJ2J{J>tLXupszlv;Wb3& z;P8VWQ1hPY2J`yM4YI#>7 zireQKy^=K}D1R~4MaWdrsib0nND3$M_c&U#*Kqg%2VGHab?GOkorHcOHXYZ@JjdDm zd(#;(r%5nRs6^mCMF9yUPb^HNOI=exlrq$LG>do2o8%yTFl0}%qAahSqFR#A_-E<2 z1j){Q>u6p^C`oP>$h^Gc4$6+QP2F!ch zR}v?Nz(Y>OTEdh`;eq$HR+E(1VO0+(bJWAw9aJLhnT-%|GeYwhEtMf9$Zu&-?+zGR zqRsAkK-ko17Lyy=(4Zr0D`$15S#6 z{OEBWl|>vPMME?v6T%NYUgs);3H(Ngkg@1s_l4FmC8AS&z~mcQ5ejYWEi6rx){sSg z$UWL{*q#J{SO}l7=l%7M1cW0B4Rq|xcE7wG6Bcp2dUA8RawH*uxt*UkI&z>?Pf}F+ zyKbg!xnr6D`o2dI1pgkJ-f3V^x2psNtrSL49$Cnpn$ehvYqhoM>yO*QZjlic2`lP| zkB+QGU$!X=ZhOtX32avc5x#0byh4XtR_r~u9Zzc3!$D&_@b`XLSX7Uq=`p+3-X! ztHwFCFraI5D*Y% z4X8#D-J|~!2zqmMa43aUclG9d0Ch=XeV^=vfJ6G}&(F1^cD{C|+t~hEsJvpR?8MHw zeYsnc&W{W>85L${SGmR)15?(M(}sKwoiF<8S=yI?TkwRTF6A2fZhSco_HUdHJqt@1 zdx)#tMhCoCB_isEH3hi|3y~Q6G&dKSJ_WoPXs>r@a!Yc!3m{qf zE*Pu0-TTJ+Ctz8C5Dcdo@XhsP5pPcT2k$RwB`7n6@;Gfu@VHhu$#;Lq+r5{O!l7_2@|uFn5Dxqpp81V$tE$%Olx_Hk4D$ftb=#C>mQ5CP{1)v;^X&WI(vp zbb=suhn7$Iti4sI7AY?OGmALKWS)gR{ zdJWUzerxmz@I-K#_~#q{xD|$as>;g=SL9yX-%4`4QOtu@e*?lO|AEtbehrXJJ%V`~ zvKN5#{)*B@`U!>smX!8_a3KCtqGNq9UkPuieBPLiiV7|gbBzOVO15TUZm!)xZf@`u z9?g_Z@WUaIXlDAbdUqib3>V%Q`CkI3ZOpEdt_OX_V{`Q*OO=MPxOJQk(I*P0oEnzNdWTlwx);UWM+tWbiN z<7hg2go}cF=!fX+mA-U!8O3ih1UAw5%;z{fOundtDAC6&t#<{I(LG5-9UbH;=*A+O z;^X!9vDhk4#_Co}#GmYKA$c%n6TUhmDo;_S93e31dvlaJAE3W@D%*{@83Wk_M&>$lIRqe|LJ zp;qUEq}79pwQxvPP~>;+>KEacYn)1qpG++DQXJL_PsK(`7zn7Jto1J)-(K$GN%=pT z;RUBa?Ykc7EqERt7uQAV%|~$l%h#`8&krD%3H11}MAC^#Nn_vaWX_=VAYw_(w)&#G zBdaGSm(o4{FH>ttv{&uH^Zp^uQ8;|_?9WCznnRD^FHiYq01wA9$Ig&})hgYlgh&pJ z&nMw&fU~c-3Sbph#yEbfK2UXM`N!=@}Fz0$*oWg1p#ASbyFKzSY zr$1E|*7Zo6g$i_AC~NHIsf}T8af8)o!5`+Mlp#c`IMq%vOFE^l41D4?ag&{6LK~1A`oBD%OkY zuO6c|>y;Q}gCwzxBYyHeK7ryd{Zhn50}#^;{q(ayZ$;)}{9iIYSNg}7<~+#c;xt&L zEdhR>chuqW`oP4UI*Q0xbSMiGhaCoeA^bKx{!VHdd|`x99+^xs<^y1b`Wv2>)MPhg zPLJR4uG8~f*yQUwBF;L2uS8)j;1}-;0Gb6Nae$K;Rj^C&S{Ld0GX=g+bL1u!1`8-kX>SCX zjAU3@`QN*DVg>Sab5jME8sav)-Gt!5pE`Db%tSZ#z6Zw*(MAcH#pZfmBc`Fte@v^* z=*rD|;32`dx**qsQ4PwFjMaSP5UIDCM#zdR%K%YGm-3jQR2h(nVv@xPLD{&U1;0m2 z)Cpv`Df{c7J47i!S_j&JxiyH+Lp>KT`6TYXwjS65f9ZbxCG zy5v5YUEw|MRg0LMl1}3KJw9%#=Ldk}RWt@;97ip{AhbtxH*Xprn#CmMsYZM#(zY?3 zB}w2k<;*^>_3|5x2kpG{(mP3Wx$V<<=g!{!)CsK8bDL35AmOneN@;kn>UZ*f!nLOL zl#ig@jW0$|(4V*HDI{)j<;r|YZ0P&wS!;HU;;=STDc)oiSmXJU^1aAU~`-M8Jf)08`^>y5by+_ zMO(Et&}hVoM!mi@6uv|pD)aAWg|Nze3*wi`g9}Kn;ag-)vFi!z3$b0Gj#0RF89fN8 zB}x8RP#uxfEcvkYNdK0wJcG+V3x0Pc_8|{0#G4})+ny+tK5Rwo@r5nM9M8g2ZaW4r?wd0xqFC z%LhV_@aZJT`v%%{Ei#%6517NzLDgiuO=DxfyX{8C@@Yy5QYKy5KhMT2T2Ey%IVQE! zEkb$4Wk5|4`wZk*`Ac#xjmlL55Ge*eSZCn>jo0o3x7R);C=+!9`#!hr@<=>QZisM( zNyd^162lAa|1Cc6eh0ofNnNyNl_qDJ5ZUO$T|=aEuYGF9Vf}m{i?+&`gi~0-8?awiR0%oKa?-)imnO z?+sDLwVJj%(BPFu?!rD#;r>zp7f|Q75&B*%RU;GAswyMue5WW$5xZtTG)DUdxg?*V zzX0v%@fF(l%JzSKfTp-!N#)&HZ;9%EoPIs#f>K=N9xiuLE+uOGniq9YxrU2I7qoe+;((7QF2K1-O!^FtTZcsPTK8 z`U(PCh$4_j2^rZMl9Yr9-%HSR$M!Tr?k6Jy6(WXdK{bf)xv3Fu&@d{re_7l!Tcc`* zIZ~d-r`%SQqs-uvYU2)RTV^8|BFc;1bhlS)|3YL|uoXq|3B?Z+zzplkYfgB#Zx8Kt zSNYhX^cC_D*n2tCq|(ZDzaT41wENsi?O`*(&R{u(&7iJZ??dut6V!0|`v{r_6eim? zRP(!mwp?UKE$Ykf`=kGeIf;6{-}){x6#5{4sh)I+BpaG41Yq(1>ZNR;*b79R=SN3oW8QWs}qbh`Yz(8Cz zpr1{Md9%gQY?U!56qy0Lk+hM%k?k?c>y(PvtAw8$O;S|v~_+LZBx$fkV zEmzyT3eHfEg7l#S8gRdZcPOlvo1F?ht3-y3;8f$t1S%1JQXLZg2%0j=b%dB`3sce^ zVL*OZY-yg)u_1dt{C3&f^nFJ05^b7rWo5%2LS%UVC(wvU4h;^-$}OVUf15#1e*{U3 za4lXhH|-NCRsP9Oe2k<)JzigZmKOArh>}=~_Ej<;G*>D;&r50B*9R4#`{Bt=r0r(k zTf-+xKm0th)6-jq>WgnOYDu`b4o9mRyJSK<=-n(abkO9u5Xc{@yb(0(#8XxdC9sp; zSj=WpS6mO#k63+$Toz*)2v_2EI&{1(Z2}zZTOt!BbD?TbS^1q|04@3m>cSq3d%&Z} zj9j-nYP2+34R%~ah7_mB72}5wq9_~pMnR77bUB`shlN3!0U;FbXMv?%oobJV zT0DLPFj+jGaZcEt;jPHpgmEc1CqM}a#?ATXJJTk6>rrQMxrLPsT4|xtd4zeTDq%o3 zGo;*p_ZJ=vl_ow39@cHKN)>!eZcwQe;-*;2z*Oz{I6iAS=Nkny@0Uv#FLc8f$HHeALm&(;up^t~z8$_?g zW8xAS>$y*b;4y?mFF1sa0_30?_|EuI;NC+|baWo-i2aTf493y{;Y;)7D}kaVzY(Yv z1KhuPD+Q;q;jSf~)u*vA4? z$mUd*kb-)Tk575Kf;}gVzy1kGlLuDmiBMM204ceDjtw~wqi}j3qSu`K8`u6Bm7;HxO}a5L8t{8Mw~JPZ7@-&sqt1ze3HoAaWzI?vEElk=E2hh zU;bQOS*FlKu^+*FqlytlIrT(*PxQ(MCqYnvA3~FP8KNuU>m{Lpg$G4*-g?3*utb|I z0lsh73hlhS7EcSWN42l#TrVP8r7nU+{B~Y!5Ecn5AiBvp-!yO*`?c{C40f;;YD2S0 zKT0!oL^n;Ep7V^lX38gjpdE4;f=%%62u+#pqkKa^MSrXj(gr6IRZ!hO9pjB|<0-3& z6l-B?Qbj^`?o+K*d~hxGMLDAtTnruyoB@G8xfwZ@iC0CsQQkQMo9O+I2=2I}JS|9K z?;eA0H)%e(l?qUPnhDy?A#=m!KwkUIpeOA5o+jQTJXjnpE=hd#k7?=*v}_G)gd&|z z>o}cp6~MyxZHVI#3j=pS!WZsafJ4w>JQ25C+tjExce74CN%HSf`y%ZaOi{Uwbmf~} z`9TFoS|*@Ws1Zr7%H)1r7k@av&`)z#j?i~pzHZ#rF>=ymppz6uiSXKP!tR&uu#GX} z)FBKrWSs~xDO+QG7$)dQ1UzcslH_%;i*gCu7IeZxES1upCgd_iFTXgiFxV&mSDc&# zx!>V*S%gaEIKptiFL5NS;`eoo-@oga*$pSsG=H$3YI#R$2*MEW68wH{aM9b#@`={bZf_YUjbt7MxEDJ^&4oD^~)eBV*I3e$Oh#M6i_Aow=AJQv!x zKc#VZ4R4hWls+~n4U&?8An!=E(XUzsEFq7uo*C>4_&@8#_-w#kh8m$tG3Ja(K7&|LI|GxxvDVWz zqh`hXhAQ!kjI_MHK5e`tlsO7V`iN3ZvQ0k3@!a9i<$cJ1D1@QSj~mZ;eCboC^~URsGTLw=y95dQ&a#xax2&L2!nhV8Kfa z;(*AW;C%XV{xg9-0|0Prc)HZ%8ZKWHa(|A@c#h`?+syNaZu_(u6_*r@IS|Pxe*AQM z;?s|(DWP}Hgs}Im>o3H8T`ix7%?PBlf%DELwbV)&R5ZZ8_aJUjz;%;G@?NcNAU>fm za8YUKLbU(rN%jT$Z#iZ|Zo(FvwJ?_Qlk83YEsnDHNJhNO&M^ zUM@9r{9c!47a9Cwad_6kpg|<)To>M>7qrnqEDZMbau}ex2@DDf`e`}{ohV&_PVLQj z2_ltE{BNCt{r|c55`z5Alw+pBM6Oai9~+F3L}Z zDF0m+yhyHI`tr{>Gs*O8sn`i!2vBt|?^gPstqw`xLuwE%|92;+c}rhZLjSw1=t(@o zG*}D(a>KCpSA$~bfB9F&fiXYpbnd;Y_y0rMTSGF>IXnFDOZ1;d#rBbF6LJvwVkVnf zROIQghD2+(>{Hu!l`OSG<)z*$Mxf~K6 zU$c9Cn!e&6nY|MJw~siUaDh3dhO0@LXZ1@}gWUSkdv7_>ivDKz3qKXNa^e5o!u5)UtdjS575?rSMkL?wfKf&SS4?LW$s^A+psT(X-FMl)%#zJEmW zr@U3ajWVSe0+OTY7q9NBcE;GByo*p%`%R@E0|E;2MKXIJyW%Az(8sOVc z*g;&mp&Hx_aq0iWL{&bS=}GbuSKjwUAdxxo#damPGi%c`a{|!u?x$RG^*J^^k{NE2|l>0=RTMI)!m;X|wck-y=i181AtdXsS%Mkv(C_L{rN$1~N zy8-^Uycv;mdz8UY#_-yic!eveah%9B*g_pA7~Dcz1|thdtsH7B9770E7TX|UfyumB zaz4$s;nC*z?X$qtNF_P=j2BAAbXLb9?VZ5ZuVLSF&Pautq4r^C<&J(N>Yq(;3#b1o z>?)&Zzdd9N_ce{d@gPcpaTMctfkJ#Wj`1PyHo6kPtMRJ|_RQ5|k=^&v+ zyuMy|y@}gq&vvE4BtM`A^H^_dtz%K93p(?jQBM8Q(M?z$V^?g9e_i(#wIJ}1|Nb4H z5|Ne7{AU(M7;ej4hn>$0rve4=u7y-*R^*dzmC{a zwQfR2(*E>4)W&(E`0>w`Y3rIjM8LeGLY4hGj!{BieJ_`EqAeuOD4^!O#(?|&AK)Bf zs7!)+oV1-A4~?ZeJ2!D?hMVl;yd5=v2d}!}h`YM!TG(i28c@DJTvg~$h(>#4`SJ{KTsMnyT0XtjRdGnp)h8AUALSGSZ~4@qje#$CnnA5cgwEPEg7}$Hfi)70AAnRy(i9^tC6Fe! zP}^yWe-<6*Ur_i`o|t}k=o?1jdyW=(XQy9Ax=0fWXwzH7Om$;thZuV6{nX}%6dR@} zlqK25^_Fp$^boFP;4I^Ap*&s-Eo+0FNU@d^3u9I538{I^CRO^$ zZEYNPZu1-InO`%%@>=8_5>S>NJ(EAfjg7Tnm;9RLZ3dw1h`eRoc`WcVvt3sA;;QTQ zB(O@p`1~A!&L*RpXhv;kkK-TOEEAWIe44aVusc5FkO72hB%kREj_k)ssbVagKV-CT zAji4CKyXsbbt!yG{Ci(0VT%P_&C^^m67I?Ng<%5d)XeW)u+&2bR$2pA>T#t;ij38S0m zTq4YNmXu?{?Pk%Y0v^ol{NB6ZiUKmfg*-Ibv%q z4>~Fzg&50XvrnEfd2j4=icVK3m+bOU&x07>tq6wGJDF!c2@=0#z9&Yo5=fKva1(BI z-7hHW`gJdD+8UI=L}6X4$q~6x8uMRt!b`0>i<+Md^57+>fx88;4Lq8Sia5c|2q_c$ zJ#q=){*^aX-cr7dHC>J~;(iN8R8+LsCd)<)11(fr!;Q7ytDXYnY1@as-d~&7ukIU} zr<^g@>ExWLq+5leOgMAi(r~>~-zaT}z*h1OyT`6_Y|0`zs4q*YM>%2ufyZm0X8*y* zIG-qp-=LGYckGQD=?Qads$3zF2sM_To@LXguc!wN4BINz{+!z%X?ho4&6dZTX%pKm z8WSh~OxEjvHT&fj+Wyt1KZPD4I499ucnSNjWS!{xx;<5h*kWGccjoL9#gDl>_p=ki zZRe~-EUqgvCuq8=;$>VIOZ#he^PbjuUvKgCvzYl(a`x{Koo|UUnY2YuJfog7)6@)K z%l^?u;iXO|pS(xiOP;o)8MWPZuaZZf772W3xdDw8M});&dy>Cb`-O*1dP(&DWbYrk zMq%eN5gG>$HW|Zs_uexNmbA2F0*F)mKya|eu*a94fwU9aae2-|AOOQ zn=L+E9it=(bmUvStuwFZ3I_^E5vE2&fO@2)?kSt=#vz?j%rV?<-G=4Ap9GF}>Fb!Q z*Ob})S3RG}W5NO~PtLsUQwzoYeS$5zYl^z6X9gSv&2&{_ZjxekWjDLKU#lQs6%*rP zTcbYXanCZfr@5ylCT^Aj`D*)MiOj!8zug>8QYRd%S;HnSZ+N7U?+-hrdus z`{!*1?^AmPILYp#{tds*O?@sTztW?13YyF3cjor##kX3sX%O3EYmNF{O;to3v0Ent z6^GecE(gTnwC{Xi`Ded|A~|IoGk*(Onjiky%C2}87k}g@df?Yi6X@pR7T)gasrSY1 zL+r^IA7BjqSnFi114X!=+v@PxHA9cjA9{}ZKNo-2y9t;`o&1wt;lPxdG`%UXR=yVV zJqDvGF#(&UtFC;5MEmBb#o7U!NTtOQW^uNqmAW+uv~Y&e#`wz%J}((lV!f` zi({)GoN3%@YHXF#{zG8DtlP&zV$?wW%f z8sUPL;`Kv+kF%?hRAK@bD%TIL$mHe7ZbkUBU8ku3C37fjeB|}oC%z;S0SGTn7I5d; zlu&X7Sq#9YS%$}l{J4S^p!1wi;r))pzkWy5E6~(p8O@Tozka$xGGAY-wjIO?kLOWP zsBcKaKE|Le1;2jxFgrGLc}x6LvTM(FPH7aS5_hdJB`~`pw!^cD(>C&8V1{eBb@$WJ z4<9MVsmA!KW00R?PWA;)my?(@Cq*qbPwi9_IEP&ll5J+??x@j;e@-;PT6aenjFYfI zwnrq3ULAg2m0^AyTqNHZM#saCLGOpFntr6uH%34!+Ocvmu2`}d^h8Vev#0t-$aSnY ze8;K0W`5y@$%bTC`#!sJK2`xPv(cmRRNH&`jQRVOq!}@gpImK?;opkMqg4Ck6XN$1 z7R`z7#FgTqFb=P@=gO4rv=Z`SrR&z*viV>fxEGKcgCGUpnXj z=N*{DrnnV8&wNT=I91zBjIn%-#)$lVHvkt@EKJPz{+yrxx`~^`PvA_OU7zs- zpNU7_Uq5h+$0{M?nycy3JC8Hez8C?0&FHtg->SEiS?jilX}_l4={h^!W|ez?8lA#8 zNhM*CKFYa89CMwy3=-OkayGi-dmHE8U!#(8auY6bMJqgh2zA2pO`8SiRBv1&c^uIo zbhCR6FI2}vs~lOtb0BF)gFgW*xi3uog}NhIfLXN#+I@VfgsEWuTIwXWbHzsVdQV_i zW6G<(O>8t3*`a4SbD^st1^1_sR%?yoBiwRxg8yNi+S`h5{r#moMtvSjThDNCTS?my z@yH#gF0alVKAjZ}!YUr8+l*zzkHhdR{??o_c)IU48J%XQO;PDjx-h{Ym8_K{)$#u5=nTdFu*KJH?u41h$sfZEs!{uuPYpHmq zW5&6tBV_dzIbLh~i5f(4R^`mU8m{_IogFjsHB zkM9}FkvtwWCQ^epLemws9x3$;bD0MOs6FCc`l+Q&Kd*XjP5${ShzSce?GX(J@0!AH zU3AAaB>vX)KUbBrruDv6^-ZMr9y_~W$^7x;J1~l97+Em~q$x8#m>*3o4qQ3vw+ID& zt+|YTq9pj&Cp!diSH2AVIv>AH)gM*gf^x{@Nb-(z!&>UX?7|PrgD)omYvEw8SJ4$b z2PJzmhZX%ud6tvRR9x@(FUovjYddH@-bw4FoqBEe(q%jo5auoBF1h70=r671ZhUrq z2}!7xV?WKh@-^6!=P*qi5_V%W*?zv}&rhmQQ2LfG@(Nn3`#P}bA?tiYW`bjLLJ^yp zRyq*;4$0owW)TA`O!c-y$aH08;GvE^x3*5-k9uXnSmdC{&oujm_HFla?FC1VEdlbT zT4pOPe0FKr(31s?4wt&3jEmcTO&?9v6n7_Utxr{|KVZk+SLUmf!Q3|p z8M95r>78tptpj9vwn(9D;Qb!-=$g4UgWNv0R!sCqRk@O8xA@m-u_ z@E+|#@qjqEnP zjRf-LUt_P}w3kMfBv_lfKx5{%EtYu*zBH&E>ZAH@{YQS8xn`&6Op($QIVM@wzV(B{ zifL{rMczi?9#bZgnlz@qltvEI%ae_U2dSd>9nH#S*8oRW^_$9cgRvD~`6-c;8f^?b z{0{@z*v!nJag75X0%Xv(jJ5JpFy_-r_l{*T|M2}sCgnSaDl1C@)0~!xnJc}YP1$CC zM{Pg8u2NC~0N{Gdn`6@uYULhugl){UvH6TrFk zR(m`24F1I%{!80Tj58a=Zk@#mnoK{;BouQ-1-E81Yg6@3!U;B8*n%sW3 zc$n_goQV(1MEQsuyh_SAPA;?J{azLfh_!puz2RxIbTqE^SeP(b+1g;VW+PP^K+X0z}oy0!n9N$c}jhA;&l z{)Ep-e$oBPa}(X zzT?koQF~&B3ZL|ZC$hqU38(4Sd~-jz$$NHib@F0+|4zxxt5O4Z-4d`{rWV8(hLP1v^_M_8{Ud zavNhNf8Lb{fvTw8;ZKF_=OcD<83pKSe!#^RJ+!Gwkt&uGyhJpuO2ny)_4 zGK_{r8$~go)?T>sNkB(b(P++^`7D8*zncj~!L3nvA=SyG*xMd$X}x3ml}}~>U8x*~ zxwY2wqVw!{rbW9lNgTjCUaE33w;iJ9(~IMJ5utU1gkk%M#b8)|%}p?XoR;xG0S}N~ zX|-y?>=SR0ZVDAx$BKa47Ch6SmcynFDm1^pH$tyJ~gUFJ#WJL0Fyur_tp zdoog97FpRbH9dPlo*D&>_KdlT9a2cNORV&n@gD}dqW{Cpdj>WAb&aAZA}UG|6r?L1 z1eD&JA_4-^J19smp?4CB2uSap&=C+2=`|GTy>~(nHK7LxNlyGf&-=d5nfv9=ojY^R z7iPlVd+pWss=vMGCo|rA2%9}%Z_UmqOx_et_aMDR_($*;Ai}<&;P))jrj)xpJbq!0 z3NQ9aD>L-+GXWYX_(@{Qshh@I*M2WB%b=s9&+y_!9I>z*kGXbGm$42pa2PQi;Sv=d z+w&x`>kWPn->r<+&#rgWnktgOU!0-F9jKXhH)&hYqshx}Z1|~ z^#HW(tyq?QxKsvNwG7SNbhs3#6vAGj%CG>`Kc zq!`wV52w^qNTyC#8cx`!+BXwcbWjHSMQl|8KdjEfp#_i#AX;?+XF{Vn;H;@5(wZcpL9 zCB>z|CB#!e4cu!2c2u~6Zh;uC5yp)>+y6%wy=l0h;q}tSuTrHH9jOxk3D79In0XnSnd$Og0%K!U%jRxHhG&41>IYjjT z;fL~el-dGKTQaG}Z0<#Z^9m?{aU0`mIW51m%A>LyZP z=^Q(e+_?>q*X6vL$!n1r4wkR9ggE&L#vw4_;sR9D#t^4w>nf z8(p*Yj!TR){mlO{*;V*JWMgeMVcqz<-=7~+1Lc2Uko`POaN9|XUU3|$iyuyVaQy+( z)63g|uK|9i9S4Qsj>YzIpi1d6p6azlPoov|AJ!^WrstcNxLSVRty*c$mvc3)m#z%k zDPD(l6Hm)_n~L^GOoVnilt;td_i&c-(G>9UBV0R8cDkkt;fKA5BsdfgAec<wS>@XJ>*r;my0^G5Fzq~b9i=TM^;s=MYJ>W8?o4+@ zPp51FG%ShzllFizJ(lmgWOiJNiO4}FzgZUUG~k3#Q2Iv>wbet%d{3IPjx6CO$XQMD zW^Os(Jd60h#^0h0&tsjnGaj>FD070zrU2^)9?g&YMnTu(gji=(RF&$Z=e^`O)1_@z zjBe6xwe{+a{N~KL1!;TCUJN!D=#G55n9?q5p3wCf1mm_J(OH)CdUD3mB zRCbtY6vnl@uXWCD{Mj*boWx^y2gVDQj{f-KPy|^wZ z2>6!WKv-B6$1@FgTT_(lgXweM+u_biDcjMH!F8uO$%X};-OSyK3ziE9LT_$dBF>CM zZ+Dw0CZFF8uhWTgaJe)zbawz$02$S|j(3IR)!Y`-i@5raUD6<8>r>n)d~f-9-ke1T zAA_L7uXlxt^G{~m>HBn%Fk5+gW1Rx8;w2Gz$3!zC2ryiC+5P<}@<(96flJ?R_`v-~ z6!N0$HBr7dghT>gNOk9c9~XVu$A0D-j~CK~-W~<&UruU3$J$1!-Xu8vQ2$--USynd zwq=M8p>Orq^nIsP67+)7*JbJ=Ejm(bO9uE!tqy)n;F|cO#_`~u8n)x69d6&+moI9s zQGNHq3c~4kniBs_@g!_=SuOg#2c7(Vt$VdtwMPfrvrpQB3HD5{dka~oI;>Rxs*Gy6 z9pt8aDQ^%wcnDoNhN333Z)!g6X+YtUxB6M@`)ZcGKd-lTJ>ITBwjMdG)jVmtc~kl& zMXc6k#Tj^04(m3Q*PT5R2610I)mdR!!NP_qMj8nZQKHMnBqs4=jUkf@>f-|FD4T9_ z;7fyMZpJ73@rHYI6W+W172-T$fOy9U3idmo!hIXRVAkdR3cF79>YGFcJ2E$GfYnU00WPD4)ljfp zulk#TADfVePO?Oifq2>s=LQ(mWW@75z*71I3rf2C?+;8DH#0Q7tdjj5RDrsd z+}qJT-F7%qp-ys{{N!8&oOcz4@*ZLK_X3UDH_OJY=WmKl3`@djxRg?2Wt3_%Vkkd!0^5nM~>9xDYU zA+uR@$R2Kvx~>T5<{N@Qp?ZsXhXDo8;)ifD*r^>-r5}8$J)YrN7%3*)FJ>v~@7+IU zIht)w_pylT(fORZ_0DgnwH)QnQ+W%g;t!wspD*0rtnsMdFeV;aq7=B-ihE+__ zm(yAl_X|VP(Z^T`HKRBY4{JWym7zjod1p~*7CF#}6GmIoX(oBipYMha(@#bRwomiI zfUieS5&Q9ordNgbE+zCA3Ev0HTWa)2$5%4cPY?hi-O6(;3~a+h$f`B^0M{J)hu={N zrw89_Q01&MgfLA?N>l#VmAZDQ8U-5t7zlssQBSzH719YKrb??j;XZhoC;ho?#bB<{Qm^1mVuMaplFqY>KO!0KflC92m^Kpn577xaZBsv-zq>!nzUehU z)F!z6J~_(*2q=Zu-#-L-enYk?uY;Re5Bo|`y+vqVkm=X3%|e<9Zpt{a`;uqi8@rJ( z->q$djg<6cKX;w|o`FR1$)x#cRdIqyQD9M0T=lc<{zZa-z z=aKB!Jl5<0*>>Au6ZFiXrkHwzxMg)MlP!=z~%aon%y>=L6>7lCR>SWi+yO+tCjUM?IA z56GVHW@Rkm4BDQX+~rSIG_tnsisojq`gVi!>D{`vLs~0$cYR)y&1n;`&TqGBz>Q#k zN%%nXWQ6YFk?9i7$m$;6d%1J~b(h&+8CWSxVuI*F+yd;Z!43ZNdxnr?dvQ`KS6QPN zokAx04G8C)_eUB@k&40akxH2FHsEagsZ#&_lbn$eBA>?SA>zCT=Uvd&!TPHQ-y_jn z=>?0~S9y5L=r)H#Pvg|F>e}iD?pC<3P~T67CtWO3*o?+4#B9u}4js-9s23d<_Cl{~ zkBSt%VHa+yzCX24&CqH9DEeaJ$0nDmWFG9}f+=>`bw9SDKVQ>nB#hGKDd5p2%qAZB z9`FzWkF^Lb@E;d2<6J8PU$>ARcdXzWn8hy%O|InSx~M}N1_>J`2K-R#!?@Ba$%Q>E z^+M05Ym=#mHwAERid9-EYQ-?nt4c3Zed3+Q`0_-9FilfTg=?v`*t*kowrHj5r{iVP zlv0;K%(TTq4Ay7a)2X+k-M7BjcnWzG@MnI5m+@9_F0uqN@6X-0;7aUJs61zyw|2*Ad9 zytC?3SI|bWVYM!i96E<7QkUhI$+IV7^}irOVzs(UPI5Wgb)6d#vFh z1J)r>B_8@L!3S6=LS~3=>j5nnWcoKQEK~Nx&{9EjOB}|x+xH<9K#|3axI8~(%o#nm zGF#F16B|i{bgFUAGSK8^F5)qD(OkXNmC1GPSed8&w=C2drl78;hm0hkJ{H<4t$7{PxZtBe8%{s5>{9N@~hJ1ECkRz>n z1hmD7B=jsw?{8S|UV`eN(uEU^J-qcyk@p`Rv@Xt}YTSKV7Tgm~+EPtVmZmv8NmMv# zW+!G5I+ba$)ou~k2*Sy47aY?V*B&~(I>Ar%is6o1$_s#)cx&oghU zOc(ZWkcAq%Hz=jN@er4<>Z1o=9LgasD~XcrA3Z!g3{p^0{l&e!w=Z}hQ~j+@epImU zgSgocXq*+{p5Wq1`!NO6te(MoVU0v})tH8IMrCKB_Bla}IayU2UpJDX0f(TS{F^C}#kq(L#Itm3)i5f|O(5dwHJ@hE@hmIay?u@6fI!U>TQ`F2 zGkuH!rXBw(_){F^&5zBp>7LEn03xH$e#2zm!LS3AuF~21E~f;}l_~H^T)5NKLfOVm zEy;{2q|q3rG=Rr0xYjpb->>h03ptBGecx&l;l$CDMyXKvJ*> zh6r#o|9P8oq+%(}#T_26J<{t*1T@V@w(=`(uYsoJ6{73`jw5SqBa{34S0-c&(soJp z4&=?{wG5dAYfU)2l{z#A|zF8)8qh$P&X(usraL0Yd2Ey5BUzEXr}W&4@0$# z|B2K#KBBg`Ob0N2s{Oc>F#DPJ*J!!i#i)Lv&@JaSlKtm@`Cw=ZqA z)X_S>Fae`+w{iZOgF(iu@3U>yS?}nHSH)dA5G;VsmA$cP#(X|bOJCC`G2$yD8BKGZ z=Q5s3LBCHK*QrWW%ojsimsKXXK!5zEXh#=7Si&px;=FW=c&km*{9!Wq>IGnlErf}% zyOp1H*rm;7F=HN}!qsN+zVoa4yRAd9g6#9YTnED2{Mpt{25iN(G9bOpXU7_<@rLtg zoo`|HRY8|yb?ChmlA?158mkiit>v_<`nF`@z06BR;ldEa^3HKdJ}V4#>D!-F+>48z zGn02n&zroe0lPvVsL%+!ry+>X!g7rc0Q0oly+oIMd+tqpwf_8Vho^C(2!^^qpL_v3 zsex#Ku-p+}51sUTWlD#}9%K;&P*|Ya+3EpE$I1k*2Jr^*CK|E=c`XY2%LAPQh?%FE z`6sTN+m9!fAI%QCJHj@xS;itA?WfTeyxCwN_?5W$>Bzy*O}5p_^<2w?Mv!WIK0kID zoe@*bJ3K8RwSOSQbXg0a0WNL(*|ouMaU^#F#dlIMLCzz3Tb@WlAp{YDKrn-yQ^Ozb zQNm*RVAQR4u-jWCwx!p3%kCSd|MynRou1JCQ3=t0K>cC`XjX(V9CHzvlgM781Kf3- zy@~nIaqDv1O71b!m*LU5dfj>Yn@)+g7_*JeRmju0t&mM|nzDxy-=9~|E-V!U%y})n zz9M*oWuA}sbt@U?YfgZeS8ZiPMWc9hTlv*MtTCMgvZ?4~n=2348{JZ9!+6F))+ibY zbkwg|aNf80uI7l^5S@RDB)a|7MU#7=QVOD+u~Ob>t0%D!C6ep)fOxj+HKdW!}@ zUX1q$0*aSGf6yoAd%c#~{GDAO5A18pajcuCniMQ+u@#&~S^gX1T1G5Svr|bEEn3`S zow=mec5=MEYGD8C9zKK37ZK;d>g%Z;A)&Lbb~N7DQ?VS>>u=L?jB#%Em;(IEukMTkhM%sf2b~Bt zf|$WX^r1M`D#d^j^F<{hSyJMS`-UD%^cvg1e6v&)ExVZ`n6b5uMbexEF<=1$S(jR` z1L(_IEF1Vl;}W;gDBL#i9xEhiSAVdV#Hq3xzP7naf3p9spXr?Pi%U*40#gZ}mPb4D6l{^ad_4+z zdp_P+84fqGzwruo#dsAW&((oXGJJfyjvij8D4cx`giX!Bbex2_`RJv)F*onW5;3Lm zVjkCn-wc?*)GGZJU!9V(#vd_;>5WqDm2geNW~k5|aEa}#>AGO~%=wWF+PuupUdz6* zQE7*OPQ%=zGm>ps%DhQtnZMZ#n3MQOZDx`FW^N?HZdg0VJFR)Kid zl9W(1*M<0Yn_b?$p&N&(c@n)EzY9Z+nKo46FV}9%)_uCuBj9gi&`U~g;zXVdVOme& zbopemSC2Kh!W=(uCRE?vj+AJ&H5FL+8k(XvL+>uE_cl9@zRgyUj8bkzSlfT`PEzso zxhYs~LgZ?crxyKdP~VEeh=)wFNKurY`u&G}UeNhnhuG^FVHoN`Dhl!nrgfqd8s>ca zIg?z?WzLuR3}J$^Fx!fIjN2{KmtCx|EdFVR%teDGaqkXeus(1Rkyy7%;h0u&byLBZ z*kUC$6`Rz{1AOI#DJ<@VH4<%sxel=zdDmye==tKE<>?v1wqHwr;U}lfs=QF$z7g-i zb|<}918T>62@A>Qcgytb*40beniM>Bj2+5$7vKo+3)EqMPG64e5?D8U2_Rp zbXDc!e!znA-DL85sp6PN>Ni_S5%sz4-mG>fcT?f()Ih+kYVnl*igh=o^?hIW$pQ?` z*G)`e9h+r|X$xG9jHS&h_z=S~)6qXAj3kub@8530t~W5CQRh0ae5e6aaF}oLkG$Z% ze2b-hE_*&|?4{0OO`y`jZ9F((z>MoL>_pW2JK8ZcxcxY#ViIF>CA>)Y+GlU21om(l z!WMI7nbB&Md63-?dkGWyY)9LhMNROfkd$s*IJ@m&)IXKXOCG;@x1hNo+Owg+#+`3t*ZCP= zRFH%r!w%^MWtF*azvUxz2FCp^4a6;_xO1~1^ z30m2oOYV-`ksqb^P=)o-iB>!zTnbLuoo6c83YlKA7^Mjk5b;U%jA3qlSNPuMPJ0Td z&SBJyKBLgv#!)7PZWK}u=Pg~+fS1hL9pP7gW6tCceGIJm<3DWJvu=><;%xnv)WQ9T z32ggQC_Dl&HbUh!LSkyv^+o%hnBZ&1xCwSFEpWj|P9DE9*8KkGL@NO$o3}0a1P2k) zi$L+sY{9o}ez!R9AFq!lV^J&LF-`WeJB_dJ9N+L!Jk{f6_^EgO%77fw2d5-BZZ;lK zfYiAJeFQvh6;xsJ8DPjPN*C1&tUVt%z5Q2A8^xl2SR3218E(LK5J~CG32&X#$J(ck zYd68ASy+m-7?72ZZ_-9#V>)569k!;nUdMp`dctRe<$H|dBYRghNW~uJE|D~vWur~P za-K<6msp%F#JFVzxX-rzoa1*ZFWAw~6D{6j;hR+-sp;=!xH@mwcAkNr2fx%raX}i@ zkh9rvh41>C=G9rFzS0Lj)s9UsL@irwY3G6$urBOPHo6RMD>u0uFJhOFF2AlbKWWk; zKTKleE<)@I)hp5$udwB)jGs&83~A~|mjmt#1Sy0@q(}s>)2Pp~fRrX?M&(vZ*lXH% zWO_N@Vl^pCb2;wf=&oC3L*lpLgD$xyeaj9#RdgW!)yU_drkE$!NDza*l%W;q#+CyC zaJ`t;UXFR{f7Xo79_3zU3e#4*!}c|vBBWV#P3E~c0-s#z8q%P-Z8kz){D~ zx$*pp!NIksw5{1RRA}V5Gm$cKu{v#$jg!(O|B;q*S(^Pir6JuY##-e&^Q=S(Qpi#w z21OR*=KwS>L}&N!dBV`+r)`hm$eMHC{#Sh3z*&pZ>?#VjR?)6|w7MKepC09EZa7pp zb^hZB+tFDu!69%i6JbGvuY-hxi0mIaoXoCM8DANY4Q&hCREx@(afc>)b>=B#{R;Xn z4R%L(hF4g8qpMJe@%Xcr2SNQw&fupfFw;^WWHR2f^Ii#ZHJ}}EL%2D!vx4d;RmSs5 z{$uF_7do!b71TvHH$27xdx+kPE9c^kk?yC_l(1-xi|Ar)K6^W%Q?!TfM?6foaZ^89 zO&GaoxY}kD*VvFuHxZmYXBen@nVbMOp0|%pzPyQU=fGDL<`bzz9dDs%Pocsbo0qjr z6vw$|VcBljxB7-Bz-B|QYr=gt8d{GGuWL5B{ekX3?`tgfRyfkhY}_&`rJbblYUrIn z#b1V>!!jVEE?n~qwkAFod4!eJ`%hJ)-X}hy(;vn>JvF&5J$zw5n^AUvBAUOp-L>E8>hDtsRu9@EDqU0ey@#Gly;3`!QIVa0GPG4m_5HzoARWCIL})ie~(C_nB#4${I5BJtoVD9o}#>A@0ve(D)O`CNtos{1{F}^ z+FFsFU&w#*oSDC87_^c2py1`$1nT$aR_$0P)iW_rQ(e2oxx`<-36}A$vvv85T9OzBpK%H^eKheo5W<*taJJ*4ZZU!BoRw8`b1-9p#zx3w~WtPvLB%l&yB$zMN#|F5AVwto=~%&84P-2(mNLMCmI zm6x9vB9&1AsODEE3a0wVh33$9bbu+|U!-jcQTO>cKa@{Tv^afkj>MaAhL*M3 z&`q=h-QG>OM1ogd><%yfBG7xEDhy&5oFJn3ICGg8PRJwm&u-Xh^z}-9ltuBYXUh|G zcv~}aY5uk9zPmxOEH$ z9~Jl0RC?S1;mW6VJwz^gY6qcir`$Q4(M92=asBo zJr)NXd&=VFFx`R^`UO@lk$7?Z5M)GFOF;R>{y;5m;pJ@n9l6y2tsm2s?x>QmbUnZx z#mO$4qm60t@WiP8ViADySH9#Pb#3|dONK0E8nyV-l!oT2Z6)Siuk|35&`;wIff zqMqU`I-bL=C8$d~nw#)HA>2aN(>g%_IOWHse*X;qo%269{9gkg*%Ha%eY(zW;uLV?pDynj~U7GgX3v0Y%u zGVpDg#?3jC@L>11I~D$WMGdixAWHK8eP)8DzVyv8$-SVb=HKDalg@4LOH4A0_Qf)?WA<#hee)F&jo`lsUvl-7fVjO%9o3SX@q zgxd1sJngGyYdkzRs{dwU+xi#%!=J)!Hm?@d_hLp;6Fz6RT6Vq}Yhl0(z~25ZRH30u zsH;*IL*hZ3__a&N-0RlLrHbT1v@tvG2tNKV$pINO=B=sE<_LMdxmPXO&8ju|@=cxL z;~hb9s{e1bOTJ>j+;ndTOPY7hN;zv9QJw7);;lk%{Wp=^jZe<;=?!a+N{@+n7!DD? z@$klu|7ZPE$}a9@*A`DXEreqnhKC3bkB;P@6M@?R+l`l1wJnQdGM>S|B#o0B10G%u z`Zf+cPTZH|+ilpT|<@tb5inU%L)` z??zlU)8dVF;32`G-1_p;hZ50$R992w-c`<}Sh&BY<$b6azEA(G^H_p6%;mced_nn~ zn(E^@t=~FEO4E`-I_*vWR_0*H>qtlp*%6mZqZ^AzW_6L80cM;A!}AEw?NvJ(miw#d zb7!@wj!@c{>@-b^9qDZHElW{~Ta4>Fe-X*p@oyl-Eh{27B{2s)-(HONAH)?KJ@lPy zxD~NIdgS?Rk*Uoc>#ru|zWhiS_ed#-fpwKZfO@(a069JHWx!PybKC<^*d#Ux7G190 zEbw4{zC^m~Sl#C^hQuV}zIoXV9nQpu;%s+L)ZclX=|v^HKz*ya{t>M`CdEd5%(gRn zC5uoI3&r_Ezk4`=FWhUunyz_&Ui+;As9Y#X%&Itt5d<F4#+pa0@qz2rcC7vevuW%Gw&l>rwLNZrA?0f##{lTB~*`x;&A zYaUcPR*W^pS?d28QWZ?S=Ji4hJb%Ujm70WF*dYD`JeD^MQL(Zt&0+c0par5lvGN~F zvEZYf`>cm3Ph0>(g8RP=4en1H9?l7q{tf=goTOKY+R=26X}j3JGv5!Qypz0vHqHe7 z8z-Erb^`Tw8KxxZ>S)i8QB=%c9LCK#p8t)3a6BUPlht*D)IaNk(7hEWU_tT4FL=K| z27_l)v}U6ZQ!q1YfA5@Q*T`TobSE$(NDC-SA%PeXCw}oxvtW4l#vPSw zT6gIF>ZMpF*)tB_6%THl%*6f)C)`|+Op}^-6f>XAtFMMYkqlUGl$3Cn)GJ>DF#U21 zVDLUIaw=Q7EH`+ApymbWLjA?y3q;ddnbRcp`r(mBg*sFuYa z;pCJ>^_O8{)hJvkPe?)0-9Nwk`^`_Hk2EjqVnEA3;`=8a=U386=lR;-nW)GtiWE{% zmya5^ee%{wc9QIQ;C-@OW<{UE=q~_Q5w+yxMcRtNyXUf|{gppT_^INupFHq>aN?9h zyqsb=+V)VVjD6jpmXky5L$SK=XwnzQ>T^9={Q!g4j4b`#zeBdBS<8aSLu6_>YYl5u z^X1mwmUrlBXxsg6eosp>Q98M9M&Ytxc$b2UB2kz3#xJdV?Y&x9tIfb%uxJRla^7P% z;2}-}+{|!_!MjrfS5WkE)?P-^c+ib-RW%y1+bQTb$g!#&J{1I9<6g2XS5h3be)ikm z+qg>@f{=*SoSL(i3705LWb1w#$&mWmb?Q^dhi_96EfHGigOovKyFK03lbw1R&jYDA zQ*CJ2{ZySY>ezHYtabwKT6lz@Va(G+A!=@G#AhF41pJs}D7$o;-P=S)^cr)}NeD%i zs=3^#X!hFAdHQ+k1de}s_gUU4*0D2SXtAQ9=I3wtY3*Y4cFKaiP$V;x+!;Bc#S))4 zt*TFPVW)}*DsD-mQ7-u@<+P;GSvkWjw|=vKXZz5|8fK~V+ua0rF9Qkh?Pr8|c(^Fv z?+wfif~>tU1^`X8)Fq2duanFPxh#JBi^?n4ACSF0X!Ne;PuMVW>`91Tq(>cP{7&yt zzFVM1yQGk>v6)qRZ8J$Dqk zmp$HMF83Knu58mudJkQ1_5C2)C4LU1)|2S9d{jY+cIo%)p2N`U*S8^{i_QD^zRgrm zPqjXbQLrqdBp#UU!~l)4(@k_)J?K$h=-Jh`V6y9gT#jR1ViksLY6RjpzD)C9Ll88w z3oPiF_Kj{iMwdR{U1rD_IamirAHmMeug`rvFw;7rXMC&Y09{NSs!uS5 zI78Z@C6iTtlqmS-dpZFE6fAVompOt{`M|>d`Ie8fEhT^@p;py@*ZgKEj%Fpze$hMIKC#~nuSaGbXnwY zt2F;9d)F8#6>EHXA`H2E*Ba<|AyqSWXjy4|nJp!Ae%4ZkV(x87ovcMYpa!i^T9kRx z896OSa=jz@tVi64tp&uP>=@F-U3>a&J#EI%y3JgA5dWPr`9=f42Ra9N_J)KuEJf~w zOAv1*OQ5zi5_7OO>s=IzAL)_#Ev9Aal0N82*^++Z6q@pq#6h70Y0NdUY@MkqyJh#g zqcu@B1w#qCII3CNpz%3V)nbD1(3Fk(fy;;};tQ_EYUAU$TO>(PJg^t3M3EoI+|+a6 z_*7okpFF>UKKb=amRw)0Uncn{$XmDCyYK!({yqkwTe0kV_N_l3kj6K%=Ps?n(0A2^ z_DDf}QNLnh@Td8q{^*0s#1Cz&^sNnwK!$$aVUDsE+k-JfUIsNyEp>}wJ7~qrk_wO` zW}3;TxKY1u%0$yOVof5 zC->=_Jbd11{HzDvGB~xC-8E+brchI!_noIG?BKU@qdtE4Wk1*!qhtR*Jj(CjdXOnN z__Dg`;is$-^u04zTAqr9BwdeRzg*DAa&AB^ml8XI!TS$;754FcdKIEa52Tn`Doo94 zAS6V1JoJBaLN`@p{*dq1$QuB*#>eYjOm79aM2FMMvx27-$>j6rkP&`w2+*oE>i3F(Pd`l^KMq7Qp1kDvT_^pWGn|0mumxxpGx zH!2^fPms%qHuh+TGP10B4V5eE%{og5n&WKhmF==>w7GMn45ud_yy55<4O>+DEB49Z z69M_{WNoPbg`wUwVhmrF4udhPsOLier)$7XKeEtmAW$ogoaX$WXIcL_h_ z35tm8x9aqa{-KiF`n}uTw#33TKec1t9}1a&90e+l_pkYceAq?=-whyAhxX$oa zGA6EKcw)XVzxFXXF_qrUge%z2`TkyQwXnmPu7{)V%Z1bQRPRmwXqNxV{7&IhcHJd= zC5!+Fca6wYMM~3~m5-L!{;XTU@&8y5H|>k{nWE>L&7w`A+Nbb67(GJQ@moGs+PDMLoihyc(@QO{Pcd?|aK4q9+^=??s< zd2ac=U7q5$7tl0G?kU+1#fbZ@KLG=oWd37mXGl@J03lp;#Xu#+6%v`Jnw-dl64Eah z*&Yvqy}UzUZN~IQ!Kx_g5FD-XuIA^sz5m`M!vn5;4KpUIsArkkRG|V8`~8)l@!!{; zl^{(dMfJ)$oNsD2I?&j-h*w33s)@bAW;J^e^^aD+GY(Wg{3*D`Rn`lSNvn?26lbG% z#9s>Plzl8sA|=G5iy>MC-Xb5#$|=YSaMTX^`ZQdsV6YKf^gdzuc!mT*$Hlq-l%7?P z^t9%QC;vgo!d=S25o+6_i-H;Td+>#lkJDZ5|D+gr37sY9%+!vm3Vt ze((=w$ihLM^XUK`2Rh+`(G29$Skf{gWaFw=|sfc6E0?8I-MtOAZ?y`BOoT}QAr;m5s_db8^GzTjS_Y1y1q-WqD(yd`Y zD)XAZUaDzc@mYG$VFRKpo}8#%ifKvJHo}})DZg=B0Ax2Vvbe7xnTJJKA5n#VPN4Tw zXCVFc30^a_$?tmkc!IWVxVkUreVQbHWd&t{cVd2iPEec_DW+*jJkONhTC-1V=i1RE zIypyhyu_TDFx;v#JUO4m`f8r-X;Lpgb_|>X_D@Z;D(djy0uF&rFP&*k0abJ_OAUOP zJ-0AM&5KbRE2*S!bc1zRBBpHk*O@A5mOcF}A1lh#vX229N#VsoypK{sIMDv=umF17 zBIENJ@>L)MjtoSW2LdaFy>Dz3V=P^*hV)wCmwI@ct# z_7HF6))Glgdc_vd<_rCIFVNLBE+Hd&^f2|c%SX3V`oa@-D9_JQ<7QY`Q1Y^{@5$w@ zXSg`Q?CDRU3Hl}IhphR!W4*Ped~bQdZ&#h=Gbnw1 zo)@c_$$g?16^vt5@2@QgRtDAzU^)~M-~KI)O@+-?$I^16MYp@k{F8G5MK9{_ zW<1|m?&pEbrv9{Xm`^4&DTI8royPFI9Am*l`jRcRcExCz%H^-lu1A zy`E=nIfQriATm)8a` zE6o?*?-sO%Wq-$ur%)t2ByhMZ`SOmAt>L@8O0Sv3nO^72lRIP*->oE=M#)4J$+CI# zp%IjNHo(U<(S6`>(f1|x*?8~mP2_yUnZd#rM;Ryvlgk5@(+c;wTwrTN-uNn@>XEB zu{|Vpzef1Pt;|{u-D8tpKdVqBW(@u^k%MU2DRpg9I?T2O20ly-H6;tJ=@TC_v7Pgw zR3%?YC}E(i2-DmsnEUnnuEV4LSmJn&#{!Azv5~{~qBUqVVa)e5NC}kQ*zn;ye|Jls zA4^Vh=8S6>#Wx=n{w$IksXN$3@#6>Fpt!n^HH5Km*Qb$3Gt`W3|26po!7m*JSMQAPT0RjF6ympehnVSRVv*rEPOhxv!1pu_xPn9ELk?sIlQD^9xaZy)6{sx0n1%(Ke9 zyK_c7Mf~vM!QkRbM)j+ntVbWB$XWLKN}4)+Q~qUjNwo)!{S6_$tWpeOMis zK55M5N59WpfVxPJuf+Fw&xzA*Jn2Ji?cX}$j8vi~XBT+w&RSYsGh*9`Z%D_A_g0nj z3r@sEwCXI?45Mi{vOYYilyc>KJGsK@TX%m$e);HEfSZjUSx;YIkgA$A>#LUYRP`iU znfe42q*M(^c&Cx5m_2@-cYqAV)}$tB80Ms&&o?{zg6tMa20x zMAp!7@Z`4*NtAi@65z^TGM*b8Z%*hLLzn(|R`~V8Uk<%m1@5vN`K|u=s8T7j!ojkB zt2J0lk$R*j?=6FF?Dq2c*PxjTlDCef*uuw!vAa)i=+GTur7{vinw+nOh!8Dje%mR4nq3 zBE#KN^cA_wm8NpuR9mJEp=b(*u;4{nBYPw=U4#qfC3IJ_7=KtKliyO+_;_~_!_ym5 zJZ{Ddc;z_EbLF!iceCe{o>w4rE4zfiCjKGG6!W+JVv;>k*N>HbjOBklmP9+6eR12^ zXeW|NKJr1~WfbIp0Xjj&zE#O0mBHh}y&u}LOs3t9`f=f^?g1Q%3|3U~ZXx8dNO^>k zsdF|Wp_)07B=jRgQ0O4IxrUWhgpv%I7K3bRFR(ZZtw@hlpj zELhGO+)Z2RID;FfiJ*~T%H#lym1^v99Gz9H{^xQ{FKs{{y~qjXLq@jIh2fYgL8I!a zLy=0;x*8NT=*LH@djz1^!nCvAb;^M&h~RONiZ~T#LlI0B+r$tuB4kRGJY4Cjv>P_o zih>w6D~=A7xvHI}opr2{rFAm>_;6XM0!1X|-3^?Br_^Nl+-?L;SdSachYaXMwj=k? z_}~2}3JPUieoQo6nvIwCjtk*pS-uO8E8udur;i`KXyIf2E{G37Kk(t=iVa7L)*j8P zGMY_R=~>_zV@f1Rsm3w0TjcB!nO6)>*C^y_1w_U`ivafvynyT*HZz*m${nS4!`51Y zz{6(8l1kcLVQ<_}hjbV`DT>z7&I((@ni^%aGH6@`siw?(>a9mjxRStFQ9Dqt?H7#P zU%#pf<@CCN!Jc>co@RcyrwyI-2hnoSV9z#(&Wx1>$t3r0s^v@!E+>PAg~%dRq99NK zv@{)jt+Q!MZSc4V6k%YZ)HV<&J@4+NzJZwF;tU&Vmue{@7=^2nG{KifqMaGGHK1=1@w)<;JAO_`58?#be{~!e91g zJ!^Nl-^t1sz$7LpuOpEUx8H~YDLBM$d_X`X!i&2M8jKCr(4WCo|i_D_uIpBz0QvwKWPpuYldgU7{W8W)HtBpf%L3O(I}$Tf{VADFbZZ)J1AH;2*u?XQ|~hG zYh*1_Up{%b0xhI;GS131ZK`IQn9yla@^F=_+|jVET5zybkX+MKM+MU8TejmSb*e6K zY#2ID(~f$3o&_UuO@=OLs1!OhtgdoZJ5aEqN=HjVc$d($p~h8a4;mXDkP*n57{lf| zw#kJPxOQl0KrbC{=W3Uhi@bpGV$k?-b&}Rq?=o(z;hp^Xhc5*8`;G8hp67VMV09dR z3qrlwo>gAEJ-c|xmLmqclOm+%IcPb{3u-wX6Re934!mu`pt0#m8l~Ld z%#-XAo=h7P@SxFQ>K*}TgZ0bHnI<=-A%jLo zC}LE+n@2)LXF!>xp%fHrb{jU=v-K`nlrm&e6hRYAgA19P($R`|4ME|K0*hfs11cIx zllrNVa;blv)S_Nwc&Tx}fdugZyH)e%{yhVIZbva;hAwl5Cs`U(F8dvJ@K0186(mX0@Vt%#NXuw{7;v5zks1z zgHX`=F(L98HPgg4uBs43DcQ$K661NPsvg6_E{~!0GTzMAuc;9o95NIYtOy*Umq#fD zUL;6qkQ`T)!*bZj)j2T*g|mV(RU6Pdh*A?M9Tzz1{edmlY(3b7_J*V68d+9BXa@(& zV^k8Vyg+YwmIw!hAfA_I!hj|o#H zs@-+&#x>O_f z=e+k|T1>dVivzS5%y#FNtl||1a`LMhbINNukNZ&=z;K*{VIp+eD~F_y$wS zF`ZIqbc(?;!9s?pi2x;y+uWjF#?)Eslmz1-WH^VTLYv)Okb*FYL3)`Mm+oCwgPZGZ zMwLQl=LrQm;W7p5uqlc9S{J5O$XsNE2BTzz)nu(#DoGRU`A;MKfcq~V zn@<#08Z4A2nWHK4?oU_pf@pU$&s;SwF(R1d(ceUxB(EBlNw)8Spn;1+US-|E;;J3l zrJGL{mev_nav93^(M957ICK&9j*lGEE1`c_dc|uxGs;X273C0Kt=~jqO#Fd+$qH2m-1o;`5? z#bVj+6SsZxE9sp)n7}w*z@?$3MHHBMhk~-Y+}Drp_sQ)up@BEh48wBOM(fG4n%@rR z{j}>CD+r#c18~oj)0j>riwFw1G_(8UeyO9=6a8IS;G)vS-7laccUqM>+t|=VoN*F% z0~?VdFFWjxA4u002x_fBDP+b|?wH=v!l(?|RE-?P>Z0V;S!f-3`EfS18}70bZYvv- zD6c&3j_WV4&gOKX7)DXn8h2DLdD#(nOa|R}N|38WrH-yIa>n&lmK{PnLc7BSmJN!P zSDtXk_m@|mWc9H)#}T$hHln+{?5I1wFI`_KDD@&Gqm314JMxl4?zrBRu~H-mF(8bt z&hZrEC=-@UI8B7x#)c#-Di6E54WesugaDNI6STQn`XVZSOs1z8a#5kD8;aa^m<)UD!xQB-3eWHcQu^AC8@#hsV=D4?puEb~>>VMQPan z)8CU;6g{);U}wH_`b*w*q)r2v$8Vk$Jqx7|MZ(j(0PRr}n!krhAD{7H^FBPo`l8Z$ z>nfx&dS(>z&hc}4f6`XPGv)Md7d`*;P95ke)gK?6U4EYUXCs6AFDk1K z=gxfPn?SYV%nx)Lbl#k9_U3T|-kCR@Bp~Sr{soJ}wd!EbiUT>twM~W9jp(kE(ImS0 z+$_iOe7~gVF})JUXLiqsjfl_(_`9%xuju0L2igL-|Ds}cxTd|dfAS><&r4`lDmDQmPk)C!r5A~{}gFl27}G%WH^!}q;AawLE=QW zu)t{p)8>xq8?Mu6RSJKX7HAXTexTKW`~ASL`?GHUcm*f$Bq2SNoeX>9m;oQ$I>X=P z0y-6?b%qrOvbUfhpsuOB-b4^m&r(4UT`b4)!uZVO%lh;f+#@zMDlAy5@^@i@)&TAY zS`E1051{AoCC@HCTwE!qXh9Uv53uFA+j_@G`ny~}C&cAu4;EJ*EUws*Q?~VFvBS-( z;y%Gi_#tcu3_jh4||3;^Ezla&P)>@7lNd*Z;xiE+rPD-O2 z;MAF`1>El!e%f{7){mFd6lHU|uN&3xv%6;byF@@I$Kr71S2gS_EdOrD@e^e=(l`^I z7AKq_p#MY$288O=R}bxd^@z+Uz1H7_1zJR>bw4ID4Q|fFU?lWYy)fW7COCO}wuv>f z-Z_52{eENk6CWNetqD{setY7k{xF_Da1q&GP`>(b-o|5vb!Mx#p;Y```WYyULn+?_gwI8Ir*548P=H>FA{J%t$TE z3;r&V-(Xm_PtVlrXHM~VLAoxrMr(FO?S^ATyYtHS7M!9fN=Zut2fJ9#;bwb829HQf z7?#p)WLmc#(P92BEbs^5exT*}+5O$eO)oln5|hJB-J)gaAHe-U%V`(x zN5V@>PmfQC=Rp5m6=eN_+`QtVB8=t0{p}WXLvK7@uzYV;PGxO=b)&`MR!BVr36|$= zE+#HmH!d@IQooe+*sz|l5&kZ0mjm1nw4C z@fJps;QsbcQBBj)(y9Z66^mBqlbr^Y^_eeE{z7SX3D-x#e~1j^!`ibE3g) zmr*26NOb`mF9Zjuf&EeG5Wv!<{76Z) z+3qwu-4r3sq>kelg%c2Aj?$zloUIiJrA7~eV`#T>8iYl^B)$PkawdrJ0 zX`Ml>kjrUlq#zf|X%uwt_^7ezNh4C@ho{BMy8H+q7lbb8e!Ja%?cD1QtY0y*cMtx| zuqYktB>+hf4i~fTMA`8ZCo(fGUey@5A7~eV`#UQv$De969WJT-Wv}!y!0KRVO6nkp z5;b1nBZIUNftvn_G1rdlJuEfe--YeibV2vqY__ZCT%V5DzjoV9#?1*H;dG?@U&LfI z*pWpkJJl*;g{|`@!htgr40tN-DP(%o(Y7ckpgIv z3J=sw?w@kykj#{*kU+H(hHw9`bXoWJ#OhxE-%N*_L6?$@lxCsuce(xp1WlXFmh`*d z+^}>7H*bvIo#-OzHI2bAD-|6D+N3tocGhX(8pK^j29;iND0nia3a9pjBIqWu$ z<7kqsZZN2#ld`k2WWJm+3)~O13&8zd8Jp8xT-&t%c)`Y_MSF|NMP86*4#07aM~Eo< z*J~7G(vxo*(?2OP*x!Y1zjSf;qrVo_HjaAqqi-IXGb%0K-2q35MZ#2Tl&f|fduZw2 z{M;OcLgBl8f%}1W0k|K)JP#lg1gtrd_tUPERYtSb;bu4<1p(->L=uFIB4dNKH;(N; zE+aW1G)S$4Il#^V++SR481d)_%O1IYR7R4^&G@@?tw9x_T($Gqz2EQ3%R%n<;S~<< z2igVTe&A2+FD&1Dym)JNaY=1cmBB(&Bu$dMAUGM8<@ibcQl<|~>z^2%6d9t|sQg{n zb^*BG>;4}eo;xZ%(Zzaf(s3&OF;DMAom&3uuHz5B^VpMR@#T+67+z+%6 zK@f5)8@}6l^w)h^PL?y+oG6TzLTV91|F2fiLz3ezAJS_~uY`~QwNmcGmeTDFU%21x zd2=qOsq^o;qR~A6!{wLt?eW0W5$(qPtTrfM%Htnz%_EJonun7tNnGYskeLnGX9`_v^Kz?|&yVA!^Zc^ROoC#VptbgQm>&O__JRw(#wb zZodAy3%Fr$e=9)`nZtQSpZ>aumQfE~J1ssk^gJ)gX^JeZs(WVf(){x3QN7cWVk2`; zRUODK9GI5;+Ff&LnrbI6|B?F-_;Tn)2e33H?%DoqHL26wtJ^ znv3O}3?rv#m7H$g&LQN*Rrvmw$EtxQE`057&S3@LaRW+EHHnpN)(pYYe6l zk9_#@qK|L5Ztex#a+h;IO4Fsm#HFhM_a3kC4}SBkc6wFOL$jgLB9)Tbc`X0wPnV&m z%-av#m=YJsa>769AFaQ1%iedsUo*US%G2|&(rY#7=DF|nFG_0b7tXtCWS_M2r~cl` z1+;Ts5KwUl=?G4r-sUy~XB`*z*MlOY?+|F^KOL*LCc?XA4BQWNaB57Jk{ZJ&8xL$b zSy*SbD<}$G`J!h|fdHi}CRo=mF?wvTq-#cGl4K{GJ?O%^f5n09=ay|Mt22dYl?ma& zuU-T=Y>x9MOl2hdMWBc7PVF=;b-6LPG-JMsu>hYU`1Js{wKJdlnLw8Ra zHfLm?O(zQ9`E3{aEeIT&IVAo5DZ><=@oHZAdF$cQ%8afN~c!3 z8E(n8L*MNGWzOay>Gc>fcl@ix_RB7K{mX4>lK)YmDDs2eR$D#56wB-{R!dv z&lk+Kc;3m=tZPM6j}zT zmB^ZBXWo)u3Bu`g+gvV&<1tT`kFGzBLcxnDoE6d64OaV!!m>rnen+wJMuiHOZx991r zFBzH=cf7Rr$?w-cJY(ctlZK&(&*T@rNR1Bv;MU9ET)E@REr*|;eaY3sGB=+ny65wi z^TrOGG9azA!LVS-n%gH1xqI?3mz$mZ(icaHYi=CZZ}za>hl?wp`EkSIJFl8GG<`=- z$yKj^`^@Z1CuDZto>TJF4;!Dq=F<6>4*I9^TdmgFbFb6Z=PkH;0{Vd~8_YMn^V6?S z-aMjL!r!-QV_wPE9*HGJp{u&&r@0XT1 z`;wu(dnCvxGP|^D&h1ZQVfvr|4LU}X#frIEjOTa%@RpqVn-};EYlVnTJZhGfhf)GInfn;a41?~=Ba3tr#~oG6+lxqtk;@dE=C z@*j5{3l7kX&rJTOjNR>aKmYvm?|=LWj}ek_1m13J-2Td)LsH`Y-!gra)gkR2-8hGGOW*lnb$&%{-|lg}65~+Jd!nTL zWYMXJ;K2K@nKHU>FN4Lpa>tbOgkYVz z!Du;JP^wla@4aTy>Yay+Dy#dYB&rpPMziH`eyLh6fBl~8`=%!Fynx*Q!WS!k-nRd; zA$?!D>*}-Je<-*3;dj649vAh@ysIKZgTVbjXQi;FabLly1BDgavPyF*>lHFe;b~{% zWTdH6`X@yXOpYCumN+IOAyCy}dbj_H3-11_Ui(fLp!n>LS?JHp_nv(8>$N*xyggd4 z^M3m1H|q`-RxW>JZneQO>d8;iVnZ2@|Lo3Lz2l`SfAkC)^yvGsp@DJ1fh^A>@8(w3 zXU0W-c-!Uq)lCx@F8=lL8%KC%Z9Z63Ip>`p{-3?`0F0vA`uOhDo$b5nJ%JD)Kg%bw%{>P11MOe*Rt=8=ge6f1V=f7+okeU9{ z-*1k#noN4_)`P_h7JW6Q$xdkxJ9XZ*szXa(o&r|k#c{dN_w10t>Fg3_wX z`*weF&UJw3yfUotdkVAvxkcXuSw6yQx_{clUg^mMhPi#gcbBf*cIc>4PsP~GPt3VK zAGJ#GPYckpIieHvYW-=;r7C zMko<~pkul2)h{HCy!?qd3)k#<^Ox<{5ALySPr=v!xUpa7L=7y}=pXRdJ74`{Mx;s4 zixP&Rk|0|RRC2WSonLl*uzL5V=kJQOnZTOg?#y}Wi`D;oWafa*NwUI>vU=B-TMsVY zxc|yt>95V1WH+9xHC94l9WT9o_w{qe4`KbG zwmw}aOIGp^YX5M0BuNn~wjY}N!HRv^xoN4%7qZJsh5dkk|NL<2&iuk##*e)9nvpaE zL)&yDQdZ;m#~Yuq0{`mYZ|)xBYB*mQt;TVE?gTc)Q4u7;_y)O_#P;}DpKKI=) zQAd&Bazqd31Hybx@CV9>yxjIvicIC-2GkzmB$9-hb?l z)VN4qfL{vmd+Da#FE3s;q-U2m9+-n7sF6ZD0Q+m)zANT0N{lc8!z;o>MVpO!l1PiT zQ6!cUZMRYs=j3>Wbz596OhG*_DP^8-*>i_6}KAd;` zxZYjB`V|h({IAx`y1d`mK3(R%`<>Gt`u5?Oq!Pye=w~Z;94lM@&pQgL8eU(q^_gj- zGm~O8%Jjlzd-K14bXJ$Ra~+N;Dk>T|di=L{j~UrJ19Ydj%{=!0H>5LFpVr}P)v)HT3M19T7zhRZ=2x7 zq}MIqe(3f^Kkqw`qq6^;1ej1#T@S4L*hpK1N#EcN@&eB?e67p#+LtT*LHf~~uN^lu z3*ht3uYX#yZs(YRy`H}HI{k?t0A&j(@b$q@fsh2YL&B->=fr*-MG7lw-}!FU?!4pO zQsZMIY};~79{wdY*tL_AdnN4{cfaJfhJy$K+=p zEqVIGukB{j;sy6;PlPK17Ovf!|L8lPcT0)?Mwyg->Uvl zMYWDi`6cUfju+NCj@LRtuLW$#a3n0vu$|+hZyVMBvdpxu2~p7&V_P5W+0^3fpZV4| zlLvL5F|7CS?x~&tJ>i8<>}JF5qxyM6?E7o>WF|&`de03giuSnwjY$K$|7+I75X+B# zc2VaT`@(rQ7)avoh0AsxE5Co@aEo5Ic=P@M&42aajfoM~Zuh-*{or0N-vaB?bAmA9 z=@0s*BrN><^~H6LYhGU5Eg^c^aL^TtAFkfhD<$^dw_atW&NZ$5@#Dv@96Nqus;OIA z94E*Y1Nr(7+s5_l`r$)2HM9TiyC%;WJJ1)Pe;f7>`xH?lD&3r<4c2)up7orHv`XYmUa=P9V zbiJ?S%W{KTzC?8<*uS6Ycfplu)R8N8<=npTm;Ko)`_EYh9airS?9MB!XmA!))-eq0 z^!Y*KJ3T&vz#o};?U=zmML~XP@v>!`_uM{t)LoOuo@E^V9N7XsH(Ejts`&(M@3NzvYej4 zuk6)1BR*10I0oF#Y|-3M+wJNsn-Z&^T0ky@i&BWq>OC%GG8 zXN=*9EnbTelB-H)LK=)fY_SMwket<^r$MZ-vOr6IH)@Ro2P}CTWZJ7madIjwn=O@t7?>$`n&NnNItLworWpzr9wwvQ3Z4Aq8I8Y#p;-fcD z4zvH|&z60+VfTXDr`|MnaDYDLcu*yUZ#aYRIkF$PnBQ;O_u5x0lA|M^zV*7^8OgjL zl~mWivE=7{M~X-F@A=~2W-Q;j@69DY_w1bV)$@0`e7|i0mHlD)g7oHs(gvU3O4Vk`!eK^JE+>q#Od+7f9rM3})U*q`~Ts@k`L4BOxf zF5g@5)4n4+j+BGz2Lm=Bf(W9-3PMtZB{j-6I3sz=;LO2YE~xWuakY4?AF!YOG>c=H z{zQ%)&;m3ofLjV!(&1;)k|+f@Uayo_2(RM~(mEWaa2(u%f&_Cc%96%F5twpd$`Y_X zilLaY3pqeD(9Ak9LdS9f!wcY+HR^N-a^7Qh4Djk$HvHaVv5XowzOzvO(UZ5Gh5Zs> zErJl;hU4k|wUkL=e~>NQ%OBpNK}l-xHF$DQ`rsNiXFb(#0^Vzs+*kn|bRw}kIb;fe zqRcUo&@HVnrzm{kp=v|1e!^!SLwK^l@H_LE^JGuNV@F4+dG+uqL7SJN-M9k3| z*c)lII=vDVrV*Sq^vc^uL1sG z^MpNMx;njn@Qos@CR~{ce=h7d>9OZO{$bgcy;lz(@Y1{+!rTS7$>Gw#Q*#5L;}gWx+X4HhoWgzy zpkE-n57+e_CwUs!!rdASBfE`2^ybj!FU6_?G^vMv=u|1_GXwzdbx;E) z5M77LlE|0mNdY&WmW3K4*kgN`o!{Zzhf{+mOMVx7beBc~Zv7Z7NzuE06iW}F2^ogb zbDi$FH7$AR{>rMYKC|Ap3!jcy{(bv*`Qf+YgK#)a%8& zW`x7^z;XkHA3unJxhrapssxQ`T8#^fLrMqV| z=MggKFwmzx`ra2ANx%C2`w9U2|L69HHop4|u#MU&`_ErA%QyN0z=^4K`x&0EaQbML zt8@FoUtmB0Tfs+jEYEX-6l6G>=NMM>(KI+;2*iC;en%vnJyFy2vsSsnp%o!FRb^}* zcL47|pAu;^B?6;>AY!a0;Mt@^T18!Ku|bJ{!tQBRc)dJ z#nnF_IJ)n6MPZ!-7z7jn!%D#o2V7Xt8D^!%-!iiIA<%8;29C!fQ<#FS?4ISiA=i_c!-d8wde4utvl4V3M8-hf@TrosF1Y#wnGNEd z`szpC1WqDXe2}j?BD1tcU?i5-3M}Z0jXIV6=Oe)Wtp|_1xcK`Bi|O+P_o5gId_WzE z(G35^>TOHbZ6{Fdk(;jt?B@j;c;pYfwU`(9f6kpYacCCHhXd%eKkvwS{JpQbCnvmm z|ExH>6<{*^SjjUVeGAC_@B3zVNr`3HGaNkN>j%p=eDd?Agh=aSv!;y5N)N9EZcmF| z^Wx$)U$5B)@Z6_s>idt}MziASd4VpR{Um`d+r0PnFMrBNiGT5LH>Jc{c>%1ZscsCs z^}k;>?$6Kcl=#oNQvv${*Yb-ip8I%dWutTbI}5BPiiHmHUxEL&o#O`_0*5~ogT+!58EN0hkJ&ARS;hUJ&a2L5DX;^SnDiga1L}AkdT7hE0NS$ zGcXo#)iP*H@q)DDXzAvHlI)V|ZHLRiRiOu`1VwpK1cqZmgf%NAZe;h=%QMsZrzigP zho`dtbX_@SoYhzM?|IXF;C7azu?s$a=dNq#j5&$@Jcrm~4HNH%&-f%>MlcxQ$1(tQ z^uaa!@qJ(wGHW!MH3~3~-TyORv`>Q&h|xlgz8%$@0_*=P)t3R->H1G3x((%y?`IEh z)sMX$(OWbE+!`ErWEe@}*<*Y7gRAso@4%CKaD{uB{VQY?*J@DRfQdxs{#?mHX7>-^ z+<{I)*YirO%OF%|06nizbDYUuC3_pR1f`DmpSJ-2N^9yLdh2svh@L)r$gByMMcb*( z*~fp{eqejfQAw0R&-?I9$bOC&L3jWD_iL7{*#TI1=e47T_R1hIZ1tYpFIR7Mdi?iH z9Y1sY2tkwp`<*`jJ^%XH;qgx&Gi=s`k#-9eq<=d}53Cjl(X*d_x54G@6dyC?@jP=J7-PA+g0{{Q#e;>Ya06dY>1>EBa=R8`>K< znmNbIw;Vd`_W1$Bx}_w9*$-}oihAdZpDfKQuDEN;*daaBBQ53#tLZPme|w(aAD;XT zEc775hd54r)pUQFBL*{DxKF4pBDxHu`c2T1aGobqu#L@K4{lv7zKj0yTO^(dt@&7TRcZ`&{be(- z=v4a98lk$-FnX>A#Q@7C8g%ano|vg2DWNo1s4p{2ybHC(vj^95M|R54Q9!BTQ}KkZ z^r5w2?}#-Dp)BA`fW81r=z@11h?;C%aW=E>C*V!U2x25v_MfMKm;dGJ9iRWQ33Suo zHigr`wVD{46|nq+Woyf88=t=IhVetP7)}7B^!fsyuimm|cP_(o;gMiT6io*Gn1NX{ z#|@9PnZpAC;D-Nd?T(L^Z)7+a4g@nzmk$Q-|NGNMu;0gTnS6Qw?l3^@q+bUc1SZ1M?| zY&25jh41k!Y9;oAZX8^kvIw}@iU$uad;-I9yoKXugTbKF<&t5Si-u@+$bi0Y&KWyr z{E(BL`SS>6qQ=Zc^j7xJT0m|ru^aI6rGPiM;lEf!l5W@y8lGixH<7(ZX=Rb#{*6#u ztm`*{=y?VBeYQATD9*-{GV$)iK;s8~`i~Yx^do1$A%p$D2sK4WTpHD9q7?Gc8$Z>M zdc&C8B}t$+eI~l=!QHDx2x{bXJhl^Ce2_2Ls?q87!>5DWo8GY$;Fjz=j!5pomE|zI ze^lB3C#6+V`)$i9+zlW^Z1esDM@wmrlO^%FJ7<6{_4IEIcuceW z_MD@+$IBZ%UP;#4OjPHD*kM`S(vxf~tND7(_Ps|+>}JEvaU+srqndXSKA7VbHA~j* z2r=AqcTNL6cX<2ZWnLYti%JFHGG(?Avl`dYjH6=P?-n8-IGodUr_~JGghf%e`>fy0}RD^s&Pt%tioyuqJ@Qy8Zb(4jvXIaq{JZ`*-Vf z_RNDVq7K0R^Oc`wfN|goFb!URh!ZOuF3`i4HvoRX(FAZs`DtKuiL4+6Xcm}af+Pde zL6XB2CYvG@5$#E_2ufcjofLuq!x2urk(G-oG0GZcHkc?~jMb>evE)dY94Xdjj+k>&*=hlzD4WN$Mx12w+M>ljb*=;yi zUC3`1z5l-O??|@^tL+jWh+px|93?rtC zephJYCsHT?ll5!0x_%Rh^!{Kqp|&Wr`3tiDBqC!FwE1~ni{nzzOYdGL)DJ-l#+!vkCgrm7ZB1=R#- z@%O$m=$b)SfSpME0M-I~3zz-XDvu?=Mnp-&u+quk4+~(PBH@Mln-PKVasY!6Bb>Zj z_{5mhf~}pr9L_Gme?3%Q1Yci-nWDiKBn{2|I<60nhg7~4JnN@6OY7+pWdGR(fZ_Ck z6UYS_(di4gLtuea@9}|a#^DbEyTKi#MM(yBgFDE8&m6u`fMIx1vEGyb#pcWh%0mK8 zcn+j^A|_GEigY3)f+Q>Hs^El1yHRhVNF9#Z4LV>VfcG|FDiXLdkwItF!Kd5NCPS13 zPTB%X!ZWp0HOmvEH|Le?E2-L;Ta;T?Bgz_z#KBnrW*ox{y;I_Pr^b!UOdFh@)IBM> zWp~|5>~CGus07Jb$5$Q(>u3!&yvqO$6W}Y4P$Q-zws`Kq zFM!QxRH7sblCxH>*?m@(0 z!{WMVbw3;rN;tj58KVuL5>@CDt$P?gi;XM|dphZQ9_hgvZL03c-0;er2@K!oPR zVJ8wHXqq!3z*k8I<{+g+8-O(pI^00ONOxdO0xJQ$nkh0Z#%`pD2%|p6YBK4F#7HYJ z8f@pBiM`_qMt%h8ueKcgZdYDuqnpCRQv+mPl=LJKXEXLniM_gi=G0+5$ref!(Ms%Z z^muQad*7~{ymjAw5Cum(Ti6)@rsk?Gd+xn|!Gb3re*V#WTXy3oE6o14=Zu~`euO`G zYCeD90LgHa9pYFqoHZI$qm`hGhQPc!;EsVS2%;>BSVX*j{5*}wg|>VRc&`aV5e}Z6 z~C?Mm|>uS84cibOgLl!tbuxOz(;dcPA|jr<@KH*%hq@T)ebMrKPJG(PQb8G zo(D*b0B0w#(8iFPvmS(>8=V|e2!B-KS`g*&5)46*;VYtqHUh8LZqz47SqwTN&T6Iz zEH&C@HR>!nU22RStOnB_AaIxsSLRIM1OfvhJCJCO%Pp&0wmW~#p~41lAfV*B(IUX{ zla+)w$Q3=)ZW__|vdmN+fm0VS9MF2~pML9od-HQ@4t-=X>R>vW78bQ~^S((_o}T~o zBQHF9U#px0MNNMHZR6%#HG&I-PK|OoLE`Y2Lr(e;(VWQVuZ2GfEb!gY5i@1KQ{1~$ zayM!aT#3;-DgAy^+M^TrLr(grY5Ql|FVeh3{z`0aq2+2i_pl56p$&@Pt@C$Z+6QdAcAmoZt)5EHAL41ndbV{gJjQ z4T7fmO;cP#IC)}oy5*B|Applniy$9g^zhM4?)@47qc zIBe-QEt5pV9!JD?69Z16`WQ^k57UO8VLQKWN&fBf-!9agRxPOc&`w!RCUER%MeTt~ zU*3_!Dak7PRbBj2+28)NmL8}f=bA1>Z=ysDtQ9>ZXP zcJQsCImHY8IwcJ1k~pw)(%9Zz;_c?MZ%UG+L#1_{<7^f^^;faK#r=L#VgEnpFL=0B z*$*zSS6_WKC%-^*J~I}=={f++ehr?20Wv%sD8pblr*vYl?3CYvb8t{#=ZeaXPg<@; zQLQqe>JMuJ{@&M+GP`ws@PP-6#tX_KrLtdDyO7HMi(PO^0x!TL>S6lTdv=z&&N41=Z)M3PUbzo0qM z)lwt@9}M6qfx(Fh@F@X;7%2irQHmgpI-*-*6oH|@P>3`cZAL07(hO_}i=n-QAc%rk zH2#c5kzU+^B-p>#=LRo zWGrVb8}Mmb>6w3CPQ`)$L|Dm|WFetSCKl_3+a_ zdV0zF^jQ0CSN8kux02=i7q7gmd&;ySJx^Zl4zYjxaM_!4uSu}BaZN~-{i@o9RQ6vA zI+5>Kg5hFXSX;pxWF5X>qc7kJ&^0a}!wVG+?jXyRIXrGZ4SFMv2XvF+(*s%JB~gGW zGBm6pYEODBdwLcG(D9(*`U^);Htl@Y~Yq0;`U3vL?cKu+r7zDnB z*>5x$eqOWXjvMb;F#nO~+a>$|{?YO*U4=u_QQM^il@g>=rKO~sNtb|>l#+t<=w_n? zC8fKO?rxBj?rxA~j2tnxeS6>U{SSWe+#l-0{`Zl}#S4ztLA>fB-g2*#%Wr!h z@Ly|`*tpMTK5es9d_k3bg`FTnif|6ZOtyT=z9soO8YXWs;G8z(OdhuX0;jfgs-A2j zI5&J)!`61tOgEld`3Q|<&&Jd?m1y9LA#7e$9fPf$ zfOclkl)iC&A1k5gT<$A$;@HzA$d$ zvD8j88HpSCz^7?zp{J zMQJTpm;LUqq2pBTzz{h3-gjtY@&adisKx``td4Rs{63pq7t(REmTA{;JEIe3L zw6)fji31p*sCsIG#EC-q`sZ#gwmCmU`*Y4@he^ZQ+jqYciLmf`BRiOFIj}auSDxs43w}XzQC#H8?=jROx+?9bS(h1u)^=^qQ2U!NoX1&JhYJ2(*WI`j8&~+^sg3Tr=?*{DyH1m*)9A`ZjAg#R7o9IpF?EodcHHIG*Wq*%xX%c%J(3L z_0h=1gD+z538Olm%?3FErVd8ft8r558xAU*GitKjBRc zU2tYRY)b@~t{zk)gfaGnP9dWJp-E}L-9M^HSOWm9lY84$*V8*2*E60MA!2U<*GHOy zL*Vn~6HS!ONV(}Xl&sloC3P0i=78)y>Z?D?621QSy@&1;84HF6+-v~G)}~hMEMZ;@ zJ|4q)ay0oKYQ{}|e?IXOc0d2~kvqZ%<~V(qnZkp#^?khND5vb+To}qz)l*aZG{VzU z>nqmpDj5X`R3R~tIiLAZ*|)ld?TKN$cYbGBHH}}-D-j@*0w`3F=`JT+{Dy^i8_{z* zBpUWMhCt@QBW z#$ET4xbl`s7tYuJMPlJNsDfzYSxq{4R`;pEX*yl1*7Ir}(e8tVwv5gFIO9MIw7&T; zWabhtoAY2+;_#L|-3M=ozmtlCtCbxz%Bj?a=A{Nb2;@zd>3U%N%&!PE z8207rBftZNm3#VjzDBp0xHc3^IPl0aPs&fdXXMEiGno^h?@0c>$a>zBE7;mFv&+7# z%)tZ_B6fe$1Qbkne6!hGUyA`>1TI}~MXj!^(sfgIt)EX;FH&JNn`?O)K;HCU#D138 z8Ay8xMR^bO&Px#DKG)y(bj_#AU-p7#x}F?jx@{uPRv|u$QbCqKnC(PmdiWfWbDYwv z?&mh@ok6dHp1g|C{~U3-GIR zyA%P@n!kI>Eipt~V_J)x(TuqZSm%~uzl$7><8Blbq z8Z%y9(hW<}r0w`~t|XHT&7y(NCNwJ=MrB+^ItAQH#O7e0F81`EoQ8I-Cj-Rbl{w9T z`+2)gel6_B2oZVIW`zo|sYG36SJns;mTwJX^hO{-S)?lu&bp=A^VU00Y8Un@J>erX z9Uctetf9rNXLQwWy28E)!GmII1MRe)Ni0xmt1{VHTis-RJ=@9oCU=0l^eI5MasOqz z7H=eX=A>!)BBH@+v(YWb26&3+dk^ zu_>&8d|f?#KYuk25K2No6kOa)|H+fopqBh%zxHR5VLhyt%9iPt9@~_;!6Qz-JW0A4 zvd`R$b_xd^78?S8-Gm-ujH>2Gv0FZlCp>u%sKmj}Ze3#BlIB!)Ys)ZgbxGvtWZA#t z0DPR$@Rg6tH6S(HCKFigA!_eQ&<5-39hf)+5fi-h+yC-62?eK*_s| zZPx$r1%^;Z*zL~f)>fdi+jsFeHdSF`m~Yx}(#HvA@5cx3^e9@J#I z;YN*_0&|lc*Sz{1V;d!4oR1H;n#MVBYbLhmSH&^e01w@(@Z78%#OBIk3b5em<3QT( zzPcz&Hh5umlWQ5oKI7nhwNMVES^f&r6JJX(G4XGhnU>%_tA`hVYtl+IHFyhO{SaP$ z94>t!eMt0?rIr01ZlC74!Ef(p-Y*N&7J=;#XkgVoB~_~Z`od z+ObD?)@$_5$+lA91=g*2Oy1!$pvg-fjgq;&1fh5$rV2%aB+)@T!T7Jm27cDsQQv3K zch@oxdp&V_c=!s2ikjHn@!84@$!4EjJtDL?M=rRNRu8(eT3aPNc?C1r^L`^h1v1UP z>s{UI8Jp7f_qu|0Cwdu{@f1{HZ7)C1Sm)5S|MDDkL@nJnC(4h-@@{82b z^rHWC&Sv!akbd!m%CP(o#*r3|lDP>Z z357tS_nkg7pJ1;;lrZfnZkfG>Zst0f75|`-x(Z1$!|=2yuH5zFKAF=vuVIF~< zeS$#3%FNMX$tlp7o~x1p`47)G{x15SDEvfpdziM$`#4Ms?750wt9A-}9UniM7=N61 zE-b5^re$|@zqg687Sj*JWdNLSBpvvyI1jvv=HJu6aB6@0loKl!Fzp1Pvxjbqz2;sa zI)!_;8)|Nm{TwogVpBS&FRd_b;{?OwW->3u0e!jWd_l?N-jx-d9ZGEauM&+-)NGmBaEm{Y404_y+T`N( z@{VV!aAfF5V=8Wzr#sk@U?lV*K#_jhzb}M!SF}t2Io*q6|2N+9|M`ENfk1UP$@7UI zJkyQCGS|D9MKQDrvicS$NS`p-EttJGe9N_^rjlQhm=lp}gya1$@PaJJoZW8Wq>V-1 zQ~ifVY^|6257n5ubY+7(%Av`35RjGB`(@aJ8)_6I33-a0{oMO4* z*&1n=`;axNzi?Q3g%cP_th{-M|N1GNI`Pna$ZZdMxJDQA_mK55U;Y(~i01V80r6@FHo0$9b8T#2{YiT1t0|$yJ{Lj%S@& z@jGortUo%u`7BhFEZ{Lo_iJ6y?mOpRD>+{O5cJMt(vHsw^u6NL{Ds*`U{j0H2LM=( z73i#pKG#JbC`B0K7&9qz?N^*9XnnBVm#t5`pFwSZD7T*c-1m`(pWygLD{|y84fkl0 zd6T$XyL{5APHf@1AX!I?<O+vB9OaU!TAuhK9Dn?YUa~5vaKEwIrp`i;_&@d$VJQ$GIjxqfcG*&SP@fI zDsDVbfR%HsoBxk)hw(7lqRpGNf@W!-Nryp3n?Jj#NaoQZ@=9==(OU$C*?~Wy_vPIaF(tL--BH3>{E_7@8hWVJUL> zlt5-n=MJA}mQ=oUY#Zs`7nFoYq~DH(PNr}cWGp)2(~W^Dm_ww znsU_VMP?!L*|7>fNDCEHd*h{}>c^Kg{kyb7K+@9XBki1Df1#h}9^f^_Mrzan|C@HG5l;uao_Do61qr zc~3Md-|EquaIxV~FUSdaoCW5XqZM{yEJ6x56PJpyvy`q+<}0}pY;`tzL_sZuqQh0o^sO1;E|ccz2OL=XJFq8G zDxT5J#E0mpW`pXuG+VIKE`12X$Exube5q5g7@Jp&qY@Z)8E&BeW8p1{XFhwo-0SFy zG=BE0R_a~f-~HKEyyh?fYAFln^Z`t#0~vLO#YO zrjl!}rIT7Yp7d$`E|XHFj=Ej+^(O+CRQ&k@xomyRGirt^dxVJm#R3yiG=Y@{VO_!Cl!NjqU%f#bvivC~U4DZV;@t`Kx*rHJ82;!4u zN#Ns)_W`@MFAwdHX*cig++>UKWA|Q&f?8_4k0>`g&->}GpnA)jg~HY-E0Kxu34rPj zcsgLo>4^jn0dc%=mNHWmN?SNhmOub;?2^-olr?7jQP!>e<`?sH=F^93T+`f|1)r8W zC$JWIkAL*)kZW)WbK5~uFl;=l$u;&bPQ(t?U1XeSKHCq`@=~450JDR_vHlYFa*he> z;3qnb>6?;LNPM!~)&K=^QwOnz*SFq*+oCe>Ul_C|F*~;h9Ve3^mKswCCzQjiOpKk* zTl?x!2)*{~d+T*_DTz(tHq_qwEP?oCR5@gFoKf^^@Sp<@H<4$qLIbJB->#lW%{9bz zr%3rJae4IerhRY2;~Ey>GL|q9?}f8zBCI=fn3ycTI)mSKxmRT6K~%Rf0;4-<(F96} z@n)gTtSs|wc=g2Yrk;Vx##XYt6;WqAIUAh^vC<)nn50Im9mz|DRD!mrOjgr6TS82) z%iKQ5qv@S*+EMdUFgFW1F0YCr7afy}+xAOzTZ}c6Kt;nd^U_i1S~)69k{z8lzP@|9 zkH5KibbNlg2`Zd>X~ud3EmD}k1#W4=7=@ksZx+`jPgs+!l;cHP0DjY>DA2acJQStp zIEIu?a2gwyh(c3B&uyuSDvdtrwZBj>Cu0dcq^-(oxei%eeX!AOjKVmw|6GF=5Y4Y) z3b~$d3m9mO@im<~!gt9pnXAXF39qgr{ec&Lpb$sNGNr(#WTcbt$Y(8AF`r0d8QGYM zDX&xD_{`I--Ph!hC^Y$O1n^P2u`;)!WkY3)Dk#6=RUooKry@Z9R^#P^b=2_j2a?Le z19x}W=)qkSP91paFfCT{4m=4UB^AQey}D|1ymCrKJX#)mqpIyFPZxjeM7vqX(gU*f zF2JP``6a(MLG_nIsB z>X5vd!OVn|g~9)oTAVdgpm?Cw=G-06MVuCeT;j5s$Z<5fELf^nLSPKl25p?c#oCdz zIwP(4Uh?Mc2C3*z)CmmYHEPEac)%hUAMs5v;)XTP;r zRwii1^5hfH+Z(3hf?O#|%c}I|H`1klH_`nQ!3F=P#=8L;x;87RT=}F&J6gEdaS4Wk zuNNd}$k6L*;>iOj=D-J&Xb!uLvD}lc+Jj1rGxEjzEH-&Oo2rIChMy1KmQ?WCUz5#? z2jsmV&6bOOH9{rTagro|OM|IsWIIM?F7)ve)0iu8`lO9jzDRy~$1&N2V_GGS(lp@M zs2Nb5@3|LbIeJT1<=jfh{EUnITpd#m4X3~T9@IFT^*F9EaddU497ld>r=*GY%|mxL zUj={JFF?Cb)^;8bxF3Up!ji3>6=&9W@l4mN;YVpy6WpPW}+E~uht6=dJ@|iB^p9LPs`gooARHk z4H>%*`-M;KwG*j-)m3({<>-5=+`yQu%Ijhw+!aqc`Ux$Le6cdly zt@dVAjv&9ZinaPttW+kxYT$xfVgEL*wcg~OJ($cv{l$;n&_95oT(^`1w|fiN%rR%5 z%S}rSjeys@L^JjI{<}Lo?(x+zsFR(;fiZN4+abh1=$ zxtoo**aC$YC#+gvYmSCe)7+*xQhN`qu+eyV{GFo|(w0Z*uX03dSY7pcNdidE5u zl+$4@di0O~>(-97i|b)wj{ zwl_^@5Oam^OhvJSV(0pM7RB^k$U;S8s?3-83f$vf&gF)0nE0g?)1B2N$L$cYQDQn4 zK7$n{Y)jfFX4e>z-()wVXv1&LNEw$Glzed;h2f*O;1}}`r-0G2$OT9zv|z9HQ{UgS zET^`L$wK7%M9WHReK$=FD&s(4W|uk)g|G2(G-8HCDOP{oy-%mp_h#)b`uMiIP4ssi zN*3pA#aFYYWkCx96Y>oixY6kGP64=gOz1(bYl5j_2Kx8}C0jML!w#P8O-q%{FGYtL zwDXDe(tLAFmtwRR!i*JWnXwl6!hgeT<+FSiOg}UtyiY;CATGmspHBEU)Q&n_|5QryU<7GdHv{G^&Drl*+T6Ok1W$UQQ(Pu$>6|4< z%*suauwav!xJx;W7a6E*D1|Pp+>TOv8Ej&AEi;oM*nto(UO?327sZqg@oCMs956hnTwl<*`yNVS|YYWkdG$Va!h{Ki<<~ z9YeXeMtMpx$A7$i9hLdYIe1Ql8?-W&{qn&B!oc{w8EPAzt`>;cy2P!f=wnx5+yBCEQ*Z;=}gvaKYC@dKb4(z^n8Q+9GCmVIqCj!gSOEFyTx`wIf=~@L_RN{p7)H2}%dt|Mqk^cb| z?fL2MDJxu$6O>O0oS3 zlN8Xy9cI5_+6li>$7ATQzDup(A=%%|QpiW&f))Npb1}v87x}ihc!RdY)%%M~TgQgp zF43Hy|G>=|P0A!X5{1R_-S_JspH{D@Ot504z2nBoNPx}EOXFFVke+DG$fkkSaV_hR zO9E7|U5}R;`Mj9OKXRzpr!L_G?+}q(HuGSaxmUhw%v*Z*&`5(~`&)oJhlyKVWCYvT!q`k1Hrmc2^ z_8P|)Fy0m-cz#DsWOBbPe*NQ;PWn+9fS!4)9Iv9a(68*4zs}!a>}sSNs-S5=kZ4#P zDNOs~Up#Fdz|@hO9pS!<%&be__24m+F<0fKu<$W6Q>UM?96aCXoZU%yQ|TCxN?8_} zfO!8;tPWLP8GeSi!!ulP?RX!zw8b;*6;107>G>cZ`8R&9 z_MRxdh_s_0vZEnRT1b{yYt#TZ3!3js818tJSht}(ycraa3Tg%=DH!tDTyT1%#a5CT zETK8sOB{nlJ&fja_CiK<^Z8&Oj)zFycrHEh4pX4iHDqGVsF}C_ISq>d2?mbC2Kz}v zUS}0;X4(te0SVQFCSf?|Z8xR)Y$F!Dl$^GX&GdHi9h^96Zo;^sVi3YstK$l{c|kbz zTsm29s}s9kWoAn%IA-Gb@iBV$!BZT%CjB3!fR|w2)8Ic5CFw<64y2ilT#e9vJxV6H zVa)AJW*_5I{?9%I#T|bL&2(nA)&eR>xe~YOe980&kK$emooo^KvK*U)Tl&V=91a__ zFIO-XR1QwmlAw#^%kGGWnlGnnSyo-;xQg)x$)Z*%J#ZGp2Blln zX0GD`6g}0@t+j&>S_JNOuupb|V16sh`;IxeKrG+D6`SH z8;^R4)we+R_VH^U2*b%t@+uy!d5VT)x%B-q32A4=-fi(wM|+IkBq8Y4abm>QHXh^4 zXu|PdF}B3ldY%NHWkrQ0ciS3&gybN$EUIpC={DXsuIf)8I-WlS*=z4!Q_f3#M*f;* z)Q<1^v+aW=mn3$LzO?t-o9-OQjjT^Yci7(T92?deKZA7+&__ph5B6{Ne=`||i!>{q zE$#=k5m2->q}&Ok25*j%CG8~+lL`SeTR0Cmh>j_ku{cFnTIKJBuc%H5x2r{n0Z`Cj zAo)y@0M`N$rjBYkYI<_TDDB6cgK`A8AB2zMXw;><_n{T2trk6ZpfAd&&2GH^Uw=7A zEsgK|S_u-e8t-N+7s$_9MXV^fkUnl61@(U*i$C%ywh1KQ_z3 zu#@$LeCtN->eb$6fr=V7jtKcM!>({_^k4M{$B+hDUX}1mqwIU>p6!pxG%BMzLyo}> z2_b1IGzImXqQ3UfgEgX#MtSR^`6lJ$Bl@nZYdNl6y7|!ADj*B*`g95lJ52wVxd3T? z5%D<9s)Q7$5JLmln9=s642=wSKmJ+CqK)#+U~h8ftrRh%6>$nM8?1POkIaqPlmGz1 zm*w*bxbE#YBbb4Rh0V0X>i_OnsD_XQnd%UuQ zqsmR<@wH*Rp^TZ2%*@~Tr)tyT^=oprLZT67MpV(&eJ=DtEb^)FLe{VSO77}E(8l2% z2W$GE9jdU>TJOpNjK2M(d2kntp9f5fx9?I_k?C1X@z$pT9+nl@^^HDEBeI#6x2XM2 zc5da=wAn3W*oJ(7;)d~S8vhopu!E1;;3?oMDyG0{WL^3Q(k*TT#8N93 zmb=}<89liD40PAGBGm3iAntm^8Jo*ailh*YurQe2!#w1OsLPdi5W;VvVa9A@ARLYS z@hQ^s9nxr3aE&N3GrNO#FE2O@8?>XMjw8+Dg5$l9&JR`O$;gCh!Gbj^{R=H2x>*8T!>83xnx&sEqwpplC$;{A08+%^gyIz~ zD`pq;b93Tn(XnV7Zthxt*4OW z(e)zwwUIN1)J%NPGdgM9a*_<#T)U3~dF*7HC#;YJ`BBI6*kg$a>fb5^|ND;<7wLAE zxW5Nct?UPFv0yBvNUSAAo9G8tx%x&NfAP>TO+w$L;)n0P-gcHueW$b_n3Nw*ku+Gs z^k(eACi3+kYxSQ^@~H&2e6Tro+P9%YWKI<=NZHOKP~WONjGe8k5kl*&sy%! z>L)KkFhv*uH64WD#B#J3to3ll3LrX}JGAN7a@P&0n&hmMZ~Sp7uP?0KRvvbnA!N>M zFoZLd&_Kb|0aldGh)KMmke*Ff3Zt#nZ)#2fD9=lnt6 zo{86)Hhlx!4;C-R(1Lur86uObt4lnlUei=i)7qv?bF5j#O^0id(o=r_+6MfR2-&1@bX`Qo0*W?c;r- z=j!ml!zk*(?9xzTx37yn8#&&uW#(T}BH>wVRKEX;%6Xxi*ymd*^mv=b zZLEAF$m_~QD^HED2AD%rXrFG^>98aX!|8Jz)?7ORoPxBLy_>5#IZJUL8M}-Ep{(-tf3!0bN!A=Vyn}znzJ1E=rU`P zzKb4oZ+o1CEmwiD)NSoGkpKl}#nC}8@Y)_HR*Fwg9qGYuPyCTuw{RK&6AgHY&1iNIX_ulcRc>n3yb99>wpJ1c@!Ya1#QKfE}`-n5%kjJ2OIEZzt*xgwNGVV*J$ zDg9{1wUtJDF&-T{k!eF2L9FEge zPUlm`g=gvMd%?}Wbg;n=zaE+kOQ_Ey2iM@wI)N^s9lQES8$jQ?c~Y@*6CLo|9)BLM z+uuvKx7j+SvBT_20MXop#r6W9z9}$t_)PI7vjo>6- zFDLZ=Jpkk0j)~ELa%V=LHM9L0;QK9mWM(hq|DKWI7TC>bTL4jt4e-eNH@it>+M3aD zvb%_o`qBvR(qZ;t=eOIfI^ULRt4bAHSu-Fb&*{2M37HV42nQJV@~3*z`UrA!l_myU zkLKAEYlDoW0aJ_EQb!-*`VJfR=62|~%YHL}NU3##xW_b5>h522f%F8-p5lXRQ=_)c zOymLuEHiqA&ayaAs3~RlwrU6T1Ma4%@|J%I}hPe)wGl zGVK;p%lj(GX4f~}e+#dv4}SxUfxK+kNpcYlLflL!UM-dxT|CeaRc)v-(9zU*BEOQy zAWlLoUWkMCQ4388n_;Fef8rdgbIezJ+#JBJp?(W4VujJQagAwz=U3J?*8UERtWAvl z{WC*`087}r-Gz1CtBrI193GKyXF^&V#pveEjU78*<6VnidNeR>%zOFY5`i?D!WgUt z=J{8vcb?u8uI41_UvHmXu?Tt0%?=sMSKmg6yT~YTKNqQ5&MWxz_CR}X=@$_+7CgK; zP=$=2Tn8DytS8a=q{!GAY%47-blz8rAJH52aZdHHP>)@z4QV-DzZ&_J~s-OKTLB!)V@aNQHI?f)>$l2slA+DK5R(>Fgn-O zUc0&Vg1WsL8EFfp-?Eogfc!?zZRPxGE7}mFBGATWLAW!lCGf%YZ=Qut-3HP+JSXES z-RGqP_c--9Ctf2M+G;}vytdb$%-S+@{PTJqVtUbM0`@Ad^25_d)tQ_r20^G)lxv%8;(SOUBA3v2TqpTCtsuaq2}MV%_AB3y3&3CE#LCLe7^Pm zmitK`f3by5pQR3IeI)<$QL=PP?b?bm4c2C1KG{x2vMutdiH+HbnJFg61tTT2=<{Tv zzi452kNH9uu7VY-F)HQ5rUGYR*J9@*-{|a|oy-2BhzDOzqh17Spx!{x>5T~)`cQ7L zUuJkcp*kl4;urE8vr1y}=z`MdkVPEItca}7G420ds-!_d&9lu}v*UmuTP1xZw5>fkzvv;2q5=5j>3PeMl~ zB*XEJFKZPnbhJG&>eWeKxw>lXBWAdoL37oT2;;g5>w2;UkgJZH3$6=Tgyys*Xn)pW zlY{Q3-vS#}?B4!sHOiQTPzsjR7l-%)JNJNeK&)CsR2pO>ioFN zSIBxEAm}$G)ze@-Mru1l0w5(e|ME|7xCK!CRae6*9*3f+_UL2;$1@C9o|K5+qd84!PJbmRN3N8b!qwCTeE&MHK3NK z)?~SokS{!uYn7sc8l|6Q>Z5AkQ_cKwn{El(;PKQRZ@Q%fvXwP=dbusN+VS~L*aJy? z#o4J?yGuHZ(E1z^m)gb|r;|mx6-}+LAoJxExaP(WwP!cChQa}fvGtBulG&PaQ^+O^ zJplh}0e<9mt6wSg_q{`v$n1?07K^mrftdzBxTl=BACdgD2+%Ej&z6k44q-#X`X18# zTMf?3_H{p5FBbnsgakRugIs}W%0f{86OK^Zn+bH~5$Y*3rgAL^Ob z@*=1GCe~9$%Brg~EEO85qUYBP{(W^c2RfEzu@u~l_}LOKI!Nj2uqjA6LG7y7gNc*6 z>J-sK0+!o{(;`_3sMjT4tWWJvfiP2KIhh$z6cqG18{?ZNA!(M9Q&HX0k8dl)oc4<` zA;9&zh!Lh8foj&T>l0&~6%|^HM+R3S6jzgX5MT1icv8d2Q9$R@prRZ@)3sJ+gKJt=Nq3jMb=B@-IbSz5|BTH=gV zY}>{o0$jiR&S{#130cI7zG}w6;9OCWm(fMP=+InF^vh;a0pYXflf0F=dl3)7OUDZC zOh{#z*9n>*W#ds5QLzT_n7~)#3YC1is)%;so?cD$Q4wlVK1@wT8g0&NxrR*eT7o{I z>s!R`X`Qw0$p-tPR*+n=Og$ z0Ie6w^4`%<)i$XO9FpN>4%3?^*mba47#KG<(9gMejU1HSh%}`vDW2TD?{?vQEq%Me zz^P+NzVH_^#E1N;Hf=e`W8C<#u4oH(-S6XYOVP%XhNIJVI8To2Ud}F5dXul1RzsQ? z#{s1Eu3c+hs#Use%d0j(GNaF0Itg*cp+#_@1_SbP%Y;4QjWl^id?#Wq>ECHbe4XkC znU`;Zn(G6)-=os{{M-xsXX@Wi*Uej#_5bW;1VxIwmY@J=+@01b= zC*MFeJ$H`v;TjKqZ4L&K^V5T*ji|Y3KSogF`gorYsu&p0thWiexl}jCyCy`OHrgJ3 zuXOB|EqxyX@PzrIvcCZQ0@?ycTm40j3QUQAvK%|=*m~ULbX`#|%#1i7c1qX9?`(z^ zD&E75!!u__4(CpP+*RNc&qS0DykSUr^DIsF=Y~uGU1hspyT%#y)avB;RG&lE^$IdI zwB=MH!;$;0p{(`*A&OcWfTKey4I3Xi6>S0K7A)Fg7w|k#s6P^^b5J~P@Hp&uwIzV^ zb!88Lw{+Rm!ElCx(u*3Bs2P3_Z*htYlQI zvJ4_rVLPGPv>kMdtJW<34AvUiwfq-#Q*lV<;h6zAS|jNikD>|n@7c)%CBu@6y8`(ECRHej`sRO5z| z%k{(Swd*1DykZtKWKZ}|CN0A+kQa_ZqcE9l4HMWWW;%BnAbOSEwMg)Fr_w~GDoPdc z`o@<`?ztf+GC9hT9>(1Vo{~G4EkE4=KEz#6lpL)iIyCNEzJ)%*-As3LYu=63xs`7J zme4>R>&;sp`+@z=rsSg4Pu$a)Lz+7F!Bd$_uE<8_3qA5SwPC!e>vJYukK>^3Llg&amss<4NP60t*}G^!W6fuGQ6LQy!xgjD#*TP9J*j? z<16&;_;X5!qq@QsdV{Cr#7gmLZf=Aim$7pg>GLF7V4Tc3-E&wxHvqi{qn{-7*y+=s`9dYU{GFNSs^o$pWw9nA zq^-TmJ@_~73?xfk;6NomC0e=IdhTEtirxcmYRK7A{0 z;;zItX2C4e9T9Ks`?p0bc-KI`(u~JcnU6{Ey`6}f{&X=E_~A8sxtUQt1-}e}iMV<2 z4_vJNB)Jth`vw3RXqSprvHYoV$ojDV)mFMxX$X^E3Tr*{lq<^gv90GS2!_g9AJUl2 zM7}NBz@h=W8d13OQq#8kwmYr2t&Pr-tZas@#M(}J_Pw)6lMR^gO0ovN$joJrpWJm- z2@azMzYK37(ZpGV2I94N4_k#u+(Jp32i}61J zx@FeB`KzA@RyHgCCb?_a&bPu#@tpeCNB$nOOiEB={jt*<1RJXLcI`dQ^&-oC}z{R*L0^g*u*J2Ev*Ql@-f zn#eKQ02Cja*;KO7xYEQ$v4Fvk7paKAWS<(=ez z565>UVnD1d&Z@jgDfr$b8L8!_ztN4vSlRpqTJ659C&b^^Ambzm=*Gn6+qXFBzPnGU z?n0Mc^iElN=-N*G?M!d}%ja+H`vIwoh`jut$jl;$j&aR495r*{xP)srPKAs(_L_|W(w4)%E8{66z6rQ z_&DA`S1QZ#&&B0)rxYMZ?Otq!;AltS?cd_a_Jq9fXGnq0>>((uvt{rr%37#e!odp%_JJbn8kbGf$4zvm<=0*K#L zh{yHzHVNM)LmJ=IEwQ6px~|(D*EgO9w9k$PHx^<^FJnzxrJT;5Bj_hl)cWPUpbfWo zB5bjJ2a8^+>{}=r3KM;8SL9zi8?aPo!=~3%%Ula{)Tq!k!ph)y7CW_DFzgk3=1$^z z>kqw@=IKW!?V}PGCQS##!qlz}OCmvno-T;}?Q1Uas)|o&c(f~8rJX~7C9;G6x$*;= zj-q;bdSYiey|eMJC+Dyou}re+C?DOK1@L|Q{BEdP*aME5IJ^h0xA?N~KkEGV5y0Cc z9=3+DqM5>-d+%aOm9b(6==p*_mN|o)4Ejfwl5aEZ>0Dd=%_f}~J6orwux*zOlE-;U zqlW)V(9wh62&eU`bUkdfzfHFa{an>vIq)|ZwkwB_Xj!86EzF%hgbSd8%l*dXWh-g{ zS3~lTvndveU3CugWnP3E$T+vnA8jB6X+t+=X9MG0eNMMYULmI)iy=BX!}!l@ogTM* z>V4Ijg3+Ib#n7UngOnG{!w1@eeGyjADLH4IHmU+7e4`GJcg-*-YZnx&A>D3(vxPWD zhp6xFmd&O;2j7*}vpWCn&U0osNuas~TqkI?L>eGoa`OVGX4&{@rUozh9iB9F57cDA z(b}~<*d}T-BWs|#4Bs$(cgh)%ULyert6`T6&p1W2e8Qm)+_ z-oBFd^e&8nV+voYvAeLtVyENxBCjcQYC2`EFGE~qi=5QwEToWlp>pLMsol2`UV>;Y zD6pvZhgd+D`Bybntdbypoz}K?#1LQ+Gyiv0H`UkIjNx*k6Ptk>Mdkf7CnpqMSDta4 z^WoxV1`EnPjzKvRK}9-j2}9o~1AKN?vc~*NKfl%~^P1AHeajOr`0}#-zNJH)U9e|< zVollRFb)ebqXS1nsN{ghf#0bH@uONt2WS<&in6`^COF^^ty>W_Oq!1G;6{LL^1N!Up=}fpwo(+uL@6R21*H>d0@9>~7C?$r z=}lSyQ9uzwZvvr6ZwjH8sHg}Cgf0n1dX01hLXv;6-QV+_|N7jVTQcVwWsUXDG2bzm zbCf{-FjJBGZHS@wlLsFvMn|_t!~hbK&da%f{WoW@#QY<@67p4rug?!BvOYQz-CpG4 zW>T<3aQqeTa*2SWtQenk&m@ETAIVK~y#0c!KR?9dJ7&)8l9z7QR0?jjZx%7|F+i&o z%dX{z4h=ii))=t%Eh3C8mlhunK7L*{< z+xDfuxVk+> z>Y^pm1<4-COngXJbz?+>rEfuGUoa`dWJz`<;3NnV_IjJ=r)rWzq4`o+nNwn`ftJaB`+zGAa4=j7+O) z_!@gno*vtTs{@xC{&FK-60nTGS)Utt_P`o^r$pn zot`P&xGc##R`jbQxaoX4{Gsx261JACO(m#3&c<}nR=1~fh6bI>}h&pn-7hSAR7^Q&!d^Y2Jtany0MaEg)% zk=Qm+UYe?W2Eg^o7tQmueYLjDlii=5*h*29EZbU|WN~JfFo%8-VpVgM}(S)iH|#b<}=_l`dbpjzb*&C^ImE!zIt7-gA$u;$;>i*2oovd>}xy`e8sDy z7NRHDYeN6xV10$z8Wa(BAWslW;(+wLIH1v~FxIUJBJHE{NI_`Zfo2pyEla0lJDhO2 zReP9zyKd831=G!c>(f)-Kcy$<$@9zqzG5`5bk=65(`gnYD~;c0*}8OPZh76_btUmx z+N=z&rZmCCN{p{bBsl#p_KeY5U$jWH?8*KX0~VnmSXp}3>b;q;$+v|b6h|Ctisc5W z%FH-NQt%E@8NLKiI>-k?ba3W!0I}F432q+j3#;;HjG-eC5ve^5C9NGU0{u^GCF1`2 zJderP)z>SKpP`6R;U99X^*th0@(4K@Zc!dlqL*3R^FqZrMNyNX+eNjGM1R)Squ(rB zJGFIxZ{ZHjsy^<9AJBwiwTSeXT4~Wcarm);*$T3r?v6b{Ho0h`V2xYbCu+YzQ!p6` zk$uz~H5vuhnAa~W=c`>oXN$#K3*D>{s9+aLMIWELY~rOcG(Bm5PaT;a?7tK69ykbS zEa&Hu5-~z?rHDjU;lJ^P&qlmgzl>$Ax6wE6%c!sA9S1Cr`IQV+S^td~JByCHd(xMc zy!i~X=-D$#>T*-Khzf64Wqu(qQyyD;u>RLq^aQ79wld0god%fzFYDJXZ_7PfD)@JWr zCp6S~Ww0%w^3%?nRT*NnLX7S}s$dE?n10`i+hnIX1y$IW&~L?~a?;Ww1EdOh#(=bb z*Y4u)C$`$fleJHdpDA6B0_)ZlcOBRX%el-wpAJ3##?GKJqupsx1_e1KCnFS5VM_~T zy%$9vRhFwYLT@WdocMb`fgxO!%Cd`_-v(iH-otG^FY6H4u{DK`d98)fIYmHnWPaTV z^Gi7`40X`N)Gi*-rs(`t%krE;wu26eklqWOe1{;nY1(Gyd+>?>ZKv2^Gp}dSLFVA; z<`_rN^|7TGLs~c?2qA%zwsIKvxZ8@Eq*B7z3AkR zbu?2{Ynd?L+gy`t6Xn%R z8pt1^pGg&fS8A)YX9uGLuHgFyx~cmyOj*o;7XM6+0N1A+q>pRpPU^(J?tR_YGayOOwEn?4R2?$ zij_S)5_$gY5oFnfKI$)YUEOVt)2xQ@>|mW4S0m>H#TG72AYLjbW?5fG@E?69EPT%9 z21gf2E?d}M`r4z)w(@HPIIJ@zm+D~Y29PI!uw)A8KNH8upUSi3IFEp`gx-vfNLe-O zb~BrjM(P9Vt~Nqd@NvVH$!;I}1XTX!T@V@c$J8#9)Xpb9KdVwk1P3TJ9 zPleJe&cY@`3{h7hI2+%0Vp+ceJl_v;#c+l@(4LCJ{K_NULo!VwQp=lKevM3xOr>uw zrkEIyJ#d1l{gjR^dk&AUs5hHob{G4-Hn-{0TljObCol8AX7Z@iWAX}akO${k0}Pp` zV^ln7ZxU!DggDc*5_HFkujaT}|C$Q)6%T1(F$eu4t|jt-M#1IyphQ7bAq-0T*7v*> z36@|ITJSPWeZXy!_3|(RtV=^JPf!2U&$nJ;?cdZoPz#!gIq#0G#sVhsOv?&WsUsO4K|hKp(nUb?S2Y~>FFc%$}Eyw2)`=C|d`$Hg>Q=|9F+@1r5=l=ITX zpG^C<&Tw_1OJK%jKW6AFJH96`mLF>=&YR3;Y2{m#H)pJEdc6CUe)0xoz@oS7qi18}31S>kIeQkW>xL(aY1$N#wIv#hlGavkA`=21?8>lQ>DzEjlaw&0Nea zpv%qAV0y843O@~cGe;Y(HYM$RvT7sUxe(}Uevm$EvfJzWz~wUoolYK6^@W$_A+;c2cYvx>58{3@%h1Fve`_hVO8<9MCGCW4W@tpmbR=jz9K+NVu>j60C&_KN10 zTbcXn+xN|MzxUuTdd(~0I(KSaUaPdo{McjJLCLhCUdABrD<|3qxqn=$37~L{`3rnkqoHFDSKa2*Y~bk(XfTyG~>*hpD@s)d;f+}5DvnQYH=r?KX3gv z?&JFBeK{-9gJN!_Kx4P1dE4{IM&F%x!-KyK!HdLZImV|(W{=$*lx}5lQfh6fT6kZH zRv*14t)yB&ueQb%_vf@v!7D`sUHG;+kq-o2qYfBCe1^n*?9Q9Z+%s|H&Fk;t4QV0V z%-W1T4wN;uFdBdGtN8IEhcVMC$2*sPTG8lZSGeW{yWQtmf}{MX;{FpTPv- z?f2}t+QfG1nGEsIKQ!+BRD~f1TL{sHKHxAi4o&s(>SIhJ#Bx=Ne!xTF{L%6Ae8#Td z7qy(3)uJq4Pd6>A&WLu{Hs6-OVDS^+sUMR`JK3KM#2kq)Nh(Nqi?M+vVSeL7Q7KoS zw}%O_ewV+#u5omlqdh48U^xYLzi1)5?Tz=6d%3nIV;~oG7oRsDKzGPaFt4mkV+z%w zVS1rRu5f_n`Ul3%^@xS@hU1&^Fhl@ltHb^geFeEA(>GZo5BSWE z_)5o+T0M?&8*+}SVSCiPXKPE_yeEFQ#`cs?=O{AxjZdmxOS@G{f=1_XT&tPhl%q$C|nrN;^3f#W2>p$zZ#>Vchl&1dbW0`^Chx8ScL zVbq@4KYe1f0+MHrgB<1PD>kQdV~Y>D7-qbV*O2PmH_ojzI81AlG1-|i@rzT@%LR%2 zinZ*bk;jpCW-H5aC3wA`#(r~1gbZ79Mu-2$l)+Hy`ZrTv0wJV>`L61zK%btiRh1yY zRSMAfK-85n&#bbLtWR!?H7FySa!sx2HGrYsi!gP)WuUU}-Pi~d7ml}Cfr(LOv8$y& zI8}l$ih-pLbxI^fQA!yJib9&54D0ipx31H?(ZUWlagEjgNZ=1(*?QM_?Gb$__QpTS z*tlhd3#9~)+-jcG(X!t5d^e1c8kjebnjao_KUY^TF4y9$X||Z0c;?8+bfKz-CELHg zUa@66PSB6oeHL<3gRWiboLD<*G$Jn~Idxc>S^WN*jwy10PVdZq!ZKSK`z6J-BCTKP zWJnZ*ZT5HEeqj)-$*fjnzw9pO_vO23Qy9TwEpGktr2b7bI%y(kXD)YFs#05(m3}Xz zEVd@0t4{pRU5>;IhU9MJ2lq*#lrT5SR|S6R(8)28poyOEs;wWdnIjgwzaJCS)%Tgr z740NZU8C-Zm|sP4t)EydTV!~7A02(a#jR*PGjmKSIap4G`B@!QWfzBXRYuBZJMI&P zI=kv;(=tG4rGG&UhZ~sp#5@g+DyhryXH+EY)MKvRCqPdo)>AP9U+CmV$Qd5Y*H z6i+x^psz#+H|vpV)E|iwU3;r@7MK@THUD<%h#vzN%nhgMUY2$Jti_jFkJ^+)@bhU2 zr@l8hGMvAMd7)yKbt!cJKsYQQ{kZB({llc)SG{5KA)hI(0~{iB#T;AqeO!yzD~ZjL7|}=bA1E;SFcPIt_Mq~GY0Biu5qXQnvdqTi+b*F zPVQM*8biEDn)W(!H>)~KRkX~m>f0w}Q{}0J|C`%VAcRc{&fL`}Io0Qe?~ZTP%v(r3 zrWSUxJr;TEEGJNx&yw=KXXZy(mP1ZyYoiq3B9RrLogZA6OX18Vu_khMi9c%|8;9*! z^K_)9cr7Bl&Ha^>%;;1kM41~%E&drYmB{9o=UcRS8p2eO)x8$FMB}W`J5-=^esL#ei#FSQriz5C!m>pH?PhJ3S?TdMg znGsvF+`YwS4Qt0MkX|R0jdacEZ&pUl+rB&o(H?zy>}UthM!rn$Ssaeuvy%3|X6~Ue zAOxqpB@G|bQ*K@0x$Z)FQ8k-bRZjwiWB!u>V;V@3YJb;+)N=MU&t%>WdO&f%y_l?3 zuz1AB&Up94$e4e+-oLD|P9Y102w3?(y(@Kp|EyKxqmZh-Za#f4*18?$`;$K(judQ1 ziuAp{0$xy;n6JV6kX!0AySH*9in*`1Tp4r$p<=@z9NWF`V?QEtQEjqczE0A{*^v6} zG+B1X#Ij^>E_4GIdy1Y3TreT_Z%7CGbw722g52o-ll+bn)O!V9O?q$44XBIlG45Gl zoU#%V_)Ts*4T=4nmD8R1(InJXQA2c$Eiucy>?8Ae;th-947JbZ*C@i6ROTQ7;L1i#3~hP}fqSK5f}Gi}|&>}nf*A~Q<;DuM>N_G|^(9lt6Ho9?R7<+j{){V^Kc4Rk&B zgD1pF+eOTC6O21-mfycQ>umVf{!`Z;4N0bjM%IgssFpbw$2(yg*VJ8o-Wzee$@ZIt z!PTlvyTyIf_kev2_WqwOg5|n1C9Wyu4hB5u(`?-?qZM4*al8mLX`$=H8E5ujod=MZ z81hVq;VlhbL|Bv(H;*mgh>JrQ5Q8;`W%Z zcoTZO_BpCIR*7cSPtr~BN zDj*hh9T4x8%g~LtAzT;4E29VEhZ_g)#IIkxT^x|0t(43QVBFlak5rrG#!}p6CmX>b zcbh!R4&p{Vn5;M)b)dYif=et+ebc_|0gu<&SlYbf_fSj`Rk3M!e2+-=R4L32+BK0r zEcyXE?-H*Jd!j3J)3`BgwO1l2;P#$p)}4vS*vh8bFO87l&!!&?A*H(9cJ;F!O@uKY{AZta@!h;Ioc;HuLQd^`Ba*XTsEBgtjeAKd3eCJN3@vI)D<&_kSPW#Jp~Q4 z%aG@=GQrzdV0E}cm>T?LUf!-b_mArC(B1@xyU&Xlk0bM=$J#vOjy^&lJr7a8lF5ES zH4_SWp#`19+_W?N01nZLa~Oz882 zgJm-Fu56Z8%0?^&2aA2X^%wnv=!3|wv;d(=5x;WPoiK2XMj+T^_VOHf%5sSq4oqLxe z2;_JM?)fj7vHfgqRfU zA6@$jf6{poO4%8e0=yUA0kEqNmH`GVGKz(fFs73qe%Pol58kx3U+AkC7mor+W6&Nz zAegi?pysxnGPOgOII(HlfoqP>R2c~LpcQtsbq%?lgox}ak=Ia7R#ahOm%MUeb!D5U zGmP{VINk_^HV0~^$G;B-%o}9+=Fdl~a0Zp;H79hls}OyyDbrV?z-K^tqrf6I3T$9N zAdr(SF#Z`J7v$s*7(fR|0bw`p5%(EzEFe(c1i(<9S_isUM?)9~Yz2YB9q0iK0dqj0 zK7ePd$s@^j3snTc_2cZCQeS?X0vcvJsX*Z{fDHpiG7A#GhphLAO4%4xdxt z0%R@zBWt`ni{|_lq~-Oqw&t6%t$cA^@xI83LvPUgJO2*XhAOwyx8jnx=F1kADB>gea$Z z=mx`I;V`>#-NQuRTYH}7SN=6K&qTO)PCCl&`LUQZN3E^T@%-(V@NK+nvehb{&M=>1 zQ~8STESxm(^9!&zojM8BF6$@tcsKJ+>85nNj{wlj5}~amA-wMFVdLuXnjB8Hb>=;> z`Dm)enwU?IpZnLfp!V`Kjn;O`aOa#iSgin4aKgGKpjiBtlMTTpn5o?C>DgGb9KVq8 z_J7U^#d|IcmDr@eK7>o|Q-+}ePG4~&PRmzVgAdUXd*T)}|FOZN-XJi#v7~{ zKw$mz;?BxfS8{f^;|u2=fTM>*l@gPLR>!0KC?Q|PRgl&DK;ru6WzSb6kHdboKcrMa z``&RYZuamq%5MmtQ$UZ=;A)cZ(2Kgu`|`lOs9Se?G`P|I12qiil(C$rKc;u zFBs~k@f)!E_Fj`YJ7CJ#3rq6=-lWejGecX+XwU5YXy6m7nm{glRO5nkubqi{9QSn6 zb^T)Vgz2TAhgA$f7w=B*X`#okRybllM0EME>hrhf(X{b&Ei(P!zj+gdOxf=zJq-)f z(>uJx00LcR4?+@-W(Qbed(F~9{X83W>Y`Ysxq54EDG^as?`oG8fy07YGi6wza!4!ykx4T@;dxH^Yq+DZjJhazl@%`9NXQ0h0 z%ynj;-oryqsZKEI8hKAoxZ=~T(Y0=x^!OF$6&q^MJs~5Z)L;yGab!BDBTHcjYv=qT zk3_Nm#*Z%;_Bmvm@YFN(4HHO3Ba{6WvX1UkcL3!|zzM^N8LD!Ig}LIZ<1|hbsEn7~ zYU(Dl_s#NI6P*<>YcHGiz*QiIq*Pe`27$~aU`5Y7j`yi>6tQP%q3QZYR$(b;wS7#T z<$UH#_hW_QX5j_^)R=>vC>Uw>H}XN9qFSu^2X)a z72Fqm)8y>o>k^;P^>Kth@h#AH@Vq9&+0Nm~N_(ydcxd#38i_VsV2FP+*s1mxg7%hiZ%s8kE@CNsM}*! z5LWNd6E3Cx;q56vGX6yvH^$tt5L|ydJ}x2|6?zOvL)&8!mRDcLSCpm-gNLTDG^ZL^ zok3fhFm;afEI4rI=hL4GZTtHa-$2-g)2If8H_)9z6bLQy56tGdokAH1akZy+a& oQ^*6|!~gSzgK!-G`>7|8Sa~Vj+tln#;K3kGHC@$eR literal 0 HcmV?d00001 diff --git a/example/ck_tile/15_fused_moe/misc/moe-1.png b/example/ck_tile/15_fused_moe/misc/moe-1.png new file mode 100644 index 0000000000000000000000000000000000000000..91a1f2d9dde2eb892ab621bb1fdaa9e1f7f23a8a GIT binary patch literal 92535 zcmce;WmsEV*Dg#;fffq1xRwIN-5px2xVt;W9fDI@pg?eUDDLjT3Iu`{cMa~ros&M# z-uvC>T<`by%dgC2Wvw~K9CD9)jG0g+h0o}yM5qV|2s3N3$7=BR03loFIT-Dqk(y#B;Mw1!fW21)EPoFWl!Tr^k^UK#&=`UK8GO8W0mT1j zRFoM2`?Ie9zNWBQ4T$1exW*>+|Nf{G(4=_6?S}sanSdqYpPL2{ArVwo1i;1#{+V5o zasX@(_5c5i#*9EG(*K=p_S-iQgt`|;XKqNqB=Fx0Q^SOdiRtT|kNz`gKHFY9;hO1x zFaOB$pD%QM|MSJ)BAP;EHTWpIs$rM09;KC(@_}3;V2_Zm{ z>R#Ai{BO_u>$bS;2F2p1%%RVYg%Q+z&Da}M?WjkGCLjJxf&T5EM|A@J^K-QxfO)Sc zQlBQ}u{mo>fiNZl0w+a+)_M9b_X^WJzdzR}yd+@BrZNJxJO|5~(p!LTf(mE8!nY4S zT4tzK-bi4I+f4+1S%3_Dhks+S8vMO3B}(`_5fF-=5exv9zyLPB%`(FSPU_bA^kPoV0MEPwE?HAO;0YKKR1sUZ>p_X z7etOT-;4ZB^3Nk?3d(aOC~d4|#^UZlX(_04=_oVnS}w(Oo^09@QD;E+nn+(kfWp0V zthCJj?H``p6HoQ@O7Yx_EKYL<-qfl3D^>_$k}*Wt zHJsG4UDb;ghJF^$9SN@$_hd^M7Y%LEM09&x>f+g1M5-T#5r)=E1YNnv#-c*vHu+DK zw#QP!T*Ud3#&NAVjMQ;afSP~lxl-@g0KojaN!T;VPlE%OsTPUD6p3_)0>hjS2PU3L zX7iW-@f+F5MsJ{04$otftB3h!UqSdeQSDcnqlHAZr}Lw}oHnu+h}NR4G)=%3)}> zze^MEF{(}`>GdZvH&K>ID|$cqE)k8up?+&_`zf^r?lqGGkV|#w^3)aq#Xq%fTL_Br z-?ag4PY#bX|E!=%=B(F3IFXlw>$eR0bIO=Fyg_JOo**=}cC>?vxPE^}Zu%P$Rz4cx zsS*tt&v*4`;JFGSs7aGQ<2~19xSUojgJjz}A6mZ=%M=qX9!%xc9-`X~Gd;3< zK=j}K(iLPNe$jfoDd>Cr;ek0GXL8pTt4NQVIs4D+brErQB6fA;M`x4vr(=Z*er1dc6gC?$y?B=MnmXJzkd}_Ny-rml{m#(II{U(RbLfepd!u$k z)y!qmG>BPwU%6jW%2S^IjHvFdP#d`yiT-1e1M9LiI^3mA%F%ZNiceGU<0Rsx4*3P3 zA$&>8+VyqaG16wiDTOMIx)*f z5;rK&LmQi-Wn^x{rkYXa#!smd%u>^%jxcYRd2?#JLszhf7{lLQsrRnRoSj!MUX_}j z%Vr{Le@PU;BXfP7s=eQ?xAO5Pt)fP9L|XJ#8}@Q&EdHm9-1Q>oVxM~8m3%f%ygUWH z4bt%vm9NOl6sAhUMLNS@yx)Q9EzV< zXUx#qisfEScUe{FO41YpT!Y|ShG0#dU<{gk^X2w&Mhc=AJVi~jZW2ZHtzu)xq@+tF zLd_ULyZ77tl+GEQUJcUP?NEfR0Vyf!jxC-57T1Oonu~_FmISbtzi55$SL4(ysCic8mArobPh8Co%N+xlvj@JZL*1N0dm{|`02E*$+hpKT zF$={?{6@1gbB+FVaP!^8bh6T2yvdxLksCB$1H zyc3Jp0U9ly`eAOY>*H6-DKHGWd!eQXsef44#2mOpv49M#(yswHH*Sz>84oukR5C-~ z>Bl7CvCT5pM&@d5O-qUMWuk|!`@nxwZW_B$D3yaW?ERjcg}(!ru;zUgS>IIa2AG_X ze2r)KSA*N9rT3>CK~WV_%haqlTNa%nx61f4LV?VJFELfSr7F zYIuFr5;PdR;Xcy_9+I;yR0zSc{N#K_$wXE?M%wvci{j_f)7gS|@kFB^B9i7Q-i^xr zs@^nd&W|*pw(O?QsvO{IBp?&*LM2Vy2rKsN|1>$t%b}ffdSZ2|p>3?genC=9{srzz z-ux8@;Lt#wL>?U8?#ZhcZjY_SJ+j`@Iu6vNO_QTf%oDUE*FWHogsg=dELu3<2(C=W z*5Ol)y_L)C1WkNNEU%?~Y>=Z^=lLRwvxR4P|Ct&u1qpMr>4u^-J^{zkh9^-|JzW#v zg%~$n|1u@)UiEsF!V9&TSy_cv+H;_MqNnWeqo@{z=UKAd@rZjFl|qZA^FHswz%r=s z1GVORY&_`+!=I0iZK-I->3F?;sUD<*?j_xC(8f?uUNuNWLB@x8sH$m(1fqmFC zTsma1mo?U_<}Vsks}T))NGTfAM8a1mz*7irLf1<8^cOPtkDKNsV(H?nZP?+$wIWZE z+_1#UnCo$vM=H*}o!f(cbra&D?QvEn|O39T*@y-Wd7|1^>6(pu^i=?kBWi z^FJ16K??wiEB)gq=z)oS%tsRvEwk-4UdedDh5@w;ogv5xpvsvbAffYem5z6;LkHKS_>PxGTb9K~7QN)$Fqgv&BkGDhzhwsru-5gnL5ZDT?r_S>)H)~S;ipN^CL;n_5 zeD)PrFTO>YbxsJby|n+g_A(?mTJvzD5e2fxgEPhPbM94ek_J#~hFT*n1!fW|XZ0HD zqXkJA89x!xog8;vSV5wHm;qiD)$}HYv)GCm3^<{xK3 zZ;*1)bYt)H&buV2b%o!(qGO9=sM@&*^?4}aAU3DS5qw8{OUFL-OV!`GO&2xw>WxSg zU9)c>WUDG;KYQfIx@odSgRQrw^Z-YuN9nk?a=)!|Bm9Be({=+k<0+7}x2Mv-7#rGJ zIT@OFh79j5^lu^rLNiwwA9!LRW~e$BTel6&ZNzk5Uf&>QKv%fex_{4R;&IA`1viqRONByJ=1;OGCwK1QHZV~9LIr5jEkWt$+xhsDpikLy_Y zRk0}~;+y!9F0y8)`$>GLOmux}J!ucjmKH=zMMgY?U3Iv)F$MDKcNh!bU;4SRXE0WZ z!dTG>8YE~?&4^74>6LccIt-zx#1n3!1^IsxR@7Q%?W(z|+u_qry=&mL&XuEBo3^u< z_M>H=T9*|Vp2ye#aMDKb(JQos2ou?&V4m4y#|*<=dyo@;^mj@RxTn`jyo7tBq8}2+ zgtWx`iFq`qKF?p+-G&>26Z8a(ru|+vR`~Ch=l=XI&>s>rn`^!45^4x$yTRKMqOP?| z4;y`mu%`9LVnGBprvAaae?-cnPCZy22GYzMnaM}yz?I}Gty9`7vIt6MEg2}uC%wV3 z3Z1NOpRCU=MuL!jBgAnll&nff0>KVhUzQZx$4_$^Vw0|#4ITwMFn)gbmo<%q5u=lmY4$CyKY@Ncs_DBNB(l^(IuYgc0jiWb)P2)!{>_Z1AGMM>mr8V@7-m)i{Zw9Do zq-I{X6Tj%#%d_SIf<|%NCfTur8%`jr>Gs_{3Z3LPkg69B5DFllE z$U=wnw=lI@buEWbJ?*fTOLlhMJ|z!la`O_$*ciE>a$tXX6wb7xq#_??nFJ~mc|2;I zta!>PRi%~>-b!~m$*ZC3SzPrI6^nR@s>9Z1+RC-Eh;VByn=twlILuP|H_Y0LNQ#c4 z!bpjH#4}uE_Psjo5PZ>9&%Hj&!%X=Q0J zaYII_c8FtQ5R-yWR`kUVh#^&9*;PMOJ)|x!f`p zm!)XT9OnSuKn)vpkl@m~Orc_6Vidff5u=b)q|oLdGqUHH!V=-z78YGyXha^&4&ZBa-WtmhDyGX_X-`0)%O8=v2|lfeW; z^s9D239+1LT357&o{5=5%LQY%2=%5oM>DT7*5_5+u$gXEbs)Q%-H;gRhq7npZFyjw zvD#Qcws9lt6}1Yk{0FRoNtZ$lG@k#*T_}r`XoHQMct_mEVm=7HlHUHjkBj|6)o7%_ z#jyDanLk&Tnz`3%GIEV_P#`Lfy?7>qVFDQe;y?WO{NesQYt zGP8nCEcHOF{1nJ*(NiaEm3T6`>ia6*tbXY@2d58JCJU{DJw>(ycjy3Zl7?KGxM7KU z$oJmY+)2>&pn-wAA=)$d%oHCqDm*l|V5%x{d*D4u0#jOu9{uqB33e+mn#a;A3M&Ri zA@Nx#X@Gc&qI0l_9(?y@S1U84^tjHv;4t{rrxC;GH#5$j@y4Y{aBN@=90Mkm#O!VL z#VX;et6Z|})uI+HntvAab32OgvN^JmNG(Z3!WA7>i%!Sy!;j8UV@gue;JtEADlaio zM}F~N*D4pmZcwe3nyI^E?C`+PF}gfHGypV&@xKyHjLkTJe?Z#cul=d++whH{d zG9X3V!3HE09yK?|<#)V8quW;+QE+MSbsuRnF0$BudLZ3cYFiMJOI*^k1hp^yu5O_o z2+Mxz7}eA8)Yeve6*4dF>2iGcK2BPsBKiPx4=AUgtq^R&?dhU&Q)|AINbW{MJQZAG z4C0~YKHCw>x%|yCMMeMpV+^?^SbW`-{q5$}yIR)G$;80x)*a*S(U}%k!eHm0rtBs@ zo&g*RGrtKp2b}iJx&&6vMdjG#?;x+2~(q)RDVX{ z%|3j`{!UuBo)7p;Y!OaLDMhfQ!c=f&jZ;h6HJ)raQ~{9MzQxsf31!ls@sl^~UN$WTVZ+R3!GScZs}jgR$K z>(J>U}vevdTn^md?#Jw37v^ zKYl`Y;lY=zFt>x3LCIN=;HNqdB`^DD zSd+i7$m1YGO!RM0niQt-2yY~)M5_yP(Zh>*MGMqEJTkAB1qJYcd~1W>jZ79+h-v^}4{qd794k4gT#Ux*QpRkKm1_bwzZIkQdp z6dcL&i{mU@sb0>rGx1}2c3?e;oiaM%O;AOHhJ|3o>F$pYPw!3Z3s#NU?y*=$YS3rY()G_Z@f+> zxhyj(m%<^{4as(8Odl{;e(WI?F|E$iD9WK3NKhrC=%wPq4Y7pWOEp)+R45^8Vv<~} z1Lp+rZ@Iy9q=aYxO-7Tg9YB>FF5AaIar~nEfU9-8>8tJ*$~e6y8td)twJZm| zV7)O2czvpM@#yg)T2fegr)Rh+Ni${Uc?r`Vkmk|?Qc4ue* ziAJPy^bR>n=k3?{i__emj?un`6LW8`qtdeqx@?_L4)}wCKdC{9;7~V+eG^L*AzKk@qJGUutS( z0E{S@-Uzf;P}RQN@qCv`fyP|L_%&;Bla@w8{@cM1i7GW%ZbHhMd~=HI&fz(dInp-6 zzK#hbqrla<*4SVm|0>s>?&e44DL-^HIB-=n?F4)sK={ik*7Gc<_>*3^xg--3o$p+< zS`lNCUR#{^hP-Mt{J=s?Jnz3NVt94@ySaLlXF(C~Rc|TvrNnLrS*x?_GOy&EkKBg{ z6S}U=ssj(lc29FPFa3m^ z&r8}!ANCY>u&TBuFKiSI?NZr_s>FHyhTAF6z;iwX^75ZM0qA0I7TIl3gFM8i&KB$R zqqBH*@Q|922Eyp^Fv#=gVsAAkV|I{RVNtc-_`=l1_r2wT1Wj;_`s+Wo zJ^gPT-LnHVOHfg2ikGDwYW*FQYgJgHBd^vs96@;Wqq}>?%W6X4Fb!#;q&5qTerS*$e04+T0hj(D*Rzrm0_Z zhv{wT=lklrzPCbYQ%(!ckC$hYLuA1oWS0xfA6i66n^*;ltd0}Wl{42mj+c4uRx!Ot zgX*m(S*nynl?1&mmU}6^g=8VTZrgNw@8|>HaaSC+2%lO=`mK3%x8#;=*@}ct&ZJV5 zQiQq=mHgNxKji1RlspphGVi$@4WSJ8%S_}e^e-Iik1oV0JNy!U-z;sHq)e(SE~21A z=H{|cW-N#i;a0tK?!{^#WX%4ktJi``P#H{D#!DaGhF@3|Cl-WTQnYwefWNXFM50sn z41H`XPhS@DcJMmDx@t1Nd~z61Be|(mz7@_3<0F)?8G$~U|HWGtco92D`i=GDdQU4H z9dDX+9Q1G2rQfS#-$m#&Z>6=c-qCGFY8R(QEse^Ye^CsSch|0#PB-94+?sC?O87#>2(g?X2)83f*NM z?@6HiXK$`6xJ}+WJvI9pMoIdIRU#u@43C%u=ik#MJ}9Tm<|EH)=3_)>ZX*fdGV8)07DR z4%W-yaONTN&&(1cp1b>7Z|*2Otw$S-%ZVcOYaM2`*Zecv?d_Nid+DhkhmZRzY}5FL z(nlrs?$kRl%J-*KibA3xVKNc}t%MmJPbS+5t-6*NaAl?`Tk0@!oUtq>OnOH}(I@${ z_2S-+jHj=$`Nl2dz@LoF{dA?UR6*~|seY86Q8F0hw3iLpzl_?O6_sJVG>BN>_09hN zKKP+w|H{euVv)pO*s&|U?~_Hv@AU?>>bEmTlpVeqG0@}9M^ix5*FkqiGGd~$`-`Q# z$rW%Ad0bXI?BsGZctJbWxA{WcNJz1za*Yui@1f8xfcb}^2xA*|??!yNzlAlF>-pqd z19IM?TUE*PTS7QKAWVz=Qq+*&=)Lh%8u@_6-btmgGe6huvCjnd6|GdkV1ISn?Idk1 z)rDX2`qhJPqg%`~Zwv%#cnzcI_1_gY2dH^GWUthN@AuR z%{Fsf%`(R^>5BPoD5`Ptv!j3!&DxUfK1fzM;`E79gEepAz@mu8HW3t;Zvv~Qsr+zi zr@2|pkM^um@gpU7_zg7&zVD7wP!X1m%O@}UoZNbeuRcpw!4b^}>IwQ<2PQVXv}jJc zuda7Ecs)BBMsPfhR_)@E{%(1>W04tNM%)Z!_B(g*U?9U(^$=(HzXO)R zZJ^$*whT49JGECa;p=xHOTW8Ai|bsl5Q@ao0gWHFr0RJL1nXu88YRqkn4#T)T7bC{ z`tVPZTp{;$n+hg-EsDsmt!CR1NA_l}a3V`{&Ca6TVuoK41O8gpyCL<2B zN|CxRN_+0JRB4LoZ;05Sg%oiG4|q8phpx4??G0uvm-%%(2JzSrT&ofo|L;wogLPes z)SyDd_#F(l%L$)D;c)J}>RZ;#4&%&;=e zTCv~iOxJ2{J#JATxKwE8w)y1!GehTj;>s?67dS0T53fJ}Q@Dk{Dd$tmgi+9?);{2z zO6OHxERkdtnlH=CeZs`Y*uB-(@(!w@g>56^M}Mk6Ic+-SedWzkFKTWxrQ)HUDA7Xv z*!OM{6>A=E`Sewal<(3(nNw??!_d5s?QuP?^{1ywje<>_gIkKlsxGE8<@N<39L84pJJab8A*((df@V9U9S1zYJ}X&qhyjrwSX$~{mwGVeiQn&Ex+co z#>8_t3%cB@>XEsoiPiQkhlv=DA7z*D!+Zpk_z5lg_;3xh%%QzWZ}_tfmnit$@L-DxC%Y*MQHzpBL1{96ji8cL)!#9Qx*;JSzp2zU8-nschiZq ziNM(K)|#JL7S|Lg%9&tY-sGIu`qnVww?(+VIrJj)ks&s20*yvRb5UY>{BFC0knHoU zqqdpw8H(GO>2pgri3}KqK|A&?iF0$L->JY?flxXrU)*InJv_o)QqLhnDe&x=z)z)| zK$vxI>uav=F__|qoz=6d%RDpofr=OB_$*=)j}+f36nS(RTpdqWXa;hzaDe!u2NEdo zFyLl=;C7>oMIZET-B8-`oh_(H$2C>M!e_Fv2LE&WYXZ+rC?SZJ_%^&kT`~nrbzusV z)gYVwe>ggXse}?M(2b(QN(R1#9JK9Bz7n3%ZirPXuS~zYrOK5970>D1JCC;*zaLrL z?H5Rw3J*#pOt1j)x)x-kzE}5PE&(CuA{7-;hL0V2BMF6nd>Y^cl65oEhJ4k zaiKi0h|}P}ZL*&a+MV!m>2=EDP|9sL3Pxe!CXP}i;a>gMUKQiX0REFvs!2mmrLg>f zsO3-AS+InU?ua~}(Vxe){nW@e0$sl-Zu>e$7A0~{8Z9trr#vbrT0a&!Za9@N#pD0dXZ3>Q)6j@ zL$SA~@$@UlyX0By3^L!A-VF~7%sC~kM{P4 zdhXVU2&u(BB$~DPdf0CMYCjx&x+*)&W@BS()EBHnO;AG`k8^2pJ#sa^q8^@W$ZX++ zwI&>myk-t(bxW!bfb_E%7~>;3>c$#KJs;jUrqM*9g#zQ3z;VL zaJd9a`9$n-`W%XPE3Yc0_}c21R<4Reck!=;^trIzqAa(=_QWx1E-mjKgC(PkKi=GF0^ zJ_nY>9kJ7zj#amt*hpwXYvW(LegvCBZpXaCe2FVFB+KcU8doPf6pP!Y761t%7y~oJ za2(}^0OfiQ@6|vCYLZg68+sl(t!LJE)XXtL4cGLTw@1FwsCL|}ZG;8o9O}fUbu4AS zqJZrahy03}OWwK|N4GfpEo46uF>E?um2=vnIt9eHV#3llpHidC+UrD?jg)I= zt4{fR{OVeZYLEVN#wUN=9%Azm5U$lG^uK(Z+I=t|JgKiscCg)09v(2r+b`N8B)6-6 z`a z4W-*5F!Z|~=pxss(Q#g_mf<9*Jk77LSK}aH`O^)B4sS4mH+iX}urKc`#$GyHFMR+H zLTgNqMN-Z;MGqr189lMRZ9nJ(8=XcoDir34j&oT?uA3*i0W*5`=Tj+0ZeZ`#t#4d3 z6^8e_*wyj8u4l$G1;WhSi3u|qB^Dq8+k}1^2+oH@S_xUGI_uORVtN9!QB7&HIA)&-Xf>5HK-AVTNfVhes(zd z`+OC<+K@6m#9yEG-lS(W2rz&r{BQo#V^WG?3%Z+rvY>!&_AL7g-jqu46CJ<2B{*h2 zjJR6Ewjr2VTAkN>L4}<6$8P)63i0LRWkOA>6-9NqS6%qLE0?a$T-(oCm1$$X zquk9f6WiXJNs&NRmtK5nSXIO!LDLeJV_L!!fzu4Gs1(`v8xr+?HIEP zEPyXB8`nc-=8v$}_z}3rUqaYMdL4q5FkXtG_o6`JM! z-%x1sLGl|jF!`EiM>f`0a}FS~;9ORvrlmyD(~9k8dmvQm@#dlr-ue{qheY{XrmV{Y zd8ZLGA8IS>IhOm6sdMg1si9JjQ##v8P zvN|B9pV{r*dv2QMrqe7#u5TR66nZ+}JNfqbI$_2y!k(o1^Np^T_xoNdT^h00yNZ=Y zoD;}Y?IfW9xAC4pg%PMIZs&mW=_S0S^EbQ^5b;k_UngKHrwb4h8AhcN-)~X5 zS|Xtnzy{BGyc);D$_|lwpErn$!h4IFSMax=?yUu3tzO75mUI>6I}wrBt2>xqC@){7Gs z7kj+3TH;H<+6GpKi5^XQuB!7un8`Vt0d=m@5T*#nZC@v?!+x0}Go~(%V?nVgQTNMm zp?!N0Jn8erb68?``4Q4Kd=?K3-N!h1@S4or@!qkoQ8m4>e53EZqv7Cu)!B1?P+#Z} zA4Vm}y+3`5)WgdBF*Ul?p6dSZeBDF~ z3w<3fWA@EmIqJA!6PcHRI7S%62^LuFDO#rnGg^$)GG>Vq!BpmRv*zB9KWko)Q+D^# zN8odP@_EU~pfnMI-}d3VVBH|2z&ZF#Z3ZDLDXd>KTZJhUb(sX%AnDH)bo0%WuED=Ep_{z4^Qi@jhG zaNZn%ihpQn((GQ~QT?re1eJ(vxzZs~@nD4XFF+^QGYc(#BC@Alg2VZ%;Hls{uk6#! zDcm^!&S$r9p8n&1`-!)<&Sn&2C*VkrXZy@C3>lhH;|gswMd2*jkBPpG>@)V}ofGm- zurRkzOwo+a%-Dyn>C6mEYiMlkmc}<~H@j=#4ew1Elh2v6yPTqt(NLlFn#A`lbmLB4 zv8%UBn%f)4FKHzz(UK8)kXcv_3a;D{tjG@kvRL*7?3Q@g?R zBF`^EQ2qD;4W6E}-k`G)ILM22>kQW+W`&=E?Ai5fPtg`O$Q%RO{|PkVd|vyo=uH6u z&Juxq%QU~Q5scR$Ws$O$TVCaFFz)j0=JPC51^4L*WYCc83PN$E^ZZGonDN5ecUN|2x|1xfj=z{wUkk4f zS^5TMlrzQFo@!(^-MV!|+PiCb#HsLPwd&Jj-L>$Iyhg9VIkE4MS_V)DkD7bf=p4I? zvU?@y9hUhCxmr0#P1lpys=1>88@G&3iA`9SQzhNl4M>79w*ne4 zj2(1>%4dI$C2-9F{P}#WE+94$f*X{gF5$9<<{MSXaxvmM^ocKSrw3%cRrT~hHp7+k z1jnQPWOnPSt%{D}>ngA(?qjqsYL%W3EiTgsWrlqjgBBbE7KaMNv)#jEmp%vGA+H>w zEktB|pxc`_hqHeE?|3cEnF*!@n$&2?b#G1(9zB0vrQC};<_sFjs2}E=f%<{PS%=-zfSp;mX@G*75okBB6t3)Ha-5s()tx;~AxD1~3U` zyvJgYi&oe)r(J(?xie9c^Q%lh4uQ<^PbmLSLQ)I?-%UKC1?A^=M81;f>vIqx1o?yz zMlJMDSotIcoUCqf_%Gl*(0uJPnOz~>&ac6{nVt24% zerY%LW=RTZy0m^@?mYQal50h-+(mfGrw85`@D-Cc zo9A!ibA%2vFyGke5%3GUo~?PjwX_UwAfB&C`If*ZHZh`ig=FUdTu}d0%i2^axQsd=2+M> zp~h&ja|3vQ`<8}y(%q0X=P9+#G6oj8KT)jrC(c+n2<+}3v6hITjk~eAavB7Af^=DW zMpkg=SbiVJy|e2VdR z%jW!Bsx`~2R`DRabN*x<$+}uS#?_+p^*fW}EL4=<(6w5s{B?eFj-Vl=dqrf|Ib2Wu z&TWhx99pcWLPL6ZyA?znt2kg#SvUzEV&P79Y_1+wE-_pR^~@-3we4244&&O~S@`~(1^er}DRq5d z(q0Pz>oPsY;v2+|T*$(!0n7gr?;gRslWDn#qSEx!52q) z@)kls%{e!Cy-=N`i5xk=`n6BzKPbFFC4xtQN0An5Edd~p)WS`AU<)v}p`^=hgt1p` zTF*S&OrMX3Bh)W8AoUj!;Ja&cXQ;BFu1;zi?8eK_bigcyP&>Dq|5%GtHSK_(u#$hK zHvp#L1ox}(3gyW&lyhX@44nHKPnO|Y_KOn{ak)uNHcHz&n6;%&l1`s{(7e-l92nK; z0Wc7=8pye^FBev-5JlZuFBf8?e0R+mnVY_(GG*<>g8WEv5qq9$tIsNyMxjGgX)(h^ zl!(}F;fz#pi0J#3*aY58(uqRy8-ADnW3RJ5pjX6&TE|MliI@O`N=u z4$5$+8Q-I&_n9xY;ET9V3&N=%nGY-c?5WJ)1x~ZpS}`%Eh^yX%3E`ugKm3>##MtFLoTPYGxt6!__WfpwczAAti=5j7OpFSU ze)hxG>;)?}Gb#Cf&n8DclAyvbmLP(IHoP32!;m~OBM@$3AWu)OWW;=%cVn^QuM;2% zpYu$Ge!-ChF4D3q5xhM14}$o!R(yolXz~#=FX8r)j=KCN&3)!v2I)v~@|gpF1pEl# z-V?D9&|K zF!2(8RPpbC46jnb%Ti{zI5TP0wiaK0D^hw9b-2O%IpEKQH(g!Wbv=YyugVv`H3I$P ztW+N&JZ2x0T~KryzZBe`vxG2P;@~`DfOSp#nQ1TY?xS|e?K;S&m#)jd!Ajt*$?F6w zfFz~`zqMb7@H0DPWTPn@AxJ1ZGX?P#r`Y&cK1YjO1H1dv!)KS5zblxUM3mv@n*P1U zGAQR-UjVR zQ4J9+OZhL-p@0HB_y#XAUGFXn#%P2Jlul&TLU0`Z z%y9AM^o*|E&00ENWzU5X&*dosML%saO{6KD?iM*2exj0fc`qy`!lNtDO+IZJ4KryJ zL#x5|Zn3#EHET$hrlzLBb`A>%2Zw~D%!5iD z56*1pg_UEBU}7~xl~X6 z@M$EY-2rrfyBD&iJd9$WxinLd-*GwtuN}Krv1f5)$NiaWK%hy^@z7>gaDy&qUK&gn z1Wg<*G%jgc3=@7aJYD3V$6Q3VC1Jo;SwMt9bdsGipTm zk=VacP`nL@gPU~D`LO{Wiep|qe1rTfDg#EpJ=J0khio{{@_vgqnoow{VbqhX^a9Nf z+TILkH(Otd&}7&#b^&aj-}L6Qa5%$@liv~#N$@6Jy($}gnEU#sx{`en{#s5b%zhED zbFvn$tju$HpdAQp)8WkPxL?hg)Ip9Tv<6;Hm+SKJ@wIuiWes5w5*FtK4{$FNa#)lZ zb}WyKRJiR=Yiny)S6A;`ulI&OK0e}B>V?ms0F$`kMMy8i|B?i9&0z1N$MBHKl^ShE z+?1aDUzR+C-vz;g(K2JhoPD{^Wc=W@=7(|89PwXfHFkj56kogG29DB0vwKz+0+y4^ z^Tl6Py}E{$TkHdVF)?p?%9an?$(Fm5eOa4fI+w}G;06&N=wkqPgwR2Ggy&HzyvWQ; z&1T?zSHZ`3;|@=c3o*4pTBSBB4mVxU6lF2EnN8IEkhyvF>Q!i1m`c0Vr(ZW45Ny}g zqay`L?Su&Yd`|1h{;stSvO%RV(c+#SabjZPcofpSf&v|F?P5+XJwE4S{Z{uiRHv}0 zsL^nI7FS>2a|B?L`@2 z;936h;|F}>Ha0d^L(aBG4M6}lF$h4p;Vq@4_~*}1t2QSmr*9~j3Q|%ZBf}ooRr2~8 ztf#CYu~ewIDz7jxS6V$-=%q-(z})t$qa_Z}Wc>eUx#ne4_l#7Od&_4rq`DXxr=A)k zXLriEyI(*w6%TL_6t=h8;1rEAlK)Mh@dD;Z$8&sST4+d9$R1kJIUNW8z@ zw{+P8P8DmuW&+|1SbzNZ(abFG+qZ6T1aYB!S};?RH3+WQ?d|R5x{YAO9BgcC<+tA6 zcge{FYGEan*S#ddKD^up@#xKUr*^Tvgs5wJj8nwiw#oc%C4R?2%$1k~EO!5jP|158 zWR>k;0oS!KL}kq;ysd(rW|H&jSR7n3{pKP=$D60jdP3e?jdS_n{FGv^QjVbB3~$%A zrfn9}hMT+UonwlTY3C~8ddHuQV_c)FfAYt(a7+U=p}tJU?)%XY__0bzw9}mPGitLM zlbl(w-ig(*J6eb<*WYp|fKgKyEVf+yNN-9@NZ8()Emw({SvK$s-NRWm|6&K?E%Q3v zfc*ab8*>=aK~hs&D-A_QMI{pP6O;*!it3|`PGZvK_BdLEJFE9IRHalAN&M{EOT5|_M%n%`y9*mN+#vJX8{r%I)K9Ot?gQaHN4qHSwtEGSMQI9>1S!5f$!tNj zG`BH#6Ko4VX)Ea^7fuYwPtA)Z48lx2B`3wm|Gk5pQSz~xgR*0`Yh18Qy<2+ed75`~ zt+u3rio#pg{i#Km8E4;TwHEd0nP|EDe2+Sk+A>ceVg}HtXLARn#MO6@{1RX z?Y?QLfW67$sw!62Wzu*)Xv+t&0k-kVK`RL zkejJ}0*XXwUY;62hPyWe4Qiw5Sb7s5@EojWvX3S+%XO`*lYmYLaF;0gZpvqGvVK%p zH#0gK7(1*#`UnF!mJ= zQD$wt#wrG(AV@5rARyAMh(mV`F@T5+(nt+3FbG2p>CiQB zkM4fo=dSL(_YZ)#&Uv0+J!cXNcedON_}%Z7`pXz7?tNTFe3)J$%jllGgx|GVpMz?& zJ$DoQa3^IPVkU*L3%AuN?q`VYj4Jn^7Rv2w?nx6XDk=i#x88HjSiDFFI(rLL7)NyV z86u)s*j#tY)hk!xQd7hG%VzThjTgnui1RN=P}0$f`R;Glxi7tqQ%(_Pxp5=6a-jq2 zYuAxtO!8yF0pt-k3R2jtu>wX_3_azF)q4>jNjWIj1y^~{o9AI0|cB9eszMSRF5-6lRLqX)PRZQ z7IvQqD|Al4Z?b%Tc}UUT|7(fP{G|VJRb{0^-;_?;6KR^@c%TjAQu1D;Q6$=*?pD9;0g)RNGSDWX=HF<1dG{16zDy* z;>zu%LRV|X^zu4iMW;+UExvujx&`OikGu9V!~Y0M*xwRy8sEj`WT>mFlQM`uGBXP( zSKha3c`*_!B1~Aeb1pDZ4Sg}KJz1#I%eCD9@gO$cu}*kP4F|gZ;LmJefhemo=h z^iRx5l!hm`rSU+}cA`1gM3UYLInOl-b<>4klj2deM672Pt(O%kvEwBd!&X^!*CR|` zuiJ#|*n74U=~v7cyo(tf;e-;Ds5EX7oPBxXksKTa8Z1J6kVJ;MO0oS_Hi@zDg)cyq z&rF^2`>-?|J6}Suo1ST0?-rM3QiDJs>gwcP1TuP2%^Vf_RL28cXRQB(+fqpe-C$mo zsi&oac4Kk<)=QTD6uh&-y$26q;-qZ3cURaY^miPLGurk(dH==SfO&j!hi&nxNZOCM z3LIf0jG;J|bSV{dM!g~3D-^{x>@3d)$=LX%Q>~iX$493*TV4NBI;sbKF-&Eh1U*BZ zH#qS*vUBQ-&P^@>xbDp-m1X4>)qhBb<=^7uR_^VQ!7MGBa(r=BGxzYj9y34w!Q4j7 zz9)F(jRKZ3l*L}v$XTnmA~qxOL-z&W2$7uJ8yrm!qP$C#ZJY=+(6nxIl&2>(3O4lh zjkSt!!XR;(D#a-_UvE=93Lu?kc7tk%Bys5aNSVd)-aI2ex@=+O$ThLm!MJ}Ur{|ND z2U}z3TpZy>ro?llZSInXOf2u7D=Az`kF1WZKcloXG#uUrvp)SNS`Sd`UPM1Tpcs0e zqVhM27DQK)k~V}Q%q(sC)o*Mf1>ixI_UrP!!Y3uyWyA0LHLWM#$GTU;lN#&}YD#-B z4BQt*R?7qhI4*Y(i60ycrwJ8`)Wwg`(D)v=a?mMmq&|DAcAb7~7uuxMsb4n#y!NTd zEpokcznO%eqxf1Qby!8#@Kib)GWERKV>sb9trG289_2-Ejr&47M|*22y~e?31XIT< zlpnX$y}Q|ug7}MhfJQgq!!3up1^Hc=L7X$;fnqE9v8WU9_A`&H@iSTqDeM=IL~Iul zTUz>Ba#S+_oP9i)t7ZUqBFrY51d-B*RuoB4-*G0a($V=lEovAv@st-nn^w6fg6QZf zTARCFF0=8~)L)f|5`FmWPs05apYQyg&;DI3c%AHkFmV*SPOd108%=)IE_|H6Ol86`2e{IUlm?UYeT1l`|fR?wnMOxDA-du>zkbHE1NhnQekZyZO%&}_cXGV zq95F-_gK*f)EUsvNP4lk<>m157u`oJcbGZ62vNgx*86En)^|VG3>}E=Ln9VP3}H*= z-!>gkO~RcKmoM8?JBp$fLPWo&xgl) z2Q~jcL>;(7IvI#+vHkQrCab5}N1>x z+*eve73HZG6^?%xwXg1F7~AI}1!Zf2vIk#KymAsg1{TBD5*x^#`kpK1J>;iuhEsOkMMy z`N(rlWse@|5OKKf)0SYg4xcRa^91v_+s_Hq8Z0?=FS08ow{oK09j3BhrcHL!0^>nV zEjlTQCjDMSukv7S+_*vX+pg^W;?JB7ntLHH zUyf8erbwSPFffQK&OE#5hZytJ^-kb2@R<5=<_RM+^FIX}R=SdV3NK-6gs;I3HmFoY zGq#+rMl1q1_~c&jH)|=^jA3mW$DF!cn1>XTEDURMPU{+fcE{e+M<9Q)V6#gDwaT4* zo0^iHGPJ*9{SYB>d9NQOW$aqq zP1IICPZJ>dNIJxFu;XPX83hm`_RsDrFEWgL=%k-72|#*_Z?oUM)LLB=Ci4&~5*(*gXM$sfnCI5v9>rCj^t;MQlD_9w|A0FBS|Z?16CLgCxS@ip zJz1l|977P7S>gd4%|GNdP!%+(mP7BI6@tkvnceL+beLVwa+=LQz?Wz*-w6mu4UA3| z#2+gy44k1^dq(2Tx`_*o*hvvWH!+N&!h4U+HYDBp-uk6!pw?tD2)NJi0*}tofJEtN z>1q*ndh)Pl_REZ)Q8eDK-bn`%1%ic&Mzsnzq9gUHP{>p%iFfqieb7Ea=#%eE4}DtG zX;A7-bdLLzuwD=xbZuh2Cu5HaKkKR z6(hQG_?n%wGhw>*vYq*#B0xn2&!f-7ayx9-uU}VFQ|l1n0-#tmUHmT5>gp;6vv$0v z0jO+$t{U?qvi7EIrlz_1+lUB~jzCg%4UMA0Le16TFz*ilal}>Uqm|iz5U9YQCPSm2 zy^Y!2T^i;bCn?O^n10yt?{gb96>?ifbt_+UUqu?Ps!Td#M5;aKt+(u$v1v_yJ6+eW zcl#{;-bn2#yd97Rll60HFLhcGg*&(%E~W)uwGdBTF5>DUSMLvYvqfh@vt(14f;>p1 z`?ZSk3b2ppacMFl``M#FyxleSm5FJ8(zT{*jNa@w={_3Yy3$Tv-kdhZ`McCeTTAQh zd17P%={MRZjg5GDTfqF|A&}u5dQ!`P%9V2XiT|Kt7HDy>HR z{d@PiUh^@tu@On7O8FlFWs61u8CD)a&Br7vN~^jQ5f(-#4tXps{VXbqY(GqKim3re z3`6-tF)WH#ZoMma!6Iin6SJO~BI5hB-Rs%G0QwJ7wL3_S-yaRCYn4{>Qwzz8u)6D# zSNpWrtYRZHl=C+*y2Hg{2|Xj5FlYTh?=#_^hu>YbkV|_-Me#nKOq#+1C4bzc5*yPE z`)Dql645b3NtKdy>>{aOs0+i}od|~;rX!SU-R9NI$Hd%bhuz&*$NNjlbhs?i1*l6z z-$OVe^TI_qI7*aLTt-Lt@c7_PBZbh+`a{D8uM(Ny_72eqB4If>XHrt9;^M*tWEj2B zwQJY-`TLujPlha1+76V?0{4r9qZ5UK1MLrT&CT4L>6W>(Gb05>pu&sb;I)bRt2odr zMf~`J$rvLeu3E`(7p)7+8@(UB4Lq&CtL|FCqF`c4o<~~Y5+{}7{YwjX_5yq)*!L~Ic5izk8LCl*sZ}8Z*thn7>i1<6`Xt|Z^e9uIM4cQZ; za{Yb?U?Ro{NM^AtNg5yEl!B+gp@aYN#{SGPMgUm91;$9eF^0BXq~hXgx`g6~7H>N} z1+u-nvZqcRI!$6!l!=Ld4@(|4i~_GugI_?PeYDmR%rAI&cmN@FVXZanYCBEsZB;k7 z%KG|)!@Uh4rzw&e-RH3`F>d%IRQ4Xy4B;UC+a~+wAMk%ODF|eLN*KyYN;Hy(21f3Q znTdzjN+Lt+?vl(;zjv-JkXf=s+!il*75w|x@d}$>7R7{{-khR;u!g6n-En2Dwc(*I zQI`R?f9CHG2i|GYQ~Wdh(f3lDjc1wwNrl+h*r?PBtInX?WgdiOJc`EZm z*m>lE%P%e+5Wa&PNR$&Uv8S>)Z@*1fOojg2R7hew8I1BqazYGSnPI(5l5 z5Ljt>v+>hgv6SQ-&Yn!;)8n=FZ{Iq-awP08ady{r_KiXCw3-X1^F3+9!^8Hd;t3#J zrVke&p2Rma5UL%~&jJJ6)j{{Tc;SMAipqW5$P+DE9{2RppdRXPr-l9~WdNRa?$~Lz znFDTcAYwf&lpTICwzZ4rX5H}@JYI@w+jg>i{BY%LRQ=8uQ$O95RvV z*tcgyr#x{P8xXN`P?AluGyv1gb#-y032p5^ij`bCkGy!(Y>WV@%6Yc#y0s`lqNtT; zyOdRO50P6gZEk2dK3?nEuGXLDy$SMsyABaQ!fi z92KiLE+1KLT=;GGMNM5@YQ4}4)XuOf*(iUaDlql>;}Yic$EItDtd@7Io>w|m;BaSQ z$V6rN%Vg-Lhf?mAUmI()s)Wf^&pDW@g7r} z4E_l={z22>Jno|+++Wk_;}f36PWMWca1Zl+WT7eu2h+AS)yK{DNp+l}YpE#{4#9QP>S zGT8MHe@pPa#A_=8Gm-vhGcjKz!{N5zJEMX0qsHA#T?L=&*^!0FZnQ|D;v5a!YtqJd z48Feop6tA|uzhFQT}L{!-fKs3SNpAZjhZyWX7IlY z%5a^@`>enoP3%IuLK+LYG@yp8W+@iEF`tcOg#aF^`qea|`vp*i<@SjNX$p}s{a5l+kg3p8T4+ZD7<^DY{jnJ|xsZ*S;xWsj-!{S<@ z>t3r|DCM~j&9Bnl!A(26aYjT=WGNp0gQ>{j>T6Bs2XE7Zk@tXQd9F#x1{FUKY&ba`KO3gEP>zh}8Z#s&imy=mk>FU;p@hy;a+>Tdu~?8v%w@_gW@HXJkKGTKJ*5oD z@w=UCmTbkZukcQVNHM-%5V8GXzC^y=y>j|G1ksEPp<7*jenS{ zj$B#uy_aOb<6gL*A#pr;`gq&f)%u8Oi3fjio-+&~uWw|lwW8sD!IO1zUdl{7WR(PU zR-$Y*UtvKD0lodu?8xF@jg&~k&!my@ulTi4#UJYM$TwLB8oFf)dq9; z%Fp=I6}x%I>l#PB6QBKW?Qa&;Z6rlabdH_Ix-;CzKm_)PrX{WbdFL{9$$QZGspmxMFJ_6qm8E9~Akr7i4ru z-1Z(566_Q6nycrMS~;GtGn=pXoZzp3m%zVNfUUp~eKOXY~b3Ej+7wSH0ybE%|uy2LzU)MA74=MwaA&triaFCMZ&9 zEY4`pS#ocn&BV_W=O}^^@D~nQqGR-J+>lybT3)R^hQ>@;9Y{$W9nG_`?eB?in1=VX z9ki6F9M58Tyu1#;3cIsB_NI;hxZD}Lh+lPD^WCOAe;Db`;kT`h8LMZSFf`)7D2X8y zJRjcf5kjvYt>shK6&E=lw+wWz=y=LiY`=%31z#36=I@ma8y?e|0Ws#uhTll`SBzYn{Sr279^=xo^=>q0kF|0p!HAt!QoPBo3WrMq^{v0S^;iY&aQqov= zx6DYkSo58s-(0VWk+C1A)NyN;T!DtVHrg<89UszcI#TOuqp7J0Q0a7Rxc<-YRvk66 zzQA}CQ?-i0Q5>+{BC_|zUhVPya_sMGve+UtDV*K7kJ+L)Em~J!(Bycy$jcguqa;Cea2$nK$eJmhN1eyRqd_TOPdF8rgHa(P4rXtz#wHXV`zcKN{@+B;xq_ z+G+*(9xdG>!N^uZe}ce+SvC<75jHltgzML?J?$?_&3yOH^7C6+ zAY&@a%a_5MtKaL(8W#=VgdIjo-EjR=xvlF2&@4}fU%d%Zi%h)60Ri}DWeRdSJGICJ zTS!+nO6<6e$$OmtLx1=FNpud-rp5ZMz%>+4H0sCSJ9 z&UlLGNR{AiXp=|e8zr5IUw@fhlQ@$0mXlFL>&a*gw0bD4(NT-_WZc=J^=IrD7#Wki z^B!KGmR3EmlodS+hwTC zmw-dkyCYc|1nDcOiG#3hQ3H1$KYsk+(`<&61{M~lS7D8F6E%LNZV9uqvpXaZ*vaSz zhlcj{_klr&8Yc;dF(oA>okr=>FfcGMGCE7l6>@{`1c*rMdx+L=NDf{wV~+$wm-A+Y zv?a}bJ_jpO1;xc&+4%!a`nF$g^ib3~luP`*}cYpa7E zkL_YdJV!x60pO}Q9*)+~g{^()#vCjsr&Hw-;qpWFB|aN>utva1AOBe+Y6B4e$8Ea? zTak^|PQ16OQ_-sSU#AL3(NGtFdDZv9E-p=i$~vDRNm)S2c;&e0R=2#o4Cp#vXFRw= z({mRuhJ}Y)s_ScL$jDaeuS7G;~yu2TV~|0xG4i( zhpD6YkQaa({ij!fL0yT7Ra7rNm2!6B;;9pc*$3lI+`ejpJwpx0r^7v!1hgaIplm-` zujr5K@8^j?Qc*^T00PXxp;#A`Af9j}EG+y48xb0+UtvuG)6~)`WACN`0i(P;@9pL5 zsM(xZBpod+2qnlX3UypdY$R->VQh>ve=kgf{Fs2=&E6Bol!UqgTw`q`K)iec2 zu{bYBLF0t}zP|U4vGZm~I%?`ig0whiHMPqo(Vgp8u3jY~BFcfmW@l&dd7b5Z?ZXj? zZi)PHiHT8PtF>sUOCJ7!_Q2HY30}JZe57q^2Sib^$-zO~($z?qhUa2$$NJavy{hJH$f*anuDVr+Vk z9zML3lE7)y#?i@oaI-$C8Vm<1EBp8$vfOmAN@Tvoz}x*#s|&J3TutFw_P>n0`1Hi4 zo9atwmwUgP9K3~?7n6A!TV;s4hV0ya^EMDr--Eq9$H_(s9-gRfKH-U4g@U2EnHiAx z*-)nwaS9I&%`Yl~xmnj*N9YLKCQ*veM%-wCPk7KNB3A2`?yLKz&1XiP1mg{Vf%>k8 z;W`6DpY;fN*(GrQ+G{6F1qV<|lD<8+o&A5jz803EISKPMy5u{Q=k3w?2D;v`ojCFP#cmbnUM&L0vZ_r4Q_&EHNuwFtnIrbUZj!o!V%N#iL&?JGzN z1&w0IRHes+B*;3g$G<&hFG{j5oYbM<2UPga2e5e_0e|<-{PWu?56W8gX!1RI6+dAP zd*??BgGmB8pW8lsh#Kf;7)9I}t^Q6~!gav;0DtgN&lh~dM|~70PDGvo2_(?QEG!t= zRw@fV#fL5@Wc*o`Sy(0#lWL=3gM+CXJ3BN1)SyoM5)%CH!Job{gt%5@rJU%(dLL(B zE_Qhi^4AAN^XX_UBO|d>NHN#BhRvNX%~FRe&T`2q2DshEgVhV?lgwR>2-w*1@$uBu z-}Qo?K0ToTe-acE6C)c&n zhi`l@x(nw|oVyV2Imka3q2tQZB{q&U&~W7zbkn={(V^VFXxMqK%N7PRQQxm$YrI5H z?_^-4L*r~~oA>$A;L;YS>fO5m@dCsJ030EysHi}hTyPT|&R<+yeDx~&VD!ab?yr#5~@yuom}i_6jx!~z$%;4bQ^z?Kz@y1_W z?}vqOcM2)C?g!(xu%<6Lo$+Zz*mzV@i-x9NUBl!Sx8JDOYm^7JySrOiMMao4V%b7p ze+=YEgjkV^quNFKtqqms$_Vc^0Jc+H!WICA1DPO+_*$o7s=7GFkH^q>U@m%iZ%a_lbIU}0JEhgp-AW_#7 z7}B1X`oo70Ow7!q_ZE_qX>Yw9vWxotjx_9DM~9QOkB_*3amRgbN88u;hE~wnI9Xi9VjeR+*r!?mzO#LI?4u5*I8F#o(o&|c{1$VRn6 z*-J@HO$8!xMn;nxPCY@ILD3&1H*4|tWo3)>DuUGW^YcgItcA&v>~D#2iC{q7EtXAH zKVR*_kgVkI)YPxz;w&Xd3q}8G@206G-w@ed)f0a_R3cp9@Rp-a8vOuq%>CrFcjxh! zmYN!q#Vt>Vkyl^lMa&MiMb_6zJKi>GFVTL4=qbp^oHj}8+~CEs&|9hNYifSfnJE_s z4Zpd$+1Az;3WfSC@=qu=829dcEO$aA2Rz1O8ob*xh|&)K3om}WzA1u~2njY-0nIB6 z30*n>7%=SnKbj)wP{jH5v|e>Sgn%cGU1yq022D?R_%F66B#v*ExjLpZeE?JCmbGDJ5Q!3CblJPC9fL7 zhU{FQJb9ay1x_fI>~4Q zU9dRu_tN^?8RXJqQ^fZ4<2|#KTb<+_92^=N8hf0xU0q64zOc=&UmuU6Lms$wc65w@ zz+W+oh7}CB;W4jY=hBRXhdSd5aTzhb*SNVO6Q>{U2a(X_7Zg~v$`Q~1`=$6hX4_18 zzY+=@djDAL^sEFYpKoe8Iu>nBw|Mw}?fVXqe%!ACcrPwQ27RUf?%bOsBLqRhIl_e` zS-7XApd_aNp4W)o*zudA@awkdcT1F+5Yd^(D`teGZ{q~pE^}hyr?PiN5qWJEB`26_nbOHVR zlq{on%2K_GmYcQ_#w|xmmp0{g&yW$V-j+V3$>1xATi@2!XGPdrjdIFLm1(W@dp7Ts z#R;3P>kZ2p4;*fL<~cM8W_v-i@s!6m=vPEN-(oAIZM{c^)ZFRi$h>>>7pA510Gu4p zKKx^NnW3e&TIP;*$;)~d6B}>zq=))x$VJBxG6V3)JghaMl#qp=3I=g7KITGXhQG-9 zBF=^lUHP`e;8DIE^tS4@dxpdubRdnPB1?#TeiNa!#^-7}K04Y{nKYo`4-8BD<&kA2 z3q6M7>>Dk#WIV&X$w0Q>upv-p@4dtj9!#QLtf7 z;qKmQ+_;XC9}r@_=p^6TaJ6K3gqKT?mc?nJNjc%AUUzurr{2T8g3ivU+4-g#H=H!x zt#SiDetr*FcH=wLq}n$*EQrL&kN9j0pjcUN--1FwnPE8yl+|C{U%b=WAZnnST>jB$ zmt|h#I&X=!X&ZK%Z2N)9MvD$c!77HxZxs5jN9nlUFtH%TL53Z3!xob8MI#TY(&#P_ ztb*R}xj6n7ELa|f!>qy@LA8^-808fHfdLbFMemse1MBtx7|eV;KW8?{GTQ zejee!_pOh1x3@=IojZ5#jT?u}O?GxY5LyKD5E|<0%MAh^ckKv$GGs5FKex2DCIu%q zO)^rK@P4BWKW?0Sx)OKQSlSY|+7iKQh0z~vtqkN{JRLd4WWBnil3#dp+LOEL*q-I? z@(JKe)FrLZOV^-fnai}`DDbBo=d%&`i!?`vSz9}Wzl7{gc$&P7BxNjQs*muT@9_*` zz|Q_eqk%R70{%QaJiWcWCUtJOQf;@b|Edf>)CGPeq}akelhH_x@r`T)*rnqr?xJGLiJwvwQ7x`WD0P?C86J_y#J> z+Vs_XL@U82Iroe^Hj8RO;^i*S-hmE8dkerEJRg@@* ztE(&ULL=@oA3RvEg}NSR#nJFs8C2P6xw!Bd3>)a_)dExnuw!4}!~IT$GA6GEKj7HS zl$BE?f9R3l6A%z!W@5tl?o3=Z5Fq~LBA=xi10}O3>zI0eW&X}QcT%I@adt)k?xRu+ zZ1EGksQ-bmoks>=;`v)-?F#fYplckA>5DrZMJwIg8@I?)>*P*K{)`%BJDlt(GLq|^j;Rn>_@%mm50)jVJ_D@CF0jm|z zcEcRG6F*Ta&^dNInDj&^Lz6gen$+9Z#R`}}3tZe|(-f+wQOoL>#`I!EtAkM`3q`cH z?OCQ??xR@9#=E<=8jy!51&g?;Q~)CdiOx(|C3ENES=rcZtgPfweKBqQlSt%!Cj^MI z*bf(GUQyQ&hTUOi4*bo%H(g?;?yzzk0^}5sak0%qa=<_TVG^I1D9~xe_8gzDiPEKu z`|$HNwf@7ajYoHtoENt)*|owov1K#z#Bd;~oi^k~#rx$QwDwTOYT;g&>#+13%q_iG zBcp;V``s(v`m@Rwigxlb-YJdD6Y&lFC^z&i?Gl~M(^#KVqxJzX1Iwfrv}=;HXQ_Nv z9gEjQwfH_R6fgG}crh^}Kea(plDWS|fuwX%aWUoKkFb%1m{^cDViTU5`?BPd74*Tw zz`($P{{Hp#_1!_Mx+JfL^tBMHc8H$tYt^J6e7^p#o~p$RA@gyW&bN(M#U8=`Jbj3h zH*bEb?B+F&|1atOACLXb>IyB7$*kB%iC3~Gk)5Vs$^*YYUMoVZ z{W({_xLS)pBcsbOPFXRb7#nXD>l-(rL}!5j8pa{BmfSE2Rx9sUSn+-)ejR>v8^OYS zIvfSg;*iQdYX3>O;93&;k8<%UZvCIag=e4Vin(@0zI#N~UG{Gp*$?G`e=Ho;iuRWt z_llr(CR_%;@!0@Gmytz1J##B7mQmIWSZY#II}gcF z!na3PGb7A4K!w|cBAz> z1mA;DU3bNVF8|HXR&J&SM${RviT;Ia6ycr5;*!#j6#i5!D5?CE*!T8$#+Vu|%mj0a z2yf%p*DVU`V%d7Nbx`;3b#dPuu{QE*%RkIL^OKhG{k^#8V_D9SoKK2Ls#O~}79OLR znj%BCwav_BqaQ|3tj9)|aR_a}Dl6rLomCRDZ?H_X(emmVb+oaa4KxEq3lxvfpFfAH z@&j$}PkHr+rSt9C!LDI#W!~HuyKtd;G-d!3S~H>+WRO0PY4CXLNjiULJmDCr8}vp^nb{(vrQlc@m!L zDj3k~5d6gwU=uMed9hP}W~XgO^cJayc05V)sL&%?PouYhTx}JFU$1U%k4Ri&x55HV z7Y4ZH$4*lNn>;J9qQUEGzn>NDtE8xiMnEf&A0h1Me650-ENyM=V#C_EcqnqU`xdQtl=P#akGdV_wm)XDKumD2 z5;peXv46SKkra`*1mbD9^XK`56V$IsY2IKm_s|>Vww(shJkM@vzU-Snpick)tJ(l< z3iO^&-ZIkCBi-RqQ8qwR`UF!Bmd~z8hSE7S-H|!@i)cJf(;Wh66l;CJz4;XL9U@s zw;z9b55$0U=v)%x@ow9u%NnM^a!g)c9{51jld0f`z<$M3W={?^ZYON3Vpv;OOm@9G z&12O2Tf>NsZ+>HZ$1d+*!A{{SmQhM>?V*4{UTbd z>E>eLEmxA{xvDQ17h}KvJ)~i}(CN!xFk#9K%3Bg#^Ap7+^$s6SGoUA_J1J6dU6V^* zT^0`LTLyE|WTCFKJUA!dl;5|wcxT=bPI$^<>^Wg+FsKu7sHvT^Gv6GOHh}U!g#p)l>BQz}R*VscjNy=aHjsUc3MZ^aDmft^B z_q%EnkXjN2xZAWq0*tNc_qhi04=a4B_F{1@jZE=C&@dKwQ}1|Yw8634nyx(%-aq;& zThY~b)h|_9I^q1SC|NXJ{;aScpTC%Mih1e)Wb~-L{Nohp(gl1%?gnO4ZHVF`14el9 zH~A~^g^d+-_xb$m*RMA>-87S4`s1Awn`%B*01kx4?5CihP=&VB8=Dd zdcrU_^x#8vd5I-=hgNJ_aF`(F+7hnOFfsUQXA2=gNM8HJeg635`56Nd*LMs(kHJ@J zsPXSKd9pR^X-vJ({8$A`e?cZLBPQkohk=$}uKgwKM5QgoC}&27126TnXU|3|Epej8 zdcvb}qurIQz)i03Kg2RBdP1@luvlzS6O-lX!wi`=C)R@5VzePg4~XO3sST%7sH%B9s45d5mvMwVH^Mei_q% z?(KCx*j@nwq=Q}wu18D1NSKX{O;j{FajeG00we$&9l2lvokI4Ae)kXIz_*MsAfLkx zoh|-)u zhVq2gTyOHp^Pc$a*g4$cBE{&EM{Pt(`a>EZe7V#dmMX_vg#hIi5DA}zC)U=iVdLsv z5J=7P3OJz9%F1dPU1g6-NKC9y2lG_&N=k#o4nD1Lw_)Pthm=>Zwkf(i$jtCsWEW-g z=1>P4S&^sVM#Z#Qqg2m&UmG}f;%~M=XdSZC=5Tv3!%6IrR`o{himKospX^K>T8gfA zgS@@Sp@BQ2fzt@uDe`7Iz$8Ig*eJ6cOienWw7RzWGT6!I^qG+Dxmu zsQ*1=#U`$kOfPpU@pMx6cVSTd@fZ-j{r!(Gd$UK+907Fo-Q@oPVZ7ZJ#dK1Q4Ge7! zMMv_G?R2h;*T18a+j1^NSPG$4Y6*c7aXmMZf}wk$d$r&1K78><|0_=TXmLrD2Lqw) zwzE|49nZ2We!`88@ym6MlNxmuKY;kCb@u(3Sqpx*WsjIyxmRA7Yl_Y^;xu=#Mdu$S6h4@;g59;{ zxEV=oa1trD|1XF%7s%Nsxi(X)UU7XKBIc}Lkpu!?XA7;XK&%jJfq&Cv^`h>1f_nyE zFq9`P2fG_w{(w_{ZTC#4(J59wPnL?6VyTuyl>TV-;uA6F##m*w1jlHr^Nr|)h#-fH z7G03btFGXujDDs0kzTHF-lT5uIXoRLm-a)$OOfdc`iRX$e2<%4n)X98wR-wz*xRUeGZ<%KyLpTM6g70Sk{x zWRS>yBu(ltP~GS(yO72O& zgt~z4U}umi;9eVjWna8tTpNw;F)tO?6uj}5qgj%MpD753S5^)_5?tK!O2~Qh=6wfE z5E_!FRiIsTftVP0gO-jLNl3UjIOrnlx+Z=FsQFtF?L~wC({w*KE8&#hC|CFh+yR)3zoq@LVsc!Uq9tqG3U*hxC ze#OK|H#tnHYy^yzo2t{$fHe&cCieF;l*xKgc9P*%K~_bOkfJkM2|6*uPyQVtO5Ga4 zdBecauq$Efzwk8qWBgC(G#6P|SQr|5fqzB7rsrYDIUeeupNY*gKqy_-sY~8Lo=0Tr zSgp6X!*d@6N0hMkk2oN@)S!hz|8EQLU%?=_E>Qg}tCeyjpzAm2mq*H&`S{dEHA;{3 zn*Yp?{M{T8{e52W8;S!??fQ%Qvofv^`;1y_5McT3jBnUSq}3IajY9Mag`}kXLD-%1 z1;~xWME*TCLi6b#u@OZE#k2N#_i+-c3JPBPxSY{DurY@Nt6OJ(#-|s;fdy8C4+`RD zZZ#^ml=Su4=x$0(WbEwqcfs|)02?>s@b}wa%?!tRga{OYjDVY%g{Y~yE>F~kS}bO8 z!ZI@hN_u9YukmNko~4brzi;qQqt7z`2cOUUn?cv_n}6Tb{*m74WiKk3+;d1V&A+?! ze==JYU~#@&2qNj#(n0mRn7)}z;97Eo-YrJ@?V0g>iD$V55xd+LgI8cK;K#D?M1+LM ztE!4@jRbU(y>mvDnDi!o83V&)MG3vp)?jv}z`l?~!R&8gfd3F^D{A8Is;Wu$WZPxl zJ-)48aHv3Z_xfJ@cQNQCkaKS3a;7Y2En5SD2yiCAoP3fi=Q@;dV4YQXi;=bbN?HC+ zo}CaE9@iYc-e*GxhLU=$VOHkn;l2mk2!wcbANrH=!<=`}PSdrV`qk%K)I7Cio(@a{6Tl~Hj11%UyUQ&}aDIbZ z!rA$9$hgn}{P!z$=L0p?4OtfHx-#s4DkZenAHG9OzbpHw#1|5jl|p<4SOXjx1hnC# zd;Z&s#r5Y(PEq}z@=+fSCY~L@c^3Eg7hwIKUlvNO z*Q{=W7bT%&Z8i`>NTJ94R&SzwSGbJkRX#l{US7aQQZB7(J@U*kHZ8 z+dXiWGx-rpzUT0$0MnC>W3LFV%G<18V3?sH23OECT{x_Sv4E1(Gjv)g73_D1VrC+$Ijx>KppSl2^B5qFbvFwa@8ct zKH6J0NsqV!ugH(1DHohsYOstbaJnGID#j8>Kc8hqs$|$T+4BJb1cr~tt;ER&lXEx2eyS#CgQttt)345Vy&8<>7Q^mk(xk{ zE-}pD#6;d7M}{y^NQm`%3i|srRy104j6FoFTHQR%p8CV+{f7{g_~VJMLc8@MJ%YXv zzZ!d*7d{({>>~aguf&UBuXY z!U+uvo9pJ})@!^b*eNd$*>nvAD_6#S_Wo2iqpMAJW;t((PkJiFb>pj_sOl=xP9Co& z9~5uA24~8_okz6>pl(u)+RdEy!wzhQv4wT`C+sYND@!=*kQt3BGvw- z!M#+&5414~MsYuSpL7~lt8j$8IDgmQ&ChcC8Ew_>)&0eD7?qC?S*3uUJLz}82vW?V zY5u@3W=n^>J|fiyifM(d##~MYLm~?mKt@wFycic)*=paF37IzBZK3~K)&FF&G;0WYXP#A6gAxH?PiFtxVimaR* zhh%&1nSX+Y1k;noOAM`sovke|ETg)i))bjWgmJ5M2P+x&xwrkXrD-tE+5(nOvZ@lU zgE4~YTvr<(^%Efcdfe#VQNb%+gj6BxX$cMG-u8}|(?pm%A(S>S$jhMw?JrqwgMS@ z4s~PN-{|##^+v`Wi7jv->?YcLZ*^dW#d)N0!Lv`ydNw4Sn%UtSumbYz(9#-YrOL)l z)IXh^R&;Bo9bTmUt`9=F8>8v>Hz9k`DuGkK@)7U|+1c5*hn-s7Ta#IxMoP^97|Tal z{Wlc+Ur^UfzDyK!oWMI^(KddfApJ1>NWoLwdX6>aoN|XwqK1X7hhSdS=1`4WP*_`v z{6pOBZHo_T_Tglf?0w1xn#JzY3 zwK_0!|jq96yx8>zNV6LT7}>un4se)0lj3-pg&dfyUv$^2k>WKr`xYSoZ%5<3SRrTY5PL7Ox5-M|Hs3#)V~J1=Qb^q{*L zsK$;6w(5_)MW(~YcUl^*Z!g>ht>h}W{TfOJy%X3e@Fv~?gP?aqFCyj09%m$X^ zI#M%&CR+(TqM(m^s~s^!0#nlY8q?urxHRjn__Sf|z^YMnVaAvv({x`sK18tozTchY z#dop`R8Zxe^M*bVZUMzx-XE+LQfLnzrpHk@8Ng_I;1uosd)i%wclXAO2GUaZIt(HT zQkFH#DVwYl*A;eDs{v`VzP8>ap-mN+@PW=D;d*281MEknqH!i_@;X{FRM`RZCynMf zvy%Vv)2ETGxfK87i8b;>ZX;a{*VVmwlkSeOdK*dOEjQCd*uJhKaGUv6>oR zF5S1>@89jjt`E{B+zi~PoUR0?*PKTDw<_*NShlf_^1OoSeWZ<*EtLt7%_Jh=EtVXV zWtx{`tXRABdpuKPx;Lxc?to&wQc>dATxcP79>iLH_RDpz4Z2yE-)Cy)hI_dd0gT}yEF=iaSNsr@BH2CKG7ystmw)a~(>Xah}>Q-(DUb|DKy~V-SEWnXhj1WRx5DgV|F&_ zz_@L}4zMi2FEIb05+;~N%kNz44e?Zs=C+${mY>siILqTZ^vXnzUn#~RKdp<-D!}Ay z%2|hPl0UYkn_e1zmN&_hPa~nv<_Ly`KBzJgMI*Zg3eU&N)`xq}HtP!0jdpyrHCPR) zi|ja(I=4Eb()yTs?GoR?{Yk#?0a~6E+ws1ylIZHtGFuI6uYHczRvtOIV1wfS!`NGg zMWKD)!vh8?2q@A5N=t`yDblGZ-7PIKbczBJBHbV@Ee*pEf(+f#rF2LP-Mk0A*RSaP zJiqr(9~?Mya__a*T3h1MJ}a?q*%_t(7q_*mJdpf9su|`3tbJdAW#7ByjXs-cfaZe~ zJT09l{e|t^zBuSFdY<*J^0M-S39mM_DePjDOu=THy|Omfu*UPg@@)x^lxVS5(}Lkmk_mWHqTL z7%)!65cvS$Nc*O7#-_^k(6?A#yRN{Qv)>~gWnEfkSi0U1y{-rUkv%iN|08=gCMo{k zG}-p@eVN4&pt(8ExKq00!A_dL#yyDG$+f}Sup4x`Ja*%A!a4}dWEDDd+E95}$N=W=sqNk3XUtvMscFKnn?+F%FV_}CN4de)N@!H$l+#l?b) zmk$&bSHgUv@&*jUkVWgc3|L9)w{dX75)&<#3S&>AE}-?6FAo;Xx0+TV0Se^f>nqJ3 zO!*&FWwAIU%<)j(Prab!_Bt~9-+ej zmC5|hcLR`U0=6-y7AE21m#jeyBI{~}5S#sEe{)xywI8c)PR00J8Nwj>wN5fog<4Km#aHEh>Yx>5Fb2WlR{tM4KGIVX_dFMw3L&3%kb#I14z0$ z&m&36mK4zdpS6fe4i1hkp;K2?)fl-<^s84vv8UuDlH!*H?%xVZJ#F$*OP2sOROlEryyKa!tN!O*k!C8x0S{Q|VuLIZ&$Me{xAoPXz<$3AT#q|$2T>!;l z6k4q5T4Fag*w;V$Y!bPDC~)1_w-+3jGe9yqxs9CMDD3r;-AU}pi|2sdk+a0A#cQE5 zWRqGb%J>Jj!p29JL#uZjL@0SmhXLA0F9k_KSpm4f{=_Z+J5;_lua$3koYK-;J~-w; zqMAa7TdJtZU3L6w5F=^X9hw?pK0ldfQBXN#ou5};9CaIdgh4i`jv8M2&D=Hkjx zrzqX>Z~V_eU;Z^+!oLDqJ3#)* zzY;IJPo2^+R*Z>=KO?F%=r6o4D6T$?TdE0=RC}SuAMOCr_h}Qoghg6H7Jrmd`JihL-I`EB~2c25|Dv`%7A|?Z?s|wBO zKzzyW!0cZ7rmOg7(>!y3nY=V+aXIW~;rv^-eOH=lAK;~?61T~-%W3m_Wn;HP+IIc9gwsb(NAV+V0f3Lpv}yc z2q_7P=C(HdgWq*PyfS*mchpbD*WAU)JPt=EL9JI`saWdo>16z<+5}*gjLde<o5m`i0> zy^15?!slEn1oJ7aJ;>@>BubBZgn9}^ux2MH&McrYT?S;!#PTHpBeCY2hQ%H+lL~sv z>V#W80IHHO16qR_AIZqJXOVqBnOxt8EJnT~H`k~OA{A*m8fIg@giCUebZgr)ylyEQ z4oLCWK*%o$ZQ@4Fv8T8!$(NJ3%OY27egwsg~gSQgdqHF1#&plSl(d0_0O z&`YiQp(6WQ59A4+QL%`tj2$bghPX5Yc3YrQOFbzt?wlIeoy>d5OQ#NqMx z$$vOB>9>N_@xsFYfh1Kq6~<12FZeKA-#<(8aWspj@}EYy0ld8=EHo6zTHV~-&fLAOF)Ge}`>Xs^regvkb_p&ahuVF$3{!K_d@((Ttw*y{ z-+BSQ%F#eW%M|@N^y=N+d4L=*+n6p;R!2D>%`KziYd?BptbLGOl^$>!`Xmtmva&QG z1BS}dLo!Cv__W6@2AS8?h}tvcrFjypbC+GwI4E?!eV%wrn$ay-ezC$4*3D|`$cYly zSa`e^EdfMVRR@)}Uh|nS_d7tblh&)bd?B$N*8aH51hUR~D}5?f>Xn z|FNCMFlgqJ`l7i(UwS2whsa^(va05`>GzbGt5cbAnyD9j8v}%bM?IQBIK*e~48R+6 zvUh;KJC=S8JNqX=G56N(QbgNepQ%hB7dbgOH8r(>!1x#aMNgUN8qHdF0T1JXAOxS~ zoyS^AJIA%Ab9Jr>>n}k3cU~Z(t`p{OsIFM}ro!&Q2(@ff#{(mwXT9qT0WWN(4N(k& zD3{u9t8l)7mMU~sE88Nos9W8IWh?FU~QUG z>Z%Vbv*FKrX@lHiwzxXIAoI{hlT?cVYgc`MQWck|<(X+>nEbQ3=vAr4;}}i-!*7GS zaEwb8f($!JEx)vsH7I^<*lhP5O?F1|aEz@a45TNdsmX?NhG7RP-V}S^X z90n{4a+G(wE=<;X2)Tuey7}V%nek%yoA=!Fy~HW&o!;w~939Xa_!f7jrWG~Qjx_tV zptBruBX$+1VvmTvKu|Ap_MwitTD`8BPx@u1#@0A5)tQ(#^lgZj)d}-_assyDBykBh zEMW2xbFCxeiIGffTJXz~gy3!V?!<{);ihWq$`q3>51pc!4{aGScgr!z&&XJ%+cfU* zd{yOmD7k6pzAdS(9xxj2e!R-+PTz0SpE?zp*zCu;foQwCT`Cy1zvO?LL5Sfkoir%v z%gD;k7KZ~FiRw&bZ$m?ap%u-Pg485%jxU45<0D3j45b0EXrwbuUz5&m{k6`kqJsNr z1l4QzZ?Ey2ZT8hefVw)-{lH|g&E0qZUn7TKn62*PT=B;porH5pbpT+~(r=34W8pKU zY)K==A|eczo>xoC)K?-$dT*MnDB0JAc2sayW98!EwHX?<`AcxwMwY!=)1x|8jI>mQ zW|dWEmbQJT9;)L7bTzZ?`)iHIh182Xjo}hDqe;&{=y$e2%RM-4=BYx}hn+QnJO{QB zY4xD-LM=m_IgD+%2+^NckeVQB=X7tFT+7sZ@Z~HK@dV?RTln5mfS$4`uj+^vC};u^ zrZ2;nA0^Z!GdI$r?Y@C8Tc_7H6+cs4{xg-r@K3)&7Z=pNn`)W^)8W^Uzh_YFD!>>p zxj`_u?_kz~HQXuRo)p7m(`L9z7Zjqv$dt$=$c6Ti{T;p5p0U1dmX{oT96ksaFr;w; zd2$8pvRklZqb|jy2S8(n*}}NMgD>0hyt%+ksK1qSb*i?tC?O_>Utg~n^_g~)%doHu z3|7P2i=6%DUYXz$WjFodeda!|(Ujp5y3VV*3Lq8R=#49LHnU8+Q(vz(*n zv_O&dU5Kdh*8Jr|fpor6zF<)Cf<0fX7U|u2x78b?iRuj~p@wpngZauvUjL)+7lP5tkdy>%?;u1gG`YW$@J&Fkj%k7gc zo5^DiE4u&Mg~O%eW*nw$-cn_E$G}12{;@L2{D?~PNC}l)VZvldv0@DoIx~4F(^dPr z3wV(~nPq{Ie>2N!#BV)PlcI3q3&~}o8?t|nA3&f!QkGpyMX%tfy1e#(Pb&5+T_(Tg1YDmRK}@BG%hk6lz>W4~qGxF|r%icR zVjB6!z4+06yksCZ1u@`U*MP^JEEEQx_G}6+BWMpZ zlLM4@&dd#P?sB<<_4}yyIAaqH#&Gkx+K^a`SM<6f*2}Pb zd|uAZ)k#I)(x?bwu)PMvLL~IbwIHHkSxeuvyM6i??+w)=Nv%$$bRFefe%gB=F4}X} zq#sq%w4Cp7DC)>D05YK|Q2Vy9xzd(P7>(XRV}cdJY6RE6`Cqb*2Hyz$QQnKoJD9KZ zxdR>iqTN#02rmDR!ZVAjic{fAbAz^-f#D5i?0_@4j_L)7P2XCyG_xQA@o_W~uKfuU z5D4bNhwLeJtBecv>DQ=G>sdzvqt?9m7w}s3F9Rlp9~Drx{GWz`qWkNO%YSNDe6Di7 z9{L_&Ax^L(ZDMH6j2+%%1W)7|Uu3P)#_*{@E28~LnzM|p(0x?jn~si_^IT#Bh%1rp z?0iwo+VRm9^*#XyTIWkrrckN?JjSWDX~hJfM4wOcr~&t5W*?AYQc_c^d9;PV{;Un4 zKpy^}xLwhOhn0_&0PD1_T>lunTRQiGt2a`A5X21L0M6WXx)eYT1G)c=8(4i0o%VC? z-hFx*Pym7Di*%4!s#bxq)3s~Ya^rPXfWc6a4RWf{?p=7eh{xu}2GEUfZf*iVKigRo zZxt(0nczP-JvSO18&0Nqc=uW?xYRZ#lAEz`p5zUzHHQcRUCjVHmgryJTv)M_xksOW zI8?-9ssXtyaIamw!_Cc2IkRDvl!a7i%~dV7<0WC1xL2eAU=z@PE8G_o2oYV~+&aW8 zv#8Q00n7cn`4`VotJ(xixkUb~uXD8jVFloS;rLp#VK^~!L+&o7sC%@Z(lfZA!|HW! zBP9l*|M0mQvi$#-B^yGa(0|k9YAqqF2{>S94`1b(DF>Kh(H7wG0j~>xfB%q}(f`mz zf^>QWyWWwIZB3Ek9{`E`Mk&=o_7l3Z8n$wtRNoIg36J>D5zxWuh?s4^fF@TjU!o?+ zMI|IuTaNP@RVqwViC_O|c!ZN|W@`GuPbo*>pCnYLFZ8^e!LZ|wIwlrYR7aE%qAkez zzZ9o0RXZ@nb5%PsUaTon3%Lqzo;iQ0i2NVH{+Z%R-{O%X7brf@mQ_wufBnmw1hnbiUZ|ItIv zzM&n3oz^a|3uI>&Y4ofp-D?G6kcRr6s$aG7KT(1?G;EGZk> zE2H417%I~0OUNlvvyhXMv$MO;TnfCx0XuhOL`}*#(rmy{R=Y%+MjI4Encb1-bSD!o zYlW&EAoscQm*|8B)U1FJnJ+~OvQ;BWx8TUYxVF?1n69hem!?zr^nXTG%0co&8Lb(7 zG?^J08>+29#KK@hKl&u501o)G`hP}|2m|Ttf0d566&5lRo&DeB&$Ui;x*Bc&&p`kc zR!0Z$t-4FyM;-=XjUEHD zr1N1wPNOI)PM8p5g@0KpBSL-T(DjLh5PgA0)zm zS#UQ+ejp5=t@Bs^CBs?0_Xc!)EjqU*ixKmJOKlt%m_|mQtqJOPax>Qg_Lf#dC6G4* zhYlxA5ZtNjYh0iKZ%W3mxngKMq%Z?l)Vori#?7-ZYy^Cz>MvuH&|y9Ax3B&~(+ zCO;p`Q(U$$1h9*|&*?|{`(7v^t6%<)C{GSEP+7yL!BJ8P5D(*q--mbVQ;}$s8a0hs zjS3A9y6q4QO033D7KE;t8;`Vybds+|%~R#XEUUEw0LnmR`dcHB7fg4=wdgeMKi}7w3fKq6#}4)Kp&G6wvGhe!nIJ^sp>Nhu zv)!gN2%R7W9vYennjEKgz@a4Le%LbyQu@_`g|Vw9TPi-!92VOr0uCJrfU3IIaXt7^ zr9RLGG*J>Hsi6KDiNxO~D5wBI+Zk0)+oO1HsXvcDW%P;lyDp2lHw^yLOd}u^#g|K} zA~OzAH@UG|&JcoTm0lbIyl|Z-_ul~&l_;zcmNpfY*Y9JIQ?XhEbA0UF{c-(0FeY+a zNcPnNtv#zuw=lK&@46xz{Zq>PQ@+)}>Rdv{OTvQIjLvRDSBk0vQ9#|VI!hk;Rgyqn#JCC$}~8;gm@Nn$Y^yl7=yi{ zbR*t%*$6YCY&!wp0G1P?QnKYA&T15>j6!Hd=6;fZm8i1Q5bm;edo@RWV^oM#U~?sT zQozWt%7lgOV-iNk`l+e;(7vUmBozsKYeC>@52m9_--k2z+P?;AHr|)!POG^}Ke`&- z^waRdFisR1Z1jz^E%x)6VyqWByS7d{Npv3|m}+exW}QY>7!$PJY${)KF@3$eFd>aU zyk(-eN)YG}4LoUX#*Q)-I`En4z$wsw*LVN8f}?cL$&L(deg4nHBBslgQ}YG=S=Mb* z+{~Eias1J@T*ljQXxoTF+M&G_hY*-e>0IO<1$k`bGI1{dRt?McHwuS?AMt($42m?ZmaH?mC=?%PdO900cIAIqpmH^hdyL8ZfHf! zmO$EW%q(pfYV~0I_{$Xie>hNlk#$Oa2Bmh^71sBhE9jYJHVD(I(ULT7h32mhvv9S1kANje+kR}J-V?d|Kv|; zZHFx7<#vNfx^337#?)q%q3x)}lUC}A9g9&e6R+a-d37!QODo}io#O%OMCz0 zN;%UJoU2s}FUQavQMNRSo5fE2G^p|)T|XUi!C9xxwmYpIy8A#__~K>N>&?8LL@!aN zaG4((GBsCVzj(YVs^nG4#vu2qcG^1yrCts#84Wv2~>?Prrb32eUK5vxl}PsitQ|GN$n*$YQ8O5Yp}{Sth#N~r)2lt5A30%o>X(O4WsHW9QoGR-}iDso~Pf15U{WOI- z{n*fk_QkVvg;qRQ<<`}yAx6ylM3Fr+IODWoJ%c>wcA0Jig!}v#k9tgC6;b7^WEuIv zzYRVbCzF}6f0ES^34@cp$?t6F4zD0*(u%h2&S=9Poka(Yg4 z&C+SB<{YRM%^Jnb{zxT%r0K}}(;#mf)iPOme{>mU`1-auu2pSv%^*|aZka7$iRr>I zeZU$kqpx8oM3-HIm_3ZQy33;V`l0j-5M4tH#~ek4w+~*YpijDr`AY{sJK6T8Odq@^ zpU~wfUru~)HAqOnhauC|Y>TihMOw*q|M9@KtrPC@WQ&rFxZv)-%m^(RF3$EkiyTGN zQAP4h-Oe@S^_=lT6);cr=^T=Y9P*df5hutpNMznNT~ z;T?dc>REt`A32NT{o@UjSU3Hq7Vh_{L=mEbKIMw;g`~q!d)Wr8;RAp@mVlGQel%~% z6S!_NQ)&+1UdR2lyj!dNV>pe=`0G#8iuzrGdii1MaQpnnQkr)~Hk|=NRbUHVF8}*z}v)H0keuCGW_QG*KUa%8QTtMqBVLwu$rpc8;=TjqdSQFFuU-IMl?p=`TMS%L!Z8)oZ1 z!~Cp%_>i_gR%*b}5Nqq$euUZgu=T30Y-`_AMks7Cp9;S-&sCbffEKek~Mdfb$ ze66EzUvs_JrKsTVUu46nu(PJuiJ7Zbdl`jp^F4iUSLKc5#m$b{^4dIj@brKNOM3NI zS$Ub^3=5?5uQRnQFQn^)xHo>1VphT*5zbvffYokTW6!j96gxaIA+hDk3CrnB;s|~f z73?3mdvAERV|$7`F?-4MM7}6eJrQ?zC#bnOzb%+{d{qZ7C)!TZpKdsPDK+Vu@t_{7 zpF`tkFHJSh!ox5~-PJpt<=6xT_3x=FK!^`hF@xhy5sL}zVOVGgKr8l^u^PQ8%VWP# zsVuApO*B*COIkvwp`vSsPBVGYo>G+VPF$J}hDC;X!i~#OOfSTg;Zj_F$`~EnjP0@T zn+0|Xr@d>vT3U4OLuC_wQkpn&&ibBq#7z3t1P3QE{VgT{GspFAy3;qRo!AR z^$f%N z^!6Mquy+L$c>RBm{pg<@$>t>KQMq<*6G|^A(e&Q+=>c~`tJc^Fk93AgTvQ)? z<3Maf;X$mpXJgKN6Daw(OAmU$AoEOMVYx|;W%x`v&S$gPqr4l_ds^d?5tqmi%cEc- z@U@4Xwv2f>2nxcmJ5Ac2wdi551%9f23M(r=Y=G3={bN-CTAtbPB0@>i@lO66p}ouO z=M(9wnt^s_YBQgn+SAxF(_N#3T7^FQ@6~qk(k^8SY=xB7ifb$CHA*D*%50kQ56)lTx@!f^kT9^ zlwVhIkLGR4uIz20u*u`m#Z+R%VA0{`UqgIV@7U(I=wE}0a40akHc7+W@>SAP#`?iv zM_JI4&qFyKc(T}Pmkim4|Cdxl8yE8 z^{%e=E3;xzMi2(HKR-G6`l2ZuW4Q6NsF~og5zROO013%wW* z7S3;OP+yY6Mo7O!?pBLu6pTq+^R=I+VVS_bi#3@ZJQ%t9d1iDsu(tAu0CxX^4$t|Z zfJJNcYO|x^;H-G}>d{36@ilX@njHGexOFJSQ5UdH&QHhQurJ=L>sFI~;N1O~J~d3I zxj1^R5kA+b#kjH;Z4(r+c>~uyz0Y5G58n>jOLZWfZ8a~ibjs_%2SP!Nu@Q)xRaJ(CC z(PV+{y@;5Uw+Z(aPBvUL2yU)1F~$uYBOzZOV9%=qw!UunO$`9zXzIYrrmK|MFKzWH zTi(i9BYLPne}}ZP>XLJH>v+16!#<+j(t4jzozLNqugC($)?(jd} zc44px=^iKab{VydHme!k{oZaEOEU^TiMZQ5aRHnB&$0Kz_*;na%Um^ zETH3H4XobJ?;<8ZZX8;w7VANj+B(e-wlhKJKdg3rTy6z{<|LW=j@Wxd+m*9lnoNya zTs6XsPo;L+itXZybK$G``EhDk!Xx4wa(nNpZg#1e9U0wUW?Do>B-ylzizRv4j;HL)%!j>o8d z?AftjG8LxFMwp8~J)7I;L6$u9t+p|Dc#)*28i?=tj=Z3X_z8?&Y+bDWq8fyo^!%tF z6EL@!_P?#Yo7^v;RphdnuBp9QwmynJnR`b$w|CB~HLuDv)K;l7$9UH0O^J2&460f2 z*QKn6Ms1Jg(Ftdww!STasw*(?-7dxxABRzYUJ567?b}-#lk4ZTGuXCDbj& zb0HHs`$C$v8a6!hRo4o)E2P;ap)ISjToVhQ-FaLX-aqwoUME~pLakItczdpg3x^&_Qbz&qN)!TS3vI|p!?OuT&sbufJQ_0pa8^G9CA5}vYa)$&QE zr&Ptjj(MdS+byoqJHC-BhC!?L@lYzkefnCFJw_qg&~h6qgX!+E4`_+`k2BvG5#-YI zmw^~lDI#OEOUumMU7PgijOS_axsKj2ifj@l zc6Vp9COzvNzdLV0^|}$i1=?uvbtVklGi)|?%3(L*4WWFpi%I5Z2TAb!L5d_Uc&ETN zw5(jP?oC-TMx`>K95UC>A10(ZR3Wm;$RIY+DtPZ^t?TXKn8nGC-R#F9imH2gw4u!y z?DhJMiF!*$C?q>$t>vGIO-QS14mR)R+4M`vj@o_`W#i#SWdpvoW1t6WflA?28?)6> zEHHcPF3_(1CwtTB(ovVyfVXCpRxB^nbg12R zd#S(naD45k9SMv_*5~EbNPUW4qxFunZYt;PTauj)|BMR#KR_*KFEVF5zkG-Pul1;5 z4=X~pHn`2c`ba%%0(~lZ6^L<4rsSW#YO)C#(wrrE)#Q(YKxnL%)<2zRpXlqzZj^uX zr7*as_MxkuU?50!WjNT`pYHDY-Jk%~DBSM2NP6YB=8PCEE#RNCPFGfE&F4Z0N#Bv~ zF7?whGFsHXx@sTQNCXUeOb%LugdKdT9Eb)p)kPNbeK;IZK3+I~^1ywO<>m%ums{My z*Gz@7IrV8;1)<>~mQ&kP=Gm)qj>6C4WRv30*YXp28EC>}t4x>e1S;g)o4Cj&Bqdd0 zb|#?D>B(M`P??p7ekGS7{Ulu`^)*R45L03X9G^MGjDQY&TxgE4tsRW%j z-bJ!=KU*YRrP(NHS4*=ez)KAX0oytLuj%8@M1m`OUNGq%X>KCFNX}#AK{|T;M_gE2 zm{6fX>!Nm#iz=xTJ@xyWZ?p^J#xzxP3GBWt2QNHC(oOqUak%q&vA3ru7B%yloe8UOGTgL4>FgM^W0@}UvI<9{O2e0^v z3!ja>eOZz~{607EhLRW9{<2(9D`$0!n3zxRv)`ar)&7k_+ohu146Fms64w*ZPj#*% z^<@vzxFa?qG*k~Z4;pDsdYrm-7WO4>dAPd^dXS7Y#E|m8`0BZUSq+%mhikP-q+=g- zFPsYBhQ{p0# zS)fO@D%T#$km>YV#i?@fN2`5NF!O*Z1 zGLdpDvM|{$^{dCVgT*m@hc!q9@l52DHS+2@@gk6KbZbv0g$73F`LnKkU&#}4($kv) zCCJ%nzuj4rBom5ysKb0u5;cf)-RonG6SI+SFwssr5w>S?0~-K<8K`BW?1iac_^NaM zjutNG4Oz8j#VvmPxQB3vJFJ+%DM#_c2fpxp^uoo*6U)lX+90`@*Q}ORPj_PLGS7=8 zn=*wpe>@UPtVdndN5=KtW0q&*f!>|Co^_&lsMdo3+TvikO_wK#9mg>K{`kR)u5K04X()xty!u3w+DyVRl_J_aEd|V zvh?L!X#84E$r5={RutB8-BX#CcPz?&R?SycLTAh_3ALHcILDRwPnN3_Te!_L2j~@@)YEa%d>iu-1Tcd+Ai86-&344r|np!M29qD~(CF(gs>a`RB zNo^>FNz3xgOj@82+P6-`hTsJ4^pt@49HPF>Kd0D%NgeyCQw*twV8*Jetc%a4H=q4? z$JNUDW*CL4gW_TfgB0KMIW{?wu;ozXh$cB??9%znkT$S9Fgkx>S?8rnu(G2S5Pj^E zVg=0Rrdk`1b1U>E)}eY}){pzt?oFY0LR`Fk%Pun}#Wo2+DfENTQLRa%DZ|M-* z5)}~4sv`)KpWVGOZCH_$N1$bPTO|(I`o+iRs^=+ljXm=fLPOd%-tm}g+VINqAPR2r z?>y*WzUo|g=foh8+Uk2;$5b96f-<19-P+n3P~@*~t2KQBSYyEVCvOtl+Y8w*jcdDO z;V7Z#P@I1eZ#;+Vy`>$MzDfF$Q)@i!GlNk&Q)ENpcMC2L%&GZ>F*c@&20Hp$e$Xic zbln{$edTT&rGJ9W3*^qh(K`b>XyEVf>n3~)P|dXyXqYXm3I@+wck&|&iv+E4o2nZV zVXEsOa@H`CQMu%)dH${yiA6>R`)5sycLQ=FOW8I>A7uHd9#kyNc5TTkO6@-#{aJqsSel zo0IF|Z7?)RHld^SM+yGLA-!VV0w17FJqaP!=Iq1F39)%NkH-YF#za+ zDQRf12v{D_yn~B$_ohqFqCpcS<)nMtOpTpydteIp9(MArtqTI^%0$j_q{IZgDEijv zCP8-Nhba;27ak32u+xl~qg9+Jvq}s^ZDy*kPd?F`>t4f(}S@@we zB+3vn+!+8SvQ_mX*l+3;%zsRnPYespC2yL;|uK-tUWQ*mlqRV z<6g1r8T%InfQ&Ub=5$cs_}ZTQYlW)z87e}!*pBU{}Y%#7l z(sKWkZhnrF-EKz&vD0m~gaDgj3zV;pfBja)5&p-OQ00*d8=KSDjAVyH{{a2Fg^#+E zhy(HW5ZqYBbJpW`1wwtO927J$4ad*UDZr@|g(x8LXy>LR_$I9b;G4sl65&7EHu6bX-q?6Uaoadq)%cOjXr?+O>O}E9x z!V7L>jxUjFaAe7+BP=+A1C&1Wr3t$>*hJ)LGFQHsO&Z28&kLUnNX+hCP2S*`evWpq zsrD_a&DXy9fUcG#A`M`Lv)Q73oK^WUAl;gDi!*E$Zxq=ka7}%t%#2@<;F4FtQqgq{ zfo>O#Ind9)+=samlIBcigkpyeQ9QqyuD(XbVmeVV!arKSd8f6b?z}#h*|*RIQZwqj ze#-e<)qX8Ds)m=Ss<^aAwmGbn&4*y7O_T4GmnY;V#G5w(3c=7{^6+%COy+^Ubq)>I zw~JO;`?`~$`00s1sCjtop0K@SK*a7A7m2>kq2Qdi{$)RK*A67T6MKF7%{Hy&Af0A_88k*=8EOheDHon3m^a$Z*d}#-;ch$l zr;_4TQVizMbXQO*U#;YY$x76t7g-pCo%kB`+|bemyGCt}T9f+3lJ%haJ;wa{l!8QF zjiYT}!~M@XaL+K<=77cF;RFQ?v)zKYKcQDBCcb->P0gw3EMvzbCy&u)Ik>mL1PQ@7 zzoTXm^lV9u#CS_=FqrqG%;tpDx7J6xl7>0EMCnd;%Bw87^l9C4m0U&E&n<%{1F6vl zFj`bX0+^s9bR^`3GOLfBiuiUZ--ZZK2GAj}++ArSZT7t>AhX~xo|kCy*bSmH%+(eS6eD%%G;XA`v9 zM@Na@BzB2C?B6doD!cfkrN+tt7hh6m(uxRpCKxwpMwNdT(4&LRxZl$p-CgeO%u2~| z;`itc#pnr(k#jE}cV+8`Y-suqRoTJ{gy2cg~Aog^o=)vg}gwW%6V0 zIJby~s*gKr{+10+$PS*l=Khm0_TU!uUY|;5dj{Id;`vpz;uyAU%0+?pixicHm4_f% zR^<{?-c-?^j|^(sT^s-pm3P(l2BqQ#?JNLp!&Z6CU$8;mo`}}Yj9W{yZHHxA zpnJ;oN{iEX{MKKtTY=)Se5TBVL~t=>!2N>gGN~|XQ-HHv-L;K;T2yazE@AT{yCIIF zZwZ~>{Y6d2e5E_T4*_!4(WhGh&_RA|vvb-T~);RN(0g{_=M-K}Q4X;)W>mHj;Lu_jB8b?UTg!zLJic)l8Oy>N-sRjfR6p zlI2;kuqqG&K|g=-zqvn+TP&bnaU}oh0;rbc>WVhvJ9(cl)75+?KJq&?qVK+!?_yIq zg^w~aXWf<&z$jqj5Rfbos$qEY7lJCmu2uNeB8&cj_cDQ1mZQw})RSP#s}2G#-?UfN z>^471>lv>^u11m+arcgZ4bAm|51xR;owQGbckyhmhmoJ zNQx`$Z%UEk-FupH{p3lVKma2OZgTI zYu5aP{M(HBX$5`Q^DX7-eEh(%>ryXYSxXFrIbtV-p+IWRht1Z}+B^_6R|-MAyq=!2 zC)GYkRtutmmEI!j)IBu`N}sn0;vTUzHWY3Y8{S)mLaxdI;rT1b!V3`B^DoQGCsj8B z)DY4}(sHGSYd0pRg+ay8bo2ugJ(HN15846Kghu6)LEjj`QxH8D`d~xq$j(m-*rowD z5$dMAQSoKAZHUJjRYOI)>oZi#pE251nDHUl7teodmI*(q(4Z&`(`|>t0r%QTJ67UA z^q`B^0+{8rLP6Fl&Hob41i8cg`(MF_aNQXwMpB=}<0-ZQh%F(8L9GRJXZyJ(y;7P3 zUc~mI4-y9P{CaM~LvCX@XeQR!q*{CC$9IZuun0%T6>fv@vyn?Qg|9}QOG^qL0(lAf z{S1Pg$WPcmvbIZ%Z)w;=GLrff+|VJMfw#d~XVW?ftBD0gAt4(F^s2T?fADoK)k%Vc zEn36@d@+%76_kPEH!2biP3_uL17u)AC(u%pL`?A^kI!W=_z<)@NlhnHZIm}mFeY&} zm{%3}j(Xr#+EYY@=v@7M?$AAZ9pgg%!%ZZDDg7*8Q7kjbD{wGWhK;{Qg>a+&9%Ws6 z@I)ePE^jo0IIFF}E^|vBF(Mp-wP*=SPDk6iU%~X=`aK(GhO~b2h6d&HrHLz?=C81x zJiYq93gXFfZl^##ye$6&iw>3IIs4iA`Jd5#imdjP)%Uv~^41e@vHj`H+-310wv0hlpW zVr}jTjFBW#!VRw+dU$u(PmocmsiSo_m5({ICN-4l|b#+bcxqrmYD2Xaq); zgK?#-eObMU0?|PEJyt};zIv9yw;$DRyb{CE7-a2^0*|2Scdn+x4ADhIkBD*^i5G1y z$h-C7AlhoS=9)RUD4S@ z@4oy!qK3S}3{3)cl{cmKAtTTWdhSMB#+M;OnCI7z+)f~d)pYQB_ux2K2JR%Gf6KIB zXXKw_0ij&E1WRlO#V6f7O)eIr6S~BBj7|0n5OS1r=dw;MP#^{%6UWWtjNh5?ukdW8 z3xVt%Y&d|Fd_LPkW9NsUa~Evrt?%u@Jz>MI|KVTJdCo~L%AXavz%an=74eJThw{-M zyc<}t`>f7vRbeJcZpU-2L9sMBNnlshI6xjyo?k_km;5jA1!Bf8yR89)Fxd7m`r|bS z_BnnBGsIAQX7_$qxF{&G-4uw?eEk8kRbAj=c;zM>;(6u&;uj`T_q}h-wMRC<2?v-; z_Y0H)5N`5#)d66VNv<1`hdardFdPK>165jZew`#Q%Xj5GbkMxo?*(cNJk>(!9%osH zebf~Q#PZil>EMHkR8mS1Xo`3hhx%0s0ifr;6ysvFjfxDy4|Kn_FHr7vkr{GF@7LS- zF`&5}*{cp}Qvv0&quV7WeS)CU1|&@*axyYU0K|G18k#an!^PzQ3Mc8S)b?66uKWJD z^Z;&*OiC(v-MEQD%a>N=ZNxNMw0q_BjZ^m6oA|E@tQ8JYgWXz%D3CjMer?q6^}a|S zoMA^vK|bhGctS#lJmDx5Ao{?RODvFp6?DKx(vZB8%lk4O-4VwHY%$_QL0MRlDE|rD z+wTwGtG~VyhFgh8>4@@sj;RUwG&Nhj(g1K>87`*T%^E0O2Av^gUJ`>sU%Ax?IV`N) z1VH^rrPV(VYvu~i*oeF=w@7tf)+W>gC0)n^tbY%vFv0pjkMd!9-+>hzw3pVCxA_9( zhgnZEHxx~bq3OnYB@`cq(- zuL-le=d}G%x8c4#vmxoHvr-~h1ZN6&DLeO!rlug`CYvg7V!_a%wjqa7<@=cXjiraI#Z#etCHxVZovx ztfub)<82c-?|9=gMI8`l!)lKb=d!TC`uCoJxm2hS2+pqqC>$hVv;7&> zvYW6#s5-W|sL1(fjjn^;He$hgMgKOl)$a1(+PK3c?%`o^`^{l<(D;2Aw3$y_Q$}8s z^y;j=M2eJ`=>5AtnZ}rku|D~#*Zhyzfl)KADMI)fTSQ= z83u35;f>N1@c5n)r_Rn&9sRF2I@!liKJ!KHr`VW2wr`95-)oW1KzKv;u%orrdHw%m z>@9$*+}ihHLO`XvK~zGzyF^mD1WA$Z?p6_LY3XhRk%kS3gaXptu<3Twv4L-KJnxD3 z^*8^|8D^Z>u%BnGb;osIcdS-4SL-t}N^Owr5yL^ybcliF+xL-i`4iedaSzPA=-GqC z&8|bS5Z*cbeK44f*=|fiPyK;30P7I9tx6xrX+-snNhrj!+-xTyER-h!j9msT?@D9E zo-ns5BakD#x%Hs2$!?dGql7E12eiO;WpBi0A$N1BPG61&zHkVKQ3azN7 z3qZiT+sgtapBQiv?tT4-atiZm`kA!9!&OY?rfH1TACwU@S*+kCV-sL((l;~9-@Ppu zA1?#HtKeKSXXe^jrTKITV~KWMDi5dF9H@wkWWvlPL%e{g`0hu0ZJEoL*p&if=NT%00BXknvaC@eU0q<vARU%BjAUBtQfFRDXvks^LFqGg<3bbX9vu{GP+!vti5kUq0UpQq9Tr(Uq zbf~D?-saxa`;68seVC?SyN^^tb8Hg2feIW7n7(W0{`@|Q(%{xA0giq$M-@3j=D$Kc z^ll{g%Xr*L)j%aO8xq!0YLFkY zZqNl669dHu6+-JvDye956~@6)eN84Edo9_YNXL!JZJBiGiC&Dh@A z7|m=r_au4mSOma6ttjNZNNk87B7%6`UlM%NaiBs3!eL&2KQPxjma){|RJQF@2Q2p1 zB3E+1>$A)4vYFE7T%XnL`DE<5GAhM~D#0wETwQQ0%-PvlvUvi31&dk1fiI3?H`g^R z%71yBQ({@c{F(WwKBRXpSFTfjT39#!i;HW8nzZ#R5O@IijYb1mz4vZI_TZ+t}W+em44o} zXRCgniajDy=qyVlWq;}IIZ&EvM$qF@5dy=?Ednj$U{c~T2IC!FFux$kIl&8O=UKm% zVB`mZ$*RgS%PDnHTp}^iWd28=B?Uh##rQIRZ|`ih8f6@Q^HHoIC#jc|?t(q4d=>Am ze?HfWmLqsKQ+mEjh0_(Mu|yCwYPB)^>{gDV_J+q_CN$-gnG@E>_ZF$!$` zugkI9V^c>k>j*yxa?x}q@Apg|^5^|9_;x0D+^>e6Jv+qf>A@J5$7qlqS6iC)}{xHmoA437r#r?bgKJsrF-1ZOYJv!cKimsZQJ>4n$@qyLm!2(V5 z{sWgU&Au0+z_^vV` zE&vn)h)6kXf!YpuvA>YeP4^3uJOfWJc}4m?Wo@a3={!?>YN3k{U2=4+tda16>9pwF zu&Wct`P@W$TH1wMBPx#}y4;`r4JO@6zk6+&@AsMD9!riZ(n9$x;lV3gS<(nV4zHjusO|dB{%kC9} zV4N&(_jEIMwkNEgA6rn}bcsjgk*J}U`c(c2VJE-qrpA(Pk2Ut%6L^&*(af6k)_r%rA-#GM<-LD zdM)|6q3Q6WWbVq1%=X9aXP-Itak<+R^uz5gX)2;>dESR3hx-Cs=pp~MN5B1yE6z}g zQ<sagt6H2HjG}Jp+UQoqxD9-nn;#c$qkJLy+jy1<=MY+z)mur`yWxTqa-8 zV&LA~uuC=_9t6~PJg+@2t%c`8N+VN4jvRr4j%hRtRcdY$tquym1{gD4;oBOI_exO# z|0PqD#mXTM z{t$Zz3^nah(FLezp;!2D?}UcP4`B_ZTv-L`@mkGAqd`#wMimX~kvciL3qA?3XgmR0CJxy^;*34lPS|NIgqm7qvu$ z;dsMk{jEmPTMIgWA+H&Q*0V78$cCi2>Fd5zol#uaYfhtIX62>=U-%W@%QOl}Bn`R3 zVvF?d?S;pX3%~1+aQZVN1+|?i5r56jHmVfYObz&gU|ZzNo~V!hJub+ zyZ8}Sv;Mj`Y-|F)kKA{<(1ZI8MV&*gO2?|Cw~<}9&t=;>#fT}h)yoox_=`Oi8+)?< zZj`iJEqGPAV{mFXceS{4U8aT3N2v151#CjR9ETK_rNZjrASjWxe^)#N%3KNYGb$SS zX%@Fn&gh9h+(L#$jj`n%t}0B*#f(AwPbNGJYPQN4?j%{|Tw~Q~2{#>&6nrZq&`+Rd zHmvtqy7Mt@R#DJ1={Z#5g3(LPSQ_@Y`Uhnbg{wRp(`Sz^Z&P;9HAA!B z_MH!u2)(|DK=}T*;3pl4nH4~eaP14CCC^j;oW59$hy?b5o zQ4r>LTu+U@9L_4HR?=hZnql?Ot>`i^zFUH|fn{mNu|^5;LJxQMaS8lO2u zT$PgOZ#*3|&FV!-QM!h}3+?!s+)qY5GvU%M?+b;j$NcJ=Ybh1vvu8I|b{;5R9(>MO zJ0WxwGg^8ajuP9`Kkb@f;UaL2XPa2Av#8b2Ez(l(zQ(fO3ptsjk>c8UFf029fZ#Yq{+3Ae2R;J^mm`xDl;9K5B#A~rPg+zmxGQ8tCAKf36W3cmD1*UN7737?1SeEJ2LjN`is}6h36vXvUg)6)^m=>W&Iv#y!1R8OqxnS66AZ%XiU75?bOBdR2T_V2hLnVRR@R1Tii9f9A9^M;gg|VQ$=4F&L~P4E8!lpO zQMm<^;XezaunyR8kzsG;H5xm2S>u@_a?$F-^ATOy90Hh~? zp96JZ_lceKISqPEho4!d|0qsvT!QBcrx7v*J77d7Kd@i)DP8wr+IQ-FUOFKkq#okC zKAR)d*)iizPJg-IBC5H!K?lnJ5+p?6uPv@rnoP~(r;=kqb*AF{cd$rLpvd?km zQb&@i6aKXDF=L5mc!}AC_7rb#vy(uN9YilYU^~(_!6k7{WOdGiQA}iiI5?P%LXh5H z(nk&cPRU*BY$NWB|Bz+)^|v*O%yMmt_6&P^^X|S9sl<(YRk|}Rqj|*p_OuOLoQhZI zkym@7jKZLT8G9^`a^2Q`}ZILs_e+sQps z&^{bg2s*|q2VW2lIXkegqa%-S9?Lh}yrqX} zCPDslu{^=Bz_TmO>SD}dRaMj3_^l8)OHa)9T5X$H`S->oH}{iE%)3jNg@(04j>@Uj zsG!UcEcWM2sq!=1d0Y2K0Z$Jb3RbQoyoPqw5W+y_4KzClGn2OLBN&1JQV?`PuC7~_ z)K#+fAUC3QrBv{h!-7_r64ww*N(CFVNG}M<%`1jXUq-ZrD28q3JcQGue^i=qdDmB~ zk!z)Bi>ov0T~+EP!j6qk zXixcp+Ew>yF8s``u6W=S}ZU`cZZ!T&DDwY}oQdw?womS8*Z385x<0TGl#G zTz($o!vxShri!iHHTaZ{eiXn=7%d#y<0 z**E|5AgFE4L)fk8n+pqT$Uz%2Xlv6O!sFc)J#0Vhs8X$V${UPwt7+udS>`6)ofF8{ zT4Ri7oYYSsm;6qDFxOx!+;Dzfg_YMSqJ#svih3_*Upsk*b8+>&j&1VgnPM}q*sn}K{Sfu^h2BQ- zNX+^sJ5JrhK5T3-tiVxD-oxh{s@RAL6Eu?Q9c@xDlt(-1zl`P?a!>8vS@q|xcf#aF#u0ASR$_YB z%h#qoPmJW14A0M&a)kS0waSb13X0dsnIq*y)emi{T2wpMpK6k8jf1JNEO(aLN@agI zjb8%x=uDsSx>#q2>S!6mYB%wkdBP`Ixxi|e%r(QNy;2&s@t(_2#5)LEX*yBa5PBz7 zvEU>0YFaCd^&-8bi(pJ?xBy9!c$iopa?PB<`5=7UrT&4NQuHSAf$9P7sd163A!KmJ z*GDMdTJ%i)#jMVAjhX}Ru=ai;jr~oQoCr_XFaOw`y1l49vul^7;336wEPF>Gy2A@% zb_4a}G=AUfT!E@+6}DaKaxr^%NAaICTz+ObryoF_NCecHQbhfWesXjJ+FT~tevUcg zo^*QnE@+13?iV%(#5D9fMY+0B7?Uf+V&dt`snO>&@a+hU`#6d1XN!ox>|V>&UadbX z1eUXeZGY|Fq|+My#*q;`+K-WcdN=4gFQS2i#}7VF436eD6^CX#F5U-F;ZkF%?f%sm zdqQ6U=h3@JggJ3EMA`+HY`fOnpGz3CqwwATdWw5wvVp@@=R(uvUSxUDk?g}4kbD)} zEWsz_)x}z~*L@z1VaTrvKImFVSPa1d&(LG8pRt%AuTsrZ z>o7jJvH%;HbQ5d{iCk(#{0B;-+=`Y0MfJNsY;ui!J_7IC%Gp=!U^5m7JmDm5YH-0$ z*`(;$vE8U?u$K2-i!x>$fqxvG1AP0zW8|fy&Z_?A{A;EJHnEzMOSeRnxI6Yp$li$g zr|ZN4REQvwb@=mbf7RxHfqCG60NpM}-)HlbN$=`?2X*AmTs50ixlg(0#0I5pBn+;Hp0KC5W5$kR`Vw54iCBdk*QJm z^{37=48Qw`2!;PKV8{$fV$~a`mF`LX zTbU2<>d(A@fX~F+!4MkCUv~c!%%COR(*1D&XB<;vVUZ<-#r$D*G4gf2=BTQ9KnsJr zXVu{hu2R64iIa;IaQ_bg(Tw0lBuMO3>+6FK*2=eM>h;IVD7QuRUm$h|xqxzX3h=|w ztxv@w;{FEz9{USOGX7N6lw+GpkuDg5s=md$PyaI43q{>OHE7*>3(0 zCi!ZWy}$VfF@?lTJ=ua(m=fsRhqjZZ@HW|35&JRHWIe7l!^EZ%mNo)iZt65mG!isR zM~+EpLMOwEZAV0piMt?>Be%qZ=~vxVBzMS?AI z{4S^4`?EWB2}Fj9HHW9FjoZ;TaA!-LoRr$zM>2gamX^DQ0VSF==XAQ@QVIXrrgato zzx}AS^>xPN<%S^sibTU_5;iTdDV+VJ`05t3-c2VdaHX2CmXfNaWk#iWm}~sZp@omQ zmw=cQTb2x+a=N1Gc&b2=?bJ+GtifG;e>iNxx;$d!Ys{{LKr3l^9Y*Ngds=uE$o`d8 z`r00m#rRqXXJqR;7E>Pu+mvDQ+SW4R9xuNH1V<1mgVt9w#k6zcV*R$amt(uw6Te-y zwntdDw$Ji+kZcJ%?H?Jr=8WjEti8Ham%0j-sQpHr@jlHr)sFE{tmWwpXaXxBL@{qb z?gqIQR$#Q#n$)oH&&;PNg+#yZea;fEf6r>MWXiXYEO+g{-^E?53O0#>b(e00~F>G0W^{R42kXxBdQ3>ko7{(Z?u_kD!~y9kI2Q871p?dSMdS(A$+>IS+D3pWP&`;|3=((|@hG-JZU-SP3x z-_5#=KS>%bsXg30^7ayPvBbjtojN?5%|6YK`7q(NGnG_r{mp^7ua@?VC_Y737S}*c zjeNI0F4p-zhvPjIAKh9%@$#5GJ_D&wIN}UjW>Z z^`36U!xv-a5`Lgqg-zWKPN@Qn(fF)9R9&WCV7~FV7y+V4!{5<+c-!~lFv5PmS*6jH zC2!BMI}X?Tm|}C*n^}MT^hN-Dq(I8o0OIjfb@1%j$7B^|Rwsx?D`pJg$Rc5WZtSW0RdY{J%*>|3&>$Ea2G+?0E zL+ZjoJD=3*L9kw`|4%n=J0z#JL7&K__^49Pmv!#*X}7)a;0BVF`=Bg2=-nEC2IFyR zsu@PZ!A$h!g}Ld$x~ugVjVrywQ*&aMySE?xU#Az%W!{X}T00?Q@(>8G%cZaDWQ&!6| zsBWmHHYcguwy3kjbKiAq%r@_d>cfJP>o1H*JO&kFqZz~o9!y-$wG`Ery+b@IN)?-B zHclQpI~Msg9z~9xN_`=3-iUSvP-%xbCCbS6Q!d`R4Q@y@4*0`bzpAcf@I(mTC6eNc zSzh)oP`V6*b>?_V3lCN1NQ=-)_=wr7cIxK1F~u+Hl86rgLyKYyfhcO8bkSe?{#+2< zhdjm><)0yO+F<-x7@YjgLote|3dT7%h4Sgm0?3JCe91>ohLp+K+0pnuH~BclCbPNO z{8#{T>{bOn-gSI&vbJdw*C?#xwR@Xh`~{dJ_h$@M+90)Y-lOV4Onr@W@*w63$E=;G z*W{UJ*h{|mOTrw_Fs<$1+y*A>Ubt4-mxcmucVLSZlDR(^vqj?%XinYebf2ga4BysU zNGG#XUb*XO+ptBe#}i2~G{Z315y~*8UyPBI6o!Y7W~JtjI2*6H+rP16@tB0|q)7S| zAI9&@u(oR$RPKR=F%0@>aBy_q(|mF$kF-*i5x)RtEKS^oXQ@#ofr$ zM-qC4#Ai!*GR#ADjBgQFSfdisg}&aw0hLzYK&{dZI#{q%oq4Qu!g#lyPATU{bftfyG%RJezsd{eS!_WhU29NXq^_rEMAgorolqL z^k^@xJj?eSn;;Q7w)~+-c>k4dn#EA=a{^lHbZgvi(1sA_Q**2uy)rS#htSm`5=+Y* zDKXFCoQ<+Vq%oc1!2JN{xAa%71ieP$^*uRxYmp8}w@|~!IY7l+CoKz2oo<}{ditVJ z8Ru-^u4eOiBh%3*mp8zK-_J$oxlu-Pwpd_e>GI!-9X&O{^U*N7j#_Q}3Z}XvFLg^O z;&k>5xNB*9^i5~iao%qQqhC(_tJK(H?v4CTZr(tNP*NkDX!J=V#{CtDYY{f_s$l2X zt5nV;zq228Qi0;Ya=nJV0&O#|#8rO-O%`+s5CXFYIuv<9*FF*$cySkis#`#_!{H|HZ#X^sUQXQZ^}leEPnLstw`p#L&s{!&@f zOHs<-`$W^<4+LDTX~E>Zjv;&L-a0o>Zx+#%p)W<$m!=7)w(wJyg)Fh8rvFCXbB)V7 zv!Uw@AI=D9ZX{(!*N>Sd<8Lw)Ts+ptBtrCBQ+}5*=y&^^CcqSoCZjeAT%Dx<0c8@6 zl)|S;QFt$T#^2+;f$f{Bzy++w!gAx_qy0&-f;JQO8;_@qj<;7dg2=cFry-v6b#)fS zyif^xT>#SR&=E8Vy%Nf5HRL!s=M-hvAGr1O89cJ)(l-`zA<67nS6f$R%Q`ul4I#J{ znw=;U*7sj7RI`?$x#%HX&FVHjJ;iCw!})9=5|6&hKecKTNo3f3sz4!&I6wB0dZ0v2 zO)ZhNcflc|{B^qcPtU2$+4WU(9i>~BM$&`tSqF+)bwjr?Ad0wzno}!1E!keIujsE# zPHAUuz4`1VL^h4ilUh4Z)o88Nhp#N0D<@!J|HP9tfhcsDeUn)dK(m>(RGt{FpNID} z5$5t`lbI8#vg|XOVLQ{rR<*2#8sNHc(hBqbthuIdLqU<{fpd|aS|bw^>m6YQ1cYh! zqs5?9*FBViO_N>!Dl~&L-?<}-$Ibsoq{LuPj?{k{(AHvI0&^D0a)JA-WkQ#1PKGXF zn=|#ZnUDRmNw&TO63(E7wE5zO!i+jto+>$<)nP9Z5tTbf$f|At6)!Xj=A`n@vP9x( zW+&5e!TG^gbbC{?-Gm&sn&%Cup(OfTdd((|M>l?&Wz@s9r5I;Mj9N?S_lk5(L>omt zJe!S&uS-uq4=RZb!}+LsxuDVMlsub{aG^^?gcS^ALnuc71-OI{KC&(kY3l5DF^7*I zT7Z#VTdb>7m^>f`TO|!pB(8E?LJi$A0+36;=sSD0VdTiDesdjtr46Jr2gM4KFJQ!| z>;r$Yj%P3MJF5nQ5B@shXo_IjG+3NWGyd=!ABtXXli(9Q1w8ljjQe z{n>6&g@S%1ISBC>4P(Q)kw4f{0U!~W<`WbEY>o6Jvr{#a0h(2pR8oi+Ng6Am>D8Ou z!e^xsbr>N0W>o4>F5#lDqf@$E=jd66M2RvGX18v44L0C(lK)7{lQB?fE_e5Ysp$?OkJn{;7V>922i7LK1U zU*Z_@K$KHN$YmX~Ma1Xyp=b36((k)bld23}o>l9A&?o;0^WDRskHZ_ZRKN2$LKOx- zz9SwQunT(GuyRXML#AK!WTs?hGRolSd@F~44d*Yswd|9OP+Q2J1EhCmQ}Q7yL;@8> zK_gE&U6@7LC~vmd52B4r&L;AmZSjsrkSa+OU4rT;Y56*$b?MI$*7i_JB#kQ^T=18% zbiLhuqGg_551g6HQut;1%AGD3J2beJJ7nd_S8J2E7g7OQjTm|<`@A!@EpI{pD8#JjW4VfigASn6pf$VBT zBwFsLYSN_`86#u|MAR6*?mS(y>}zSFCh28rPAQPzcCjMVMu}5e|#Mk@wkY< z+1%k_ky{4Sr?&tVJ$-5=-N|?3W<|ALVE*oHe8Q0Ksl+wb(U2hr;-M~~U?F#g9V5YS zxVv8gEtn10nl;mbsY!`|Lv<7OAeB|z;|FuZ|4fsesuBSoBZnbr3>vL&B4Rv;2$WVp zNg@16Nxg=xt`hZSw0s5;@P25HxPBb30lW?wj@Fd?IP;fGIYyua1{(DKAJp1;da?Pz zD&@)ib*eV!oKu|!1Ty!5;}LNFv>HOOewN{bdGCrk&x3RaJvK)AN?@a+Hyv3vql6fbTdfBo5P%k^;oJvG=(bf3aYapX?w~gAV`_%@Ydspb^RT*8C-3RoJ$v!k!Igy$r*r z=}#rtl6bbHruw?4=Cj%Y0IaAGR*(hjRsZf$ps3652n|7WtNbKdd?ns-4WsR+~1z8 zOawS*D!-Rf@S`bdJ|>Ah$?ZOm4s8EFQJ^99UnuZuw10ZAO4s#m1~le~N6w)N@i5iAV+J!gRv6BcUdzfho>)uum;P3RXnH~PWw z7Y@{WxI3i42By#k-V0CkKgIV~J4}s<`C)hoxJ%}wsy0B~7Tx8sP)RHcc0oMUTerKa zsx|4-@Cu2_ft6EfXs$nRU{s^+0!B(|z# z{$caJVXyy_$G-8MZ(|z~Iyzhnu7mCj{D-LuIRrOV62ilngYE8rU240#6yN4+FNTcH zkiIpt@%(Jzp5%?A*BLdHU!1?_;*b^aNMT$9TgYs9H42VM>SucfM#~%jLtNJhKtbIB zVAJnnr=rnqMHtrpBsp%0hq8Msr=ROnf_OyY1%|W1yK>oIC=e(04-{y4$h68tuKD(n zy7cy>bD{fQ1;NSz&%pHHX($=M)gG|8?Cb-eK$i1YO3_Pscg>Pppz_=$ND^lQzij*e zuJiN{imT?)F)oaB*4X7a+g$L=lk+2cc}3sEUFFZh-#L=(mk2k|!yE1(gIFihNB}hSRJ! zWr4ZPP|m`}&t0uewKZS_v!1Rg3qzUK3@nhMg}7TTczV znSJj6+^@Rfd9PTOVgT+ild{e;q+K+7`i>dn2|QqbnIUF3J_V zbhOn;+?MR$3|?jt>*v{yZC7k>F9GZ=1$Gg2DZ&I``Sj5r{Eh^{gBqlqT*NR!cTV*U(GVaWkl+iM*T-^g} z$my1&`nq}l{9%vpS|Qf|;=rT$zi?n@yx9wg(F1@b8MytAMx!ywMkcG7g3DB0XaT== zg`&deOsH&`-^dH%?4p@a$FG1O>8fjwhCv+CFmQjt1$Y3~Kj`g0Ad(C9zR2Y34J^o4 zljm6IUta@*=um>+o8FYF^*u8n#N;N8vrq7B>f=HPG1JSXJhU@fVZ`#ns?MAcQw`U1 zMz66AKd8|kZ#Ei2r8-#k=3>&l@l)*MS!(?qjW^UnB%pJ+QUAbe*7OhX z^h6ooUdud5tDyXy#QnizfdhD=DPyi0j0@N=3@5&I7ySG# zHa!1Z@DpQ`lm3Bn?JqN=!_0G)&Q8&a1JXQ2X`oFyMSP!fk>TR=dK@-y^J> zrw2O`^J%yuho=UNcNSu%-oyZ^jS~?l9(Y6O{U*e@y?(>0RJ>?4bj2P6jPHqtNDALx z-SyR4dkidr%XaGP(N^tn1UpSuD==S+(j->k3kf(4ihuMZ`j6(w&=1pUYn9J&8cTcW zOj*pME4})hlK>k!s(<`iaE>*g^sY})slJfeK;|`=rCQVY9ZT{zK|NqaM`~;vLBay? zHc`|kO%(F7_NtH?nU^oELl^5bCMTH`MG&=vvT4#JI^|Wz+e5f`ZN276>{S`YAfATB zYx(_1)~eqzKhlm%Sbhaz#UmQjqslgH4R|Ymy2Im{Y;L;AT(#O0zgi=`D~$ygVQ8<$u1|h93}g zV~Mmz>(2+@7HT|yeqlXU%BnR2kdyNYL43lofOn0yR+nN_Ct7aef%+kym&yL}WTiQ{ zrHH;-F-6EFK8Z=dHZwgq5SUgyR4J7AOd%72EuvGb%)&CGAxr^2n1^oMJI~S<3r=YP zL>2To9l?E&PQt!mWZXe6un%hb{jAYWn%%UMKyc9@V&J zvMp{BnjH~ukjj50G=*uh5HI`7p^1mvp7SCnRX51G0om&TEv@Uu5C)uq{O06e6>btB zs2(l-naun@QPU8JVf9oT5_Hf+2>)OF*)QDmn1Kit_@HKtP%-mfuV^HTtIf3^a+JK= zwNotDxr`&3k7ofq?(C3Kuf`=7oRkM!TEYL{{MoQR@}uqoa}j$Oe1FAo@BzUkXnlJk z$b_2k3pM?3md~CsB;+H=u*(G!{K=u!DVzO~>XgjVehC%7ds<_68DKy=Z+t5FU6sss zz|Q61;F*$FBOw#7O7!vz+&8yDWGP)J=z*-z!K#ENVzr`+8ym}{u0qH>yXprIjLUOL~qh;g|V2F@uQ0R)G;8@ ze1l-pw?Q&?B4e{3oeh$gLK$kCP!@P!dK}{~75TBLZdpl zm1WmA-t)bXzfGn)hx~AYGot3zoe2wRl0cI6Xc=_r;>gDo@6cjz=i=s!j+dh!bzRk5 ztuvZVPEys^zzBSZVt%6sjIt+jo~56h+xwW&yF2d#b(9HFD4&(33M7ZrJ#?JR3o(Ua z>inqZjTk{Ct!`pVQHcjHRTndpfwyHtXkvWLOLi5As8ON!;rd*(5`svk2Qo52dFz{8 zIuF8Z{(9mhiahtF&fXPANhij9ihLgk{HrvUYFgS_6tz9^+CH_^yM6K}>VholiJ01F zp(}QNes-oSPJKZDhr^TrVnDC}&Wmo{IqV6m${iQ{CC;t~+F z=}RR&gpIdc5i@fpAz&zO%=GF>ja9qZawHwLzLud^sPkLt_Ld9&w2V$2CMK*XL&c+0pOcoT5jo@Ese&HHvChXy{HR z8iD8tl|bO69HoHU`}9_w!&Cq6y)HtDnnQ0*fgm zKt{Tgf&(pwCl%gxY5DScs=DPKN_Hw1ylo(^Z+27C)wLH;e5Mj6L$k9PJ?h%e+q&@C zp0$T2y6ry#$Z(+Kc&MGTIP@6j!HXSK`^PXk9QP7@&1YkuWnNwu{m?bT6np%v-sUE4 z%40$@T>~z>S0gzR+L?f<=kEKx_>COZp{rJ8`G_26 z6RTopY7@Guo_d7ll{Yk8JviR@!URr1FU8Az?X0S*1p54&)ZxQ+_SI8)@(L=%7RM|P zQZKj5%S7n&E_D}&4)$>H!JrwGPiz6}$!w0yNu&$XI${SoNu~!T-9w(%C>)em5m}G) zkrXJo(r#glhw}Nczt;HjI~l^s4q$zo@1g<`CuT6n49nUb%<+~HL_7?DVq&O3Q1N4u z18`7WIJooV1B%Q|ezd}CAKym|t*xue&Rc1ln``T8)>^=P+aMFQOSy4t06CP8T^vTH zAZvvb%z)s+(O2I(f`|~(%73zc=V!$iN)-qq!Y|nc&lL)Hwqpm7@g;koh!>I%vjaGJ zgg;Ln;|Wp&iJ%R3K9-BzmqIff_Fv-R!#seTll&jWe0Wdt=us&)!+t*c6rf%TEfp**x7Y@|yF%9@(ia$C2;6KB+LP~p5FsJ)&7?FlWzu*v_>>mo#AM@y2^l?XWMJos%eLN zEw=yh&_>^ju$2{|ND5C@H?hvni0aYNek~1Q_;PGsSM&0Vm+zqwJ7%7kxwrU@ek*6O zh)UM7$A^dRE~z;Yu%&X`S@mNaDQ_nyy6z^mDj9*2$^##rDNPp)TNii#^qBo#%i(7H zr5MTM>stSmw2Ta)mun;rKz`DM)XfiaFlj`mr$puCNFR`IE-Vn+q<*#7_|V$QM))8~ zhm2UZ;6Tge?v5ukq@kfs!K2(a41zXYF?R?3rHZC~PeiCfTue<(hpv5yGC2LOJ`NYG zsVu18t!gtJf>l?O@9!t<%p1N=^!dX;1Qj{YwI#@1x14Xfb;s82T#pnb3V?2#=7L)H z%-R!na$8l^#g4Go8ba$Ohle94bJMhHC53s8n)j#8gy=gst;c#ze)!y_bUWTAKSs~d z;Q}7$X4Zr&qY(@$lxYJ05(G8nV-SszVPs`pmT9rp>gbL?(RpG#xV4$w^@fwHJVp3S zhLp6(g!e1pcePkLsbp9pf)U;q4b(2%cggDLr1y3==u!}cKpw5Fo3P{tZ?H+g=n#IV z`Xx9Wk7?QLyBqgJosUK>XBNL0$ziF2?)$jKm9@3~?|FTF2L!4Ni(m#kieL#)jHPeK-Qz^yQTXee5*fE1J!^)Lq6Z0DyW#B_oN@!Cf@_j7FR zL?zu=W9Y9;PYKG9Bt8qqcWwqT_6;kSTifsHRaunL9>Ju2;}aQiHO19tP1X)MIE>+6 zL&~5lk?#9C5fP(@AI$oE-nqpf{^ZSD$aS*B7V-6h(knqMFfY_N-xe2-Qc|uFvFgc0 zM*3!|TW+fk537@t6{DgyfA6YyoT}P;Cu+dS?tQKKweuCy|FI8EMEmk=q@>R4{(~n{ zG{u=OtmL}#=jISYX@yZkhJ6PMT025=y9h`s#@+?`=Qu zF#0ZI*j&MNfrl?J=4V+_wY@iK*DoX7hrahsYyV4cxm^0lL& z(ixwa%b<$H)A-!-b-0#rO6rVId?&X4MsBF>olmiE{ttlR(I12!+`h#PB^i7$IL?;H-iuyGgB{Vwk<413)C?@iB zo9u*Miy$7ICn6Q9@?lmL96z+**B7qN9XAB_S1=} z6bxmHx-iVCzWR;fU$A`_Gf%B9p9!xP4)o5ATmei} zZAc{%SN@P$QsnRlli2CpJb}M|vP|N~j~<>W3hqD!64+XF)A=lpw>6LBLRco z8FC?HYmqA5eiH+`^I!OER%UF)ev>ghLWZ*^bliZ9V9x?Ilz8LTJ8FixF9;ph>@-Dp z41ySxcBSaPwZ8s_+troe^kE;ij~-U5dJOG&41;Y?ib>zszpdIF^9hINvY;Oh``U3W zxOgtJx?)(rJId)g84zEty}P=$_Usw!+Dw??_bMCdlr&%L=S$!qET(S|n7kn<_>gE; z!c$e@6wNB}s27n);_oI)Xf(k9TYj*g@O;Pvv7@qUNg(a%Yba%}IP88J0u1?YHw<(U zq+q3=Jkhk;{E#MOXlZ3t5!H%spEEjcpUrT;Z|YN4E|5yCar2`idq?e1vTSKR3hcJ` zc;c@J+3dl4x&nHPbaAS>!(<2zn<|Zc@jRYskCyW5kW{!_!neM>f`Z+7!`N7^iu_et z&k8+om^#eZ{w$1)%uJ2*&F#?`apC3vPiQ7r`Uf<_*8dMQ+kR;DKTFYn2OZast7KSV zWb>ECV1fp5mo923JXaDoaXs7#B~0=BM4eV`Do@I`5FO1S^P|;5t|Th-s;bc89Ujv+ z1TooAsA-wxNrP@$VOd#{*7t=D#)d|*#DY(s-^SAO3Os{U#VX&VS`DfbL8?_VhLHow zP4e{`T$$M$8PWj&Tzq_Wv}_^|H(8RiqbZcaO&vZm;RiT6>+kF}wz>~yv_^&qq8Jyh zj6RE>1v%e9(%&p06mw;Z!z(Wp3!pEdvT$u?SqwQ1NVw4K*3^iQP=7(5K4H{ONctO+ zP6~fZ_S4)Y2^i<$_W1?dq*lV`2ZD7kNu=RM$ zy~E%~e<2a;<-w=rAP?X!cUd1j=NUdT71@!i5d<^ThfJ;_dkg4 z8o2I3EMui;LZQ_?v-4GH&+E5%h8o}?U7r?9dFa2AEnrPNLzno(JP&?kIB1xKWQna$ zc$t$%tx--CT=BF3xtR}Uz1Ta%_a*H|jvkCttIV4Nqsg(*@ zSC1#$E61u$*b^yC?O16!jOvh}kEMn;ZhEOzqA^hu1v`>GJ$L)I9|TdSpzEWB!M}*{ zNRNJjMRgy0D;QhU%A3r2~Lv#2y!^42-u!9)S4YYO|E(0E=t1^w&6#CvD8enpW1y}Ecl(yQz3>72D?3Z zB2M(|)+5se3O+{#VhKHqSj!pAb+${{0@)N&^7z*2)ANUqQ~1MO>#{xMf~il}mm2M^ z%AAUn#9`;5LbD}=H}GKA}lW(qinV@CSWsl?d=T)vswf!)#5DSMG`wRXHOpvxdJQ(M%?|((LwBF?vU?e z=Q_}#ryZE|Owaw=6-#NNZz4cTE5^*qs≫hVy`cZK-}&N)XY7G%Ymo`E&kCE4lr& z<$kicS?BdPTaPmE-ja6++0Kgnf7P93Sd?4%?h!7uf6rQso^NoN zKi6b?SJW-CfSg9E_VF?fXMo^5j?O+|hd<(ZcQ}o` z-QIEto0i5?xFSk*^;d%cK`&@+k)1KnI_}o~7^7%mB0 zPy%D-Ehi+lV7E1VW^sP1!3Q6JL$Z$^-S_fh({6-7{8Jy7aDR&>SHMw&77yFp8Kf_@ zEV^#p){!XT5X)-$eO?Mk?+DNFn~m3C9d55YfX!kcT?$qP{jHSXB)h|5%DU(j0?U<` z5@4?I;B&vk?B}6tVb#7edidIItVi2BX3u77v~^CoMg>}LCA40@d8DOKDTN=FS@h+D zrV13ZK^w+gx&D#oXuqqzG!+c4(Qd?dC+mq8BkJm$MJmYVibF89>>A6*!4E6Y!F&z$1(m)XD7=8iXh?My}jy6P5ahen%C+jp?SfL(+{-3@<0!A_L5{H1NqrmSceJ?Zq2pN6eTkUZ z5XyhM+_z3nS=n=?*5GUqxyTb~^&y>>|Kc;w6Hxmoe1~m}sGZ?z~ z!p&@#&P7D+&$=$<>Qt>64#;BkeddY8w76@?=e`!k#iiC}^DHA$^Il(1|6KSfHB-mk zwPL>EGL5;T{6V;zi5H{c!RmQq+Phw<(9iv}Mie!}P&yky29=VX>B1TtnBwx8w7``0|xA z93fivgyUi0$vXth&&}oFu#1Vs%CM<^iH@xZCa5!OF??AH31Uu2(Bn5S6XTCMFDggo zM2ar9=RDv~6$q+!GfWmtb*S~$V%Nc@vmvc{bv9cbk?xIEUJDKLGBC5Wf~~om>Ty!v zYl`320%gdPwtJ~wj|@CKX40h`Szm{Jozpb*lft%eC~i4}`V|x<6pkOcB0R>d=r5z# zsbeWI#cqD;&7t>#HwArp!%l1VDpzl~+y24$d*{j7ehg5?QSgX}l&7WLHy#U0N+Qcq z+zXwZphZ*PE!t-Ga4o!-5BAX_X9!k-rdSV;A3H@H@(3=(g5j6yb|qb%@Q8uKW;<_> z>d3FpX|HNK4Qg1dmKq>>bYVPuVEKG1TuvBRmyPe$)zKn-^%-+}nqGb06e+v6yiCt` z2Jdh-kI_BbqXG9r@UYj-lX-uymF4UIT3$Xe^+9vt*M*W&il0?4P2;P@r`Qocy2}LxXJGd3G1(x$A z=q0$=HE^8Z%xT2Ll2h>Qz*!<)Z-;i+0lrdJpdq@deSRgz*S0D8y|1`*)e!AX(lD!)RH2jK(}yJnJi zT&yw{M+%?d4z6K}XrLy;Qc{XDT99Zl3m)dMt~y*CwUui$r!B_8r~|@xt*I zOb1Ixcgnfu=Oy9vFT5FRa4z|`Rgi?ir?6G-dc9T7CZ>_D0uxO^2e7}7t5ZqqSNzf& zw>Q4yGE4~XkjlZD!|Jos;kt68Q^@MHDK)7x4sne(YPNDUEHlUH0sEcCKA1Rvi8yCT zS-Er3ti46~>y08NY38VNwnEXDG((y)+yoe#$}~^PQ**H}5v|^VM|vLzbUP#Pi3F-G z)@@22Iy156R$f&_FcmO;d6AdNmF8AUJGb0D$hge2f^;0gQj(YDY?1T`LpP)C>I;dC zHpd%`kgc~6K|yiRpbkpamx14sC*SXI?C2Ykwif!xEDn={<0s|xoE+EN#mg*#QnU+E zG$~f1jQ$o!oW-Pv-PFmjgi$urAVWDghz}?K{Gq;t0Yb)FHZs8zq?+ee+fyA}K_%zq zt$?Wt>@6&8yVr$YBNT4=X?zrOgC!+L8NQ;>SC~ zp!Exo$Po^V(jCnz2aB>L18_aaN9%=8B=UC z|BNa4?x%7XbXD3ra?*@AMl4hOstUjpSy;|SsL3f9(l`}uF?CP^tQIw?rebN&Im#f5 z=eHH~YC{fJ@|j~hRHp^#jBZl+W+ox!4_9WaOV5#ZrrQpd&U3xkF0Jo*jRkh8`*Hrd zr5gt8zNk_<1JO+EA3Wyr#Dcsg+?oSgUc0@1HQ*|?zclA)8n?YgH*8#=+rWWETm(*! zLNH0%JFs&JDCM}3hbKW8{TV6Q>)%vL^eqn;wXF!63dWp$u@Uug$-+9W+NPmho5gi~ z-I19&)^QoeZm!3h(fP5Co?x!};rInGNV4K_~|V2USg&A$`5*Gsy*2GAX(OJN4QN%r56SWG-2b!{H*u$F4V)z{iMIwdF3BuI1UTPUs}3sZ$QCd=zEzK z^3@JR0R1Zc0m-+eM4p&x_DPLTuxPPxraqx$oum=Ta`gDdrYPdoOCW;D87y$vBb_hP zUSM)`r=SV&7v1&SgG z`xM3Yn}30y-Hb{4Ge|LuVldQQSknfzdJa%Kdlp$tbiy%~TrPTjG0|^cytZdD+~wc7 zGa>rQkd>0qnbQCrd~zQ+`ZcE0@}=9=@ZlrMT4KSL=-7AQKQ$2q^BW=8Rt zrSY$OBdZNnGICFxDXpvp*2{__)JKxqT1oAFQ0o24jnAp>-@bx6X)^Y1blOLktoESU zh-Y3k>Cg7;+{1MGv{)#at;hI@*~3gON$EB^icW}(Tq9(HVLxGO7T}8$M7hREIzu6H zM)Xo8^(thG&(_kVoP2`VGTOpOM*k*nuG&=i2dBpq)wC>+Wc2%Z54YSBy|?^5S1Q8y zr{@x6UyVAiyGG}gpq+dHO%41r=hQJ!XwFRxWzaXYpt_zlMEIQn_YqT8v4MjIcD<7qZ*j*x0fN?FqYxop}2P$cF zP)VDf5=!2C{;Ih-W_PzDD%q?hyD24i=3q;C*lQ1~*c?jvDYYM@%C|EY=JR^#6st`R zb}bP32AD%8_v9>Jz_aoTt9Cv?-EU+hAbij$7yG+g7i43i_TN-ia(w$rrt?7!I++Rr z;gH<3px6@o-Zw%>uT<)vpQ{&AF75gL`09VSOy^EM;zN@3P)fU%1+qA?Bdrli4}nYd zw}=crN6-?_&Ar~)sd{Xys~ZnOg1CN4E=kgap39kT6?Wdm(8n#|)Lq{`V|zKjO9b8=%GI~Z02gP!fi`QISdUgvV@R{ z9bV~XxsR_}Mo({!P;A=v|9pMNIPNvTQF3 z(?F(m#o_>&Iwv2Oj~7yHZBsKdgTry1AFg8aDw&iyeHa^yg@l8E2RpNOhgtCXrZ7U% zzyNG~Xh2+56-m@BkNJ)A@Zn|~y?g;UQ2rUYD@42UrU^<(cmUdV8 z#POo-sNou;JLS1xH)tvO*b%F(LTr*k4^UV=J_CahD=M<)DAkgG&Unh)bjrDEMKxpY5nW&f9J-M0p zeS8n{Y1hpw1)PnGab7TNRbohojHBehdai5dL z8=;G&J}ntr^ZhFsYj%s?TB->9#HoL~#CxCt#4~eMEIO-%$H~~_!p~%^icT&YnT&B3 z#(6S|wEUl?VsHABD;95bCAnWMTHEzGE!WWufH!x$j_gbsxE@$U&x7@MqkeWT{eq?n zM|q}sWRi`2HTy1irOw6lNj$b3ALT_h0i#K-DgM|+BfdEB*v%zW+xXZtN*Fqv;|1C3 zeoQPWs!Yx&B?QHl#OqsP{L#&t;)VQL)t7H;}xiK zWeQo60Rfp~K#jgviTAMaX%p4YGGvW4^r8rXrzcuFzp|(w*B%JCU%zsTkq81&RG2S{TnpGqvX(EtmS+hzT535JfhzLV|z@s;#zllA)H2?*UsFb%dGUkD!cMGCJl#mV(mq4U_UG z@wbsvmQqKb1)5f!X_b7S>M^<`nfk=^C%HtPZkvOQ>^x~_424535fvf7P%rXL&KE4M zZKQg4w!jQqnu$P#SVp!u3Pw*$ARM;1CLf^wT8Dq-CI1&PdaXKxR#B_i-O=jqbh%8} zVG2Bw>kE-v5crmB;`3T#z7(?3MT)wu&OXEZTBPsJK6fF8%~YX8*o93*#QEqjAt@=4 zhljm;?uQEgLN6ut+`?ju8Qbm!B>RZceR3gQLv(l6%zct#mE09t%8dKO^ibWYq+iz;k>(`kkg7|lWcF!g3yB2ih zB%^p4gQ@j_b9#04T|+A0H43gdu=GqZ=I4)31`3@y7Qa%QQXW=d0|+SunNi{DuL~_y5k#9!wiYcWgHHHtOB-E4CfH(w#G_Yy})Af9gf{G^^x zVAw+@Nshj+(NG9{?U%a=$3P4PLOhfcCh(=|rCHN^^yf-APvw`M0VxWUgowuo3qu^_ zwS})wH-sca=$aP+NABgbILA*a9wqih!>j3k@@ze-+~VRa!0B1c^cOQ08|g>I25;@} zu0S0_JZC@z+}^&{aO+=y6Yi`y{;x%CnP>&rQ($UbpBZi6GgDSt$b)yb#@*6RUmPmW zd-^pQ13936mu|^1##w^j*E&{G`0aOr%4-f)WMr#-68fRzPE2_Cqs4ygVv<;lW7gy*#C30zq@%-?TwqOKDqmcYH}g>NDp^1}?xO zS$JtJf`_<)|Bs%SbT2#2>>NhDMuKyCOf!1V+&*!#ghz^7dLhWSxeA7cWFfOlb6~gG z9htK3f$G>IA0-x+$5Hc2G;7)FsT>de0Sk5I@Y(Ni!0fx?Gpz+)N?E#%yp2W(#v0f9 zxA!%_gR{9}*lK$pxalwonw)m;k$5WJ)RbacCgeR@juCmx9Ey>V=FmSAPc`!SW_BB($=x21P`KbNU~W z$UXVT%<3oa#wt5Fzh1j zDTFp99M`FOQm?zijr^2Nte9;2fkZ_`Qfz;C?1xN6LrgR-5eJcD@Ax&lx;3E~c#q#0 zzf<16xv{a)m9%qVi3V-VR_-06PqxU$##~@3KO&uH@mnitZmt%f^6peVylSzt|2SFB zEDYS$fln$t9IH}FA48&|f{wYS&pYnDB1mye^hO znz8W+!6_3e%;&YwLrmPI_BaCOD~!Yb7H(0yv@t@xdg(-|4(j*&>1OCPUM(~=F#in{+6X`Oc%MqXvjF^#

d*HLMALN;HftG9OiS&=MMZ=oe&i=I?VQu^ zpmUJ~5&qfq5I?`%zCKPN*E$KNz7cAw=4m7-bdTmQP$)C-N_v#!=oK&DZqax6tN16> zd7@3!X+V=Q$ezdFQw_KPQ@(2zM6gcIp|%J4$} zZF}Bv1{0X>w}Z{-yE=s`svt~%8Rn0lI`%qj`>NR()iT(_tMBppbkBzm)`NpWZS~!b z<2mdVC`kSs#SF$L2oi@O$kwxk1wKZ3a{JP2_-rr3z#&@2p5mgbSf_m_u2f9CME<5l zMkL=+3{BSLAn1YbPH6H>R`=&Pa&fn906xp6Y+_+BQs4d$O3tD8OO0*PQTREneJS{^ zPjn#FzWND1fw8PbpPQc_B3YfE*4>&E#7sEz@CHVn{#!~z1=tm?VB5hj=aoszxwCLp z8x*O`@z=0tZ~|vij~?^!vFP!3oTWIajQ*Ms{Nd>DkhpUBYgd;PW%(K=3ioS|2u!(0 zn~Vkm6ooa24#MVK4jKW3I{WSWmJwv8I{>ec)bFCV74^V}0NkK4tP^DU&-L>O&{aU& z(KhrIOR6+HEO48k-z+O`;Qt?V)<070C-U^8g90#2uB01}4_99)`SNHIqR+BH%i-CqeL%qgl9wC%|x(<+3LM27D|(ipFn zaK6MQV*dm?R`7pKV99Wi-wBD=j~;76p|g*rrRS-`9@j38j5pG!7qZ|BpYYrAx(1Hx zY3}QH1zB%XSF8nJO;poz5nB@z15s&&_s0)V-dx%j@1kNJurlsYvERqqKXLN@eovf- z2}V&|-$O$k%Q(T*eqFc1pJ4#deyWDJVVew=fu9+zGHUkLqoVAjafMHoaLPbv8+{LY zFc}}o`V{h0CI2O;dVu;_{FWyf+54^3J6N)sO`reS6s<_hH5kxSZ(25$=}~@I255r6 z(fm*SA+e}BR&!cd2K&Y7@NxQ7mU0^^2;QBNObPS#sJcGgxTGH>Tm~>jqM76k++&)eC!esKo9%_tOwlaSz0cmu ztw{I~qpkQ)V)|cmZ%Y~!%1A`UQ21~chQy_bBewx0 zo4%lM{QMa>LRY2r2E3B>QD09T?TBgjw1B77)q25-yZ!jSW2?`c4Hz_?J<>KBR&wtN zzKkCc0vf5fW)$+pK+RE_90~iyLPR@T^?@A)H8q4kunHZUVeSBB7k&PZsCpLwTkX!M z&k-S{(+^9_6&(k~bz|zb4oG-X@WcRs!>sK!>?-*ai;y{dK|b~(V^6+!1h}AZAL*Yo zIq@%=oKDTfZ<15Md^ZS5&YU*tk?Jamh35v>^rs^UOqBp8hB!U7lEBG#@i&SbgIIDN zTyLB(eGZTAJs|CDz!z3&L-nQld@ux#c3fLSu1bKT{?i_t5O!I7K0SR4U@k$Ef!>!z zpObk$!v%b_=9A#fxty=%6|QW0++QPh-Ij{h9HcU6f^AX`>G$Y&6dy5ecGxvst%hu! zP)C?zc#pVmSc;l^iY)Hqj${)hX|OG&v#qy4Np}KR(({v9Y-yb z0&L*~2*i?6v<-gq0q8=5vY_yH$;)28h4pO1x7HHJE#O31+OpX+=Bv(8TN?i*w)QBK zSn4Ec%f0*1;j-IV}Z+D@+m^4zAn zognb6R@Hw?EDfkr+$^gJzm8z%E?Ob^v0xxU8*fF0wf21ZQu`jN+sTABLX<)LKKkvk zyCI_iZ5P&TA4N9=+%pVuoXuxAT_{eVHKKQL>X`m!ZZ!`y*$x&==tl_k=i0H#B#C&b z*bQxGZ+Q6XMI`GK$JE3W$)zH=CMtpaMMMJaD)J&o04BP=md#@LxMgk1wPXDp!lq5G zJ&Tkzu|fr^Di1PUn~`<{0OUOrf?4C%J^?!yNkYsjNe9zRy@HO>K22e=F!LDV zqFQc-k?!CVm*CH-OB*ziJxTjbTe(-&%18uv<`s=fQb{J2o7|EGSV&=E7oh#%Z-{s# zzm2{cf1;tGu|4LCJv%XIN0}BT_X%~x^qyRf#k@mg$Qp6S@c#Z#d;3(92d0^NFt5-a zg&|)6xO3o&+Z9;9yotM5di@MOC<}!1;<@ajcz!r!B`t&#=tM?LVKYulxDZ$F69eS0 zVaFyZwRG<~t%~7c@ICV>E2)BUlKv>LU4zm~JMH}b4+fOS9dAjMpc(7p4Kj9y4+|&mTU>D~m0Nsb~FIiTxkbkKMdSbmTEhSHmwi2%O#@;~z$#Ug82}Foe zmQ-{RTx5KAS8+ zvQ`2T1$3rL4pvQ>VwS9=Srx5r5k0I^y2Mcb@FQ3QPRy6kKi2>;BKGsVK=BtU)}Hi6 zA!tO>>8YgG75B43`W?7UsV`Pa{7hrru8jQ?#$L2Nw;giIs6$06iQ_ai^l~PupEZn3 zOsR+n_10;{I~yvc^=u# z;jszTRpHPy^hS>Z*6N>{=+CU~6P7BV8bIR$qfkqzqLod4sC~?+{7n~^6z}=hwe(6! zb-VLrvXKvv#2(8xI5W7r=NJv!=Mxg~!5P=Rb;8KTHhN_j2BDb`@jl?3&IZ869tzx8 zAGjFZ0AP{Y!(ET)=#I3@`~Tso{*e?q=o#J8H~~&e{InoBM~Mk~*TV}>LXP2>*;Fp> zf8~WdviA1qm#a!jU%ldRNjN`*j0$@@Dfn04Jw+D-(5eGqTiz3`0u=^rVQe96#?v4C?@cUzK5G}!g? zKkx40^z}7jlBW0%eFPc4$Y<8-At!NW)yzNsJN%602qw_Qh^&r zs`D`fp&Wwz`0v2*a`bhs z($bZ=rK__`A|IshXlQ_QX0Lt5+hr0FiOzifrt!Gnj_jX)JB^#$M1m~RB{1{|4nHRZ zCRj>_Um^ujT4kk*yV@_SDU$w(ftot{gEW6Nun*^=NVAHAsMXW#+D0S7%gTfkqn#`G zSqR*FMJZCCw3o@^*>*<2WSJuT&F6W>sct@4+U!-3NV9I=XYCw&e>8tu0ZCPrv}4jU zl6Bef<+GLkVv<=FisH7pD|_$KBv(}jUN$mpGk%pFUjj{ZOlV1_!ROIafD@ZzJzpYc zB2OK1GXiY9@T;wk=C`)GX*uGVKcrDcU3*6WH(oVN45Tq(m#1$|J78QyqMfgA00qR} zOk8>z))^-HJJ!i~mBbUOWJgi*;jW$R7J3H)^FP9)xv2eSdC*(EYqa61dX4c|0KTqp z6%W1we|Y`(jBrC+vZ2&3!HV&$8po}Gf8^hB8(l6B_ZGf|apq<4#GdrgnNz)EBuPwH zx6((C(uTohODppP?%Pou3m^}UO{pmF^;;ykw3!;RyO59Gpv9`KR^7@*OwLkcaf_^q zS%H+EsL1=YAqtwpMjV0Nfe;KqvWhB`D$VX|oJ2}LP{jxg2iN$dNP*`JA1Bw2g@=sy z5iM3^86j1>Ht^QV6bzPc5>w0HxwD1k-gaLr92e4!H7+*@7C7PxY9>Qok0Ux3^x#4` zaZBWZ@TQI;(E9?`mOS#JDwDG@ouzH>KOg6cesGIo)*}N3!=qEvFI-iJut&J0vJ>v90C`kUC4pr?iJw|6Xcwz;vw!hcUe0T~nd~U%N4NE%6&f<~0+GsK z5(=x4SB0?MCs5BVGRB&g2}kH6U8yQC$uwv#521*T|H%}>dGr?6_?(HmnyKS0VAkxw z5GuLKB$VcMDvN;}=W4WE0x4&of(ENH1S_Zen01MF-;<0yw`9MB(ji*2d$z+JrL;jm14F?2$x5yCcQdS{)t`qg z5dP$je)W_;)uxim_vrql5}hD;cjluupr{wkR+|RzegB@wkVs-_+pnSqBJ8|{QoysM z*I6}tch_-I5F4tM%!SJ=XY~TWuD$War>!aEVgwAM>(u%+Z{rja@~z)LThK0q#{P7m zPABiI2h!8u1HyBOx%177W-?GUf!#}^66{9w6F?uhhX*}vXB69dZ|p%D7jj!Gx^tRq z+Q&l&E@I+P0Ho347W#jDvYHDpcFutuL6(2o>sSVf0_3!0MhkU{<1zNSb?>~5LZDK# zlM@{8aDW37PU(*kTQd(^{L>OCSj1-}sR)U_qk`i8mxe`BVYDcz0g`~z#@_dJHzjfC&x5==w&S#ceK088 zcBWz@J&7AQV%Q9Jd*s<^_n0=sC-49)_Uy)%FX##byDOl8-4%QjP{dyx|HGNM{%glu zB%sUYv$w8`oCUiis1PFep9e=NQT#uA!L;w}$zK&;eFFcC4&LUa8#oFI`8LEVP|$67 z%+?*hkcJA;2Ae`48NTD6uW5)Q=21^bwc{6_VKEqj%@%M$Z|?EWPgKsp@~}Zb3jDs~ z7cMZpbaVEB^N=zeZ>5idB7qJ%_rM27gJB>4g7~ZBi#Yl3;3y@If8Glmn&C@Edi(%mr(F^C8#-RTh04bnLX(k0DMLpKcF z1Ml*_ysziFpXYtPAK#B}`}o7Q);4Rc^E}oO`?2rG`l75TjgL!#i-CcGFDvs}6$9hW zF$MK;?uushzwoz$HUgA66FBk3(aHZ<7m&l})oD^g80 zSC8rzv8yr>gMq)G7aw2W-CuSVp7nCkuC+%+FFlq1o^a_SQMxg=+qd18f`5|xpedK6A5B~cA<16JqFI)S6eOb^JJx4J6 z2Fu@fzVkqor7^1zBNY?g=JsDtbbNG1wTtopms7T9RE;u!mK@}N6G1*nSf4#}3mUZ5 zg8Fkh7+COoiS>AAoHMEtEI8I*rex<%h>PdVZqyeuR<{jtyt~6mIgdl z=oWTU<1CY(Z(VfBzuZ8N>h~u0K>yy_Jj-AE`#=f&zw|>c5bS~UI)6Zea>6PzvZim# zQO3Tw>$i^a*L!Qh|Gn83l;dv`EWLo6@R5D`5%<%mO1@qpy)vV@aU!p0QROPSH z7gluO2|bamdsm7of1FtG!+-4=u>8EZ`#BYzU0vs`!@5w>5kayz8gvlidhni{I7?){ zoRpUh=3g)EKh|_c)#8(i0y6~@$$8wUJc(7wWqxtd_`v%X198%~z^y7#@id}f_HatK zsc)&@K9|{?>HXUjPgcODM)q^XSba1u4(1jxId-mW3loVq{Dg;{jb1&WbV|W>V~a2H zzXb<%89s}|W6sz}C{%Ia%R8zkRyYZ8^;b3QJ9Z-d^OAu*&8!JSg9Pqkwwfwj9^8Jl ziAbc%jMAjK3Q|;B+&>pxmH+H{F$m{ED#l-iz4gp|$yqY`LbVXKJ9p0j$<;}j<~rOQ|) z7%uL6P4SoK*7{{$g$Gm(W!tC;bm}rz{Ce>zR8r|+F}8i&z%S)jVG^AFo#s2t%(=Ya zZsxa&bwdB;NxLS~s6-mkOp%}gIeys(2YlA*n|KWh2XCG(-usl}iml)LthncB}F`06hDozYtV+`8NG*&hV$pq z;C!Mq2RX?t5AO`-CIg2l$>=DyDN#tu!;d@*1$3mfyroRxkurbrS zAuA=qKF8NB|1ENK_bJ<>m(*6XXx7&6V`sRWW7ls2kLV$nDp_9}mu%b!OH?s4^1a2= z=pO~DlSm7!H9%6n)6a|>NEoZH9o+WT?xppv3!>$gBGZY}G>dUQ7_c4cL=pu|M2+vh zAI)<<=9*WxDSIKB|Zb1)X$?-2E{k&p+vz78Y_4|HIgGwR(Ew}2J+UAu=C4Q zU^X&Jzve`k^AU>kj|-GSx%0Ds6+D-SOHH4GpIOD4P5(Rf3iY7gf!|q%E$j#5tVWDu zdEZU`*;qL%yDb(hqoiG%@=LNSk5u)uxNe${K)*5515i_Abi*&npZD>2mN(t3riw=L z$E`bR)FL4|@hT2|qVdMuoNwo!AEC&NHARIJc$_)R%N}_MoWy+U>Z(hOxUhvir!8$U z5??LMfaX1m+qYe2B#4cS?uNLLPlZZ$zdcUu2ySf8I>^eNGF7;^?kp|I({f69DSPX$ zYW>^eE#+1-*7;J2uV7;;*!ep$9fdjP zvXcv1AGyZHu?LwV2}JVhLGi&o6T})tX<*lE+LeqvITLC_ncVi6)yU*SODro)WWV&B zyAUiU9_(FAD<=CyQ+FyWf>v13&Ge2^X^a4hlZ7(d%TeFQdG?UeV*gL(Cwa06CXPx? z4f9U2)34WDK6GWi?_x{k?n>>Bi5JwW$wm53DUqo5p3hVq_v7}!akB4wVLmd_s3`bgeoj=+_a`=4i)PKFD;`#}BwU z8kGl+1qJbshn81%lQ@gVUJ|P$PAeX<*kqz*Ac#!HIiN3vSSt1jJbhcd?Eq`9rH+OT@a~ym zy4bP54Stiifk}&B$;EFTJt59F{S1e6<-tFSb!Q-wTcNCAPDeE}jAPYx4zV+;)^Efk za&TgDTl@!vabhGX0`e@sMQ}8)f20!`%fCx^zqp{TcABqOmW{*%Cn*@>sGyn~!_=ra zS3Wj{qjml)zp zX1x!>OP*;H>pzb+HPuaf!Wy2(X*VhlzHy%0cl&OReobPXKWIBX*YM(A`}Z8QOnDC3 zi&IWDqJYM>cYe8WYa}x3(;5zP;m5x)@ds9+pQthoJmy~wW*^vE!%o?P!-R8q=rlhD z25{3L9@Wllc{EI%m{%jN19k!@ZJf*7al9-EL1JEx4?stQAtjS$q)Ma}^?m)z=!vX% zZ*TLKcC=T&Z}i5>Zj5gyV?DWJYQy&5te&e3kAYLi7is&dZ7;}s@e4*s#Bqwo^hOG7 zTj`)TFCL=Ntl!FX08gIsS{uu)k@=P^Z|w$@mC-p7+Peiz#bX9H^c<*P&ayiMiv9u-&WLGY;tm5#YgxQ`Bga?Ub!@JS|f{wp7MR#*&2{NPl9A8GuaD8$*wufEAN9KWL7*lB z=lfm1b}0=~6ry0K8UVY$k-nhgXCyZ(rTJ)Rivro?%?9d8=3JNf&s4_t6Mpj>W2#A^ zbO-H2CTbVE-uoPb_jSs{f7x;bi~Z=CN&Eg8_h&Dd;f{bDgk23lI9iGNPdR<{BI=TC zV~k5awDWO?XxnZdFChxnNl5;~*K-}=k}2kI+)`g$7Z7l8i`LyZP$aKA5er3wWZ|1LbS0)@}9Ur2dOoX-pK)chVv6fX5Xvri@=)zyk{nmrNpI zg2RC`Fkdp4io?fj7S0r(NUUXTz)vj%$J*0JW`#4GJzvZ}OkWpG1Ug$aij9lsm6ee4_66x1nVXY>!s?8-xnHkh zvOoV=o*nu|h}NAF(H8RLU92R~|NU;MaG*GjX{~D($@{?bRwTg``jI3=` zJd6_252pj)ol%`+V3vl4MGo5b*%>{yEPo~HWf~R$DNmd?*ROd-y2A7)r#kV&GkM9P zdE;ossR`98yk|^zd1Qxrv#d;BdAfhdef7*!q)D@QjM2q}zVcwvxvlIH)L;;R8y8CgE=XORHy6-<2 z9iSyCQ%jyGeouMaaxnzqqx>huM^MWk$QB_42kc4o6DORzN$4XGjfwUO8xKdFt`L`68qB*MW0&U zB-UHpueR;f&aL6Kwo&bQg3sB*Dr8Cr$4NA~4;NPiuEGImpg9|Y$oUjoN4qMkv|T}| z9?-!uP#5`wZ>Hmj$sNNgq}O2Hl@~~c|1ej@Qfo1OZpolm=Bc#2`;qMf1WV5xC*E+# zpz_LjJ;_sL>TL67Zt&8#nhP9J7e4MU`4aXbqAqU<1D5kUn)jI?Jw0L}jK`;u68aq= zZaRyrug6>BPwKC_vh~7Fic1s>zX`9Q+HcDCe+8dnqracaKzva@*sT+e?m!ttJ5d;* zfA`@`^~mxQ&imiF1e@8AsIs4*>I~Zqkik1PZ1LX4oweMc$6MsOg8{Q?EVdfV?&ao; zX7(^x{m(!MH_k{oQ}C>m`)8VeZtWpFrf}%F1xpjj_c|}dx?Xw*cdO*DNss0?a{5M; ziQ|$|7)e5|#E6*@br1^uytG`%Oc;5U=4pFD6K=l8`-)@@NzBP0#Th7<$yPbC^H_E3 zD5E*2QLaiebL6tV_ZVA5oL!!U)Sx-7a^vV&c*%>>(!^UQ=N(h1f?`_5*77hoQ?0^9 zwmjB~i$G!E6;J%};WLITQ`TOOy5679wpJ~V13w+H1-$mY8u!l>T!d1N&5nu)k|UJn zYPUu4^Ccr|Kvr7jCCTctOl1z&1X_Mbr~7|#K=asdPC=_BJudiny`JEM1_->;gWIo) zFkyIJ(D-%_n!yfebY!lHqs+N^L|bZ=`uxTEwh+ix#f%zeN|l)Ogs7JLM}(D&pZTG2 z`3unHEA5|T#)k2j6IU#rD&d%aXcnzK5TP2cgl=_+=x2M?{n<_{Ljp^~j$WUhp4r~H zFHnnx_~p0R@Z7~8F?}ahEYGtnWa(%Vq$U;Uc%<5*&siIU39W~skZoM zgar6BpCO1-pn%vDZR8w*9p&I=C*rfE5-c5t8eLyv0I7YCbN(X4E4lGk2+b4&`l%m zFKqhx7q;iSwsAVoNyWNVo5z_7Ec~rY2N7j^{nx&`B9Mb|f9G2Ur_sYxqo2E!S7$%_ zs%GjXmEyQ@4dS^-&4Zr+gxT*hS^M%Knn0IH6x?!mdWpC?Pgs{I3T@DjE`F?L5us|- zOW~A0nydfbuOfzRaXT_I`ID%*eZntE`g>7H7Ge9>Hd45u9HyBICy9?KA4VbSd;H!A z9n8S6I@aV)L zn!2Ab=}A8#4uzk}eXpPSz(G>_OJ zSLqOZ&V8Jj4HAHgULCG9u-v(<%BS96HB6cC6;gjj6rNaR?%Hrs{H=8&t=@fpb5^3G zN$P-QY^ro(J4AjUPdlS;4LxPfbSRKk2R-UQ_t0KrC69&P*Fot~M``uQo?xaRA|Ws#q41w$N;uC0%o?T-tNZI#)TGATuD zjw^q_>>0DRYaK(>BTQ*&GJE2ZDz?<+st-bL?Z~Roup^q6cl@Z?_%p{S1N-&(GhAZE zRTi{aj(?FC#y9dhj%;j!qg5p~$E?@;G+a7PIo8&WH@@qS_1+<{(q3B+s}*igVu#v% zc(?x3nI63M1WJOV&+jJVGUft(6!VmuVxm*)cwtIj|)Mo`{ z;uUb`;;M4PQB~@sZ%8bt@7+0cFn-@k8>eLM(nPP+HtzXF4ap(n(Y@p2E6eWi&pgdb zAc4gdPV20pt!`en`9rJZ)bGwpA99V~SLm#uZ}iFc6^%D>LTcE3EYcHRo$MlYlhv7# z>BidM%Qw-qCfDx-^B~3c$Au;0P_y`1(m*!}?=Zii?O?dFx z+Wa}yWBcMHQPah%ewejZ0rB$b@sy8^IGNn0sPXD%O+CMjL2)ZhgJO>93Y5RhM;00M zEDC5;ML>6;Yi3--qRG&D{6B26@@u#WDem}!72A8I0+v2;t|6B`NfQ|!{ov=9uRM*W zx}J&gMl!Cc28Kao0Xo>r&LJbWYdcV;AHNPe`VEA=qU=#rkC(u~qw`<{7NltFgs4}i z>9ym^WA=YW0RaTVcgl5rz5g=qqxax!7!9)Asm4?A3#+>G| z6r$Ce>wkAZOrG9m^CJJ1AP6`O`$avXX-PuppZV!MBfWQKC0u}^BzU03LdrsK94D1f zICSB~B(u28mK~9zXL|Zk4S)iL6ESSqQ;EqjTWcJHW9ybv>)Bii1#&XI&n}y0O#>~w zh#07vg;bZVQUqVEVu>lJ#acZHe8sYB z08jp)BqrBS&-s`+l6a9e0RS;*QDWhosuqt_XJ%#={e0Xb=d~3=L+1oJ8D9B|=O3b9 zJ_a+RMy2EuCCZ=tp;ZA=dXn%DMOrcs{`$Dsz!yXtCqTjMoDUH#S61f#nf5L3Tf%+4 z+xpw>jUy?4KqvD*)ER&xdcpih#C!5D+39c26(iR8&jm04Kfb`kzmD=Wwbos_g|qy) zi#d3ha-9&{tyH7@%H@&LrOA(e&}nFFdY1KNYcVf{wX_E8>f-Knces9&nD2;z-wWGi z;m+1Q;s{c_e1}fLgHu<|jru82!p;E zDBS}A!OS**daWFlrGJPn!M{>uUONG;$=Bmzd|1rs>07&@$M|sh-&Ww?{MG47YP$#N zOMLgN{f0i|bt|5hL-@^J;m$~|tDWjv-A2k0dG?r?XdP|x(USbCXA_zFP6?|&X2Fi3 zc&B@Pc0&Zg%|%3r11f|qX5lvbqy1|ko#RZOeq+tIg~}bsi?cW!*-i)X#9wtzB~q9B z!!VSa1GP%W>6iu%G}jeyfDd?p;w6Xk4=K?PS^guP)8cC_d5!ySDZ?W8=d&Mmtb(15 zg6&Xv=7@wyjaga687wcnx94nn3tR$kVC0z;z8ctW8=FIpG;+Z!c1On$(~3{7R^4EJ zbB>!j(-aT=QteAi0Kqc*w}Pw(ln-T*|4`<_@yXy7<^!Az%ivFIB{O7~*Q1Y*K>lAl zQEW9xl%5w@Nj=i#45lqz;GU%lO$ zJ+?>vtlV4j(Z={(J!>bCGuSXbP&GO3UJ)=+{kPY)2P%!$LV@Mzt)qRQ788$6VyzWw z93B_LM1LaNisz&&S~0zoHQ0TbIaD%vi#!~ZrR?V>CC2$9?z1kfH196Q-BPf@x6<|k ziWVFt#9iXk7G#waRMJs$?pP-Cbo-$H6=y(^PqEyR2;W}h4^%BF&=}M{-2EsQ^N(@te|QrHQyg)*E^hE&8n_7- zC3J0aMVFs7{`6RVTigvHN$a5A6#ZlVeLEmV@9{u`7Mb0m*Tp2^_GO?Mx}D=TtekPA zS{vsVt^0V6O|{5-FV8j5W-FZQeoH%jGbaj^(xo;k8}6L#C%X!3k^GKYbA6MgDDQC-YzOy8u%k zDspGx<`NykjkbOrJL|X}n(}r$$G|O^XD63Qt|j+~&u(1o)va!YA*oEo8+FCKb7F+- zO2oy`8Y6G^QwO{^LzP3K#Cb2^l8P@SDEUnN3NNXgWi3a_s91dc1ysMVC`lyoNV20E zY?g~iDYtOM7e68qju%r`VixWSRK=#}F#8cwG=I5+ute$fJ!X>w^uTUN&8c3v_VM-m zXa-ubt-Y$r2?`CGv#o;@e%5d+q;l6z4+v~n6Ihi4^?wsMZ&oy z28(g-MU3z?&5{6ba_#V(_>Gw)=rlbf>(#R6PFYYTO)M+^d1&GsdiqEVpgflQXJhX|zwpg5M`OzW&nP6u!C>+I$(o zbB{r%F~xH^4IlTguLy4ws+&)z;m6jL&Nb3M-E@?EybGk4T56j~N=4pXhz_EjG(c|p z;Q|2aebLU-EC#=^7`ycMnO&i}u6e(&i|)Pjiecrl9p9{q2@1Y_bq!~uK~x*e?nNd| zho5`ZqPZ$vN@)`Kid8dgnpkj3W-gsZGSG%$g%8$8P`TL^9Y;l^6w2>`krX+>KS7A` zLG|8MxY}P+DSfu$E}6)|ULEy9t`d{XRQv*F`bAm*YPi#l2O{%TxQ{M&<;N+Bsd9a z2x~&Xi*xHE`EPbUYZsnUHF=29|0*?1={PP3+`@kzrm}ZNm0CVs`39%7WBeW-^+@we zWCwpEu@buy^B%|-C2={n=>xPcPbQj;;w-RKfhmTwM#_9Y{}X)ODf@s@HgM|#l&&y! zN9|m0klm!Y0N&8(C56(@B!iFGbgqjpXi2|@WxI`j5Tg@`ZQ^h4V8$^lUOT5Mp4*!q z&twTZX)`E>&Ujclm&OC#X+q&i<7^Y;L(nsEcAJiB6PhDi^5q>UH;2=Y<_hcL64gj# zR*AxF)?x*CX>CWC{o%c-u{@6x1JOD;fTHuAakR7z3N-T(!z&`VQrzeWM7|_^01PKN z2m-S!ACOMC01Fr%atV{R(gt{{jw`cx#Tn1D!&OP|kt1pY*xvjXSPC1W@^A|MQ^WV# zwheXl-Mb+z;WCk$9j@CLAi9Ub-;^R!Os&eYLvQX-jCvfd-Jxr!%GaS45*J1%km$nG zgc5T$@PSzh>KQfKH3i*DSWZ}4=(@O9fAeT)QA7}2YM=S=XGt#9;vj()j-KC}UDn<2 zeswENfWjnZBo0E+tn=c^a2RQOhQqP#CpS2rK7)C-nos%`<% zqky(1vC%5o8`?ID>%|GVjcEX_gb8|r?;adZ`R-0tc}0YZE*(!wM;T?L5unaH>wR{e zHzGOu$0(Z4*=~+O5e^*^50e+q_7BqTwF79AI)@{Lw{OXgR<_rEA{P!-HG&yxwrJ#WhY>3(dO46Tu9JtWzyuK+T%vq<4gYlicH_k9 z5%{r{uU%8!;CL=HcWr>%MYp(kA#D`g*>#cuem)9{ieuk_TbR233>X?41W?S{o&V`0 z^Wg955P0aiRJGvO`V)X3jj0#dKxXBXHhi^UZRT;fB0R*dv>7he-Z9kL-8*!pnJ65a zz!T8KCl z6s{zj)@wm!Z=1`GI+^*kQ$@OQRH3Xvs#(GOGr)Ox}@wKiw>yz3VK`uJ18;>_jPa zj<-)C>%0yzK1*0|?OvyGI6wI$?C2=wfM!(W^FC;>Jx|RB^)=(Qs?$m!!z{)7d!8RK zVm0^Nu`u-T&!}=S{x#2+h*ortkG5F8<5`y!LAnaQ?R|an=&?JjkE1C|Pi+Zk_E^me{L>Jpl>xCD4Z*PgrSjD@|Gkxi>{a!A$D z67Py2W=N1T@Gb5?&}Ymd8nP*-uyniW3B*Jb*zs8(W9Q3Q&c>_GgE=}FqR06Q)Z<|T zCC5X|1|NqHU7aE9?0K6oy=l&LiO6M^UT?Y=^!rTHew^CDjP=Fy zf+J&^F9qMUSO&`tRkFKKT;|d@^SH1ShLRa&)j~eCkMQ(7mjmv1?NL;M6Z3k!e6S8?5m#FrZBjc4)T>i z9Dx8l0Y_Vi*)5Y>C0jt+cG}-(rHtkAe3NJOB7%QK8l_N~13j)FPzF8>sH zUm^y&dYIM@c$l|>PlvriL{;`>>PtfhB`4@@Lp*~i)E-%`iQ%0hwW*lL(4Zd%@V1Ax z7ck1xrOZI@`PJtMvh8yx&%t1L@>rpnPbCv{O#?u&%E#W8^X>t8PZds`x&-sbLo*;I6?0iw+2pDx4OUDumVy244=0aFBQ zUBH-8jdP{CpsLmOPSo-xta%9uE<(~;qP`C4&FPA2-Gz4gHR=)WjhHo%?n>BXPIB*1 zcuR~%MU5u%UY}psaAsItTsLk5yd=1u|FY3gm$G{t)+pk1om0Q?qai2U>3#o8W)sv3 z8I52a#7Bzb_dKX0KR{++0;8j9_o0kphXZl?l5#yL+iBlY>7&4e6A+`rEV9ut@I?#j zHI6phBCL*>_o`TF`xJyNa^07cF|Y`>b~PZH*Bw)YyMPY{KF`}SS|)uKJXMaXaj$j- zT=IR}fzaz7u9Y^FP-Mc^t$``1cV~d^YQ$q>g{}TEd_n=#hEC1q^lY+lS?Zg8kOTYV z_lvU?pl*1_#HFY|0j-%_x{$1Fu;cz# za{dui0#^g`rRKZBV08b(A#dnC(e77%^HA@_);P-Qn~NkdXs$MV=$>p%Zu3a(r<~R% zE;V1_hZ@m!n+1V7K zES#>RE2_CB3$^=Vre{>suGi)n5D+t6&=qIbH0)}yab#g4%pbaT)$|EK!FUHJLf5mm zFI^BfN71p%rI{VgzXvw{kpY3;Qu0R7>wL~cU3Pqn#6{A5Ymc7Y_%(D zM3fG0YJEwxhYtf)&efd}PUpvOi|{O_Ael2-E!ny?Cby+q1Z!pUc0vIQ3OgQCyZCJTdQx4o{Lh8#D-BmTyPT`bQaqZHNx(Y- zC&;Iz@a(F5WZ{g^`X)aFCzHxdd8Hyw-{oHLHk*Z(q1Y7F8p0o9k;PPCER!z7R4KIl z8d!iW1aRzb)?-iz!PjZ`>+vBS1jA>bYsCyQlj9cvrcn-eo+vBhQq)}_xOWq39iBcruc)2 zlL)*my~(8`G2srp?WF+>X>@g$&gNs|3}**}-e(1Rnzg2(T*+d&?ZzxZ!;a~tuH#9O z1==5`OHynm5CUJ`xgrlX&#seRDub6zTTKZj9fPy!>Nc!*&H=xc?duZvqzku=!9={u zKv=kRQM+^;o)Xo2UxFiRE~^KnzR$+ssU{+6Pgimkq!U3h;?eRVW!efMl%~y8sm}S@ zsB<^^Lj#H!y@rh(_$DLfPnEEi?C&6a2j#eWh?P(Mqc`8vLb8QlNgU zaGZr@d+%5%I>g&l;{0ulY`JtvXMHdI+h5O~v(zCZxud{|W*c2@^DB7i&#;+;fKi@W zian4bFb6x~hXmwd=3!@b62@!3e47ymY&W*+iZ&DgZ{xL=!HMV01!*R3>o2tbYov~zqiLlObVM-p3A;ZiG71J5Vhri&U!et>i&ET>i+Bb-a@gd>PAAG^F0LCs7 zYjV(U6X7o=yeV67lnX!5I-^RU6_X}8A8>diB6#im;&M>}^!u%wK;A%u$e*_DlJLHZ zC}v{okURTC6K7pqce^4J0faAJH&tpJlG)K4r!?}c=PjWGqtX&bOeB1p*`y@23Rzyw za(`T)xQt9Hl*x$l!C}RZ&>~tdw&$lL*t|NOS9@HmyPZb*&DsrVO-Ceju8F3kCXhZZ zC3(LpV{x6nJ)Ychd!B{|2G^Mys5S(tTF1Xamg!GRu|{$Z8{MAS7FF;_SDJpgq{b^Z zjIx~t*Pyfrb!r?oR(8#O&djf3Qt_YV;_s>xxnVJ20DDrjS_Cd^t)C~yIWc@OSmcc4DEZpjqXJip{bDm}Ses~0Wy%bkdeF|)f6XZaSIPCwPzH>d-) zMkPzR*=W-|mv{gWMl?<11@xQPavz?=NzNzV2th=P;eBeTNlA(vZgW zD>kLN*n1ysZY-0(^$qf%*p_yT9-6nuF$=)k?&r~N;N35LD`)OoM7->fdT2(i*;Y&4 zvBVYAm8d%i?4^?GQp?+S8&D4nQ? zRg3hpix`V}*FHL>y3wmF?MqnAV|$eX{$*jKacPIYYO8IQe^ z+Gtya?;~yDn?;&d04=6Qf6d<(up4R{eQVI{J34ZttDo#GT=wpbpl4ckR^>p>&OXY_tt@#$kS+r(2gIYpzW(4?0yPCCcTZD(7|t2ll> zXBwk8L?ju3BnY;Q%PDk>jxFp(K-mFES`xa!=zw&#J?G!Rrlk-Yx>mP*nUpXnf6aXH z4(~VA9evy#X)!4>^2AsAT`_ZMpcP_6-s#KEO-;ISq;G3KtJ&>+i;2_4)r<8#5J&r9 zN`!ZldD>gHVM*Tm?0Ewxn&~jMy7@!>P%m(hCJJ zV&_8RGkBZVc$9e{?`=q(VW|B3S9H2$-zA`3aRgpW-g$pj&c%L3g9lgM|C z#7;UXwi-jBi%I53XH}v1O*Y4?492L|9JGHrmap&e#^0)_v@MDlAmtNKr-BecDdxt! zrEy(+SEr3QdZ6vUL;0x#VRV3^RCQxcTBo&)RTj}KRJv;DGg44Fcw zj6S7o|KvPf*U7=_M#m6phy36=kI)C3huMYbt*THD#>r);01UvRFM%^J zJCYg9R=ElWTD!O){Q|{`@9r*5cxbZ5Y9|t#*MDtqVWs54`q&j1&}sAvq+Qx-&XJZc zfdB^m*tafpM}xW5NOR4n_{&{PR%wBDlZ&|waT2AS-&+DwngxCEFkrUEsg;Sy?)`P2qH{26>r@wfmmG);$ z88oIJOXXY7X6*ck^Cejm_MLS;3eoWDz3|*lq|vQxAm;-LnZ91s14u2i%x_{x^F=PW?^A^;2njU=flJv`M zE&C4dvt!O_-^)gu7?D8*WF?BpY!_|k3-4GW-)3t5nK4C(c*j9+jr=a$4jCW)Iuoz9 zdTdY3Z|9i1=upuHe1RzkgXNY-y}8+|zZkZci2p>Q@$C%!_3(@u2a-LjqVjn6$m3}h}5th5!!N&Mtj17p} zHxy5LuGSE=4QU? zc)ZQgJk7z?V=N!hkmCM!8sAQj8`Zh|u+=lr_kpF|tr<^6-N`JufPE^;aeVNL$h*KN zq)tK%b6B5{PW*B)NTZK00HG*xY@99U0|_s~HNaB5;Qrk$VKTl$)QuS1YZ=|cebQ?Z zbsu2ni2ae>UG*-IwX)~~5QuPgXUf+{`7q>P9Y*}X}vnS)_NQd;PJ)+EGN^fz?f_W-J zvpxV^8u)z7 zew#=QTcAKfEi$Txl3|Zi>pnX>AgTxXQ$J-g2W+|^hKF#W74W-9|N8EMd+8(U;U?G@omv@jV;7d82vP zXQSem6df|$<7f}G2MYgB+8H1CIEoB664Z#fC|JP*1V2gX6w2{y!XYRtg?;|;7&}Jl z;cs>k&64T@Y>I1tB7pUj8gz82RmBo7;d|YflA;^OqQS$%Q(_1=C^ZmEdh5=7#)ezbhA zN5kXzrwqIe-DmjRKQOH4xL$j=m%sV)==~c$l?v5+2)0R3L z#x-Bvj?S$W>ud&$BHm%BtdrD(Nf#YWLXIZ(%Cy`~Y&)Z%WK|<J} z|AM&4nRoU8#whNd!5$xjUS-TL7S~ElpspACX?eb!t(m3ZRN7FGlgd$s<;aTUXp6-9 zK}Eg=H9q$q86JtH?@J10)#A-fKHrJs)EKJW-=^5j?tHzOS_ z+lN9*!Ipe^Z`dWk0d5}w;|n~wM4(U+)YF;l=Xqn+nVFeOi;G>y$TB0z^MjQnQ+ao4 z>dU1oUSA*GY;VM%);2IC1sytWb@Ob?Pn3q}P$3@R1t=88Qch$Tbxy^#x8A&?{ui+R zPZApYPG@KN!?!_N>(<(X`ubbKD*#A*az<5w1p2+4(v(KoxP#TqfOD994Qg<+Hl1ziE74+$oT9dQvp`{7x&z^xsD&ly)2&v`c?wvdkr#w|jM-HtY0Ye!orN)+pU0j=Rhm6AuSm;b+ z<+ff=_!&QGAJSztLZ9)r_sf;{$cJ{D__jJD$dNGE{7kVIeusCx{S=OmuAuP=n_jKs z_GEcm5aG>Qnr~ls^%Jp)t)(S6@Xyc<(EZLw3sB_*`y<*#eFHR$Z+6 z$OI!qNv77W_?~Nw#qu)2#>_Vr(=&8d4D8=cXl`4*R6-Bn7#5+gGd;3*jjRJdEK-IZrOI;yO|mBzI|<(O zSe{V5oj+-m5m8hNVAqvPSqo$8eM3h%Vk6W_;_PG8FVW1m^wyfitz81gQ|b|pF~R-! zVj-KF4$zgYjb@9BHJxJjnOY~$CMXcL66m9mKzzz-TLjUA2QQ-CQm!Pu<)(&&4ZXU| zI42ST!1AzSygKwg4;JxPgxG`W9XydieVS^EWF}=n)}1mQthMCPFC6oR_M(}9h6a&{(VoQhKXl&;X@Sx zb~RzhNp!cQXNQDr3f#&NkPDY*yGE?w=#tP9;`l?&biNn?w z2iO39IQ-_SY~dFnC1F(|Id`>wn>#=0h#^%f;ebJITwdt39SX&(9tr z;^nfJb8LmYJ@+2-=7#T{o_`kS)PHo)XIeN4$hp(q^I+in0=)mJC<5*^L}>|I$j3jo zKAuBND+Ry#$nxX5O=pa1^xH|6;VR-YoY$`PsjmZJHIf`rtWYesqQ>hYe&MqbBU^ldN|n;^n2ashl;27Xraq?YtaMOjWJ{DqeFUIUoi3jOIA1sW zMMYa=RLt}W8nqa$N~)AJneAj3lIb;QKz`j%^8-5PW&U>m*m0=|m=674H)?cvexR>A zI7=~+vG5`1|8Vx!aZ#<^`!I@%f`Edek|I)4Dj>oTA`+4c(xuWZIpm-qAd-TVlyrCJ zNI7(O4c#zw4E5fF=Q$7O9MAK8KR^HC4`J_p@B3csTGw^0wK=gXB-IZA=b-6nci(Bc z(17@Y+lAdqfywBXBSRlO&N=c!r;}GqZoxOkmuRUjuTTey0kPfPxI0A+hVYaqlU7IX zWyGLbc9Wc3zhMt@m6LDGExIitwvC5!rKggdS^ZYpf$6ZuH!J)Dr`GjV#y{n_;WZ4T3lsO}6Rh07-H7?Sm` zS@%SU0#I~jum&-Tt9+5kR{R6ZV>8EcqZ@gg_wo*j6lLe+ejS;ImFX>#rQ>NA&YH6P zOywLkL)}(AO9jb=z7pMpbuz-JQHPb(qVPMaL0sR}1ADF0O!#cx99DHvAMsP-4pe-c zf|-yOC99h(9LNJzIYX6psRV|2{hdnho{%G7`XV7vREy*3yN(8mw|_j(Woe1_P3qP6 z;nhMa_#}ESkzFNH;{9K@?+Pl*4>B+B`8ZA@1LYMJ2hj#d+i|`v_~JJg#D)i5?u-jL^U3eqlA=ul z53kkH3~b2p@rfS=)a>m`d8}`cvPhqS%c0+&ygT8gGqT-yY$~(NV@0BG#u6cTf7ju% z$DyCk0Ba}CNoF+%ZL*|o6$|#VuMQ`;uP&8m}sLxi+P&w$v;rn zV5|q*3^#|Z^gulPFQ1Ar6a!x<5eC$}c=dSz6`!CZ70ENZG(c8a88tpm9Y{||@aOpSI1cQQq@5DI{b+RUSFubn+K>DgOBvF$UrF-a{!o-HYeSC!GE3-j0SzKt;;8 zQVwB=Cy3|)j&9`xX&-5cvj?fwD2yRI`}^ZVX%!WP1n~;?hq(}I!>9Z!qfQU++}T_i zk=mYe>c~7Qw<8<1t&yF!b#RDJqRQSsdF)llKOZ|59-4DhKj2;v%RonDvG#N=Whr8q zNzu&s&_sMk*E%}9@;UNqNT!BiZ#HuY3$0~^hCdBGRiXZ{)$zI=t@@V<*ENV{=k! zsLFkZmEpFZekiS-5{)CH`z$ObE$_Ip(Z-~qg`}Y zK&IH5IVcN#VyL{y2D5~`L{ygX2=gB9QB;gq93Kv6yvz`emdU2a1177)_zBFb|`S=>i0gl~a<#{`1P`Cu;2 zPb0P(Op3`i=t^|hS=mFQ-90>HqIkm&oP=j{OZmZT0D{kn`jtZkb`k^F!Pm&RRuH22 zcF27BuPEV2Q++=h33#ZsqRMWhtkl|mWPX0S8y;6%G@P1`Z-1<+q*;j)BE1=)?xZVw zbKsNbz}({E-kn*r9}<)#k|H7^Nz##_9A!4kxi72Jwjspy(h=->zC}gQ8h8Az&d%0I zh-e`pC9hJNLWZI_Pue0je3BekhfcI%P`aOS1BUS@IbRZ9nef4)9O)}J{qNs%7E^pC z_BxMLe(5j~pz(3H15&soe}7d6QzKBT9{Jq(@IzH$!bNGLnAcdS?%6 zxMX8xq%6bPKRzL0q}(ph_#6&SxM;i5K1FVlN`YZzW#x>wwx4ijf4`E9j7)&L@m)H) z(z3D}b4H=n5bK3}$32V0C%Q7vp4E<&K{Yqg<>yZA;~c6D=k?^b(fYrJiakl8S4|+f zbpBz^r#-RCfX-aY5p5#iwvi4DOAGdNBkGNc;PrX&V^;3kz^b)k`EZDyh-p;#tMsle z!^l>I)QbPioI#ww3}`tFIQtymgx(#=IP$goAiXIB<#>EJ(UZ1gjO5QV>eID}nA0gm z=j2?vdbNVb+A!Cs&lu?JD5ye)3a9x*t+)5(nMpk`rp08C1ISm@4Hxi9tv2Ug(vi1K zH-+qPFFT+{D?WX?xE|7=Q}}3Wp=W=k48@GAO-w;iVZUV<@P&`s>sE*5d}qSjn}7tL zWd8fh5$DTl8)?sa?&Di?SLk{1b>**RxT&H! zeAgA~XO_9yjU)*R(K*EHBV5|(#;iEVs6#{{%CH%95e@`0DaOf*)w)rrTa67*dPnTo z=i551ppS*27x7ke)Jabc#sNXwT|w3Y{Ri6!D&)~z-1x7B=;(;f(5RT5>p0&M-d89P z5()Eo@NjFf4_ROol9h!jx7!%Coy7k+pN)xIDhc92q>qg2CFei;_|PglIyy>6^GjLe z=-^vEz#$;7-=zp(rlF}p)BDTtlL%SgbmM@o46jWh;I$RD)PD&GAX>i|gFcz3;}qpI z-2Mh6)5}{Mn*mK9&+&$m*ffbI86{ z%Au}4^_bp;@?mXlwzj*PR+Xj?EjU<2X+u6WYq>DjH@DCr9m7(~p~s`8@7@aJybvSv ziL&riNuuB1$mh|7at{Q?aQLj`+!xI&q>^cs72(Q~k5jv~Asz4l=#-2huzYRNIB}hQ zu8gqES9^90Vi^Dnao2to`X)pas6FEis{HVYO$fw>obb{b7tZjTYDzfp2S}cC=%|A5 zmK_qXYwLpDqF!dBv%R|;YHVO{Upj!|Fd4L(s{f)bBPn^AJ!By%8iwAT^i#}KyG21^ z!rHq8?YB(F4ZkW9V7<`Ya0lzfgJ)K{ietj4x~8a$uq3g6>?JPH8vvXS9hC}op|#yF zG~2?o?4!!a1kBpz_lVKyp!IIqlWFCl=aV1wnt@()`27~!GqKou*JX2()X;b3O15ia zkQo9>Rmd{)pxnt>Q}4_8pBw^S()P2}_H7JT?5Y*m%*;hiA5<{67QaIdn9Q2l*VFRS z$k1JFqd+=zy0o}{>kF26hB~T9fIXXQ*`;quRk*Uc7qUXX&SUKEIo3kZl%&n z4;YDL+|p$xlR-DzsO{#|koCU7;M^oDJn-xKPVEsiu&Xetg4^j8w3CA{>{&`L?%qN0 z%Qv$tw6*-w4n?9Hnby&)#TPK)OB$YM*b^n_${&6|M$-+qYHE^DfSJy@EmD>A$hbSa zg}CCKqVq&8)l9SU&eJc@BZth{T?$XcA>V0QRVpres zu#-jA$I5Hb0v5gB3aAJMY*u<(L>p)Zd!3D{X6z>S&)G849p}{`g&xGnOVfC9Y({G{lpGLs#a21Xgc4L=Dm zE{~YDs6JC@zoox^$V}|DyQkR&xVxcXn*Getp@W9i2hQmzg{(#+pI95sa&`yAnz|gK zdDQM{5|_8>(EG2vCSpmT2PHnrx<+Vpan!M+YNLo)?udMcRE`QjHK_+}ZRq(VT7iS? z$22tIAyk@nYZF;Z5!_ou>P}SAh2=MtA2ltgfw_J*GEc?3(IH+OOv54PdF*LkdmavFb$>*v)7TQU=8+xGm?yxxCnDzK*Uw_<|3GYLsTqPFEl>c5QZlXUyWeo)t;-ZmxUa z_-J?I=tC6(6-Kq^+8#Y~ytA^^CjD)0qpCyB8FElpn2&JIO|WXg?%*HQ*qf$8?eRQ~w8kEN{>^eu5<^dea8(WJ z%+>Vu8Ps3bX|m8J)9o@8hshogsE^kX_QWv4Ue!p{L|8bFLBF10sjm0PFMxl)0^V-% z18i*9He*I9ANZ`thAxW#KzJaD-g&1>G#b6!>MR3WjbnrJAnoT1k}k6>D+V~{Z4l1; zIonP&C;3RJI+soiZ8HhjEskhrlsdZ@R)_2K92c<8?`5F)citZzL1vfIAHq6Ud&1lJ zEIt^?unnt+Zrg^^eL_ymj2sIcbvW#;=}4JS9Ir>yLSRt`Df87hqS6(62Q)I3YqN@q z2&CP{rW8sZZSQbA^qGoM0J;_wc|D$BPcb*CsI94q37AiyRE}2bPbFlBs~yZL2Ko^D zk`hCB=a1Eo8cEl)pfk?PWsKU5UQJq5Z6ilX1!>}hPP&zyTZ{3>^2gST@IwPs8QgKO z51}Vjl^L|dET-n@v}1)?=bs0SHq`sCke!9?k_}U@5xx|Hf`Y0=rYtNh3G=qh8%#zV z-zFv|ii;oH^rE*GwcfsERV#kdn9`NA0NvelVwiH7N-Hfb-Q3(9$kPi9Kpr3M`1trJ zM0V^BQ_0xPiG$%bSYrl1P-5VRPXIv}-JD4zVOdy>SV*UDBKj!|r7 zc~K9Xx`MY_akymgS5_8{4icA9Y{wtVI!m?*9drrUE^yl%QIUhYC20aQT0{b;cap$#scy!x0JP}`ncx61R-zo z`h6etSV?nvWlzBQz(Eg++VuZ{9nED5n~4eU4g|;WSp0Z7WfO^laiD!ggPVt(cD>+z zQM;3a^25`udZWyHBY1uz4_Q7&5x^%$f1o}US=W@|jLeC@(B~k{%gyb#PdhkpG*l9^ zd-dw;YiFxq4iCpN-Ii~qr9Ae%iBjQniHf8??k0H&Xx0Q@om7gR>jx zZ$Z^^RomJ6M|_j|NCrjZY>7ljrp=f2s&QwHkmfMEKzrPjE}H}+`**K3kDu5kqj~$J z7i2cKOZ#Cu+ui;js;aV7nz_!)Tw2{YES$Wzcf;bF{pV#(^J$f~lo_)EIrsRsS6kW^ z${RhI$}v@Xs?A+OXX&Ljso}BCuV2r3Y;Fb4+%ZMNx-h%<0!*}=1v1L_*mqK|=8IC5 zK1XhOQ(PHyNp*7=yGh?Ik9}8<%*z$m)K%W!Kw!+WvhH$LYKSgjH|vyK z3n`Q1!UC@a3@t0BwcWXbO{mWJw?)Kyr!@v01~%#kt1HwOC`)BA8N?Q3kWNq=#Xjyw z4Z~3J!WT-MiLu6bD)hY7qswvNlbNrbf^Ai8cd`lo@f)3}%^j(7yAR0Wo44+0({TEbs6?BG73mb`-TjY9LZ<{(jo!$mSU}Pj6vf; zCVXa|jmeA5;PM`Uhphp$ZM{8_?TfQtv@EU=5xr8}k$zTt`4avBUk?}QVhD=_dpJbQ zQYE7_aR$#KBI@WF?8x50#OG+lb+R}%Ha0Xgw79tVrDm(F=4)5~eivVJ(2;4&Av=}8qxAml94egrSPo^7H*H7-VZ3k$J{2200BXsD@UVq>e<-9Anh2hz&~ zxMpW(fBLj|+?2NoiPv5kE>Xz{juaC=rxzwkSD+%b;O{y!GxOue4>!CUDmH%G&(6cy zSxriM@aQqL!=I@V`x4d|Ibe8w)nFb!`)*#|2@ZkNCRA3)9hR3bD=VuqPhVBf*V*Uk zJd}R^{JDsT`+#vGRbE~mm+5eF?kXgHaG@tHEj3l2hcd3C*kZcz=TDJHY0;z3yg9}T zu5Z>R>bSYNUGUZ!nVB6AxB419S0R(_!KxMSs!t-8q(lN{6IMMri}e_E5qr>hA!EY) zYET=Yh1Oq@=hwE~A0I{&%mu_4W1K++3TL;joB^X4a+yU+E->6FOhX z>k-eSB!Da@D z-h2r+xPMb=Mi=mvkq?qI34k%0lD6hbRZ zQ#q~!#1%wcT~1zamL?T3vE_L6d0FX5UwL><8%isVz4h5uLZ|%y9-x0bO4JxsWT+P( zMQIhsBpB17-`v>vb}?gpW5eJT0i)mN&qPapnSF~!Mn>~RM!o4NX=!nKCWHBmoSgEU z5m8|v&y(Wv4~B$>*0)Yz!`V&$$CXwg6D9*|D^=#?O}We!E*<*Xb|IorW#u12ihl&% zzb^y^eYimh7sq0lPMFlB4+beG+3_Zrb3Y3b;XB~e7$z!r`1HXe^V+ZcnV4{r=w9DG zqg;HjB2rR@l}=o2So2u@7Fb#E2R}dwAFQ?q{e2tX23$NV-hr>n4>a>y;p)m83SO&! zoVNP22fj^b;c>UeptVEQWggbQ{X@+t?`SEkM{|5e*S@Ny9YXj*^QRD#-TKExv-WMU z3-HM&5h!RtV4ihp7#Wvte^1*3=`B1w zrnJ-^@Q5WPC1GJ@k23FP z!t?;Y45B_yq-kpe_x0=7iBBu) zJ`oYoy?ghVHN_tLLhRP>a$m&5^ZHbL1m~523yIIPMyB-`4tT~tdl%fGEt}R8Pr$8j zYD#eVvRZL_BhRx=VheE?jGy}vsSj9IkhKaYJ;}k2{M{x%Tz`*z$+NNm_5O)YPELCI z)PA8uk|(u}IDt&NY5_YZJ)zfk&^iH(hoiOFcTHL{_hVJWv6HD^`Yl_cX; z!0>*uxC~Vp?raVs$!{Na%}-sJu#}(z@>VAcs{k2mi)vrJRC(ljSb-g)QYx^o=e)T0 zU>h9Cd2m~|B+7^m30MqFOms}${?GJwNkL^`8$yowf@^JUEiIjy?t**?W~}}t8@y+> z1;a&fU&u(SNf2Df%vV4B{<^mD%#Hx0O5l^`<9_6jvc}`~0sdf9%vSVUhdh72MoY^b zFcdHa)t~eEPM}(Por)?tFi^|Dq+U==?Bbpc$Zi~revOpynkfP=Yzw+SVc+S)L%0K} z{33A>JJBGh35M#gK?147cfLK$x6sZ`i$fmxKQ$;8AvjJ*nY*kDS|Q5405bK8}bJ*u!~X8M=NlyS&E^ zlydk(fgdDTh|QGmR|)9rKfl&zm0`mspfQ8~mVz>FdaWpGCg-%%feUk<%c5zqrV^Or z;_8*)%gCCSh`v>XS31(hX-&yci1rG-M7lJ;a3Feh9&0h?Yd;kjY|j9~$y%YvDuhGs zA$Na{mgtKY?ZEjXAt5nu$k=qyNa~Xa=mu6+-^c>kAsQ4BcsE%6Ef%cO(G*qdOn~e* zmKGN;UCb=gA5T}#i(s_posD4ve-rEdsXX?Cg9Crol0oto0sanr_9dy4<#YwFU;!*MB`L-o zeH(WB@&;q-n;Tt73!Ca|_?-M7({xk~E)uG;3-??uS&Da{jH8#ejGsuk(*u=AyWrqp zFC|(ol$hrq5k)L7gwNUIrPCROH2(oQl7{UTD;K#Am6;uSS~=~gO$i!!&JFM6@8l2M z^LoZT8Z&;TDN6pi((5`utkT>a7QvXM#XNqPsbDQTywqSIP;1^bmU}&b`bE+Ghb%Ig z1Po26$HV~DocUc`dRL4br>6$<9pc=;X_+Tws!V@F&6_%|o?Oa5+n3~F3+May!ck@{I8>1$i~d?F zmOq7#!$q*w(+?Iy%rXhfGv)#>=lM4X%lxRQi9wNE4m?kAVM;oQ6WgqgWsq&Mp2@B@d# z))J`m&}LQi%HgD*%et$dblin!oaqwvNQlO$bW;tHtbQ_H7uOA@V?+B{PT8lE@0GX( zSVbC18PaYj;v@Q+m5z8TEcC6;W<0?yE6H-b6DNk)g}ZzWgy&NWwqeuU=r~$!${=N> zk7I_rz97KhkbmNb`_eGIT3QL0@_HAv@t2S%k!oC=$mz8=FoZ}nTvJ%*st!%IaO#_^ zXF>RJ@M&WiA(ll#BuY|w4UhC|EzzO(wC{lJox+%o2TL?w8kn8K!SjJK@4T9$s8GR0 z>nF3Lhs$9CX1Tl}-KYo7LY%nAE^qmA?9F@LEy*m>JCqHvq76)A72_(_yN=%XCb`Uy z3c%<-5xwRAq;Y?9<(xeaTdAj$N@H@xLC`;a%!E4Ny0nj8MPo~9D$>5=#Ml>|6 z1)LyuwSySH43;8$(5dvdWczp+b$_;Vzh^CK@|d0acD-2@_F3hPLygHAytrhd2`{2x zHRZY?UZV*XycFLzizfbRSFOtXm|QlMM+l`Iqs&oh^eL? zX^_ll*(p2RibGzN$aGv8L2B?jZTf~3XyStQjcF%|gXTPT)i@ZglwFY@4+oqO&VANF^+dg;B&`v)l6nA{K@#?$P z4ucpdy^=NSS>tC!)*R$j0FitMftM!IWnI1j1G0 zFj3BQTe^{rP{q~!MF%Zw>&=3QI0Gx@KBW{}@lO{<9HZ4rrj7mPlaa$phHvVv?b{qh zh^aRKvYVzLm6jX7ylsWAR%#0{gBbNbXt`;?5t`-YV++5&H{} z{JdF0Q@rt3ruG7dS?f*++O+8&u9R3aTMYjvr+A5E<5~JdW&Z^2B5=?ORY?|!)Gg{X zJ?gT)mzqc3P>t59dD8LXwtX{+BloS-<&_Jjys|324!|2Xx4=-rVs!kojHvZmVnY1_ z!v+)l3Wu-%gF93gkh^1NX9uhYV~?eix-!(udK*YC{Wj!K!t}he#cp{#LlpCCsYP>+ zuI5cW_P?6P$%IG09~m3V*XoI6WE_CoTR$>Lj-Ajd<@X|=vtlQtI?B3w-A^7aCnu-% zC>;L+c7~jWhUp~k3GB~OwrP$h%Tx?a%I((?7|v1&Iwma?KZ zHSQ2pIXgGE-LBq@j*bqffVkN|{LM$Hh=qj(;BO|NY&76WOig@VM$6|Fd;1$GXK8{uhRG??T26FI7nN@Ydjj19R-C>)mtPE=kK&E zk5+_R8Gy=&BNzdQ-``a4YAq;RFxb)beHohKnNLC3Uk8TzUw z<_Ykb0Ya42{8ot8|Ga^p;07Yx^fqSR=9RdTFIuq|8QXbA`|8CP-rb{st&n{{b}VkK zt${Pl9o*|q4sO8TY_BXF3?NO9nm+d%-t;zJ_NC09b5Zqmjd0lB%%Ex%Y{0~ZT#*Fq zGzL1kKovfnT9XczLDH)KK_oU`J@jqRPEAR9GutUL{%;%V{Vv6Tx!3^)_&Ch8g-6Uo z3HW(@j?Ft}a=#CSylAP#b>i$o0XURz@p>w@OoIxy_#M0@@0G=FWXJ!D<{J4+bJdUI ziS&I{c6_wofPf;MiZU~$U+5j6{E`5(2I%v@FuE9KZHd9$DL6EaduPQ=)u=HA@#g-# z(4uNa9ANxvhk%mcYHCHxz(MMxtgLKYArGh$pLFVJK$W;0c5K-@nV7H~+O&M*1oiZ0 zI&fhoBqpZ1%hBbT6kY@_Bne4LZM_F#d%(*=N=9}*34-7Dr|5gS_5~{iQgzH1$kdOA zwzHG7B%Aq8LVjk)d%fYodj>YW8QdY*@tQ zgzxiAglT>;3h?8bF*|4dtK70FcX&sR$YqwuI;k~u9M~i(OW%#VEO#N6j{vuc{g0Y( z463c?4V(3nY8zc!w~@$sD_`(QkoKy&mavxPT#k=OhQB%^%kKwA7AlkI@jIjc1r6M7HgmMU_B#)&q8#*@X%Z!W+d$&hbR|wJ_qx^&M6%uAb)}b1&@dc)qCu0ePRWFK>g{iSAF zPHInuY-F~i5)%Q&`lY58rD>UR_dr2N|3mIOwfEi3*CC)%h2c@I;&w}}ZqZZW4653< zkuMq~Cm`+5Cu0-I(%Z^RiMC{4m#NUCfGgx7=pCbu`TSDcTbh0(kla%M5E7?2(Wt-y zhi6Dxb;=OtHBljyD7W#jzo%p972Btj7`QJ-pVPs20LIw1@+%YBoOFj-=WGP}y6wIz zEs09O7$V)5{pmudZ|yHzd_taCzZ3E&`K&$yxq6lTE{FEcD*@{9y3e2eEKl@KBSqv- z{+`*#naRlvbaZsAtft*J%z^j%A41h_yvcE>WA&_*%KGX#l&8yj6midpCV4!fd8U2$ z1{|`o5y{2jrsZB|_N5Biy80?6$xHms1%c@n|9hp!%7T)BP?HmORj0V8={%TIIb%O~ zhW{0d=OL?wS28bYd8bGmEIw-?_YBBZw?mN1xl;qBrNlZqW|RV>40Z5+NKaNKf1%Yp zsl|N}xZGc8?m~?UtNG61;o;_HNe7g-`aANS%`O?yXV0c)XZ?(XxJH4|TIZMg{a*>J z-lgV{z^mof5^8y`JU1z(7s4@o93^V7*%Gv~ey#^yk{IPqy5PE7dUx#Tn%VM}?Z?U) z^igmgaiV_m8nqhBkhqYok|e%7`&4e=ik){@Dfa%Z9?Mw6 zrR!PyiZ@p`l^3ZeG>v@79e?O(tqp1~9#b09p{L0VHb*VWaAEH!c4?@9VKoF~h&> z>7kr2y_0ij97Mttf0YNkg-D!2_KJto<2Qgzb$m_NQlA>Hc|0pES2Bh89kOe*OE-<# z${VV0`y)4oF|Fdw<7U1!ng9`e43=;N>!0hqwP5TzKDxeMG)I)WwhBo_J+&n-;Ai@+ zTKx(2jpV;*b5WyUj5)9&ULox{g+=@DF!{}!#qY+Sk{u3%Zc}M#X&amENa!f=v$kZi z0|U)(U={faboaU8h5dc~@pGEar9VHj87W-^UpzV*%_M|w10tjE%`MB_88^@C_7s)W zcaGJ~nua_+Q|nqPg(!6sy0PGDaSqt1T(%fsa)8n7`p799xEw{aHZC9V`w{F?aQ$e2 zA@tKV3z>a>39vol|3s^CjnyddIYY&)%+%D_dE5U5SoJSe(4|C2M`!gdBAt(&lBUtW zMM|K4?k^;AgxAP)(e&#?)Qw#vqGXp7s>9vazpiVdAHpn!-zCI++2a<~&nFH|J%`R? zA@^~#X$0Lp8RC@@J!M4=j9B~Lz)je`EW_Ek~(O}l6RP#AC0!`UZ3(+g^73py3n zOZczumZt2n`VnqG9-*6i-zk41{lFrN??77`;MIY&2)@LKYbg_agS?3%Z|&Lp#Ystc zo*6Qxwl-Qdn5i3xRNB|BU0bz8>T?Li_q}n85DoAo$c*UZs-A_3LKZ^;Sx~J zgsBJbQAA2Z(VGT2B#b~u$XZxfNbZWc@ez5;K^5~zM2G&^S+^)jdDCGB$!jqW^}%D< zCfo0tt=b-AzOMdP5*qBtxxs|0OkXAk^e=* zsc*=1+i+kiSw~vmD~2GH!?{44)l_C>3k}>wFQ}QpMsNS?D?4(WdA&&K)KJf<% z!?|1$|LD+r)Ub<$AU#i*L~*VKyS+kWzay($n?G8Lc8#Q{+;Dz|?cCp^WTA7K&v|>% zg4gfxDkk1p0MY;DCYl-cuNq5plh_&%cYqY4vJGTsYVe6wp0D^BaiAT{n2KNjcts0L zu7F3p%iJQ`bNE1Ju{rDE*H%y%*rn|DvZlL@Vw~$Gir8V_*n#cu&WHNy-DN*k2E8Gz z2jkTiD@rSwbq{Bf%bm`J5NR*ud~?yuhVQ5>Ym{=Ag8T1z_v%ZCJ?SE)z*zl-I4?A4 zhXcu}3<+LDY{v5YkM3%%sP^TwesLy4jk&W-upPU)mT1c(DU z1!S0JCfmWSikwm&v`T9S2T!n3`YT@AXRfA|GMU2`g$S!|GgIb*`%}P?#-BIrH!NTi z1l4~$(r%cYWHi=U2QWogBOuTuhqnA2EULd=>CC_VCGV!Og!MeXJ~)=cSayfLR=;$p zZ;WB&ptGHZfkA(G+7qsi~0zuSBZ~9Ofh`*$fOuV%*j|3Wo#&a6kEM z{-zic?kzVV|*(3?~zx{Fmz*;5R{R?$u8J!j~rNncUNWE$X<+C?yL{JHG|t7rOz+Ce{OMn z77Od$%~N>3s^`F0@x8spgvxa@Y~~`+pNJ1W3u-R$3;%m}ouWyV=((n>K}eQokMmsSkgnSW zdgjvWnZT*gc>3D?=M=GJo_YR1*djq`hYj?hC5jPn8kZp!;-42wXbuvmp(E~&e38ujd1vwu6_Vx8i0U<0`vfZ*Ur>Pa;+ zth|i?WXsfSR!A?Viosm%=flz=NS4#PsqP!IY?SDArX0SK?lVBwT`UY8GL`mKW!N!0 zkM*AV6i~%f_L#5il2mj{Ynb#X|>n&>r^2@ z6(x>#%qE9=0!=I|BXCPS-(`lV9GNX{U#5HaK6`q4dU=s;39FF+_4)qKj+R+hnk0UK z9jUzz`v-ll+L$NDI`!nh=z<}TZfhMyE-YvP3vfe=dPvw_`bYPra126y`BLUmthHn! zU8Je7tgPen=e{2c-ADO=hvKuB*ZJ!%uCC8ZTA~&HMCi>ghi?rjEfX0SjX$NYdx3#> zIB1FsKA`)dTamU$B!`D^#Ck}`%aq2s=S<*KB@Px=+u7gTPaK#N)UWm;J1uo#xwx1C zcu2W=lBxuQV|XlQcn5oXJ9=%NJXPh|jr{hFG;hH8{!RVJ^>yGJy{QNf1YV1vAX)}q zo28ovzBGdtX*XH{f_2n?;pgL249?u%=#Z&MA|=-QQ#L(}zbO}V<6WqqhUz=|RJ}8^ z+f$WY7=rY){bJ>1DLaH!3$wJWQ=|R;U%Y(zX76dW`EWxZeS@HqlG3xbCs|6<_Z zFs|R_bwECsmk(0&f`js_+IF9UpE(&}_;zM5FA*^cvy@#r`-u|kNAc-s1xNzS$@i~A zwKpSDW3Dv?0QEs+5en#r3sS2~446_W;LDdckHY5>OG`^WWF8{mvwA1MRafMvK1Zfn zNp4Qg%cL$)m@wOau6nIhZfg!~!(%R7NBPunIRyohYu7eAM5(w1nh1e}0NhM($kFRd zYnC^f{GxJ6yWX~wnLwQieIu-Tw0nH^b=zk^-tfq*cll;!P2)F zKT@qctDM>P*2(|q+D~wteoMiL#lgByc{;iPD~HCJg$7R}`I^L; z1)TGRO#7#;21$#hehS{D7#H{Zz(Z;4tsGIxJ-0^bk~oZsbZ6lX4nBTafeI%}{D1tL zzcA|8pdjGCiUEoLu8ob21Y}zd#)PB21G!8V>)}u83~e^y60qtsk**ZL!rD0<%P^)G zAGL`aoL?UoW|fTEVA2TBZR()LNLZ#p_I|43i{!qqF`jd;ol&3^78wDySuTJ=^^lc9 zo}L1V_$&~6>5QnScv^2syhUL;J?*KfYQ~peN#UZ!8Zhsr9@V=qQ(&kJO860 zK){~3R!JV$EWCgd6A`_A6yAVrY!nZE?$Mi)nu?2u*VNhyUYCwnc6|%S&CAS|d7;Q0 zZBXEX4ijKV>NIOr0(LFj%t`3Y!miEupGY6fIZ!&Cz$U?Wr4~c~%eOpKtoHkcQK^U@ zUGr=QqNEY01K=+9knGnUxA&-7i#=0_9l1)ty_TE~~HOyV5Q{IpL=+TwJYXP>kYEZQ$gY{jlqaQ=e4nerH+*+EC4fq7gsKSl0==rr1R>sh#cHzD} zM}JSu`c@LfqETiN#h6e}Ko1osqrF?vBX4rGw2>jC{pbfSd`y%vceX0K+4L%Ua*$1} zSl>A_`szU|7%&6$7=VF&C!Y!l&7X$B6HxGOfjv?1C{s1Q16axe0s_`xxarYxFcOif^BfQR`69M(u`7f%}=CPMP>eBkJ+GcHLhWWw2xX?A5)1qAuv!`W)%`=Vss8J z#7-ONN&&GscmBM|NNEA6xDqFGnzPb7^g-$aqX7Q!6xJ zU_jNMj`7=gr#&_4HOwkW`~9#&aj)y07bC&0ZN(l6(+FE%_2Rha;f$Ndo$r5sX0=aV zIkEdvG0FfmKFE_iw<#B@!4Ae}G58Xori|tFl5`1Gp->j`IlcpYf#JfxbuEzw2{p#W z#-7`cTiikq4yvG+zi}FY!g)>^Ol=UCl9IPI2F0pr@nD&AUmCs90}gn)+s? zY4b<#5;p1Vuiho#(YbGYS4+eNzFi=d5?#KOI0TSj?#IxresAQGfeV;94PrcaQdv!Gn68tff7lF|LOm z!lfe)3-`B0!AJ1^Fi=Mzg8<^RIhy!{Vf)1_Qe0pH^N3up7-3OoPhszD}Tf^;4Z z)L1asbm{Z`>PKsVSfFr4zP!`84r#^602nZouusvr*DJB_8-YF)Fk~j)d_6*f^#tn_ zY(CAIkldybvvnNlb__vy>`{<8=^7O?=S^KS{$kkA)DfiS4G2rFiz#J9{l2f&dz^!! z()DT~x|5z*eLz`2uvC?K+_aq&ui^t=8DJJ9+#WiMb^ot#qV?-_`1b~@%q1tU>(ai* zDUCyV*=1fAChwOB?Ld;m-KYj@df%+E&tyCPWKQ;wpJXdv2Ql~hHF1Jcf+M$UDkFUd z_KH*(QbQKDj&GFPdh`5#mLDh>xSlA(aAL?au`rl|9bR)qqTTf!CD!TZ{4M?9-V<}; zCh=I!2K>nAj4~kn#jG#m6^)Zcue2R;_YaIijzznxv(s{>8XcYQ?l#1ex$;H2nB6+y zRL;++T1i_Psv4!tVdP2m@iJQj-`Z`uIA2?ZAD)!iiQ((YtZHP+! zMQb7AZMEC!vwCGI-Yz-*W}z8QAZ_@jE6q~q5FdZuq>uq8)*orO=(=cJGN?JYYEL8`6Wp}y!A zpNgoq;0KRkYjEab=Hzj@)ETa9ufD(m(V)9 zZcM2GeCFhRM~zJ9n@bMXL)!e|veNm`RI`Vi^IL79Efg1ODw1TP&5X;x_qC;EjF`&& z$8NuR1un!5TM2gpU9UN_LV|k-%3TJJ{e~&dVLiq^1rG0@cqt^gTwYK>L6bH03Wu(ZIvGH@-G4DGmn{$zS18o+XFm5;T;Ex|~fEx(5 zUAdWr{?#7vO1^^P#tqt%j2NgNNMI8aZw1}P#$L+;3r9~+k3;E z$`JT79TOl5kWM=?^;M7SP!~gg^@h2-kI2fdD*e-i4>r)IAuy^_Q&V?zbmW`yL}nHi z+BDw?J_pX(s`g1>>wk#DZZmlHhx+MN0R%w#+3rB%zEoSX*Cl|sFntetq7Vf)t}pK$ zqL4;Li?0`x?s9OTQ($k#9SjV(OG*|WvPTyrVJMs5*!H!qas$<-zc?I@jCZz-k|h$pFZ;=@-*7} z4_)zN4@1=ZeYXFMaoT7)u+VUGYcwjyMS*_7CCv&RReKAaAn^4SV4nsE=@kjr+oqx}t1P>P%%p0&-pK7q2ZT&(-PyYvdf3}vyB;hCWaKZ5T@s%@J zYUHQ(Hx`@*Op`KwueO*68VsbuKTAmL{Hn`A-7Lc1qi2zXad3slrQO{J^t=+e6fjt* zsE}Io0Q%ot4L}4dD-QKeorhudjHEs@tK&6O=~Giv0PX+$`4fzev1bnq2uMa!_W)67 znyEd4+j4MlFq!O52%i(0gS)@bL=}vxsE~PM^-DcTi!iI$L`M;3s8HU0ec{Rcsi1cL z5B>yHwXbSa)mYR^MJ94nUP^+`jx#$<`-*#l1=N7wlq^SlV5pDtR*&L5=U#yRHNt9*6Is#Mq*ivLypV?y?0vV>#q4Wh zHGJNyv=hH*{afsbxgQ}@TA&>V(EvOg{?UOO2!H?YTZ?i^5+&T^cUA+sKE-28@G*E?R#?(L_nZe>R&TqGu?ys8zC(C+K<_d*F zgjaV8>?A6|1jiY+MfE-bLrb^!4{#Cm?(QyV3z0AqB6#}rX?>AVU#3oqHTIfW&3S`Y0Uos%9_9%j;PE^ShTN zKQK-G|Lm2+yYUAZJ%Q<41bj@t3tJD=s1j^uU2AxTdGRm35TTJy_Zl3qCxWMU#UG5G zb^~1NA?FvkY_*Loen|vq(hO**0)GJ5l?qKqR8&;NyVjIu6nphK>gwvu&CQq6!Q2x> zZ7rCM`r)fwl2#9Dbz^^Zi1vo#azUSNYs4NP62N9!lOaFe2W%MV*cQP7kJ|_WpfDf| zPL}%*R`AIm0}G;O?vq?2_4%2e&RjTpPe34AWZHinNYzzGbIb+=d#4>M9T{2Y@Upg- zhk(DHL99S)AamiTT9b5%K-79W==TIgBzR2E*D9$O$DnEym+$^b>0scl4x$VPVP3fP zxd#_C!fby|y~+lexc+nYm%k)32fGRTeJ+SW7zKmprr3!+1Cu=d%i&_0@O;mLJ)Q90 zAmt8d^|8M1k7*R!MP$a}sC$f3u z|MXJ$-M#kOYpd$fiHXHd2-@V-x0~^Md@#2*DxoSh*XYtKRYm zq1>F)yRwcMKkw0@pRT+4;P3ccuEaLW5%JK7-|VeMo0VK8D4vn`bi1*#v|y2r5K^l{ z;Q#ONoB(@lkyo>sRS22tPSP(k0rxThofj(Z$*n*H>Ztj1yrV=*k73hD%44$5ScM8@ z1^}UI1{FO0Tz767x%7 z&&WmKOTel}gam0G-*XaNM~76mW0PhdRBIppTHkbr4B=LwIN&!p==7J`?qSU<(wxax z-l^tX8oq+$5XM23|6#V~=}1*Y=ndndr9*`F#?dCz_c}3TQ9dY^my^wRS|0&=X8NPs z+%Zf1fV~IhmeV&R(gSl4`?5ODTAIf?E58b6S1C}a0P7wAn6_R}aD7%+qTSrH>1-$? zk_GC9fB>ayA({y>i}{ZrIx!<&p`P{Yc?cbGGG{udbXWD3@P}2-r%4CW7h>pi%FQ_( zM~AKG8uFE2i=2>kyeMfhBitSP?U+<~@QuHy7ucu!FsmssPf5wPu)Wy|AJ5NndUn4zC1p)qB7p5qONi5y3eo`L-#BbA z=;v6w)G)(&7z~qZ0JV1i;z79>DLXu8R`bnytU?pmNe=dpKyK_JN%5Bd{|=D_GeRRY zqUy?}Ks9J;BB{w=`1l+ff-3RmDNr`$azM1dzih$9K~bZBZ$w?o68eVXe^+Df6bsQ} zVq!*kgB+WBwOd*JFyZgYz2^Oy<6je$u0k`BI_>W+$GMv5UYTzP%`4H&mv94b-%y$51>wFF%c09UwFV#yzqkt ziUSk%2jtuJv!sNo&(e9(s>8~FHqsztMFS|ohRWC(d7Wnj-(^W}119qSTd}G28|JJp zU$t`j&ujtA<;6-0AY z%-=zOG!b3O4rV_ioRcN{XXdiuKind;r6Y|M&tsLbPU8(G3=E9wBimL`5I_;QZPr(* zqzwI>XVc*)C-9f41Nz&qEoai}{|7P4?WRx$NkI7k6u|Wv)An2mC+f(|Js=UJzSk8% ztye~hGJW~)Jsa9SSV=T0rCB_l!K&!)YSYhad+kgA(L|K`%xeDflywceVI!OI1IS5) z{C`qSs1p|By^X)`n=JnFkzgi;$7DEoGIA1RZ95&-*uK1vWe}@o@9Ax3H`(e0#W3Z{ zpM=>$Nd?Sdf1LyVKB^GdLDR;Lo)WF9;mM*D(84iaR zsagC-6wcsRD*j*63ByiTfF27D^v@}ZI$Q{3ln35%0_GQNDagoRDMTOh63PyMxMy&{ zfLd(4+9MHeUsRE1mtgeDZL7vmqEk<5tt-t@b7sUNj$&akK9<)?qe4Qf_?FMX)5-_q z&tlts;M!Kg?ABS->-@zw|4EW@&WYTf{NJ1i$C|HZU@|=KUB0exl)Zj#O7}AyyycFE zu9Dg>s#>9;9a~o^D)p5uUuK*!0PQs*FMqlmR}KyRI{L;S8JCs8uX{+VAHVeWDZYOR zrx!@SeZLLW9HHJ0n}$2Nlel{Cu&iTZ`R#EE~FUexP#g zOhMV~V2r5TX>C7N$-YwSM0Z$7EC%x8y&*O?HHA!19IakhnUy4rSozdwe*Ic7xv`(M zeNyr~YCESSP}|BxN4nPH(Ik1uw2QjHoBz$tbbwq&(NL8=OZXqE4ldDGhFVl`pL1E8 zK}{Ga33}Std?X7=*L*foHZqo0{ZqwbV@4HL47sBUvouGNzLIL{>Z((eO1=lN!0)19 z@cD_7l2ZQmi$ajPQv{hlM@L6coS^>ZspZ!?9i6CX3lQN&4t!!WIz7l>l$YCaJ2~}sLt5|;`xQgTDhTNUjJgDa zov+ALRC6mSi)zDpKRs*y%@r~KxWE-9llo|+0IGA1Iay$UwR0IaRbw?MJ4dSD`g#M&rFBVX-7n?MwV)pYyPi{$mKx$8hs85 zz&!7T$6^8|0~|7`Tn+43>%eB{3}o8k(sySxyQ)`+;`mP9t*&;-Pg?##V!JZC%xRN8 z{&4rzoGAAdwpa=8x4i{!UbmLCG;4*lgiNcqY?h1ImC;?$zbl7X#tAqyJ|9nmIXdjQ z!9?cMf^t7x&B(~Ec0c`QIL$!umSeVEM^gA{y@!{63i=cDPx!^1+B4ks9z*(*U*2ek z9c5>v@GT^Zc~7{eU@2t`mNFJHTAWrobrTi1rN~U3tm7W{&hQOkc-#xv)+-#I4v{n# zIBLJH+Y){yk#7|XsjM_OQ~Ri@S}7{HJs0j``N7cKGw>7GIQqhlm%o*lFF!H`f&xGt zmMdjMtR>O!0~-M@Kn&!DEB$-nI53OJ6wJSjNCiPR8Yn|7$X;+ivCRUC93<}qFG9Ia zx5j@43quTV%3<8sW2H=2V1`@m_qu^e(5|ujnv~tPRmeg~Lf^I5YUB>7QqoRW zkv4p`H>A#zg^PQ-m^y3xl?Ka0m3-6{Zd*4bCG$fspm|474%c`1{jlC(=WW=7Wygh$ zds&khJ8@=|WLx~q{4yXr)m*wIex?R%-{pS?lm%Rbch!;f5;$?fpM~@L`jD0_wKV_F zmM!%|y|~jyKlsN-&sD$EQ##;5!UGjcXB#CShaTK=!xXXw6SK=CmG*ZvtDZ*MJ7Y)@ ztI$xI1gAMqYQLEiCS~W;+j1LrpDUWy3wTp#yT2(pgpq>r1ieeJn7Q8VW8ml5sDtm% z;bvWS$5v5|6a^8Ip7riHG$&e1ZN)r6ppX`q52%l%&6o3YdX|zHeF}$(zvp1Ikfp|KVsB@%X8r zw~l!KY|`AEUf2L18#qjd{4Rq}zmsIgnb#{yR{u#fJ#WJ6rH^Xs=t&B0o9&Xx~Uq_952eIrcr3ELp+M-tP zPh4nqF>X!SaGnnfq&+D3qpWmmPpU=H!;rkpdg)ZsOD|329}YUzV@DDt!()Hpe&P z2JX9`sGA@B@I5i`+)WfTat^s-*4F?F`sGUN4v8Ypq+k%e6g^=&;eVUpqD7wE?Fi?{ zJWoCpLV$&l*f)q7&Hq{KbDZtygEbu30}4{7$8D&ZZ=Y0~vY4&%*)_uZ;qv*ilC)Qs z5<9O*Dec}YA=~)W_@-wz>19`PT%RTUov<6tW3?K`>M$45jIor-;gbcL&~x0XomAGImOQj;h3IV&R_LbUHOP(TIwZP0XFSv11)N!wdP zUP**r6d&V6aY9^9MJ1a1<&1LXUF`XqU0D-aV?$yBK_xO7)zxH%f&AFsJdG!}#Int& z<8tHaX5)AqE0{JiTC{zcxAQMA7Z_kH7mR?#xZrq{WPj~D+N-L+eqpw!{eJ48TW)ei zTG&gaxa&*Ya?50HR=3%20xUQw-&-=HzM(krZ34X6^a#EPrR4eH)whVa3dcFD3OwPG zyb1Nz+edhdm?fuX=}uxqFyS@)GOJI?B8%pwM^Y`yRav82)mkuiR4-C9V+DfV=SO4@ zOzke0eHPbs#bvxEe^l_Sk})vHjmtVK-U4^7^)ETsTV$!Kw#E@O$gJPAcXzyIW`AD! z{p%2ld{<BlsvAiU!yde^VyE zi9Ep3Gt%kX0T>W#ITyn_Iin9C*5F`=#mL9|^2qucP!mc|)^V;X)7*U{+{y>uOt*=Bn-L1i)avE- zxV+XMfW|lNesmkCbnLss##S+aE+dH!b?q zLxTbGQJ$WjN=dd{fHe;b>mC>olc9%P#l_{kj#C^ShISsVu8O+O7dbpdcv#O6GdVjr zxS1Y}B5w@kOBZ;rzZUUKg?VeqcUUb`*KB(kq3Zzsd1w9$nt6Us_Lme<(aWoZ%l!JgT+rP*uPC$Vt!q3;)B8VY}&qxhU%w)@R z<>v4ztH5$JWbDU&jHk|CWTEb8x-^|m!~GLjMhnwmq{y6Fd_fAg|99B8&1UYAn?v&Z}3nv?CE<)tN{fPQrD-WEhjH=rg7xn~033>E+y zEabOGUHo>yxH?@)Wn|1|z(7+T%7)3_JbrQudi9vah#&vNxyKzz1_O^XP%yo@QEX&r zk(TZY2-Q9H;wle0N^?LPUxBAEZ108F7iBiSW->B=|n6TU8u-SyWn zJN>Gj?t@R&TTwO@3B$)4FXlbG`L7^&9Sv z(Q&MI{jVQw%Zj8goEW7I5@ggX{M1qQZS9YD%*-i^b9{xbC@Bw90~MS81*Ek{I0s8#M#YOkGw(5;k?{egAC+8>j3q9r$DoLGD(}=^J(L~edy)h%hXr6 zUwj@Xc(U9{F9d}7cI|{L=VZ0q2L;N0lBy1&_IS?Y8Z8YHDhg(Rp{(zIjJS(@tufFr z|7)2Z?hBesHBk82Ei00YDICLiUZr;&clc@fQ2oSXryZSj>F)Tc@WZFFkEocH8!>5qq5w;fXpf2s`M0=^tWs`W}yPm42y>;GQ<&^6zcP)B&7xBr^tIQ;u)*8`w*9m^p zP9xt<$E|_I@H#p%@#PbYYL}gUE1i$zbCbJk&NVxY%Rj8z5i#tq8DaBP74xE8Zg`KY+B_{lG%X>>fU>1??!KA|pl59K8VnXf;;J%f%HjUSvL)k4bb3nU~XK zOzBP>&wC#8{3jm}gWJs&x@QZz?&L;>`cXxma}WhX<0?{*u-Be~f=QQ7rd%LHSi z?0Om^d8`ork7BiSJfn!M>Qn!?`lhSG?Gp7#4%yUqrYZ{eo(6Q9Sfpm1+GF=KxqIaK zeYR~Vv5poYbJhuR(hfJ2fSmdlxLZ=Ps?ZOvW8zjU8k>t{|Cq(wZbS!yFA?#7#0)UPJ-jQyFVWf zEb5XflybCUV`7*KQ;g=C%`q)#nLU`% z0tWIyTp=-WK~-uGU+AbJGlMfhq_?_Sl&g1Ee-Y)`W)8a?H6Nx>wrg0Vm$$c$RaG7e zbLiAgiwGQZe#cHZgkU^4`1uJZpU^a8Q8r&HN|2kRAWkX6CcVE<8%Sjf$ zh>q*YiU;@M#^NJAzu_A~^i`(91i*yT@Rm+P3=_41Z|WbFqCQ zSQsDfo<^;mHP%&4M@0@N<(B+x1&Q*BR`*~- z^3d!k83m4Vj&ffKN3qlH#;V|FtD5*7s5#Djfpi~r4vF?6HT%R4xk#Eh^O0!mSkHVm zN+2aQn1E%a$L1mpCNA`%vg_Axh=0?!G`*dC-(!$wWT^A~_Z~kcYrzI0G@I#qll$7H zw+CT%&8i|6)9Y-OL2K<8PEK<0qKra9uW zJOMRYG)(Y7#7f3DmZ?45KF8K&(`o+pw>Ag1TYcKcAnQvXwb0yRyF&8Ts1^6g8hsiG zCF$xH2s*nx-^ar%*c48;F~;j)G^ZP3u-(fJef$Zs!_PPTSqO)!2%pYc4{e-mZA6MuV>-rWjSX_YoaIKH!9S9o zcb^o#k)W7jj}s_~QO z3YIG7)dStbvQMugU40JF$EXPjPl+6p*EbwC=Oc48jju{V46UH^Cr*`?8nuMZEgEjbSSlOX1UA^1BOWL-E92tqNN`x1yC%6)*;>8L* zTA{h)YYa&&<5_pAH4YUW>KBUOMM*4|1C54O?Om=UcWkpI-fc3-p}E|SA%P&@bp0*aj~Cdk7wf& zt3~F5r`*Z0Pi?k`)hI9zD+Mw3iZ!e-PX>Q+oCJ`F4jK`A#5BSCJ6=-hSG^Fm+>0IZ z>ORJUvPv1jDdDGcv(yv(J97S+lMcrE#CK?HuHwsNGyCCnEb-C?*yG@r5KzW1i=Gat zX%CzI=(xS+Tk?`Y4)654`Sq;8?PC5<9+oQy>%|Y6fg=6V8h@;(QE+=L+nI~}l$&Ck z9i`CJylEggpY?HNbFMOWDq%32q$djqyCA_b^h2G*3D_qs1$6XQ4y6{Yr|l?IO8P3Q z?XVr3^>76u1WZpxOi;v~!C@eLjYPkp$}hU8;Xq8R&EAo%aYMyfmX1v1u5VEbKVf89 z&Z0HFm$9h2)cP#7W6ZR^S*6?Yx27;D#ePEu=w;=&VAits;O41s-f#kVcac-mjk*-y zFL(Uq7w=t>F|+F~>Jt1wlu|I$`q)(=1Ug~i<}~$+Zdt5L0J^CR7z5N?;f&4dwUwXu z!LzfM%#Q=wYZU0I%Sr#FIk5#J3)4#~|4D^n+b|86*8!C=w$0N-X$8kG=%rjPZ{b|0 zey->BVDTz#WL+Gr($yJUDSLqnlza~FPg7xEgiBe{4~Oh+3OqTxY_oWE+{8M+3=cML zE(edNT1*9tX14EA$LR+89p;w6yE1yjUbd0U9_W+fPN)CfU5BXbwEY|zIg`uBq&96dBi#9);)NAbgl3Z7fVQM>x;k1S9XIsw(_jJqEN_t|a1smzNqbvOk!S z)BZLPVpQ*YFfT96%?;h0YYW9XKHN3x&nlFam*40V!8bH8U{=j<$h)N8taC3@y(AmB z+5)E{0RaJC-oXRj<@PJ*7F&|79qy+iKEdtEqXHg|KRP}6BZb~3b$gnw?d1~>fER9L zeea2$%)JOkMY)84*+gY}q3^-(TL}<_gqBbHtZPK-9<3OPeM>k%nC4LBWC?36j}Ffl zr+i}$&LLE%qSDici_w#1-8fOqo4S2CIGLxhCHRJ=I&4&u%J&^%MPrI~@TBCkZqcyR z8-ywm_5AS;8$*jsw(v5#XFf1r1k7|#&j{KM9Hfww(u%bfCa&z`vr|G$`Q z-?`}skflTkvu{W#r~o3-F}(6oDBp_(WkBTg?WwJYQA7&AqfWR0C z2QJ4Tj|QWwpe=*z+`Ej9p3j$Q4H^glOIU~N;q}ipW+S!Z1e`a zTaZx03Cj~t(jBj6TOceYA3@xh-Q1=E?RUtHBz;$OLJJs7Wi{{Q4dI z;o@JAa?tIIG0>bHwz#xbN$@}@B5)6H zhjJ_9s{Fm;cW)L}Zr;PQs3g>$B=yI0NeR?6S7>wKzAf5OCQMbMr7ANi7D1;=UA#_> z`yp0zVdl2I&0T8cQ?6gAu;ytb;IBBsqbU{$V%at^Qj%!x2Dh$xWus>@%+m7bk0_Io zMG$X&`TF&xkk)TBQKJ(z!efqyS=K&_D5KWI7ekB=i*vfYwrQr<$XY-7P}$NYBsbw7qTfYa;(p zxL8pBR1dLs_Hp{Bb_fDs=+ht=4JPPOoS<^u@%Whqk>)z(&k`AfS3KN-v*@e=MvLT2 z{01kt`-siHtRgrAPmL3vfp`B9Z)X=9LFez7Cq#%hNV_bfX^uin{bY38!x&||k!rd$ zO3yGZ$(K?2>#*}X!o#b_pAQOL_t806q4b9}3zE6{_kR^@XW1`Wm9gLb>hyw!wx*E| z4TTT{EyM=SFNJzY-SrVzZZqE*GcwF%;zms1?abfCl7(h%@vyKas=$Z~n{xNo3=E^+xRv-ln2E&T;A4A=GA za<)V}xR+3C2&6zPb~wvOCGEUoloi;&qUJe2<0jw%s?zN(OCYWY<=e5SXxl zO|Wx>3i)lvSn;sQFaedgoVgdMD{Ms*i55qwgSlg>X?f#DUaY9w0`RQY$3gQxF}W zpSNVL)nQOyzI?fIe=Ztm%|8S+Kd7;_wqBcSi`LiI-`9jkuteX}1%8!e0}i%)kOPPf zjM|G+2s92Q({>9Pt+pB$8j5e+JS@F*MmwS1co=FEbrxFJ z^aD?bF*~u0{Pd%Rb-08))jqh}ffJU(t3X(D*UGgW*6py+u8oZVb73yYo)DrVX+1p+ zOUv>$_v3}Z*Ldam0&S)jC<-y@vMF*qu)`Kd=i@!9Gc(_RpxEXmbg2VOhD5f2Oac$Sy4G^IF1inYz1u{>@jSb3~y5<5a%)@ z6=}8@MB?DOF$T5&$OnC|Z=kiJ-V*U+<(kTp*|jrH0dL_JQ!&XbG34`ub}5Vi6gW}? zO9#Cd!U1dmUV>ZV(x}EIy|kQ1=a@F(&=0>C3;S1%*{O8ZSI(Fje{*~3hJA$cKdDl=qXT-=h06MM%Uv6t^J&@t%7eEICM5*a^ z)^^%ixPCq^YtkvZZnR@pD=iMex$H?K_=ddto6({GT5h^63eGX8#to*CQm4=GJ;*_T zTDC$-kCJ~5w>Dq5Q@(;VXlF$-L0pq00U{&v-=F*h31Pn@#wXdv)j9Mxs;Gp#NKqs& z%y;S!;7{+fALQOsp`EMDCcKm1=$5v)MZo%~@)+%yD74DWB_QeL-fGI%g4zD+qc5TZ z#x)pUF@Mb^C(p@u5a?GOp{hkktGx^Fm)I@riFVe`a|dQy=aHVE^&4swrP_R?oclZEIelQ#t(tB+b9=;k z&DzG|ow6~f%$Wo3!?Yh3()DEgRx!eJuDcCdN%~hpbl0mGtNXw2$euQ#hdzm!lQde| z9oG;pO#3W4_kM6O6HobR7H-jFHerf^z_*OdVTj6gNo8B{pTwEMfOqR>38*g|Xh~YRim|_VDmX_^yv~GZY zbysw*Tz#chw}PJi@UxQ^Y-*B0MtTd7CZ(7WNG-PLst&Emjo!W3v0dSxb_3Uyj0F^% zxW(5e$kyj>!9r;50jszyJo;^0VqBZ~bLx$CmXW#6r&J&5#*_`CNQagIH}CjN}I0&3WQeS^MQQ^f1>*yIPq`0w%moGxG|VUcQK zQT>Rl%4Pp26;od3B#nSYcyu%cp1D~zU&rtyxn0AYH^#Oszn(}Bsj7(k{?^U=4HkIR zguw?idtdKqg;A$8(zs6(QS`YUu?=bU^~>+X7bkrAxEMj*dI?;gC`zwTQb~STPRJ)X zy~3%xu#tdTdQ`$s(T_(em}fxrTH}NmA(h>W7pH>}bLpG05EW@_GZB$|oWe;LKSuPa zCf{zoYWt#vSiIqR-ABM%hv^|o57tX=Yfin!up4Amwl8B+L@yWOU6|nT{a9A0fcOTj z$~YU6R5PLD%V@tUY0{2-e%%hIeo7=0~HTNT>bokPZKs1 zttGyd*bohmpI;zo878P zTy`V-oF6_NBYl($-}Acep%MEmd;DL?OUz~Yom7f55GP9NIGzbHFCesR@m*@?Mn{rM zzlR1z1w%vmb7+a0>sv6LPten0%ecT^S=Z?fMzLD*$q!9(K_TjUp z_ujg{IAVn){NFe3H9p_P#7#2()Ys$_@rioDk7<2=X^q$pwDelR$*F#nxy)K`J&@`W zbX_#uwi56z)dqHdTpjPR2EW%L@0lK z^e$Q>ngI5NkKTjeb+mV{HA=5LE1Cjc+@yeU;-#QK+k+cLpd zjZvA;#0?{LPr7*4_&*_kH$Ytpk0wAT(2o&$vw*Wr!)Zhtxww0N<1JwJ)(N!RwSs;m z9qq{-O(UEm-E^pTG5-Cz2j}eUE2(~L_X7r48@B4|g4u|sfqP*J_Q5Dr*C^#^|N6&V z4FQauHM`&^2a)r2cE3hVzw)o29OQO{2ilI|OkEflk=-%pC==y?%%uQ~((C_yT+zQX z9^&h7ymwBw@YyaO>;jhk71o7o!}pBp?KB zgdU-RcYgp5(9-z>ztWC^TvlW=Qk3vKr|#7eLc#EJ!~u12dGe6qk8QU84~PU+9Xt); zHE?VI(B>N{cMo~raeE!2J@aq|<#EmC#MNq z0x1yutkSt{Kevt>X+!y0l+&%;tk!n$7w#|M?c^7gEw(mC;RPqM*Vt{P6!g2>zKR>p zCB`5GbfxBybG1<}Vq%`oJPjXMSEDS|Yj{`ym;u(uX!hah7?^SI+|Cn#%47vDzE5*d z@W8n>3oVVZ4P-=8yAJAA_t+nZO}EIql(1a8MKcIbN#QYVVtdKA82W#UpNdzaKFc(h zS7=p_b3bU}Un`~zAK$+27HKeW4j-X={x0LEPdaq<-%yJlnLyI!=H$jIcCW{AA|RxPjc|&Rg}_-wNe$gp_z-%x{V}m{i+;g`Fzhg%%r0X!vG)?(?^9E^64qqv@y8m>wfuE-0qEf-6KKYv0uILoO&HvjBE3QlNAdLI(M>(VFR$$rS+k& z>C2fT>^&tNv6n*kJ{^QPG@b6B0N$1*kOkytU4s`WH`vgQMYpFxxAV;>uUL8Lb>HVr z=pFl)Qx-`HqPpw^L#}~To6_WYc-*Die*#5+lN_{Uu7fn}OVHba2%U`h{8Eq(0?lFA z+rin!LjTv<4l+H$8*lD$N`Ge+I`xyUKbUx3Q|*ZM=|Uk_PoGD*cLibG$}>0Y5?=cd zj5)tur0>dlJP)9yC%vAALM$!HpJ)H5f6JW5aKFHz!Pzn93T>g_x2PZb{Y;efWL6$7 z#;3mMvQb=$;&->ME|BNhGKZT5u$|^kYmCS5)SLtZB>e`;hR~IKwo5TFtV;$4xX0KL zv|w(2_VJe1F{IX{7mv_1^iPO4ck2w|J$4P4@;1K09SM_1q} z82&Vtc6;xgdhqMAaC2IMyQgwr=D2O3BgBlM;%&k)q%Pa5AX|psf#!W9iGFHPxr8J( zC*5W$hXmGq+~aOhUOL#x-UF>=L+_5ev{j zwZ0G5B=H5hoKw(w;#M^_qKPzMN|%hyTwqYaL372w6g~fb8R!@YZVd_)aF$XoL~Wv+ zFwsyJ6|Vt=U-y>ZYxm1?I6?H2_qjPgT=uP5xn^l$EM{QZKvG&X>F>!rZLjttGMluD zuI2c-)q3ttCRJDS(+&C-_N#3qu}R&~Li_xV`p^u3Js`b*w4^Gm0kRF=0)0 zr8W_(1qPA=Uu?W~H1<93~D z)FWX^q1@()%-eEFt$?;^~c{`dwXr_PI5V z7J*b>Wg@G}DWeGPzhP@3?sggjv0%@v=-m1fJ@E$i%MNxP5wrUXr<<>HH&ikPHUO#W z{b>sB;E`i;Bw8OhbPJ(129i)IRcarcn;$&Io0Xhv)*)Al;8xT`$d4#w@(kH#IfLxD zIrVkqC2D+m-A0W~czrFDkrDaC>0Wr^0Pdb_PhK>4*54nmJVFC6#N80iN<52qv<&ok zlaotN00?f}99^@f%v&#mZRKQb7gaSKy+ieIGIJvfAD|W&dO-QiOuuSI6Z6Q;zO|-X ze+@F`S%Ue?#HZmF{1WK8Nz&l#JOL#Z^`*?Ahxv9bAlcq(0}f5zljs~5p97zq>-tqT zN>Xl-Ah`>$faCX(B9Bo;-r#T66l~huq#DH75Q&%(x`cTP8Sgy4Os~ZPYE#KTa3c!P1^-eF z3qGA=>g$o2dslnH48RTqjQn!;&Ckp;-q9$fHEU-rzzR)11J{?4!1ipNocwoT~H)YinK2Tnd(hkxcMGOEMOLSpvz!q$b*V;G4vA(Wv zY@a8BLwgA3#fk;ZY#RfTroNi{&z@lvDc$UvrZ|k-Sl9iXUTI$34`Jhrw*<;;xi?9X zL=uUtM}GUB@c4@%$WFfK!wxOsb?p5}B%bh}?$;oyn^zp+YmD_kKVvqgX`m}IG% zD{4pEl*ip$OU>tKh8Pk$lCP&8GSdq3v{R9uCsiQhEpYr4X~Ie|^+ce>=4a^X3R)-m zg=0PeF=9hVSE}=Yh2S08Kc3Sn6{nogv(e4o1r$1OVJYS;+J9}ay zXdqzV1GGyLX=YKCP(C}@RBFza^*ipq!Pt8;g1O`up6E6HHzgiA@oCsseqK^Z2hD9u zQ7axhU8@9lr0LTbKa60ETG9&?)H&rhfn|2A#i?2B;@DUIqR8r&GRv zgPi99uPPWLjc&+nW8>Sfuh0GSV9L6a^I2Jj*?U*fMGQ@sK9$)~OG*UT?e8L6$i8-p zR04u8M^)Dj zl!RPI?sa^DHyCyNOUVBAprw5QV^vR253ndrNJuzD)BxpMM0j}KhEcpbDG7;Tm1^^j zQTT@&&I^h2{0etn0G2CvQSu9fFj5=^9|hIPswzV{?=InKd1{t9t>@3M_7V5$j^@RO zgwQ>HjLLrllL-{H;(|61dFLXQOJmeWe@2;9iGXsfo*p_0o~0y{Q3)Q`~D1!nofs=-THCZQ8nE~1rf&L4}2iwy-z}b)Rgu3)hisaE85dRmh?k-6*ZO6Ui0{ve@skpK>_sl$UQaH5=1WC3IC0jrW%_9h7IxW zS3q!bgL9gUy{b8(*vNfop@+`Aa2&1sjpJsQyj|M7U!E51QdXN3OH(Vj^wpNyD%tuz zoky;!S3oqFSe&-eu$_b5IGy;vg010``-crxwVw#f2`T{}gTftJW+r^-KLH`r<)AR? zh|^S20j-Wfm!d}!jpkg-_T}ekUhg+=uH&HZSb`o|mX?+;U(R(UOLljG1KV;*&+_6yT6A#G;&gqa*kafd4I}l3n7DXLYb#-JK<}5GM#}u#pXa}a z(DjKp#VkO-e>WeFi2Fcn#PKO9DLx}RThQ6oNK7nY0AcXbQnlRP-DbUlI9vk5!!rWi%lD^32(JX zUqFWOd>QqMa%Lh~o$z3E$LT3didWYwY#06#WR*r#*RR^Y!juJcblq@gW^h^OLFwg9yPx8=xp%aRTMm{}-}Pph9FqkZ|$ zY9onsPD&<98JM_`*p2q- z)9)Jb@1sL@{>ABcys~pZL=%7OK2H!$Mtst3+kxp&jEc>P?b$5F0D9$R_imO{pp;KN zpj=Gsl+5%+rt_wXgYL@s(Y?;R>G(`}$9w@vTTa`69I=4lgJaY0$ zrE5}pdbEcv%bhHso1Fj~NKEFo80cvyh4+0&C@3oW3kHQB@vp=~-Q3(BJ@TA2Zq$VO z6&&Gdic21Drz@sKIDAI{1JRGqi@sPjI804bGv6L~AFcl3Q@M}k7b%s{d!USUAG{=W z6>_&^fj0GFKtVFtC9bAd>WlUBZg|R7owthOF5$UwGWu@?C@2Zw=?(F_^Eu4q+jg9g1Nfb z%?kq#yEekm?|cN6D9)R67DI*p;SPM7CW~#Sj^u$&3kwSY0a$tRc^{ZHx>ZcFM8fq) z5q_eNK`S1aSiasikVqZ%JC;@(%6S`5eVd8@Rvwpvb_yl%E$f(XbM*?nU61$<5zPW+E15o4Dl*z3|{26@Noq{7I9Hk&zy29d)~2=P^8h zt-07h)G|asNxM@M{U37{i}!3EkUhhT_h&^vQ(^FAu(1f6ybnpV1u`2@S9QnMH{n?J zTQK_yXq7k=)_5%Kn>Dx$<$JfQTJ7fS;^JaslV|wM=zjYKcqDl_w&be2B}&6pe7RKC zkND9=lBIXet*ucM3o0eIEWogcc~?5h-HxX)jK*J$=aKEK)l$)1rdLA;(ChN=#4gBQ zWeRus9|!%%gvgPh?^!;CgHr>4z2V`=p7rl|w))$$O7-jqkn(9h=9z%@8Co5zOe9uO zS6WzU@9;&EM0p6fDX0xL!G0|*FXTK&evjk^xi| z+~IVPdeU%BqFFymJ8JZBt2@}o7yoMo*v=Z2Ja=v;bbumyZu^B=aAERo@dM4+;vzP^ zsBs;l6bJxjZL8Dgs+BEQxkW!4241A#4Phy8h)ihZ_s|2kut@rcb6?ZS!kc{$Ax!#n zuBFTl?;1!kArP-Ix5amRI+?{9kMTRctUTVn!pn7QId7J-V1qIbNS`or`w=&o1&<0+ z;_{YBgQ{RgOYCfNcGa(~z_ z1^$V^sWig}YHv>V27m<_~(YEFOvDTaF%3SQG$rmlQ&#HBUt6pD5EpeAyOU!Uw8sjtUNp$*a)=e; zT)hCm0DXt#0U5dpj3BE-yIvFBN!u=HSEl_o`wOo#iKqFAxb6ut+ce#XyA_48y=FQU zs0@O*DzwL{iMpr0E3`H6jSm^sGAxM;G{eFVlpjwO)RqI&1r~r`fDfWd3rT2nbcIDdjOGRW6C4-X04j~*tf$+LHe$lk-(9rA~OE}$r`nJpQ7x6IAnFt>* z6z3h@{O8EbX5cLs_B208AC{6~)$&}eIhzjvV;vpsW$vM_2iBZGX8$b4cpnIEv^Hj* zfT+R={m)q$DdUJ`v0^@rj`Vni7!YO1_vxpRd@Lb$<6C_g6ZYr&0+%2^T1b7pMMKm4 z!OxoFzWyDGO?IrK$*gq^UOf-`{c{a?D3oRLu&lv12MVrc_1yg9C&u`0I%dkn0@a{Z zzD;>zM<|LLOQnEdi`+0HObunt4$_b z`=Z@^Qp$7uRcfa-HfNegT!w>h#tDEcjm3Tm<5;}mEGeZr(U4%!jI}-ds}ByjURBU_ z>--{IK!43Z#-)p9=Rit)zjTkz_UF>>m%09D+r+6Ihz5ckqJG5Q4_C=SB zi%NbKnaKi`)TtK8r64eIG<*!p92rWhN2YYCQWPzB2Lrr!MXBw#75BwZ8I(=87?m z>5t=yv8u)D^fuZ~UAfp-_G4d>=eW2qKP2z5^aupXq1G$@5tlB|e4P8SmfymdBp*BU z1b7|cF}N}OhJS9YcOM(t@)@W%Eb71h3-pPTKP*jH%;X#jcxqx;0|ZwyGX^$CK|Hd4 zS1m;e(?u9bEk;U1M#pC7Lkq@3bU1o(Zt&MN74igaJ1CpGIRx$xPoYvmU-}<@3-i$g z2ADR}O@n(mJxFBl3hpWlvBFH73qeRN!rIA^Dn+#QJ5ll;1c|D|@|!~Y#6h=FVP^<@ z?F}&64QpvQp;9#oJVN%y7?@URshOUgC84I)mX+1vRimV%1Iv=?UFx4#d?RG_p*F%T zS8yA8TFv(>Ir7CE?;6-li4OpoFo3;dl!j(^TqK9xC+~b#!(btC30>jOOdMXadPKe0 z$ZPo{ynjF(2xOnR(5g5fCQnc)X-0q*=w@_BINW6k&haNIkXEc#3Rc zdz%Ng#WWDC+hi)c4FEi!gw@xV zjEy-Cdbm&nI$VDBzdhun+9rvKNLFU%c)l4a>jmn2GO2x%5)!^}^>-Vr`S6YPb?@!F znY-Gth2%6eqlvAFha%v0F5i?3@}369B+kSVeP zIxK$Un8+ByK2tF=k*2h&@Hj!B4#vmQ(qI5b=n_5A(j3n9^HaRoFm(egGyzy>!oMBm zfd^1}KG3m`zXyf?P?-B9V%ia12=F6N3|{7M18V1j0{soI^!ndQN@iCDTieH$i`34+ zyoR#1ZW+7K^Lu*)sR~CVE=-CZ<%>YT97gXgZS@;kigQWzJACuyQHlK&ukBc5Oblir zSEKF-C`vC2^_CFK?L!~Ife7Ei0>yEA|5FAgvVKI#J=S(}b3rto4}1RMCrl1HxANw* z3%k%KHz1RSiWG*4N5aB4zH9w#M`b`n&EV^$?;2RHu+vhp>kWGQebkQ=vUu?!hglnFY=KOB^~Spz6MP z<0)bC=@^gp75DuG>AE_CK3%Z=IP_w!TZ3;!x%DMrL!v<`4Q6+M0X2fx34YJwW@3#D|%4E1Qz#kaS& zD~cW!BqUVlHbPfnF6A4=)@YM^{od!c^mmucZi=qw1n`ZdKhU5+=@(aFWCWJX>vOJi zGM!k~ci*AdVwyErnA|)W*)^{1(tlz4)?e_FhYznaBf$SCnhqgQM-wXf(5O-&IkSBaff89kMnhH2*%~`>0#1J^a|-d{Vc4$juNE+sm!>Ol z%BiLrm+0;RdM>AU-UO6@tm-LBsOqcW0;8m>&Hqc|;)AR6_s~7n_I_K01E+W} z5(aiCE~Ods#x`Sbc?-tMeNqB$BeJ^M8*fQjNAzp6FbKP(0KX=Vt55IsONe|h;2vz0 ze{aI6sh`n@eMn?On_7pM+brGaQAx?KO*{;=?vaH!N}vtbV?Ixc<9}RQ1nJUSQkwmS ziDlmHH`3HQ-@!B?@Nz54*%|}jxW&ai&_eaRY$9Md@cr=&1d6e}=atE+Eg%UGFh!!urSaPUW^}=j#>E=(!$s{+MBn;Bd3Y z%)`E*KOHd@1@;-DHmbi-=|kDkM-z>9$apsalej+E(|D`zL+4wbhs?vS*e=AO+z-Xt zcU^53hy(Af)Lcanq=9LLPIclhKY2pK%uIjzG^yQB zdz|E5#7+aS)!~dB04CxRK$JAiNA-e3z<4#fCkhIkBZXF)#I_r-OE2FBAU{%;?|2p4 z05(BiN=3fi=v~J5#ogF9Z6^$7k5(27B37!Nf1B#+D4uZ!y9W&J#DE)o+<0XKwGxV( zkx+HVkp6Ef3(9N z8YhI$6*(!X#tN4WUYn8nRxnlqOgoz3p$-63Cazw+`t+mG0=&dGw^z$24|Gm8nSVWv z`46d$PIG4VNIF>D+U)4)NnV#zX^auwqYQr6IJ*#sShRrq$|00%4J4x2%&v-)xLkdN z`zC2Ns->!hbV6$|xzZDDe`^eTuKh)SpAk-Lg7bluDo8z9=h7`#TupGpK>n%rj|26) zB}#k-#B}NNIpt|c@W+p^=d*H?O;pZ+(O+))fk>?zxH1VZ8OxQF^<9{anh|wfq@q~H z%fFCzU1#zqat0Re>0`;prto`GtFRRyx;$Mx^1tMin%4z5Ue2`hE0PlMU;A?b|3W#r zD|4ScS^=w;!_B0sSNPRNlX(sH;bqhh$BcrT9;oO0V|!67oC5Ula|f|)-(MGRc*8y$ z03(7SAFaqDWw+$b=EtF>15OC(!{uSb1*$_Ntmc@_fO9B58Sj@gsmRd+EK}w`mx*LA zjAg^##%4KA*j2-@J3&k^mgB~a2NmTHWAvjcKSo3}O@JXyGyg)Y^4|Unv8q`Qa)R&2 z_RV!lD@NIJ4|cw=UKbXH%2U+gpVUkB7)1~zS~r1%=PfB!yPj#)7UaN%`6P)bhM0(V4Rv6 zjC`tWqi4{>uGMV`8d|Dm-j!XkAgT^`@hquxmyHDBm@JLdyM0o76$%AdU@t0rZ_!+! z()il4=!sD&yUUx?$Neu9tim-9(1$nM;SsEZL8J6R%a4B**Xf)5>Q*7YSevB$&p1>@ z)d7a^@nhYS6Q>_XnrgPWxjc!w=QZh&6YgctManAqXLS_fVO@w3Fw?H1YHT~v*DKk_ zhEEMNJ-gcfR_0hAP=MhYMj+E2z&Zq@TUz#B9J_sKJx@fGmUaA%9UV{p8r?$(lmlhz zygBh_^VrhK{u!E0*dbpq!`J182Z~{%YV0I&Rc4Cr|FbgLdd;pM7Z!Jn25CRjEzy~U zY_&olG$5e;koEK*UjvxModJ2?Updz%fXiLBG3vy`qNx#}kUKt{le={966&tQbSQ)L z+pAZ`gkD8`zL2i}*#0$IveRLk zb#n9bXMT`(LXUKuoR(WcVd4+YHXv@@ecslayyWkXPF8~2*9Yi%JRK|1WMNS>Ght7% z6cfYAXs7;c36;7)meKS$X{6WmabVoBUjKzdWXk<;lfr-AzM6Zd9{vCNVK$KZsY*!* zCH4hyu#Z1bXTA&hUn+(m_DdNngbXrg*7mEN!29|EL9v4$B)9foDhhBS{VrFo1}%8} zXZD~8^rwprxVia$?YTk~Gefsu|2Xn)wHpF{R zV+Up&kp@-tNLmEFg({#|G(amy%^dydNs%+z!NLL=&fMQbc;K(Zbz>#pw|hu~#`({PyXj5!V7eUCeu5N#=UaBy-m=7bfb#3xLL3PMl3*DQx3}JbeJBLd0$X5t zYBe1`PhmIGK?EdOqos}t#iO`+zo~DKLSXZp{dXz%ZZAxB{_|%cVJkJzP#B@CGAJQ& zA<%92FW#84*`?-@%)Fi5ccf z!2MHkkXch~&Q+yV&dgST0>bZ@zhjrXP#4fyfJeIuTgu~ecV6wDAe|^G=5&p1^J^ES z6S@TKbXCj+Z(BG->-&aK>?)_~{0WF2Nz;bjQOg zT|-X9O?q?OTTvAL+Y!Fm=kV}#PLE=DE^>~Oz!>(@U)jGAr{-WAd1Z-13uVK*U*7(8&jMb5pwibRCknVV`vTFfdmSWYGAHanF)ARki*eLqhySfPy z-e3mlJV6YwadKT&76M4riq6<9oOaa6#7-ZSTk)E!4xozQQ$M271-XgaeH6kM-NTl( z3cEiJBJiShkMgzOgi_Oh7hwWZX#ppuk3u|t%%|AO3}(%~e}Ccp`AGnhPfkv%rdR@$ zoC<(QCRSjuj5Tlm=mNX5e~p2RoE#e~D=RR%9ApGfsxbp5+*kr|)jq?));4Q7zF7~o z(Md3>k1qN01LIP>7kin27zK0MuWi64VFRdP`wA4Nn5)o1DYg!NvQs@bTPCP>JoBvt zT+7zjQ;g)V#FMv%oIUJs`t#H63MSrgQ3vFkb#bt_RC>F{%5bW(bk(HaikLvnizWuS_kM}Y$M-v#a#GBb{EBO)lQfZ6rIoCEy{vd)la939Vu%(j23pV}3RVz_x+m1}gttP5|4 z^gX3W2FQAx-RI)-o;ISbjDBu(AlL#EqY7H8>YCz5b9Ad&v@CIclRAODOLiA=xC=xz z4AMvmNy(Y%>GpV0k6)eesy9>X^k!#fdLUmw@#aJK64=$7@2rdm2M4pNWi)qo&U&Wg z516x(r}J73WY|^h-4qbeq(w16uH<=>7|gau0mLuXWO3mlz$cxCKRRPr4v!uY-x6J4 z1$>Qu*Pcpe_!!7jBX(a5(sW}M{cl}u)jGm*3f=X(`aUH(vgwlMWO>K~no}se^*iF0 zL1TIx9kCLA;b?A9_x_nPD~qxcVygN{;KEP2XxAgSH^2(w8K4B+?7v-~&6h!kQ46)L z88ffX`}f^+wx{!}BO{v;>w=E{Lq)(1;G^8l?VaX_R?~1W@T#7!>ESUH)VuEiQk|70 z;`~jp)`yl5%C0{tiUw~$#^ESH}wwlpc7%HyPcxXbQBVA zAStc0#;)4Ayp%Px>pNLjKkMt>-pe3uBY=bgGz?TkKHyfM{#9pqy;A9Y@^?n%*Uz?e zb;S!hS^6vlxD3#!rFgfds}-^nGa-DGIaOkiVFQ>gm=G+`$U!=&7@c@GGeuNHAV**H zx$by0i5&qRzMHm=!9mna(M}<@4>wShPt^7)rdD6{a(A}fKzxzyNt=kQMBo-UOfopd%r{eya^<|S-+Hc?49Or%T>XkSUVEo-z+ z`%Y3nv;JPAtOKa;f+SjLb_$G(JN5jlO%U^P6ohIE`~45@J){k+Zx1s&wfqADmX&6d z5b;3i0lWi!Nj3Q>FJ!ofwtR?;e1Wax+qp;r75v80!(aybeTLTF^Pwk>BiuK+!Ha-m z-dgC!`0Gl80}8g?#R z7HbS#DvcVbrHxv@q`x^ru>j+Kmup->|vSZ zZ-Pn&#qaQX%78{ZV)CF2GF#{q4%veC#crHAz%xw4=XakwGl3&uIM2i_Tg@l)e6cEW z+Adr#JN}+UNQ?208nMHDfOj%ZIo9pXL#Cb`pRy98X-;M6h#plHZFRn<=8=`B(6~8i zvY{ve3H24aiS6Q#N3fIaQM(4^kSE`ps&hT6fmTg56<28k~f9?6v8usj3F`F;d8Cll|f^-)MgX z?nbAmuUw*~wP%KAexot>`5SG+npXExflbD+4l@}0(LZ$RR`FShDn!>WgL;M&9*Dk< z{1Ef_M z%$#p?yaxVkPBG$OSDQxE^_^u3Q>DrW_3g>vt3MiFKL>k}o1aOS=eK${_|__HyoMW< z2QC47HGM5aZ$hT$zi+O@I|TsTT3vm1+it96Z6P%t`SUd70M9+hvK?S&a3wVu04Yb$ zO)}rSDa6x|kK&g6dG!C#7_Kyd-r9%euPvtM^SSHol2p=O;-cT^+GxoK+(#&bJ_(J> z=v&W*O>V2btP=%v)1rP_Njs2X_VRX4OklbDz`))%3?*)(fkHM4qA;9^YYvgU3SL z-kZFK8Uo3V$5ivgK-hk+7(^A#aJ5*LP?7I{udkzXawlv1Fs zAbqzyNJBgSeeo|4{oa}SSJJLvIFO;KU2MCx$t;9VMU^NV_(Df#3?%Li+R4f)D*5^O zZ@^^`ybC~2N^0uc+c0QR3aIzg5wW}iPW_=jwRh+C2BQ~c(Y{MUlpZ@^StU!8t&~NL zDJ(9BGq$HbLe<$Se%#{-VJ!#lApr~uH4GbbOa^%W+-cIpX|5jR$NS|zGi?#6si}{8 zX)ZIufd)d~b-56Lt-s`zk~ypY!huqBP(};3^@zDai%=A^pmPQ$15hYOJ!OWX@d)(g zji(k?VY&#PwpbE8R81$FIxY(L1rgmgZEc*CkTC6c!hc@GNOI5`I_|@LBfmD!r0u3$ z}zHyt1OD6U-T-qSR;nJ;iqlaA3BU}7?vX$?ao4wiw!&1G?>Q|q!|FwakH(E>YB z=|w^X1=_9LDDLu|o>VLi zE!{)7Fr^2)_lNZUxy@sg`wXrRzG&FtGatk)j>pCH?$Nu1>}A6A%^oXx=MceK&KoF_ zWzdJl4PVyZX~9v8bi>AC*`p^HCeKL!aSa)J)))-s8cCbYJkB+}eMP%`Ha8`ghmrxlrN+x|&jb)-7yf z6)6b=s$H$5U>4QLfXfx7(C|%#$9hzIr)S$I?-M~a=WWzBhhG%8lsWt!EN^F{Jlv#% z8hjR@4-M^{d#^}=mQ-pBnVfWU1kS*?K~FXD{y2iUUwhIE0zSuX!BMa%?`XY^Q~u`V zz_1)LKy1|X)&YDq3P>}wd5x86Np#;)l5gwECT_U`-siv7TaR(V=geA2nhudp?TzWW zE1`KADi&#rH1lTvD*Xt2RoLR)mcl^WFz8*$q50=tY`Rpbr~gZ{M5h7_2`^DmwcrwH z;B9kIk<-1b2CPl^l1AGAP%}1gcWYzFy=sCYBQ@028r5olMAxs+w0Z3J#Vb>5HIcUi zJ;d-APMedqDo}l$3dLJz(`#daP#k1;@P;DE`s1(varVLYb6c=%;7tttSjuy9g1Fz-H#OOfm-YXta?j$g zl=;QLm{1z>0F8O`Ij14&a)tBeU<~lQ&;p=dp#0BXwblj(MihCEOS+fdL(CO%k)GQj z{ouW*lpsH_=;1L*Wwn40Y?gj2y>jUo0O=;|@Ie>7mzn{@SI#c5Lu5N(LZU=HjzP6- zJ%*S8KFOyltTQ+);bElzOurdS4|mtVOxFV1e-i z?RS+G%-}8ksVlN;_}UT8vjlvq!B92wCaf>+oBAAb=asP*Tg5U$I;?;a@#gO#n=4`Hn+PnipB-o8*Bs(yiyLwxj8n3w1#t|DicmJIn z6~DUD=Jz)O7zif8#WDF8%~B$FJu~#LI_i;>6n>Y#Tq=J)N?J{e0m#*#Fm3aztz&() zftT1*n=Gxu)~vwUSLjkZOeE*FErtHr&cf@a_>Wj36hli&1&3=NwpagTwCS27RV_HG zfnXAn1AGYS$6j6`1N$cTk9|X{d!tE;N|+8~qB zlhHsf1N|DqafE!f^qzf$%rxRh?jTmqeYOG6gy?HWqV^$EK-{a7Ke^0N=&y^I1LEGD z2E|Z*20*v+{-t@s`H#!?x2nmnSZro?wlhXRW!L^K^btzjHLnD{ki+A11Od+V4b;X) zIY&uA_uQQ0Tj<}TdU;RjEbA$HvtJ3BR!35b$W>5S13_&#oxq5R^w~qmJnzaHj9{aU zwUw30=L_x|aB>iCGuP6|XuK8<$T7Qgx)BKe8Af&gQ%=eLWKQz<-K2iNnYVy1riMia z2dZx55S%TgG~Cny!1XW-`2VSNd^6KKTm;I%pTuuIE7o>g&wl%mz%CQ`B>Me4WD1s9 z|5;a6all^#d78e*_d8&BOP*h1=e)^|U^H!|T!MpaPhQgs$tqHN{^(6d`VL6CX(@Z* zhYWBLE>3c@v{PRHzFgao!)kf{S!B2J%|C8oO24Ya_NKzZLcp;ao*G>)ID&Qu!nKv# z_6U=0Tc8PaW#!=sL-lJ~0qPj~m9`Zd8XU}JSbJ{Guy^`BNUpEBOp|S|Ck2W;wOQw| zuGa(JZ>(|XZ0-1brU*T$3&aOqCmJF^8(3b^^Xtz}cI?t`cgC5`fZK$LST9F%g>V&? zFv>p*=0vhKwRS?@QC{R-=iOe`4m1po*XIIjx+mmjLMEj%i{8f%+{by;C@Eke`X6>;O=#{9u)TaV7j|fHI8H-E?z->#1T>kv`m0S> zWAc~6=fn(a{r%I|^+wU?SynIHNX~?tUgT=EvuRP0>EsC_$)(JMrnZrUH~Uk*q$wWg zU&ZNo=pUh`9U(!$aF>_N+*PGrr_n*D{)(9!ZifLSe*IwH$!jpOG=B$D*M6sN47Rg# zv^)8Hwo_m(&ebeyCk`940h`htr>*DB(Y{U8jU|Lu1TjMPp@yD_CN`~1US{UBerwDj z8?j_YhWtQ=cRL3XVh1emwKPe6iguyexue*jO3sP{2Ir~!U^e%?EWcsrlKFELn=0E1 z2QRJ@QX|n(bUBvytKGlVe$E~{5L-utpp@0+ABhEjZjZ!}G#JBLZmK(;DCPGKqEnOA zVh7Zu2(oL|V4e6!kh7SR<8K?<(P~lkOuY?+&W|K<_MwSh6e0CuC+Ct zfmr*G`zTzgn=XTg-D$x$#iurs1_uXu5z4s(edPp8@Vq`hUe3|{t?$Z|UvV97vD`J4 z8oQj4@Y>oz5*~8%>I2?yAKtahhHR{WJakZXGf*z2*~j;m)|>IglQ2B zs<=Z?PqG`!X~{cdy&scgp$h$R_@QzqPV-T^$q{7eMpOfpp`$*d72l8*B%#gAv* zSo$-%H^%c_2mN-!%gO|NZ48Q4uYlhYd~sctb`w^6?nptq+p-X($N&c|!mqb5tH#pX zFbWC^-n!M)pLw75UD6@M+U2^1Lv7i@p%Ad#_gmR&w89*+G~<#z_v=q)IS*@*iu7)v znl%~*$P!U&-b8cuw~^nQw_=m|vU)>Ino(Jyil)jCtC^TPQT> zoD2Q+K&C8P$Wz=Q_N#dFaBqm7%q4;&z@=iQ)Oygrjld*Zs^u{g0zK70y zIS{#=Piye6n><}#23jg#-seTI_=O?N6bD|*k};T9Lx))Qp{XDi5`{iSyU#amFs`-a z%{7X=K>aU|KSCj5Q9=cIpMAlW=?22>#b#POYXJB>M;|J$(~^o-akIc!o_c4aNIy_t z-d0lmQBSv?gVDXJ6_NTx0=U8zq&_SWBqgev0!!VIR?Iowa5DNq#TLHC`W)n z(-<$WT92JG!xPVrQlyJHvK6mk7fo+uz8Y74{0bpG|NRZ4Jvdb^Z^F)sWe?8H%oLVE zB5DkL-@RMg*#X0LimM#7k2)6%idBj#vT5(#cR5M;VfU~({19cP_2mU-8)n9-Ln~}@ zQk6wIYMebjYsFcZ|NRdB$_!Cda@rK z7dOsEQJ}P8VPkSl6ff)j1SA0+%>Qn5y=I}HKEy&1Y$=gHcNZ_+qHgR8Bn)PM@I>o)qEF2V6zVNLqzltmJ&nx`EhYEp~opaCd=EfK>1jw%@>-- z=(~zBUo|dWTUvFedhhgls#Y4G6kR;!K6KWHoR5zP`7PRgGW6}o$qOj2625sjBSe>c zoGos=Jj5pyM?{T;SW@TDqi(gwvXF#?gpQ6*OiU~%88rCbxzjy5`q7jgyMJ&XPMsGl zG;Jp%Ba>c8bpb!LMZ&L6`?IfJcHsN>-^}k3kkT1HchD$xFmZ5jP*(0`2o&PB{Kjnl zil>0JkAA?7}j7jGI-2M8Wb&n8=TUq zsRUY+x=hfTrJQcP&+dZLst9a4+0lcJMn%#{cdPXz(nI&$Cma4`gPwoH0Nb5|FKKBq zUw|)n8R)T&iJ=8E%#^=?Mn)v9z>nqS%)zCWazKYz7779bYJiJ^Mp?k*F^Yb9Jc*+FGwIi&mQLA0Qm*cu;G=PJx&JQrrz5RaP3H^MYaFGwxHeGpCTVkoa6 zqJ3z+<1)Srq*Ozt$%B}p_Y(0Rky3XyKOWGHW_CoWBU|);Z8{UTNMF{<-Hbr59C6|y_y8K|fTeGRD zLT?`(@7$f$(8S$`7N0wUJY@yNEWCA+z zRYsySe&g(`;9wroVqp9o@cuochXoyTNObhWsT-~?E>r z(B^Uj573dH@YQ)0hHo_fO@NsGCNgxO>#WCTm`5-q-98Q-JRXb3yR^`Pay1?*ZM_|$ z_M$Z)<>e+UGFXKs;E%6P=B0T%exqaxnyH6L$J1@qeIOBNsdI)sJIqGl!BMrHs zjAF1tKAo?&)OWM|fY}TN!f`Oqy|J;emXa#$;S((_A`tL(^F2vUf>5jHy9ggk>YO{K zR(I1K*mj{xj~qFdr{HO0!)p_b&1TlVGuF~_w8zf3_MpS#PH5Rr5n~LKfm~h3T$5jJ zH3H;`8k{2_oOAzv@#~kN^GH`n{&4};FTrUa>CBr4`Zob75I$#JZ7uNsC18M`A@pE4 zz!JjTP`J`|@zDiQS*ZtgFkDzdk0QQ}0^r(<9FrXN*7Xh~G4yKgvR@MQ@TN@FUGutN(_z zA5DNFwe8zXgqi)W>2YIP{X3&nA?wm_lbT4=5WB(7=HUyC78U^NL&D6u=$*lsf!)56-zE>0#ksm_XJQRwU zulNDzwXamm)GBoRi9n1ic}9@QJ?e$;3E{7{tMhc!=+w0bq>ICN_hYw*2iMkB>U@UD z@h<}RcU+u8+)CqSKg!0kvWVkxLOJQ z7e+|>d)f7mQ4|Eb8e1wjH!l8TAN@+_2dG_ZJG=6f>RU&QJ0AL?_+O5Sp{CM)UPp%L z!we%Pv)RDWph!V4(V4}9qIE6tzH`dX=SaXL$^|QG8rusb>q%72;fs%)Vw6AMZWecz zK88Rra|`21$v1Er*PySp&?+Ff^?=sy%0sqCv$ zti<3QkbmW3n=p{Re~q;9@mHvRBx-^!2&C{#nB0wBz)8nHH&&c{hfFW!|5iuH6*%QB z{mOCdgiAI@k4$a_>Lr9RKnRW?o+UVc9OZrF-Pl!d?)wrF>_hKxx0KKNNCn)ZqK>~- zSyffE7V1|RHzBNn3XN4Oa2bxvB#5Y0f(F54qqy$=h~v-9d=JsPOeTtOd3nz`l(j7M z-j|DcQBW=k8QGzzh&7Um^u>#OH!;^YVV$`D?dJ=$Y@R)PW@>7xqZ5-p`t6Rgiy@BD#1Z*NOq*e$j3jw7w}a)CO&Ux;;xlNVFnsUn=zlCAwxqq z`kpJRli)l&hGZXl&K~!VP>?2ig%%ruoaVmZ2XbcKd~P|vl(Mk_hN-cZ4bo)4t}@(s zg=A5SNMXb<(imPMH;Yt!gdli(7!+-o5Om=PxdA>Fz3qRl2_V=1R*4uLLD1tlh&k{J z=&KRpA7}?g?jIX|6;te_)BK#Z7MX7QPb3pX*bxRD96357~?+ z+ws;!U8eq`_2DpMn5lr7U$kxbzdiRWaO5n;;2j2VR^BmNufL_m?vEfKBir7C`o65Y zCVv`D!pS)kFXRV=WG!tJ*yM^H`t4iUXW;`ghPK04cmE~@7mWe7x9i+hiOJ!q(g*to z_g|PU!U>;C-^@?edeDEJEGz^yqKl^I}tjoX-TkLK#tFp%k8DtBHM1kJUd z&#$emfhh{s=?0t!+smPCGpBUdRG_GB`aHl-tlNi*ZtQhz$!4>^3oOhd+=og?fS9tz zJ3`(siPY=q&(_A89gcVm7Js$V+ne`H-$wGE`^QxZ#<)`nIDD2$di_NaG<$$C4W|Q4 zRaLKiu2~3*04X&1-H+TR4}r}LP!wDtl0+a71_s|3`_u1obEhR(eBq7slgyKpmX7DQ zzrn$Qgre#oSI+Kkx`32(0;=ncRk$>*@9ht zYnI(9r-OtB*m(_M~Y%}^>(-Xwpk69JC?criMeZs*-QuSDZ&+LDczz{rR zlES+NPpX@nqkTrFXJ@6Pq(V^lNUzAs%O~3hS5{O2cBJ&0W*W(%0+`RrYR*=VW?^Oq znl&BpkyR9hB~OJ_>^dbuqU}v*xA$Tr)Q*{0wU4&qphnUnmhLQ@Fkj#MAhY+=IP9Du z?HaJMTxdh?IAa7$%;KYPvCHGh?B!KzJr!N7}#eb!?g<3$RV?FHA74YK#8W}1H1S2COD{$*pw){41J!3`D z!LWJqC!dF(o?FcDrDu`qZQY68#=9KBKj1Y^<}z&aOmUBp zbKs(8-=#=uQ$s@KI?~3nIv&1GJtlVVJ@d-OU8OLCB17u~S*?L@Opo>UKO(I^q*qkb z9f+45;I8AG7vFUi+FI`^B7JAM(3gu3f}=5X>yDCt!RIDi2f=&z3KUs^skmDfzJOyR}Bz2gJyElw3><@Zua@@>hZi}2$zBY3T!=xBj8Io=WE#tqB{>}FlR zLx&4U5OHvJf+G~WXf9KO@qH|Fd$F5guA(v%*T_VxPL3u38n?k+b)BS*L*8=kE(=s* z63#FFsQFPJ>cZ#IcBGrdZq06A=X1F2xFK<#;47)JN!3nj5=4dxrQeUc$)cOf+Q6}= zG^IcM3`HMnR_s?hi}{O@#GO*koZWXi>p6OlU8iFwpKV&T8Zc}T-($Bj9J|c&fV;q6 z<^yJ>o=H22y5{_(wXx{aZobW+Nukhl&7kna;i+ugzG?P0(>wM-|6;0m!*EC2;IJKW z2kxZ~%=Mg1U|r$D(D3I+JsnK0RX}JQfLT$ViMij#NXqt1Pg|Q9K@cu_OZy5@1E7i* zblCF?TUNf(b_zfG2`JQ<-|MXrvrY%IXfxH)Tgs=IhR4Bf7Q3$bZhR?e6aeZH5<`5V zOByXvdsDoODV^UHO9eACD+Zotj9hsvPIWbA8|i%4dS z87b$J?!h(w7Fo%|o_LUVv5Qh`)XLhf!CPc!d*!Ta>0_5z2v(m2I404WT3-BUDn0Nj#5@U|nA&_Z0_Pgno}-Yoi`T;`AGf#9M@!)p zd2cx3<_9g0(G1uN@!XL&wchc>Lz*R&&l_(*u!_8%RK|hLm|d1_MxBwYTRwdU9&U*) zcc|N@g7xO%pD+BzO+*4b2OkfwFghSM-XzgCtgt2e3G&U8uC14qK#sH3_DSV53(=&V zKh0^MaE(GJVqZldS|=>7ah?;aTZ8rCs*wl$Oo#>%EU

7} zeY(Z&#q5@J=gxVz^=FD8<{GoVy)B@Il=g?;=gm*D(?~;GSV#pTgy~}2Uot5KF~AZWno@jw9FL(#jxxgPZHtRMg|y@NO){Ov zTV_wDS2at;QMY0gqiA7UumLwg)th-f)$t`}dzU|>i>n@cjhMB#{BeF;5FhTg0)7MH z8QJg}LGWH~J!>!VF@yZu#yRQi>};FE&rdT`;iLmt|@?5)QW`DL>hfc<8tsYC~CgLo#fx*Pn97}YEiLvXcr|$oTPGpx7f661 z&mX=heccUJ7tChBI%Eh9dd+@T?0Na#PR^ZU&+s4T)J>(c?^DE%TnxHN9$uc8!o!@$7M*;X6mY zag!rC$tri+6FlcIOUy#>Csu4D1~w1@ftjgkV?#q}8GATr+6tGDmY$oQE-Aa}t6~2E z{hFPVQC>lVFE*aw5UJBsIq@PO_7R>7sM-pW@8p6yI3y?3+FMp5#!!TP54!sB^DGbS zDX8~KFwMgmeCusIb6+SQ&>~llATJ08OrE8Xt>G4~PVQczwu*wjuhSGJ`u6I5RH~7{ zv|kuIjQ7rp2O8JuD~ypC4lwzW9+``eH4Lsuk4CH9(*@NXkq0t_awDF9qKuLr0=LC0F6u2zz8r zk=GxPxpH;D5%NM4(RG{ZTrdLERs*&1pm!SU*9M+mU;uxeRE6ZF!>q4G?)8fj!w*ro z!xO{BksRHSMuM2&^sX$|8({V0dGT&~7SfmvhC7A8w;{`eA@paw&cu53c4Z&!6R`!< zJNH;L#i^PzS4c1j_UYCU27}a?k#pgaf|HNc{3|jmt=gid4D@TUYvW(EAC&=`5=N>BjCYSx$+u~$usyz+) z2~?W}-HCeOK*+o;M6_45I&~js`oJYyks=Q&wbY$`8IVKkP9-zcb?cS&q4JfFhY*++f#*{#|g*x{Ean9npr zXWd;D#Evb0V9NDE_n+OKJZJWr{ix4Yt?FZ`5S>C%J7r znKW6iRkZHilC-b!wGM&igE` z;{0@g)>d{_t|tU^95c^NOdegu(YEx7zn(yJh-6O`8D>V*W5AVnB51 zM9Jgm-J_fda@Z#Vdp7Xqca%J#4rjgO`Dr*kUTlvnybUiK<4&8qF%bT5< zSXmzIIQUYfpesZCCnWR=iZC;3EXsyJxhR)Dc;Pc1a+v}OCzH+G2LPn|?)zv5$2rcI ztS^NH-G*2>I~mR|?Yx0|)Z2o@@s3m@#yonN#f}_DxR=<>wbBadG(&c3V;DEi&Od|y1Pl-`U0W}BHzVz}I?bTkzVb_IO<`1w= z@m+w|{XbpU*+)g_Zw#xcx)4)()smn6oQS#Ga|ut;8tZ3qsTz3`{ekwT$;ad zpsCxoPtZc~1CC}?oBMYnO2W##`axl8OAj}On`ox{N0)Q`v4QHO)BAbv-Fxxk1u!eE z2?zaD)z$ay?ChdnR}j~k$^1X?p=YOcla;?He*yjYk2+~k zUI+sF>wren1_1d3YQ!CC5!Bd8;KPXdtH0_eccT@G^08o^Y`zn#Bby}#!SZW=J|r-o z1MV9@`2XQyvO+2ViP~CPxT554hqbn~EdhsLu;d2XHy9ZDEuFNqMu1KlyobZ{=duW8 z+@{)9D~3^@3l>^o41f}3VDMseULJ7*-E1-6;W34kYI>;Z-yK<6iZ2?$v}CxkVb#G^~GlTGhaUcN~c zvbFsdJ*c0X_aXBq{GmyF$!U(WmmLu96)xkJRu9+!oHs(a1ta{)c1l$ID|lDzuZ!`u zU?NBmEk5z{0^@fj-Xf&a{>6A^ncv~go1NqZkQe|Br#U-tB>^g9p<~R+$(ht6bL^k< z7vs4tN<_Q{|Biwx_eb+sB@o_|O6}7WkZfDo(*%-)!x#E%Aj%^+^$e1Kypz4j|8=Q%IhSbESvf$@NriwNnxV zc*+<4&sM*SeN0i1IrmfG1Hro@XK86^ZH)m`u#8r+at~!Aa7W*q?-7he77Dnb#CZbV zjI@8HZA19jyc<5!N&&HJ+(FF*4udKvRy%5V-~BIA_}MfWpeQmqj^lE%rAwG+ev}e8kt2BlLdIFjwlK zfO(!;eoy=7wwIgpk9!&3uxDzq|Mk2ahBBd#EG8;k?l=&qUTL$72$FZpgdEenWZ`9I3YxUcEXyAzRvKI47YH17vsC^s{Ph|0&@W7 z0(n~9gMA-=a`v;D4CUpj0=K%kca=iYGBe*ix}a8cL%X>Ubap!knzdTn%W z6_+L;u%BSENjT}x$qmn}soDj)3Eln~);b?y0~HqSBgo8)Jr*>=Jwub%KOYDwAK|P+ zZCflnW!|`6mRYtbENE_FY9@9%lv^PTJb z+m~C{nrqH6=eWlh_qYcU(TDebLuiKGnKGmQsmdZDVH_VHbNqM^E({&EN}0{oI2UU# zJb5yXW_|N2+3m~}f7a>*Dwu!eY>9(EBj|XwHF}#FeOs`ldDPNuVc*|mle2 z$hO{+u~6IdvM)&2-{LeeFQhpvphTNq1paspzLI2=(+me(Z1@|>k1^y6+Q*T!Y0O^} zMRGEtVPHJ=>;`?%!nf4>7Ma&*_=Xv@Eo{h3G@)-*xJa(u$G>`e?2;^~!s-SHV9ozSnv~#2Sv5tZ%Nb)=>r`XcbC7 z^hx>w1aSJenHz5(pn_d!eFCU)3(>f^xrHw*F+RD&4T+LI8q>Be;WcvhGw)Tz>6310J{!)xA%h_aX{1d1BfcXhz~n3QmNyR&Hc#gMWh z4LiG8v!6xjtFerbWujjH>+1#!ner(xCEXo7mw=Z_oOS@ou)mhMvJ7{bOT@OEJ*xQR z_ZgP7!&SZ7sBWQhW0}39cNA5JDnP!z4AGs8c4Gn<{k*0PWJVHbb$uR-qdki;zd4KW zTy%zGquRo)Vp)%CKKU&Ly6JkVYQl*keQH7hm?6CA9qJ9=ZylvWK}H@l~e{HMHJHb zTSNhBrdMBkejh|)rJi|$fjUUU*rI3^;Q1wfkL5hwyR{SOKtYd&a<^I3`v(|X@a_V` zJ)KpX`0CZWY2SHTtbH|8(#2i{ZWNFjCX;O5Yyb9ZcJR^3teDdw3ij7p^{>I7loRi=i0-puWTm^XD zOgW@%85IRP<_U|*Ev?=-g_ovk-E0dR(b9OJ&>`^>|G^W?>mEAb>$Q#`s*%Evc5ihN zLJaCbk93)yMX)AZ-1Qk*t{Ilvw;nuKLBFWCbBt4V`{HES*z?=la%A~JHftD<2zjjFQ-_u6jYD8+~7)_ zk=8dz=O$z+qd#Tr^wbA4sXjg zOZaMk2(&E9-gej+aNP?%!efJHW^E60-7g$Bmj-inRLp0mNiZlhNtB}s8Z)xFw7xop zGBc^__1zM*O{Tl5Z$8)Tc+8gv=)1XwaG%`bNLfOuI=OAR2RT%i$jQjc^5Y+=x@4UC0f?&_hZdIQy@oZNT*dM=Nx%$!M}26$y6Vj6#K&bP!d1ti;t(XAyD zO|{+^jrYV`F;+Q#mB;*S>;PN?#lTc|A(3pt*{+;oSqphgur+hje#bFwTJ{E8c-$lf|klj`g zbcm*0hEy+FqjU`1Prh)5Hn?vrr)O}_aIaiHC4W_Id%TgT?abE+arZx742XYP%gmAV z75d6PA_Fq>Bkb}KZ$ZMwT`{3^akM(iI;kO$^cuF_(HkNNe+)YgY9N>GttsD+e=&|+ zJ?vJCO%i`E8z;&%_QQo zUx_Fw8KwweJ11?T@;TGYCN95V@;04>h3;IPmG37tyC1n*vCN_hSJkaj*IVf{V34|} zEY_3e3R7~b8mkq8{8a#jz%6Bh?6&i@{l23@|zqS2TIe`^g>Le6a=48|9Lc4h{|`5pLTy zli%vUeTS-e4C&7KZu|NGMI7F9%G6(t92tNYl(8+0$X!O3{JdVX)`<~P8M)|8amW@p z3~aRkjEnS0c3mO-vTY}^GH}{O;jpksotEDG^r^SJq@)u46~*tC0b*>{&e>UfXr2Bmefj zm*DzY*;6JFx;}np)J9W$0<=4VX%Te?Z(EQ-4NOX4j!t;mW@AaAwUq6?Y@UY3G%4EM z_V0~e>tB5M6NbMhY150ZVT<6>*zwFwt=;~fW{m3u%@CcSPO#yNLnm&+~F*2g&X~25b9YNB*6GZtt?Jy0VlJDX#e<}kmq{Ldlrs$oDru(QL zny|>oFot7ZLWH{Apa_`5mnQn=4rq}tp5St{g!Clr-R{|7!T1g0-rZ6m?bQiidcSO4 z{y2xXRS+5xuN$g+>brCK%=%eT(}zo)&lihx?+~;YjsBCf(~}6nAU_7AP08^isI}men9NNiLbZRN`?rwf> z@VxNw^wj()HeNx4zI%A+6O;QU$f$O3g2RA{!9rHPPb<~}P<0~X zwSn%W9*eI(^O$MZ$+o*mS7b@4F>7R4n1oDbHuK5JiSKUD2&^Nnt+USvev!)A({>?n z++4On^b}X}Bt9q}Vd_0PNwX#r@d%np3H>hWScuqd>|5AjldKGim>hoO#gkjA3^`we z_FkqKmo0{ziX}|oJnKh5^VxfaB^=+xYfT{zI1Wu4@i9da!!DpUm$1v7K_-u|mMTIm zg1N3nPOI&e3q@c3RJUbgEpao7*(^+deLK8zFz0z>^jK8YF7`0R7)kvz;{G!Sh6If! zs((0WrP|xnx`2(Y<9YOAUsbv?R6;hUeRkf!$t!xR&^0x$E<$1>h?x6(3YprpqNE?W z_oXdp1WgI$!Oof|keWO^bWr5iDsTdV+`tcSv27Rw#}C@CP!Labzi-_z4wNlhl+)#u zk$SCWMtOI3sXgUirvBhh9Ox(!5)*bPB!%>4@{8Dcucg=i~LEOkM8SOd}&|GsO+vqAO)}4h%KXxQE-eMyv9TmK-X+FSz(UP zK-lLs$hb?_uKt)cZ$RwF z&(FW69_OG5`U`R~9}#xc4#Yn)>v+B&3cXZ9xST@Eh%sghTiC^;6IHLKzL6>BJ+o1Z zs@x9%2@Zd0=*?9iZ|b60X;ns(BGE(|3CCP}xEd1X)Dk>1(`Y*nP7(HFPtbz*ce9JF zefrF+6KDaR1ay&pZG48!!3(d(V-7MK4bEdc_(<2?+{SNKWSSUonbsoW872KYRT-}4 z!Nbob%hrC!rW_T$#oW`zDDjum_|`lsKo;ptI`Dx6;ySqt<0Fdb*itQxzV)n? z{Cs3cqUYnImb0YB0m{A{a#UKi;{OCpAXr`EwF7yrIW|fEje@F~v=VFaC6F~E{SIvZ z>*$;G!5qaWX=!O7P052fJOX6!w8ko9Q}hT-neVP_x%JFwpi~dH&1(i$n7S zA3yeciORBYD>{MtN?9#86!lYGH-^I7QExq;JDpO6BGn{4W6vv_5%KIEefMVmLkJwr zvT)H7oJ*XHg5@`r*^!zggkQsNW~GT-GcEv^<7ayPhQrqAl?`Hi~DNKh<2vGiD`@C_rJdy8vfq5q3`(6`#}O{V^g@o>e%uoi_XUpDSFDvmIT*4T(z-XR_^UmoierzV2jir*IBJ(?n*xxwV2{XOy9OG5#zsIt+FU>f ztiz}4MW)W@Yj zrxG-QgK}I%gB=%ZiSv?h9Oa{H!s9wy41eEY3%8k5a$Astp`j&hHtIde7zLa2--Mq# zrG8U{YjH!S3n`AEO3OzMfK&wJrv*D}4y? zv}LM&)~(pI3c>Y?M|&mF-C9uQ*r3+2d+sHrNWgB8%d-lad#6B98c={T2`k4Hf9~Y$ zaX$x{h1mj=WJSR?4ONUw)4mEi9=mch_~n*BfgMx9<+z`&9^!qe*}{f<%5TvN(a`gr z<|26yI@D7;?Spe}i%J6QSMTo=vl-nFgaOh8B*^W3Gic^5!=&uhW}mC?ty&9s;j zizDF{AG#1h(CnqGJT913^&-_IcElC1GPXnpLt;G+_px=3B+MsSU#fX2XbF^jNBpmv z;S&c3*}wDYQPI)gYyYQ=dV<;&#;f;J!KqkIr-|6jW{T zwS55#AJT@|cf6V-u-JlL3~2i3n-3p8^de1^B7x*!Pw4sjY1J0ra^CO^7*r%^!t{_F zGA7aA%iJ+jA4B8TDe&N$;9!d!C&^9ULjF9y08bnw`ojSA^6p*#t}cgauNrq?_HnU6 zPZ{1Y)p>r6_r~mJtp|Kpc=fusL1w&P$I;+j>+S#EFQBLhh9hwj&k!uhzj;G1B__P= zxEc4&mjsHC8uL0K*mr7c7s3MR+2MH%%r1!VhZ+sE6Jzw9Z=qz`XmNYxIjJ(cc^Cy% zfA9&{b3yjZ(b8X4$-LQnn7Tc)v9Sc4MqL`+nV&vk+#JV0UqL*TBVwhp`S}geWD7Mc zbvBINc=7&c{BO8?sKhcI+fDx`KZX?V|JFIAKF06+sbN&v(l};5;wL8$AQ*-SUGqpR zyyc>rz;;CjaD*VxQDv2Z$E}V5RGY568VtL~5`KO=z%LkN`u&?4_#6z~P9waoE_N;! zHuEhwg$_fOe-|$=KyRMjkZD(xH1zG;th*~Kz3zPU^drA0h!~`*35)>gzi})*P)HI7 za7hkO1icENy&u6jY+$H~HIi!Dwd3FY`BEU~q`qvkqY2;{sD3<855vGn6AsmaME9mmYyju> zX*DP)zN~a~R~cX~ZEwxaLj9Qs;I}OZTe zVl1Ys&vJ7=Uec*+ZLjg!4*)e8}_Qpi3x zi%oG=6)lvNlrQJ|_cmNSE8C-*$ors*(}MnA3@6eu(+VwfwaeG5CYn$5jVw!zAB_wx ztXVc;K*a#y=&I5sf9m_Wr|=ZwAn_|(4(eBu#I`A5M2##fYaTG1W|_<(*kZ`A@PN&w zyl~^;+qbtzI`aQ-soI3Lx3f}muRQjbc=zo_a{NzSv5-N_m&o+3|Amm2!RXbl23nDw zV5qR`fS)`Au(x586;FA=Y3s2BkzILMT6#WfrJb2c$_KK--8z*P2TdQeU*}+SB=+=V zR96#%-J6?RJ3S3JJX{WyOAY<2MW`h?`L3nSQLdx*|I(50CubNC0LU*O*B{wsMn76S zG-NRm9#vs7|J@J^Oka8PreUI-&OhCZ&q0w6G{pC9fve`_A4i1B{#|mp-pw^%gO&M- zX^xt^+x2Aykz9@Z30?4l?SudMP!cO|bw~frxdPyV!uS`{N=k}(z&CFK$O+)yuNl<9 zJooEG7_=YRSn%<|EyaHP+Scz3@YYH0-^Y+kJZE!xq@Mih1kU99AKI<{BEM`ppYHe_ zLBJ>D;(KQgM%w42FRADH^(LGdT5NHkq*+wWmI%`7lmFLB=dhz?-4~A5fp7)h(hZ=^ z-32p6p2CZUaM1ugma|jYUoAoin~jK1iM#=KP&qg0U_V^!w`PDuo`5Gfe6A^ZQBmpV z&u2;uAd^Qm|7`?x_KRcN?`-Kp8$v7UHtQ2#{iIqF0zKq+SXo(t0J?+Y(*#im2M2li z;5_Yj0S}Se)R9tg$jPxuNp&>O>_Ng&c^!G)<#aDfYBC@|3KR(x=|R+~uOInx#^C1K z8X04Y7@xa2fA(V_nw-2&3({Dst+#QXM|j=k^#!@T{b$=62e_cYzYLR3#d*7J;2AN6 z=Xrf*{WoVv$IqcRpjaq^)d>h6;PE|8Hyau%RZ*@a4mzSC#p8NRYr-g7`RS8%S1%(H zyuS9m8Q+{^`Qo^q9lXDqWdvOv9qE}v!!Z{Z7oxw~i>M{wK`uDR`n~6PuBbX~?GpK# zIG7jD5&&8FU+rWWJqnKK&AOh%L+!I}nbyph>QyLbnBce&!x%u*NH+mmmfF#Mp%5M2 z<@L#7JCSZq^NMU+c{K=p!T-E&boMK>&5KWmD6h89ecm?A6k1{LlMleE1>*NHukBVb zr1aLFmb@KY`+@#Hwhl$|+K7^u>8Z9+bk$=gHUtCdyQYTk*jarJdh+OSdgu<`P0%-sEOD3c%!)yO2{tTJ&YKna^ zTHwO{&n8VLN$9w4ep;uOx;*G-#vmUVAMF*N;_GlVwr$0O`F?(KWJ^pa$^UQy8QHVc!s^F{{nM8hpdW1UnEz}QN+xuJBZXV& zHLqlBWF#KzQB-ta_u#e^8u`dcvNvcQ13Li{jrvb{giyg&NIA;1{z?TWqS8SRm%0*t z)i(;ieWMKEAsEpPtK#sxbnKiN1n1IkJyU<gnJoAMCQro*GEmbJ)`6&!#IT2&X2B(t^Y-@sF%>n|n4tamQI%C2sxQRq`$ zI*XOzP`@fQKZIEIZ@g;NY?YAIe*yF+a}9t%O~ymzV~_fEa=9{1e16V1mJYG-&j&sZ z*z*DR^W1H4GK_+?h7=4qzz0sW0>(F zCp-WbXUd2>Z+p#bZRhyZMZla&^zGY2c}fnK zCuiCoTjhDg3%J+ELp{0FWZ)5)Cixc>xV>bQC)k?n<{Ge@<0RN< z+20`Rm}dEzD#@vJnhYN-LNO#dQLEElO%vat_$rqviTfvv;Xe3*p24^k%`w1+3)j7gC_&$IgZoO7- z#lL0h!|5Sm9lqsUuu_laQ_?H3Fg*G_)t4_)#9W3E+cfc_l;fs?hc_^+b&$+a=4E7KG0SZWP8N}r_Yzwq4*J$KP&9BQDH?kj-RoVen|--wGD!G52hy<6%-WZ zoVNV<5pe{ji%oV%TMD|^JVhJ5T@&%;3obslSe^hW2?;4Vd00{dNhuw!-7h_%>j27% zZ!X7M8O6n{8D6o;&mKr;&W2HYOSCT1{MUlz6gsbuRfiqo>mIQx@Gl9rk2(#&Vwd&K zCWR~}L>`0)v6#Z_@MYKFT~(3o(a>zDvdn{P&$mtlyCEECgE#UgymK1JFi=$ss8rV z57whUk;1E@RQN+VMzExdNy!JZg$gKZo7)>JC@6;zW8Q;^`MrnhG`2UMpS&|0My}C; z!30ERU(c#-*L2fhJ9tC;XJcW1t*R1Ka*C~9-4VhBBeX@WIL69W?(X~e2GKE0r>Wu7 z(qx1NvE2L2-qT(>J1?%{bZqwu$N<4H!#`KDt3G?3TlkJ971)4|n= zij>jgc@H`#Cd^ohx&UiB-N%#*7_-W+rqq9$KgVGvx?>QxUd=1GhX!tLY zLRgpshr=nwlF-pTWeen;jwT%y{>whl17&Ik>oE3w^B=x~ZG$8Kbrh8FUrzavDG~C@ z;aQ*t@DY-(Tu05o7eoL$(bU`D{mmgE0l^qW`Io$ul+SVxKiKet4vP#6@fI3yZ`~&k z0X$%3m1csKc;8eZARh;Pj(|H-Vf!8{S^>f{YUC2$L#JkVeUv{}I zBez3oDf9jw5Ksea0z|stGBCMovc^6ej#EQGHVu?=yIudvH+_65>B9Zyl zjjlSkcz}leQ#*KsfP5$2cnDv1RUxnoa%sIkUgoYoc51`ly*hQjDCwsrwOD0r;ZHr8 zSNXXYlOvM*jc5#|ty?q|jo$gkZGSKmq_vQgRNpbHTQpYljq*_PDISRu=L_rbEr_INr-E z>(C7k+FIKP(xwr}NbXNva_MGFtr$defEq-w(KC%N)jhZ?PrIgI&HWOZ+pn>*VgE5! zq(JTB?9A84XLok!O~vMT(fBW#Hb_PW1_eF(sP+_OTi`C;1VLd04m%vhuviCE#fCF1 z0MP@kRYx6$dJpzIpg_a*&wheR1itx3V}~JQ!k+3tD6k4v>%3g;*H`kVahBQ4qaJ0C zF-9=+UX3k#p7k||RANk0E^C3@DC(oyOodb-GY_{9t{ zonsXjk7y3&i%nQ(FjYH}<(Hely#7bufl!no7~ug@A4SmXZj`>!Qi&pA1L6==BunaiqnxB@h$(>Ib^zO@;*(MnC|YmIkAudVF?q zQCZK`vYY}2p}`>>u$^stZu(yk&;9DCrcm%@%WYtE(?|`olXb9xsXAs&gMEK=2u5@= z===c_4;)Sbs?5*ogj~XWPEOe~p^ z+dI7SWqiIXKSnP4%zLWe9r1x7AtAxR#`ax17&ET0`V&3s8f<@G--md3m2MOuCDJ+{ ze@2}CfHpXBzj#?0=#Q9sb=wWKD`gg)IcRrwrsK>0>iA7g;(pBWCL`k57i zl7IuEnGf`&WCQkc_qV-tj@2r;oa$e@w84u9X`A5L#R+{GroJMml#*kT8qx6 zmMjM{#!L-t8nkHNEYN|>7A57o23ZA#5TJPWTORC}2norgBpaZXkpLNo{(ftqx22<_ z1L)E;2)OW46`g)+if#pQ{wAp5H}8YV8r0@yhfGzK?0vO~DAnFB zk&(@8y&6G>4Lq=wn9T2PD;!>~CnurNbBH|Z1B${q$ayg-Y47#M1~s+pN`#r1D z0V8XKwF%E7$BAssc^b&tz@cThr*nFExWQsuuY#Q15+;=|1X-$>n>cuu(mCJH?|Jp` z0HY12fy`hn_GID~)LNB6@vOlU)<{5QXP{mFgp<32kG4@*}x-kF7QAhrR`s*}W@I-SDh3fHbvjWa=4(U0)qsjjs-&V)g}-%3BJ zVP6f8K5t)Lh2C0oCb_%md@Fd_HXrEiu?g^=AxJ*O-3M93UD@tE$h~`XfY=5ykGnpy zLx!jYh|)^t_J?ayzztguwj1Ly7k$*x*=IL7E+s~|lLaV-<6~ay8w!mAAUPoa2giXQ zfrg|a+e7cRs`xkaHB|eVaN{jE+7Ipo|3!KAa33}mT)Lb~`9;)Rxhzf@SEq{_rlJ-b z-7r*7*5d|QzXh(?3PuVCog}{F+p{yk`_mwxw>N(*f_vl*`-o%`GM2_{{Kz=Ag{B+A zJ-z0m_}R$hNB5KqG@%IWH4`bkTTYq>U$xb(r}cjfD9o3*t>4VmsAiCrSh|Yv`HC#_6sNi(1xhnYa<2? z6eD9pBut;+%v3@GPM}#UDCFkpX}L)I&|IM2z1WKuQeb*eDxi(Y9fq&C#wsZsu`aFcxySjKRovrMB-O2#b(1|>QzNfLIpaZ4K~=NR?gr=zBZ zO_U^V!Ab+th5vO{jUA0Sgw;hOV`I$<_W@h(a{rbcEjv3MlNFM?`8$h4TGWm>QiPTDuY z;Se3MD|L1#{oB_s#LffC4%x+tNHWIbrryA>>iQ;>`a5j{yYR%de0D6=hwc&i9 zfle*j7s4!}$X9WQ+KqMpFVa@23t0ddkZ6nBI*FUmu^*;yWXr@g^pLtA*}fq5K<4Bp zvRfb5h2Z8CJV4|`4-(?MfzeX&*E*lAcL)`cT}@MZ`SZ7II7D%B>Ne((xbP#{iY2Eh zBn%X3?Ck)jGWdSCxM0|G0|t{xYW7ES#Y0+b6oWjq0!`#^%q$wpsQF|-5^e{;AvmJz z1o5_C&!2c3H5MO>+oqe6`|B^_t=}ayOZB`-L-7~`u?mOY%A$Ug2a;A6381{B4}hi7 z*cC| zfJO`Evr$pK4<>l8*T22)tk(?`cX7t#R#$KTpi*~O6&Lj5H~3CAr5vk~&m<=`A;tBb zh9S7RWXsufm$Yj-tP^-*APUwr5b`zZuJg+uey^LozHk-n9a~Pe1z8cNkAntdi`U6H zll+IP3IRYPSg_&3C5y7Zr>NnZR63o%l*XlmDXN!0dILfub3XVaupiEMey53+oCNv--Se!pp6$~#6aa8#goK3o{pv}(Q`LnK+X8X>q%it%Hm?NC|=OPL4+P9QXV1Y6MP_B1>q-jr+ z0;I$4ST1eAJ`=M63)(-=x^N81Ua9?xuD+AN=e0>rYLxic(X%z~($&9v2o+Nz(3Qut zsjTn~d-8EYa_jYRXD32`eOQcY{Ia}q(0 z@ju)0XLVB2ZmHF$5W9#ukbF?sdBScYNJtDp_X;Q;g2i4{H2tbsu(@xFP66$0L=kZx ztpc5jH&OE05uE*bQw^T(l9DJjCt)0jTG2lfDeuZ716D5nC(s`Rv@nO}+KtT-qMFIS zA{cn%-oHy*2?~=MU_c4T-G{%*$j+`J@{m~Jfg(uz_LYbKM9)XS$zQX`rT6T$|LSO0 zL>3yP`$<1@mXebnV)#%>IEB2DK^EQo#M=LYy6qGKvao6}`e0PjdgIEsYMrM2r(G$k zdhjl6JyItZ2tS#9sjZH3{t0Bax(`%;O9=*tz>D*r=m5}CQ&cB(iS_ttfm~Zz=HWSg z)CMn6Lh`Tge0Q&nMr0f##ppH27Ah1I!xQpYgRd!^Q6oBu{6$wy){opkkW$tit*uMk zY4&>FzCaBaj7e|T+MoQ!|9|~929tXQ`i^EpTLp%??0A9H%*=`GOGX~-vYe(e^WDbd zDiTvbic4ZJX*2!9fKsks{n1$R#wPKs3$;vk-V&yD_P$Z;Csz>>u~b~w9GzOOrhZrh z4(%Zbm4JUXzQU-zf8Td0nI+Xs33_^pqR^l8cz850kbi$auZWGcx-{aS4?DaK9HZuu zxP-sD+I2!`sHhbgOGgB`?w81*fLL)WRmD%!6Ay6Gm-fjnPP`#Ddb6OY2)?4m{}C8e-cO8fpx9{f)v3jrKY?<+n1lar(T40$o104)r?g zYs`Y)k9V8HagW9qUZS6pMx@%>opS& zov$sW#$XuGhP33VWezLPJ_XExL_4UbK zcy%hmxk#1BAL*>3*)@U<{Hj*!qK1c@Tpjt7w?X+AkD|Y)K(KfJfFp#tuIr3*WGqfb zI{3BlvJ=Q0Lf!$QtmKmDgT9}VeEr79IeM3(8d4KO1S^v|xaW7|!pi0Zkn-<*)J;R?DM* z)t_z{oWQ*E$wnN`_Y)uiaRQbG&*J<1(X@R}!35e+;2pD5=pb|i-JrIZf8ux-ZqY4( zTHaGgLOvXnVQB;D*=?k2{ZA9O`}*zv3id%t=mK6MKpE--r5&I}n)Nhq;ZJF2`}xHw z2&^2G2A#37BN&)Nd;{+O?KJOsTMAp;1tb}i_tl6%@_S*o_cBQ7BMuNboUqGnen!gy zHR4@=0>p3>s-hyjT)PabH`ae1th*YBre6!WTO&Lib!7HXm<{2O!~O1Gqr@77Hj!dh z^7<_Sr|bKBSQcrlE+!+7Q^R$9REL&`25%=HGKJkbX8oAuaoFZV#3=T@E#vk(4&xV1 zAuLOL1;T}V2HGOC+^>a&ho>9Ne_mYA-&Dhn6XqKOR}**>4sHq<#=fw-!wy~_Z+UuM zKu@GpkJ|_d_91q*J6-FsbjlCgXUQC2zXpnVReC1U(hF@t3NlOd%(E(r#RqGH+z;Jc zYkG~j3j-!~jW4MU);Gnp=?r z@67|FB(A4GGM}eVL1_g3XN3yhqv$H}@4# z+)Z_IU>i8avu0L;WO#VtksO$nI1d&wett^xLPk}P(H$ zcWNko;H9m5<{v&~^2I=(@ZKW|u>7JV*L&xJ17;Y|b*JQL+Bs~MHMKKU`T=OYi|tY} zGn18*xcHMDFyN%fxw!U%ALeEgLJ`yL_*l>j!wL=4Tfm@=WABiugMj#=rLhs~jFgiS znoipRGbR0?n$dWPQGQ$ABxaK2Ar3pLVDP3I?h!=8MN?C-6*(8-jPyN4Kng|rEa9OJ zqgL64i#@GZAe+L|Sty3U=SCPq2wmJm`-{RXk4p_Dm)(MJvCxM0tB2jbnp-c-k1ed< zU4=5xY(+aK__C){pRunW9Wt`xXK69^NAd4Tc_MFpd2pWi=*vApIV?)lCFdoOtG2o* z=ifh1{Qmt@czBf!gn=O=O&{kNOs}A2E*hCP<#j*yROKy3ch{hdMJ=1yHiz(iOKlF+ z<>aViB8c;CxZ3$iu6P;T6*@bJv*cVJEJClWSD<9RncMm~`6yS_#_$%`X%phR**zCG zfB&Kd$vZzj7~py^aZTOwXMZJxMahFRK>O1zrq|t*a;}yZ`Tz$^RsP}5`R;CKe zJ_f+pw)}^BeGP>rS!BpTnI3f)|28)*cf^7c3b-Xr7o_5Pu7-yYQ$ss@_NVk*U!IS* zxvwsG=W2lnTX;xvjqMs)!oI7p( z*eekp)@X>rF-mgUeaZSX`UGMzHmbq?Jr-$dZ{Ge%po^$CM>&zsDMxeQ}>=&&TT!&mEKK)ubQcvTSFhYqVnHRYfoyl5wAgoy)U4rD+!E z@Z|br-C0mZhARX0al+3X%#^&qh=T9UVqgkgWd{wNZ3ispA1?>kfZ|uV1&k$=%C^%f zP;V5m=dMud0ETBZDld%sJ2l^-6I2Sv2Kr{`vn4-}po$Gpyt5BM6u}Y^VS;{eADbfY z{ZaWc0qII!a)MC|LHA(g?l6O7d{rJDv~^G??nHayV?;6HYr!EeiQ>4@p`uMTo~b=Z zQv*)doy7Hs3kAqlZ|N!Ngh5GAS9qMvN8Rf##hv9ao?}9wQcL2z z&C!wZJsGE!@XW}_1($`(eS?`!m?_~t#O*y8^UEsvEW8^wM^@{}=w3jb43r3hxpr@- z#Uc?usC-N!|NVi_8MuoWe-`-C%VfaaQEKV4MOM`M)OS9frcd``bQES)qAe-stMu7y znc4sl>qEr>Gd)V4on-bg+Fo)@IIAa{Ho?M20yDU^;LS&A;=$jeO=GZ}L;%$p1hT$NE~xnzfTd=qc--S#}w!iPo1w#{6bXZ~YUEP}$uu;9a`^ zejvjC?3zJUk);{$ZOf3xDLuE(Mmt?mnBayr3#;ry;(UL@?5~X-XjZgBx%z=#pLs`bJ4Rx|V&CpVUd+=&$Y~No)6??P zMIF-ftn+2eGb~C`CyoW@Oa0RPqMn|b&@7A>O zSw*@|$v`X@IDB4)&W54Rz)*TH9uxx#tHOMY)N|m+_YbNTQJ=fuf5;=>g-I0U=bOwQ zU`WU)Dc$XUp>lvK>6fRye`UCpb|7*x{UUT$>EStjoOF10M_T06U0{)q1X-v#N_^0bCDKGC+k8{#I&Wens8rwzrd(SALVpxdS zxOWa9L*PLVzOb^hjRU3^IRKM1iDYu6dW6M|aYkALVvXS~irW{Y_v&UuDn0gFh8vTb znx4L&vAgvMhhErALIIc#A5yO5yNc%vskbufJQU2#dVyQD$j)S(s45~L#*FXPQ6GOV z@ec2ifTw`==qv}lAN|u!&B9NgM<^1pnW^es%{8>sr)Hf-^@#4)NCCjzY;?{1o-KlLtUR1peQdsbrkkx0aj_ zfj=oQ2>m2pfIZLr`~KcJBc;;ofIUG!&bdc*3m_zf-w5fW7Bv+M_y$A~_#6EDuN>uk z1kfuT^Eb=@O9^PCfG;H#JXUZ}md=5Tt$=u|w}0r5qkd5F+y^>v$p8l&!e0I;06|Kp z%ZE%Y^9@gXQ2jdP9z9)FN`$|cmLYFGGBM$FUhclwmi@uA#s}@szVyIkUwsh_J6vDi zY=s;gSP~MVX4-RdldI<#%~-fmi_N8qaOg+TO3UeoXied)=?% zOzWc(`OgKLQ|evcdY@ep8>_j?`gD^up78}=jHnz}`8Zy;3UCc(W~hCMwR27<8VvS7 zew^fZ(*^q$UpAgv;9%IEVZOO0;72>hz@Djf?{Eb$rjAKNVuY|<0^xRNWBmFxH;F#L(kqC!EyZ;^Tt zB|vNPtk}9B3E04YK&Te=1_KfBDZmBeG|k(HKYPZgRW8Y=>F9WL0fXwQs&ZIK6JBBP zcNb`?Aj+|Oam1$|!a>`7xepz~;jJf&dLhI-5t}dWA@+2(3q{J5Vx#4}ztgrW+W-~g3OmvKmMK>;z zbIf*j_8*>L&Z|((8lq~wuqd*XoL%N)_rBhko|bj-YcbrS!;J{KFhohb?KGfwt+GzE z5BW-&KHGHPH>%sga`!J0E zIXwplj*J?Yy@fzRc0KKH9K({0xv0|3MJzJlFupl%0?{ono~Ji^>7AGc12r@!+3S2A zC{}@SNg2%IOwh%>!?@n=)AgRqmuMEH7w3YLB?ieU%f7STWxV&32sBGPB7#V74YxeM z-V4AbP2pH_Ruy{=MX1-Tszu(QI5sYE4_xp&N`pWI;<=$=ksKjLEpeMp-dLmpIJDZn z{{Ae`W{WVhEyCAm{0r3>0o4Hb{S~N5F~JRg5j-9g0a8BA^Bvt6up{ZG^z;^7+OJ=~ zmbzE2_##J@Ld3{uIaehcw6B2;h7b)$Q>olt4Q%tmlBv$G+IeL#ntjL#t_Jv}`)SL$2~K(^|3E8T82yX9fM2vR&e zHWro^Asl@ok0pa)VU|KQ7CUyU2=qzyg3@6(SGW)11V>UVp$O3+;<~&yiZj$m#+Bw; z;J)H>|MWX5Bmqcw`#dTP8S-jh<~tY=+ms!D<6I2O>-+o0(9~RFIO4${iT4zbQF$&p zL4BG8qi8%nA;E&^OR)_`)laLFTovlg+TLWLg(OPUaEM~#>ABVyq@(~4&8{PE^gZKjI>3=h>2%i4(?Hk^^iYD}q z_&B#&KXirE#Jz49OHFfhB@M3c-$uH01UJ$)wF)L(XAKg*I9ToH zj_(DB7&fc=QW|V%#LG>7-1Jg5H zdT@RF0j-EZ4@10ikabQA9hsm#>*Ej4tLT-BNtP=E(h zfxE}ru}?{Lh67*3DMV4vJ1a8owZE8z#THf!?zU9BIl3wI$GEy?4&WKw%6IAH?}-0W zgGffiiqB8Dvid_*)Jni%Z6eWmCM=@tyGev)MOJhqf4k5d{bMj}ST4SFLY{%Q`X+e| zxJdXrA(Dz9@IzhK)7Ertqnla1_NTzYu9Mmzmv!H#kmL_FDOC9QtIwYOoS9s=1<{?V z#XbOxRVlm3zUZ}{mZodC)2=sHLMca45C3r-hA(x$@O?)5m2-)>Yq2Oxjf2+%xG}_B z7J8d=#mm;R<3+{PyAU4Q&nJ+Dq01vk^?8V1ww)JpjPygyP+^F+mKHk;3y0g`-HYRg ziv;%?@{!{$ikz{uY$`+B*E-cz)yB$R?jE9`GhvkP7ELuN-f4owmKZjz52Cuk?cxpM z7T_zZoEn9PwY61VUT&Y>(VS^T*&u3I{=^tEXH$;*@I~=UPEP1s!v)=2Tl@!~>5Ief zu0H>)MUVdXi+y*DNXcZdC^A7b{QL*?sKsOHNxiTc?q)F)y%HX z6Tz&kM2Pdg`&>ZR8L3%{5*hd5i-fktb#(MZszh6=M>x^|rE%UlhCD!26Ofp_?efem zBjRiJ^M7Ej?ega%nh6ZQzwzxLw{clbbu|yjW|ox=FWUhUpB^H*P_1-v16zX@s(R;Z z-5|J$G>@Ht&BcZqzIvN>P=*4Ww*+V8ni9@8|G#fY4ky*+uRtUM#lKleRdva!RxbVE zD-gfKW7I;lB)3=;?am6D8xF%yNFNq~BFySX1L{d>DckWKhUCoDE^Pk|y)6X|4FXgQ zB2*r755}0o3b0J@nZa)qL*h_w{cY)KV*$JPaBnQjR|mRGGajLSUommzPZ}6dtlY8F z%M_7LeuN0X(sx}Db|V3(_dlUdpJ42KXgXc&?7g%0G`(b0R$ch|ob`KS9BX1l1lI0++q^tG5aH#3zkj%` z#$^@d<%cIH$QR-s!eINnQ_nQNieT@F-vW5@uVmt&DFQYD?r#XXaSD#Qaya`ZF}EP; zR`(;rw_p!)7M(#n)!R$U#vUUHy-tV)hxk_(=|K<|z|Wg-h>LjF`%2#J+--rn{XYXl%|;&OAP@cC0IfQdziC1?OkkM8~(Octnhcc0+mZt=Sv$~}4$G!CgY zk}-03R~HvIbvW!nCr+aJoB=%IU(h=vcs8|382o_00=2m>07tULr|at*H?`;Iw_8-I zuD_%|C0!$6cy(Ua%2rkH8N%n>9}6DgptnE{t_r*tAZ1dV7>ZS8shuRww0wziso2sc z$jZVJ8C`j+|GpOoNG1>jx|>n_?MFevQ}V-ycTD47)OB)gYu%}+L{*d@l$FIa`7Lnw zHNs6B#2)W@``=g!RJiPw85x}|DB+n3`};%v(62R>1`OT#>zkX^7Z+EH5sO#Hq1!)kKZbV=R5di}UmF|!bkWz*mB&0*Sq(nlbQ9@$q?naOyq>-U>fNu}R z?fpE@`^EbBhqYY8c+Nii?0sFo+A~cBm=&pP3K)Lga67w_YhR7n>WrN6g(Z(aEjI=X zfJV*B-SQ=Rv~ZmoFMd%4#c)D`Bi?RFVDrdz0M)tRR95zE-5}e;wJ%x@P=z%ITg=qb zweC+iJO#zK;r*eReL~Q={eI3^^UzO%m;Jw~(e2fiw0CvoW{$ootMi6s zV5&uPSZ^v4p})3O!E@bI^wZAjbV*_1yZ7I@?;;)d3U%T#a@s*c2E!;xTdJ<37-B=T z0)v=nDXrSqY!LLAFJHU}O0mMLK$pY6%St?#($N9R{iLd@M}FyZt7CgAbG5?YENKIt zPNTAd*I$}*++-S7$Yn~D??pzYcoGF7R6wcs8%M#DkvU}}RTW$y6EZ=F5drWlJPa=%ma49+_L0|{_4I|Z!BqAX>F`I1Nj8uu%C zV1n3u7*hLFO*Ub3omKJ#YBDwV@pFW`oj++krVl08>opJ<^^F=E>j3I1qPFH7=o`De zH4@nI3dXwk`cjsvn|=}+ANzfw5q(HZJ3i9sqlCJe0+(ON_nQu+lX$i1mH_NW;_@=a z%5ZdQdEG96@gnIVyLduGsFRXwqJ9{|tjg6ID105H-G6#2vhW)6P znTYj_PAQv+wjfDK$z6$vyY~}X1dO^~R!Y6f7vRaHU|^_65hEmGvY}2~1Likt%0-mK z#8duN2Fk%H8tZPg?yptC0m|SvjGMrzGy`d)%p2D1TEAfAK4{Wn%M8=yDr2jV}By@H6 zGdDM@Z8Y5^0G`aP)$`rrs(TGUB&Yr(A5fuy>(2HC7}fe3hbN*)OMXp3S2rkCK7{to z%^N$CyN+AuP@xu8p*NYS{#iJ##v<^B4VaXSjmLi$1pNc${t;{olUvJ%Y>ehGY**@I52qmXTaTKex{fXJ3)>(&dNDI`- zqoF#wr}Lc_R{j@B?!)9Q{#ilp>A}-a3{uZY^W4V!HSE+Q>T+oqdArL)OA-?L^K;P& zT`HG7Ig=j|ISUHL*2=KVd$Bfv74>dwo*yxN&5-tT^k`}7qbsjn8k#*}ATyB%o=Ik6f1{Keq!Opie z!q1Ir&wHLAwojq)qrY7`Bzh#V#11b^$5htub^4z=ftzh*-qb8(r zi16kWdb+&~%o!mDC0?P*ernc70%_=EAB21^IS>QRo_ucvnnY|DTW%l0LQ`SnWtb4~A)`h02W2wzlm-;q7W zAi9(Tu%%xB4!VE<6}I31tSd;RL1l@Cm7aLJ4Bk2S*G7v*fqt!eok$;)u}Wdj>o_{X zYVBa!@6af0*FnrPaO(Uoa0=%4vrzf#3sJ;meBcUyeXr6OwxH8|DLORju*@RcGIFJK{J^ zUsIUN)nIlL$J^mCpopUG(8CAL5Gd_+n1WFR8etpUfPkHIex4#e{=uSd(IF@;eq;OQ zpFG#8q8H4rRef5i-(eOf?+mQ^u?|NGZO0^fzjQ!tN>B8<*KMRd^Xg}zM{tt!5OH^pffMt<57Omm} zw+TZvl_ zWu>WE;}q2d;1Y4v4XgE#OaNN3R!>iV+}Lo?xG?IlF{AFKqkJ*bv!M3-6Qrt?yD4a@ z5Lqf55TFD}(j_7)7+A~&>&@>1@R0#b{;{`bFen={9;9T2-^x%yAQkw0{Ja3F(t8tZ zU*;_%x$It<)Ka&-YdkzeCC}q0tOCZZceuo_=O>-XR6YK!!QL%@u5Nqddk_!6T5kwK>E!ta>u=O5Z^q!p_mA&WpUv{IetncXShzJ&;T=a< z%KX%Ew-}_*$`luIBsKIEKZ~Y^t8yPsf)?%fP)7vASlPTiS9{+~0?-s2*_&oePycQ3 z@srTh{Uke?pELFwODvki?V5k;R?gWSdE3tWhs{4E*|DD~4H|2zLKqI@Hq&6%r>{To zGmzVawEt@PdzF0id4pca#0Jjz9wc=u_RM$LT)oz1OkFhp$6vG4(A`*?Cf8DJr8NoA)nwb*b{LzPnUiIP z4UaExNdgt;KxJc=JSboQ_nSJon%+S;M0(0*`ram~0t zUw%R_9azf>eQ>Z&z0@tEH(a3~3Z-3jMv1eW_B)%_JCN%SmZs~kh@@Xd^Mbz$f|{DZ zAp#-Zf0ZVKP?)vpxrP^$gUW%1r2HmK`m z>L8fRyviMGr;||Ypc!xXD?X&4W@#R#(E2g-alqB3#%DncK$tAHO^?+;@hk{Cx9kQ1+k4zp*f!m|U{n zn$O+e7mXv}u(z|*douVVDTyI=h44j(`yo9ny0a+tK$9aCjH|*I{A?^Vud)dtH#g%W zr71~*(4!bcfU*O9GyDNRS9KHCxj7Gs;}JmWLXi7e@)50@=7E6a!Tu76E@>V-u*Amp zRY;$`f(B^z@lRtocMmV@3O_ZU-~?TZ# z2ZRJ4pvzLZS{SRi_-p;^)MSGf=Z!F55>cg^{F~Ifeu7oYBgXgwNs)b&;yyfv@DLd- zQ|mDwMlwa_(_n)dK?^_}m-HRfWo9h}KMRwqe@ zXkCR0Bdxz4-*at};_lH|KQ_=$Pz2f<&rEC#S#^^-^}EN5mba%lS=JJlMNK6G~Yz(|1 z+%E{jp9{i$rdjL9zUhXX<2c72(YFmI#1y3Vm zZEWyTMK&SD>Gbfe&U~mf!7O9-p6mUKifB$wdJmG>N6N7O0L7=^~L7Jm`4uiH;hRIk2Q?Sv0hqx_#_odNLX>ay`pn@l0aHC zt!zSP$e!GW=)I3NlgO1xNV!23i&SuO#WR(vGs_$Ug-| z?|D!!Srp1YBt6eZeXO|1ee-lX{Hhab;h`61`KIFRf_^O`*rzfJAw1NW_n;*2WPOFm z8jy=0Js^mqiiIm~xq4+Y;rkkD{!a!5mNdTZwxvLmy{Q+i9oLn^i z;YGGX`D-_nDc>GqIG8nPzLW7{8A4fI)4za(y0s_e-kkL2ypxB)xY4`;^@+xK&A7C! zaNJ$S%)fePOl@~CUFu%Z`rd0eG`nW|;;r`&;ohz0=rQ<-!QpB$3<0VZ6HvKlTrgi` z2Xf!i(aEj5$H-{9nn=pjQ`r1jrr*Qk*5oGQ#-k*l$RL;$ZCtp`BDe!V%$K;WidD|v zz=@juy5VqQTNoa0yQenJ5=R~>ZIRr5LBY4ZEo+0?^2^Yj1+!^g(u*N$^9LCO1cHxk zvw))Plltuol2n4KNBW2SoOC^j>arTdR_m{g1tgJ?<9iH0#CU}zwUWsectLu*KUa+3 zZ|d1|!eN1a__b6acVN~|VH7|m&|N^3@0A%ae|5G>o;R*CNgfk7r>?88{mMmy<3z^f>g^#ZsgrYzs*M0x zTe97g)Iw46z9lX4pE)L;WRZRqHdTMQuj1)?Y&-{-ysbXZwyvU+1r_h7b%gI8q@3a; z0ui26gd(>0Mqe$u%b=7d#f?w0+i()!+1UYk6c;__Y()T$iwOa2;>>JIu+NoW z|HD2Bh!w-*E3S%(y%~8xp=DfL#dzG@MLaWrY=rNB&G4L-k+sD7_`{zC#I2&XW=Z_# zLYER)jUwTfMJxj7qw?S^ai-TqZf^5s_VK0QEOf*6arkHTjsP5heH~AMh&)N#5zM|4 z+B#IXJiR`cu9bEo2SGSybw%ExpA}1{CiW|=NxBmyHbgbIUuBV=*JgOBmHrQYntLRJ zCP}@RT2~d`Y;SMR&zq)5#m&vlfmk%Jq%6gx1b>8RiT&3*p*523Sx#BpQcn$=CHgBu zZ$iaRKWgn(_RW}|>UP=Hb$=*x&uZLwJQgO)`~D0>h>V`L5|m<;yl+u*Q5H_8Jbuc|82NVrqV- z1-z_EChkvNY9B|P;|hc4sr)DSd#U*D%;Woa{l`nWl>&^(u3J|zCNbg8UEVuhj2v7o z*U1M2D@9XSq0GSSUwfdL09svSMiI*H8Zf;zb(mC8JKt}*uxU~}=@v@#B zo7u38Zp}k=*O&}EkTdxJi6|G>Rd$PHiwhDOvd6@ZBL@<;M#sl~LtkFY-RXeI>y>y< zlp{RDq~1z6fc3>o*5@Gw?BKBRn0Ry7kahKRVj!p>+iSKo`uic5CK>!OSUj{5N8j6T zdQd4DBu;gVu&zNB(zi+{rN#~p4uV!c8+x%nGXk~2EBzQ&J~&2_Mpj9ORcK&9JQia~ zF+<7NqYMZ31?YNgc+0bgFGcQ0iBKAX$xiBG#swF2?g|}Q(Je83=-^N?X60D<&BDS0 zv`WAs`$>Zf3?ewACD`A8-AQ~s4;%v}LT9eG>Ijqx^zZ%1nVvp=Op(Bpl2LauIvUPq zm)NDIsfkAmSY7#0x;ks7n>Wkez4Mad6u%FQ(0A{)78$p!!B0=@WW&Id79Da~8*3Wh z>QED}=6|?8M6KR4(!fGdyTBW02*HA7jSN6v?k9?FsBoQMJy@Q3D$T=&KTtb8gTwyqavl|Ts7?hArK`f_k003vXa@!89 zbU*bH|4XZzGEjN(M?QsQUnWAOew7K<)kEH6Vzh#sdcx&7Lv2Z)LJUP7@E(~V7((Zc zR-l)`I`(}3A;#MrgAkhLknSGx3EY$Fc~MoI#z3=NS{zm}u;sTEm8=i%&q1AR@m9hi z zk>1`n;o*H>zkUIdg2KW=5-AiP56|=;X~6Z%Cu zaZP%kO?!4JM{Q_A+MQc8MFhp{b4#9CiXWr;bw8*PJnVi1a6NQe&CZp}#_mi|)=+;N z^w!gNb83;_=c2{vxXpTM!X!*aeI;^9JAs6G_cQ1_xx};5reE4EN;r-H)oIQCOb-ljk;wClNT-V-v3-v>?C!1!o?N1chxwag3Iq-Jdj$V*!XD4>uzS!>ga1$ zrR)i;(i@`^WQ@OwY)h^VvMxexHxrqJEw1pZ0H{^=^?|0p5`1=tIA8N|qZVug2aO+jYjZuH%FgH<1pB8@k@}=6xODDxbaT&R}!FfS+ z3=BXHa1QQDo{@%FT3MOJ3TZrqgoK2EcN&~+6(+uk>y}$kpkMYpNu<+HUJmXpm-&8o z?1}bF>fW+oTLIyYzWb?_oq1zk6=Ll1WL3#jZKAw>y)wPz%zW@ivmI0(5KTVM(Fkd~ zUH{zEnvFE_Ao%j_DSxF#$~tTY{?XI0MIin;fC219YQr_;7 zU!JHWBCK1Fz1xS_)k9QYvc`hIQQ#Xt=&70>tE1Zz#<6w5aWa3a>NxE z7q70aeyfBwlr|Yz*x0n#RO{K;+ZT`jn70lcMe;plQTeGSXuNv;dVM^gJj0@I@xTXG zmb6WI-dY~4B|Ada>Z0wGCsGjy?`>LY=otxwLgutZflL*r=TeE0Gu#`frA1i?rO!9^ zQ~X9l`1w1&ttA5#ijM=QvW4xr2Je@WffDB}S}>P1Dk+!}! zd2|d%Jx~N3Xjs->EZi-txHwsCDDWLUJe|1nF4seUZ@9wKT#v1WFTzx-&IWqh+vAxt zv;~+m^)IRs{B|dvAIRA}W|N!22JOBvG3JmdUoR&kUsk)XfI8OX^r(ov^z_Rv1JCI& zLb^;q!s=ObHMBur98XOsYNvK$ES864*65~JFH2Eu-Oa#_>x6rSv^cUi9Nn19?Lb_w z2yM{WpXb|+AmZS?>Nuhw&c54;>s@bCX}vt(r3nDXR6MucF(aTTPP}>JG(H%uzUjSrP$~8H z<_hBr14mLA%@ZvxEj2YtxiuzYKjWbs)y$Ozxd5POZZiVY%Hrl4wD}=dETT@r@<-Ub zwKk8UH;fhdBhowzR1oVHl<&L7hoTTO=Db};+^dcEG`#p2E#TgccpWO+Dam}{2o*9w zeymQr4!Qx|kSHG?riB5CcIA}TUMHUWJeA()`6NRi99p}rWcw7O0Lmv|kH`J&1Do6y zBvWPl)nx69=<+Z5j=sIDZ>?-SV z|64#~qvJw7`{L^Ttr6rCb#r%3vNW+F*v&;d(Xa`<(nJ5y}j~`?wGWKrk#ZvQ_St1J^3>u>b^EB%gYKZ zoN@C-1ySWWFZOoH-a(25x1<%sTj(Wg%@3=D1^zFY!WACPEY3+FA4 zSIb;sRRCIVU^G(FEPd^R2yAPtxJ&{1kY@nr%?5ba(3?KAfV9i+-E8YMAb9#_d`9Mh zitl3g;dEx-Py$+~=KtZ{ZrZW z_(`1;I0Nde?qh<=347ygTS7Xa@>Mkg-*qG6`2@wt9k}l%nsq2u-1t6k?5BvZGusb* zi{o~f^yK4QMFcBmL9Z9akot+I+N286uk9xdcsBlx%n^fSv}h%xYvHmg1{Wvh7>2#-)Y{7R$2T$ zZ&y#%^eA<=8s zdNyXst-hs1b29pd{p4W?UPo|%Iuj@W*?D3)B_&}KCMXnF02c*Umr#!V)k&AN4X4{O zGS|rYfav|Q083~l*qe18e-#p1zhT9DH&MNcmzGs~8pXa1@t*v+GN6yNKbhO{-60TM z;|~c;J8YLK)Jfz}>j@5RoQpqT|0(pp%s$Nl<2Kztbi3 zwe!|&BT!aZ1%@WGo7XCg%qVK|U<-Ky`vj5xbj!pPVp|L?d`QZ#R>LbIwuamJ_O{ItK3{}gF40UT~yrG@HreyLiBkC4$g zOJ0lWgvn#6o|Y$$3Ag^{Z&CiQ-+#;Y`8Ih+U)OXn=3uyC3Yv7^Khxu@)#^ zeu6A=3IXQ;I{jN9f<`ZV4L#}r5>omjTf+5&O?%hPr@Z$RPN(Z*k(nEQ7 z;oGwYDXGkhwtK^jrxX(homdEx-3n447ceFlfP%UI|51}=w0@wufiH-{ybH|6NFAY)oW{iGeiF?97PTp6~q{K7`gW$Sh-qIn)=O`3nLjfRPuTOc1 z2rC^?Z?Bd2LMu>H;Vilgqf&G8iLlEO12r?u7qy+#Y-y;D{Pd2pUoJyyj)xsM$7}Wl zts2UPR@B7F83{%%jGrbPeOKb#Ymfya)A;~SQZ`*Wh_cFnvv%eup@^JmAV7J4;n{P^eC3vN$%^{U8O=Y|0{n3w-n-Y?t%pq;h!DXLMCnmHN6!Ke z*Fg~11=<&{BxQEt=Wbe8)c=K>yyj_~HV2`XNh6TE(N6%G@JB=ahtU4gbmcmM+%ANE zZBjl6kzl{|5NW;7|CJ~uHG#cdfjwWU6WOB|vELphpWLoN`i6m5IR8*@(A<-U~n!6=GU3m4eZ?a?->3$FNrkFPt)8YkUb-EJ+7(3vHdd8^Mq0{qQ!fQfqwkN8>-Cz@83XFTM+& z2lP&g4q}y^^V$FQSHJ-KJvh_U z$G_ffe}EyQUUT3I?v08;JE~&p2txZcaD@axE0J#TpmZtWoL5+nPv>ql3@a5%2Bm;DefX?(sM+kbd`actvhN`*Z-1AozF&9tQ)| zHYX>7BG#ikuFERr@n$@`&Kv}&V?WZf5j+7w8hQ8s&9wIfdR-r%U7}~Gp<$HryY@_O9s7s*rQWqHb=|5g6{Hfpb zgxNs%oNKPnFi7*LNffpB#PN46Z1eb~KON$L*yv8v-ImD$mR<-YAFzB=3l^{MXnf$!>D5S4 zq*6XbeB>|RR=)Qfb(dg+6k8e?8La0LZpiiu&B~ki?g57i>n-U>dWsUduP(dmni?7^ zk2WdaLn~c1csd`cK1I~HldSrg7#cDIkp%ezpIV1`Amd{{lTx8IUpSmJ&tnS0-fBc zgBPcPa-eBlzD{P#0u#GkNX$7f56WVD6ITweXdtMYo1(LmJjy!(Z(ufK+F~n+6fB}l z%~#S?>IEA)jU0yGlLc`wz%$>I`6M`rCm$Xjf)j*7qfHoe`^pxC+#Hd-8TA|?fTZ(J-(3AK4i_2>syQqJ)bVpqymKr+DYPl zDR4~brvq>3O+kcL&bQIg(iclxYL+<-DqhhvX!A5_WS#JPuxNtA;4SW67rx(n>1JZA zCYJ+fg4fr9N?a2#m3T)|C|f^!%VW}iR)DMN2)uR!rc$}yxacTxb^TMkRXn1ki1-Sh z!yt4oU8qOsf-jsFkS8xVTwJ%oTjOY9prMQ#NFvNkkb#$i{SvQ)IZn^^ViXgFzL92? zk0517J{fOFk5>?Ieki(q{8#4S0hAA#c_A|hgH)-1){R5)>be7uF@A#m_AeE&MWHk3U6TKt9vVmx62FWl3Jjch1&IRa_tp<@4G$T_F` zbTP7}S(5BbkjcooAZX-o!Gxmc3vpbHzly*64Xu3py>$K>IK@l8YWVFf2bzWFt9v3@ zJ!AsqnS)oMd~vb{l)d_eY?nWgkbkYux=P-O@)APfw{W?Pf@*5}VRczAd~GSszKHvDebrToP|T0) zAJ@O|c>g8U0-YUDaTGg-z7Fe;f@Sa*BkGtE*l!{}J)I*TyxAFm8zT`@NCLLTHw*|e zVhc&fq+@+XhPt_!g9`OU=xBdOc?urW?k3mU*e4z+rGk_EOM#cT+PH{Zd(=<*U*}>D z>>v;BJ6Viu7k;p{Y1^pSSgMN`M}O>o`671L zh9*F$qPyPzf%kDy5sTiCX^E;@C}zRX@Gyg@J!YBb#Kc5kApS`%uwPA$gR^vALZPQS z1s<;@_7EO$jjJ22E6~O@fM%}8jr7}JU?9T+#&#J8IOwfco7&Fr(^E)@ z@0hwghI~=(s(xZ$;3R<_?Zy}hhP@%h%Se_{c8%{Eh08>4t~T49UxZ_8uiT>&;F%Gu zxaop@u>*%O?Jl{A3CFWsCpux)m9_W*H|(9b{FKqT5vjWLCYyT-9xS;0XK4A>;D^Ux zLJkcLtqkXV%**>cW>p@dh$A^XGz3!UftLuZ(>!`rB3xXn4J1;*K|$liz$)MUq4sL5WRyJZL)U z#(|y#Kd6nCzn=Kl2^0;mEwuuA}tRP-_&7yxmwv8SL)?}WA??(XhYE#V+c zU|gT=8F(;pSW3-59K&wNJ^-D2U*4^c)qkv8rVbb7rUJ8`$mqF+5dR;v%XSvxCn^+( z9gKhEx(c_MO@QD9E$&iYy!FHBrwC|Mkh*DwQ$}{YN|Y0Xw#<74>gc?sy7bc3EDOLY z7~kvx&Y*wov8|3;9l~i=SK6PH8S>Za`cv(Vj!n2{wMqfNDp9&?bv2cTjZ-jrf^{41 z*rHJa_gKb>N8I#ocSs|NpYT0>`gCk;tP+a!wGL{kALiS2scvg)dtR1UCTaw45DK|t zA*mYr2{GN%xL}6`s7E)&%FES)zZ!7^+T_LEADqn_m)tZ-90^nJy{3H2{oRnp`fuTo z27xa*?mNr0PkkEME~wUGoFxw;86ZRXZ_J<$7m1b*A(%Vr7l6R=-8=l4kD2pSZxm1j zuXJD3J%Y^4f(qX1Ym}@#S`7cEQvr{}Ti`OA2v4Z}I`eIx)i(${F|SKW81rpETp8a& zjsm_rgHB?_b}xo*f7^V-dF(TpG$xghKzUM3OiXn2>u?BeOmrHkC9yC# z=RT>=u7#DA&X*S-oScmZZd3rI4|}{kb+go(<$~;#=4ep<*YWXdXpHmw+*}>rTM5^g zu7%sqKAcDURh0kMC;c-keo3?68sIgC$iP?^fwBTfYm7L>a1@R=q^7;V+xgmuA&$1% z?f~N|6!-XlAd+0BCGjb`O$7yn2(T{2%gGWRJYJV{kDyFnEQ9gIDv~!TkAiQ`w@= zG4Ay!#a1awtfNfwNbfXgkR6au_8WvHFkJxR>)#$Tfc}641hh3cuw#zKa+!ieO6tLt z4$KFFAsozjmVz7aetts!#xMB~BU&m5f*uBG$ULRt2AI-|(RL%E#FRo(LF-frDH$|E zM)&^}U;v^I_!g;xnQ*I(Y=6nZ6ZHaf0hW^OD|UvDPPkprm!@6iypA@U{Yw>|kRl7< zBfAP6;B+m>$!TwE!`rRU0jE~7{)(P@cz7xD&Buy^a(j9VfU0HAL-*@_Sd*RDSj}D+ zL%T*n*H83U*U!!#075fCIE8eb;>k|2{j_}&JKN~B$*_XYs%tgWe;HIBwIF*9Qo+Ui zj>phLrVflf6_#&m?DPuL)C?Q5`p{nywFOyM(0;V3_Wb$0F0|`1UDYE^O%m4O$r|}N zhy~qGF(6KkwuZ*tNB!cXia*s5S7P76%uY~4Gsjtq$HJ1ZP>!-*0O|jc&%*P@ z&VS|>K(1VQVk>gSecVG&j;8Aq-e2Sc{8Z2E|L4QvZkKVOd(}ujN8g_^aN@xaYinC> zJ9QFXI=L=J(JB1KU(1T1u5YCPIxV6QtlX~>z7 zVgWPjT=GQCp83duyHuG3IpfE9yIFlV>wTNk=?`4EyF^U>Odv+f)6uOf;fi^}28+-l zh-I3kNKI`E4bUXEmT=v_!-DZ&^e#~mpm(RJBT%7-=m!jp)oMf&zF#p_v0cAjK*JTw zwx8Zg97*QT{hPJUZNCW;vG|2nHnz36lz?sit@r_LQT85>xQo1tM(sIAV*=8U-e+yE=sKobCNT%jwKV=i<&gXh>hEpoWBV0Y8 z<5%CfC)Z>CH6pJ&n%pN7@Cyc&I@;P7!aRTl+kR8<_*iB?s`ciWm*2gb@TM_3HcLKg zSNtXuNeJHR$0OuvU6HYej}B#r4tS~P{&aEi%71rpUGftK8NNvi-Smwng+%PYjF=?u zOcMcUWTUCDa{=n;#rRj6#8n{J1KLYxwx}ocU|zO$`bBJ+N)EgY zF=Uqcr1!a>Ct$xJ^uekSx;N_iiq?0Wke%Qv=?$+`3zs+5)2n(m7{t z`p6v{n8wI_L>%^OX6~Sc?#a{sup#k`Ko6~C0526&@wkir!YWrk$a_Q>sVo@b-G1CU z2el3RRuhx=y*;b$I_+14rzi6XHOat~F>@le*DQ}$j5`&lduhoUXo_{I*64E1Mc6CB`t0Wo-UGkl~_=lJ1IkY;i-_!*d z8f?8aoQLG1qj2S~;8Oo)8GqXNrpQzbtobB__Bhog-zHTPrkn#h4?b%6I*ud5 z`>a4vjsDf8s9pJAT}rU?UI)?une=Y?3E0Pe?-C$nqK<~}PPTXU_3zOBvbB>52bIFd zq&;5bhPgsO8$0ee_zr(MAch;Ts6T?)hRH|`J4jkz51%H*KjYqhGY1hP$ZV<0{_C7QSVy7DjNV7| z!9w7k4hU8nM9lCT9>Jn)Z*D=dF9Qs|3xLmeHmd;MFP)GR6IpQlT|Qfuj~ENzd3X9w zu!Qt6vwu)HZj51S-r{JD-l`FJ>0!1Hb^y*yoOOO7#vmEb6?U>KZ)nm%FlYS#=X%_ie<&=q60Hssw$XcGmW2~L z8VQRgv137R1waXa4}9OhY>@h}3Ywe%r~OFPPhE*6GqVc@JlU;3W5Q&P?JK%(l}%(J zrq$a_M9zG-7IS}RK>o5oJ{*xdWphH8KdC@M>8{G<2}InI3rG73uL9H91h084c}RV? z`Y45Rkx4YjI!9b)`INg{?gxi2uJdAtM`!=TVe)@hCm($moXuIq7 zqHM1)sl3u|c?rgr(>+mH1)&8v3A+Dm?x?mE|0A^WemFA#@e@r>_HCjLht?jA62Yyo;-RALqt6}RbO?F;ui z0ZNU1{8sbfOe5vIlCe&%Vp55CbxqNrn@XhO7lGfq&CE>Bn-AWK#E{pqvw~vSSZ{6( z--Q!0kTj;`q?&BBLqI2w3`vgPQNGaOC4I}7{i;{52s+CQfu6JIR>k$N%?^@IbW`&3 z-E1JFVX#vUXv7zAyO3!=I@36_qrX3OF(CcOO|`=58+A%8^SG-?w@}$W^MTKX7m-3+L{0&inQvSrq1dgT$i=cO| zHgy|Cx^pBb;-ZN2ZvHn!tqdR)ND>bLi`Rmm zzajb*&#>fOX|9ey3^H!&i=b^i;8Q|wq}_89-M_-^sOEH02Z=2?L*H;I2{?CBV!ipL zy~=^Q`iAg`6|j!JpG)v@XTjzEV+4_Sb{3}gD{dBqsnzQPRMf4eRHra5s7Kv>tNV79 z0-@W=faOoUx8be|*Q-We`h^g@L*lyl_16j}$EFH`!*AJBVln()RW`P?;;LN;Ok!}B z9~}ggE!I}wRi2iGLGC)#XCNJI*I&EDE82*l2-r!)yL|-G!p@@Me!3mNVEx_q7!j8ZB;dB#!!aO z8?UY`$5`}L&#lBR*F+krQ?H~R-W1LDg5FM#pmkWw!n4Ym(@6cT;ZvyEM{jP`yhoml%hESP2x0kEj7`J22LB-4~?#^$T z7#&{%!f?=DTXZ5NAlq>X53vL<^b3}_+5a?Du*>-}MZNk>SfgD5e-YMa>^w_PkHE$K z9Ci?exGrAdj`VgglL{BNQTuKUs^SQY6yQCO3a84a@L994w*IlSR5WJgb<7j{F}1eE zb%TS0<6U}&*yT>-TTRi26$8&L#eYbCJKAm*`i!JXJNHMvfQO`x_ulV^y}thO4A9Jq!lNJtSxxXr7zso$u%U7WsXqP%U;)OLPqVzngoxm z1MumC(AdR)n1@|EfYbsX-;qJ$aQ>a_)uP)1gq{{>YKQ(a z8aaP{ezsln>aYHEcl%%cDYH|vu9ZL2r(2emoTwyzV{Gi!voNwX*&xadykB`_=_l7D zB^d@3lEhS3bucr|Gla1U30*91P>;j97TRRwyfdVJ5M@L|OB>6h`|Pt8J+a@th+0&4 z>fZg>Enc8jyP~5h(hA}55}BFyF1-&A2&U_Ml25& zBy>i~_8XwqiHMl5bG&_fqSV@%yaMDO|Lp!!VoI5=J^ju)GXj_FMM!;rQ6p1y8_o0l zUQPh>TmbF;??e_59%Rn+;Qk~?AE{c&r}p+%2lR)Vd#78L>WFeQ=Z#vAR13UHBP*cYrJN1~|fDFs?ODTkuCoT;x!OS!u0@`)0 zjQeFb2&ZNbln2QUV80>T_{Uj|m|gmvNG(k84v6y>-$0=$3%Vq0WZyv%(%8~d7cxCv z0FnXcACCbIngkyo0IN@7IBd(wKz@#2t4{onVg|&a{K}V+^w=nxdqq@J1isO{$k(sH z=t0A-0d4bwQ63px_X~HKrdOt10``A3{1vULKXqqVo)X|S1#)LVR)~qwlzq&}5fBuN zjEn>@Q<5U6q?y1rXurH%Y38$;UN!rOdQrog=G_~M0(4ITU$$SS0BQG&_4|FyNs{il-rqR<661x1NC zaQ9ob6%u*>ib(_lwO5G(dfv@zY{W7;HG@w>I(7LE*eG{8eAr0WV{-37MU&A_5U(5`@G(;LGhbH zZ39~TMw1!;G>jqn`@=sd{FcFZdZjb2dECG~4O#>~*Pqhx2C=d;$2pd3X%&KOyIG}0 zBku#TJl=~ibeztAf&;9E0W7{6cv44WQoGx30XhEjX6!&N_d6%3Z4uhr`aeY5*~b48 zZSNBP6m2^Ydb+;^IcXw}t6^_CjN%_`m4M1jq!%=9Cak1S4qvvFy9DbP7M*V} z!-TaoT7cQ)mh;vQbjb<|9v-J4#HuAJ{x#^$`AkefPQB~bALTe*1c5;RowEI7v~1Sc zGqcEzlNIEHPgNM9k%qW(YD>AeT1x&Xy}q&5PG{stMi?rka3k;ZXx<;L>91?!k zgqKxjfynIO_PRCxK-%G;wZG5@$oTU;oW4H6G{!&h14?-jzzy(H_~iP|5~Pk;<-^U2b=86k6RU>P9GbqoG)dGrN4mLJg38Axyz#RFiB z0?j5S*k{`(_EYL4gKqH^BdCdRs-9;@WEeK20Pj7>-{w28N3*VyTLiqTwjj}lJf@sL zA_gH{fT_DD&v1`{q2>_i5<%c~)#WQ_n&z+O`z+xx=6GQMxRQtcZCLq;O8Kq2`LMn4 z>)o0+A&b$e->Y5AuCpQ>Hoq^$>gW4j5BZ6tpx-7oE-o&1_Kk`v{`Mpg9!W_>brbxY z@qBT2KxCd?vj}}a|Alp+rkq}f&eTD!R~(0qep`o`)J!*O$oGE;!Uzk!n? z<|%KzY!tv18ui+QzL57uv{I?ivTwsZ!{fIm)n!f*t<6z$r*x|h8;QNz*B^I=SHb*3 z-0}Mj_C}x4FU<@CRy3FBpFHmc)rY-Ef7i`S%Vzlsswwh+3X-b@2(n7rN#w1>zJB}q zrB!kbS_PY;PTh*l>DASzBBW^JMBBdVOz-|aT`bej0~#W+bjYZMg{{B7I6dCm+uj}= z92B~7W4ivPxixs`R#v$dCNHMpBkF&$D4{?vukGTT+pl1p{dRr*6kc8FX}v7AeGXRe zL*yio0-}gpM8R~=nIw)u6>oK8_x_EzSEb`x%EHB|j;c%A_rGM)XsOb?5wj5Ly7f%; z3ze4jw=35*;zoyaFzzS&4yv1%MX24m@WnCUp!nYUo*7+>KJw~<9n_4qS@>l4Gz(Wi__Utn8 z;ltub+*y82F*8@=kZ%O7D$IDtQVEXNJNiEuj-4E{oqt_^qUd?wxTeHqXZC%Wf|(Y7 z8BWfYU)jK#-Jqe0lzl`y^qt&F_RvaOitu>6j{diIuidiZm=JByYxXP|oQSj)x-GbM zYR62kozTP~#HG6E^A5ilqf_^zx9>OUPt560xaPoL&G@+CoWln@RqJCV30X$%{6`=3 zoaZjBF)zNbFfq*xh zl^d+(g38-t+IW(}L6Lx3vq&8%-SX_Bw^YD()gMvObs*9WWp#9a`p8jcoWnZOH%6<{ zLe_Q9rQ*?M--n==Agk2zR%B-Do;TAP(NM03XW zKYIVT+D1QCQQdVZBKbwY(Gj#K9=V?4RiBpr(#0j&ZN2W4@3I002C~n_?106Esrn7_ zBE@&eLH?cR?}hRU6flyKAmd_U!UL*%_hzM8Nm{_awU>UjW+J-8ElX3nw&s}e2(~;Z z9}@C1DM@>w6MC0+c>IX))D`bJY?!kEW!n^Zy+5vR!Mf{ULjEqncgQ(qe>pk1ci4TM z0f{NV$RLn4a1SFjs;9X#e=2gHFn9k^2XxUpe_n^kEEVY<>mFfM@a>ZkBAi~xaLMAd zskg{50V$FtF^Tw?` zCih^mKJY-2q}(y56@9@xNr>8_|Lh{+5pS~mz0Kj`fSxtYlMG?^+gVB_-aJ>c~56?i3b$Z#Q#)XAk-OS-;#3 zgU%mzSw(e3Wx;gE;h)TKNF>8-SsmUW&)p&ma)+s%+~r0D^jPdn@@of3G-sEwab#=| zG)oB{jB8^b^+$i<(kXl7>&9yNs2e)Ns>-RP0+j1r85)hL#o3+rmX z@Ey{Gd7xOs8BWXYp~X@?siFD7U!U(Jz9+f5_O71q1~I=^_bLl-@*YK&hccx*!mGhtQjXh=PCtLkmTk zA|SoDC`Ed)P$ME$Ktdp)MId|!{JwkddvCq<{(G#IKR7EnduH~`Z+?59J#(b?E}y61 z?iW>}NdgM~%%bQwpPw2S7UKK+8)j)aXDOaDPWA>2Wwprqe`MJf^%KzV38>N->C^-(vCIet+Cl!s`OZON1=j$WZN2!=F|FGqW4Z@gu%Y4815iHT>@mENJ* z*_r#L`b*q1g8wo8n?-r!65qf*ac2s1bGDvDpx4&hot^kiZTlNDE32#Wh~>Rkx~}8p z)`ydh5%T_SMYVn;5;3Malvn#t0w=|TObz$&vFPf2z3GUT`;HcNuPwvmY|{{W$oP<> z>S!!03eU35Tpr@GqQx0NF=^Mf@#K-H~WVHOF{F3AA`7<3+IuC3!;TuJLp-3DL&dG)JRoCT}xkLxcTHzJk~tq z&!cPOCiSw%_SjmI582t;-66+%l)(DPF#*(=N394 zv$Oe?4!3+ivA>S+o9p~(0Kw~J-`d^i7*@yebM6837hav!``?VgM&v9RZ4BGDDqD2& z7&|2I?mH_pqVP2}x28UQdRGgtp#^m(#C6rl`#MPzGqpvhxakoJZE$s2yo(o{5b@`K0acbDRpqa8-p(?O(7DqI`;Lhgw9^T z9Bc&LgvC%wDtLO#Of|0e(?hyjmE5tUck5)j;50*C5O(3wR>cH8L;;6kxhc&dMJz$# zESsIrre0AAJ~oWxlDs5xmU5`3ewRaDT1|p8yp3^`{q-NTIErLNk_qNlSd}kgi!+#I zlJCy?#Ac86Q81E&87=S;e<$A|+&=i>S>-k(o#*(4W`;)Or!@SC^aHxgTwLtnQl7Qo z20Et5lgmk&xP(>!vLxxW%8Jcb6Xp|vY6jzi4~g-J=o;k{WmozFo(1=V21E8}W=4Vx zB^z={Lpj&Wxq`MM@@zAV>x{-Ni9^eZ>oo_VF=z*=nP;NrG^xyD^+eEue0mgsK)~RD zxo=84d~a(hn6_j4>eafeow_B`w&zPm2u)3Lv4_;Z(bVR*cr&VXV`f&bI-@#A$8~nM zrt$tc67-)z4T!Q8Z0Ab^9YlLB?@|Dw=+ZeCS%qzDz;qzwd#)a1$fX#0py+D$mD!*S zNO%WJX@2qK7JpFmJVh+ZBZiX|F8H?|IIW3E&KMNKZ|sBPJ@>jtgdLnVLIGw}l?-i< zWm}T4KKA^a>4PSlfffL?4(C>hj%^KFy-bEQgscG|_ecRQ(34$M2yV1%C|APyUzU5k zHcMGCvz|?xtVm1@-&)<+u)jGeR7ayQv-=XE-tUf;{uVHUv1$!YcKaL)BpkYbz$?H? z*H>18k}P&|BF+y6IDy_DoqEUx!LoP^iDo@wk{WPq3+;CCOG*YjbqXtU!^$#ZX@CSc zyqXIj%W;aN@Zo7CLEZd2h6A$Kf;L1TsUjDwu=_WTo5&q(9KyX)1(kjgW zy7@RQi*$6P<{PpHhr0eG;6s0~%lqF_#XQjvow0EMW!KMy?+&CB;piIp%)!zj&{=@v zFNr897Sx&uoUcclRjT9A&Rt)vOkjvdvYLzX2VJO@*jxC7 zR*qT*i2`7*N;UrrYri};3%tQ{dcLS5FMI)J)GPI6Z6@n+$8~@Qh!My;V8L4h(UUD}ZjEk|WMQbT1`H@kOI=ie3DV{E*XGV`Gcut#$p2gto0u7It=zCuVUqoE51rE30I3YUX!niN+_LGT8u4A^9|aY`pKVtI0dn`afr&D zk5T-tgx#-(xci4~^Tx6Yt}S`>gU-{@;ij>}C^nWoDmh{QS@_iMl1Q;K5(H2m+FxKx zl$K`GR_jb>MN>A*2z3PB*9aQr)+JjW~lJpwBsO9u-y*9JP=_#O(q z+&~}@9CAKjvKTL+vStKGWyg!)SuCXnE&1$eiXL%bP~bxmU81I#!hG}h$_^Zgr4ZCu zD50b{{%g?w>!%IKlPeV5F+hnl_wkDQ|LaPD#0*tT$EkC}rwiZ9!4h*nywRIpDE_7CUeG6qk>oL` z@OfHf*%|xE2ZUtn#d;6sH8QEV6yT99ECjCJqrR=SZsBnaDA&0&J*3{dyqGe$1^310 z*8N_x!od4(41qV!@RE|QdzN->Z}YCzxNaJR-%OP&`ms%>Wn3Nz(g z)!scy8$p1-)BJkXVnpxRpk*mY|7{d6bfT8>3b4*oGVBcyX6}D$QD0;JRyFtHQ#B&z z?I8~8Xb!sW&Q3KuyL`hpIyl~>ikJDZI(*4koK+LTIWyHnK#FXn z-zQ%Bz`pR*E3Ny|=~>3zJV8uVO>K%a-Qm9fLx zwNMR0wX45fNQ;_~b%3;NW2BA-#OaiKKlh7HkIQ~in~sE8%jZAPN%%Q zYj*EVO7SVrimgRu5B(&rdTu1BKcc43 zzhAER_2Rx}n9Egkd{@E0ef{!pLG46Y)OZYZb>sEZz6QqL7Y!sX={{96w6H)2zR?#z z;HGMei-k~Hs5LMrp)C{odeFhu%}pm%N0)0%7O!^ak87nDb*X_;uKrU7x^ww~?Z^$a zcd%Av&;{+NGVbI(kp~Z$1l5xx1^XWbshH``*|}6%gKTyz`e2Kr;(imXv1i(;Qf&Lr zvwUhvyrDBXJ~OD`UNK^=e-6m8{lX( zPxl=jBtTeI3k`t>b9`3$UAC-Ou2@&x|3grH=)N@CT^jqTP@9oernCIx&u!rCJ>;1r z34u%0RZ5qrH+M~7FfaVFhs(3z#04MVh&5&gA2ck)q_rd;r7!5CW|`vMZyN;RF%fK6 zD+{8{h|Ul)W<^jkv=oQ(*rz9Zqdha21=&pb(%GU>D*Fg#7qG(|YcB#)djn1@+6z^N zvm%?;9#5MtF*^5+3bx!oQR)H1!%+T+iuw8Zr#fm?lWxW(wT+D=kCcnq+1XrxURCg? zFrn7@_YT!TKu?;=O0oA|{2X*>9X=p%ztII5t(qY<7Jx1Mf~Jg&i~<))mh%gv#xU4M z`2!t3A}Le|6^mWj3T6N!X=^b2oRiqq#P0ekf zf;)xYxs$QXLRALcQ!KXQe8;p`7&fMBaiAkxc;|KHvxeK^>zNlD(&4UK8m-?}dR=bw zQiCPP*Lv9${wPP)$~dFj$CSZoeIao5QhlRWGFApLC70|8urh`dWtFLpquI38s(jR? zYc4PjeI`mJ>b<^+2|rSUL6OM8gM5O;`QOaCnFW-VGM+o55g!n06AIX-02UV3sxMG8 z=|&fm-`JR*esB8m0h8R_gR@b2E?_JG_H2Vn4$EH~yrWti2+VuVbZ?w|e3CGL!s(Mw zV!x=8!Q<(c0>3oG`zO8M81ga1uF1G>N1o+T(dh)idp+i9?c$Xc?of0d9kEcEw@bU{ z-?&Bp^4fPDaYcf4q>1SVjoyfXwN<6aCh>W{{M#?Jia%)fzL2|r-Lb1yHU3>rr$9xg zYi2sllFX=#tC=rYJZrh8RLk{M-GZe5*pI1FBo4O6rR>SH23qgrSP)WE~&Xw}>q{P@pp zIJ(T2D-K?tMRIO?< z5pQ!g&}Z0?I#+$I>c`_&ZI47M7#>2_7$(xFkv!1&^_TpXT}7a~Uu`=m9d0E}9Ps)) ze_-P`p!zFcSqSl>(bNJw$;Mx3#D`Jup&o4)a(%}PCyPzS+ndUPO5$O|LB(fWx6X!f zrF+__-y6y3vlLadv1e^0P$#MhH4(ie{_XLPjy;|$(b3j6lFl!Ec(lZ%q@;wDjLJ*X zd1X=WfwXgRb)6Tv4(KP$Ow(U=kSkVMX4$Ap_rwSk1MSbN&U9+9lVlrpky3K=X`U|l z#H*bGwgRQ%M=BhGgKu6uCxFuO5h6e5sWQ5e+70#6&b@7mK)`bKWjPX3Qkd7pza7*n zwyO!Npmj>p2qdt7rKLceV(J50fhdRMLaP>TX#D)=dY4s~{>)SBn|H@q^D_(tSy+Af zxjwksal}7MDG@VetxFs5v)^PMNO=2EAoC)XdfWFGYWg=E?Tu8F0z`w*Rt{B32E`zW zpRA;`H_m+A_I+K?WH;G7bix)n0dH{C0ssI2 literal 0 HcmV?d00001 diff --git a/example/ck_tile/15_fused_moe/misc/moe-3.png b/example/ck_tile/15_fused_moe/misc/moe-3.png new file mode 100644 index 0000000000000000000000000000000000000000..77c6d9b6e43ea2c2ef9087eadff6028b6af3f113 GIT binary patch literal 18655 zcmeIZc{tSV|2I6PlFF7O$r7O`E?br?St84jJ%l9tnz3XVOZEztgt0YbWCjyr>{Akj zVHo?GWb7u)Scc*LP}lYQey{KSJMQB?j^{Y;=eeKjFEiuwd7tO`KHukid%e#1Tl!jz z$Ic!Dfk2GfH#LkvAi9Sj(81?N4+6i$|B*Qj0^LMtYp9w$99kSbo~iG*^Itc6&zJu#9Z0mwxK6E!HNBtAFdw$e#vcNK?yr++Z7eTS_$^pn<}!mE9fxWoK*rqLZ+imP~5snrM` zZl1Y8LjyLL=lMNRcX^BM?0)0ob>QuwWEePW(y6flhEw5eO{IwINr`usSWVWfT{+ux z(`;87X*$|2EIq2WC-`~8NEvNtdQ34hKo8uKW4RPanTH53=0edx}J?mEva6wVD8n( z@yFTn;HAioQ3$J}U=WC}^DXnMr?86K;P-AuUVZQQkG;VA9)~l4KtCklF5Jy;|M>Gk zX85Ip`@f%;zQMP*PQ@t1(LFv8T)x78w8ja6D?phxQizMcw)FzQqO}1+(st z_K%Q%bn=g#{D0%@WKTO@3Ehblbp8G@o11r`>->K?jem~uKgan0(qnw@=Jy|z2r0sg z*08>E_;@?KEudk?`W0-3z-M=_MWK(kJ~>;;PMt_CsuS207HLlL$U3Pq*9c2sYptbC zgb%3ghrGD+6H!D&W*xR2hyg;%UqJN z#RK_5o=T$HpYzjmeCICp%jn?=N<>fViuJb@0yMSNMvGalvXvf<7Rwa0!98hkOoBJaIu2HbNGHzpljrkbbnKMECmmODN*gyVaTzt@T1-9SN@2^@&{AgaDS*1?v!nQslN2d5>Z8CjX3jLXY4}KINWt{^l;9DnJV$y z;21U-pn~512mVE-vypA{)wIpO^lMFQ3i!tU>PC!E3PaVC{RD)CaI7{(Si{+~^=qv| zmx1Fm0TDLu!~0napVWb|!?h^|@9-umgTbLO$`kqy2gUx} zkJD0DGp^U=+N3PhjlgWZw5k4%2G)&)YhEKXGRQObfCY+3|&|`3- z|1DCMG!LfP!b%DXMgpLiR|%tI0(^5(oUlpR`^vi%W$SdnY8NB|=D%mf(kk%YJCj$%$DD1Ng-sO8KUsKqE#5-=(P8`xeSy^To7YGkyFp$3bpvKByzQH( z$G+SL7S>uKN`OOSd(8KYYD>XFgt$aI1aK(pr?6o}_I?+|SuI37v zVc!Ydu}MF2Zb8g!2)Dd0M71Dr8Sn1q!#0K!tVwf0sx#4fSELg>lYFrcy`3;?$WwcR zqY1m<4gX9)Nr?{`{6|%26OAjXTJu&{JJv z_Xsvq>ph(LG_v9D^f;Yg=)LcAw(Bqlx_+cGDl zFle0Irczx%PHd9fNf_woHImpaaVfsMM`KXZLI2b?W+~->QZUACX)W|X*Ns@mIKl|^ zdzeW4f$jEsSg%IP#tK)b^C)F7j+?e||7{*jH zm0`)v!%0pvf=tatDwH zzN(9Jca~r`hIxB5whO0TC7cNfsyKvMXNu!b7#n&oD~mPN*`C;DtKk4n-nwF*E8Z&b zQH&!W{gpt@1wjdrhu^l9rt_EM4T{E_uAAw~cvMQoSFZlGs*^ zs+@@n5t$Uk9?tgM`O=jsKsQ!v0G@aF&L)DCcRYZVz2dMpG1tkpdQ3vr^9DC8Ouxv) zZmL{;NKx_XmUl_e;5)~Puj*9kLbg8KyX%*4X4WasY@f#=m3nJp#@oHs0JoNISv@kO zobX6&wWF+JayqP}GffmK?kP`QLwyCNgO@tS$Q#l*f<-mLgF7yg~ z=S^79=zx{JmxTfE&WRJa79-N?@T2a@^m)s5H*``Y^;qJA8oyh~c#UOfwZj&IdrNZ7 zp|)ON%p9S|i3#{P`O0IvRNtqUvkfZ&5n^hn)XG2rg{aEikzhwhpz6ao@i+&VDPh+8 zCpS16PW2SO!dLlIR?nV-LIUi3onaXfCR+P}?(G@#HNp~w<~emqQhHpwld~pd`uR4S zE2V4E)q>f{kw?*okTy%7_+zPE;o(o`?DQ~G)O+o9G84orPfB|Cn&(8FlQNn8ey(yR zXU?r<@o`1_(+`)w3rk^^+(IeU{n!u`lTON0cBILf2OW8?)Gir;9<-|^+Chyla`!&5 z>LKSvc>MQU@tZ%MMz>w{^b6Js|AI2v=>Zgfr+cA)_oBFEYUtKL0P{3ABp4h1#clk} z*d$aT^@411dpzlb=w)@l2cc+C_>ap&!+xN{zPf8bl#HwUmJ4)k6^B>~E)j z-xB0Xa^&q2cz}G_gJENZ{e1L6D*l=zbER8jx`xxTaLK7RW(FsqLWZ`YWgcOPS;e_> z8ysUUZP?+nWXtD2F?V4r&mqR`qbjMm9XrZcGP5UR-LKFk??3t#m3thex7b=BJbmu# zaC+dkyCnhoStLs=bYZGF z6g(TZetPrF;$e{DM}_SEyng?)K`9BzU8*@e76TDqU&Ad1pz{rnp1(N8ML#AS_)q|w zj+7*&DZg}mOx>Ruqp5g!S?$?-;~z30n+6(RG=^?F+oS(d&k2D?2cetI%Jf&I~I z&%+9P=CLA-iUn^bB?+$_2*;#3l_W0g#tw^lOsy?N>Gla+v_B1x_g32J+g@H(F=zYa zH6vMNWrYc#2eEKY5{n;xUrdf%I58LWF*dH8JbeTDh4#eE``PJ=9q1RKddXjVpw)C8J{-qT(@Ot<4K~P<*Cree_ zUh)>j2)N#dT=4iW{L*%jJCQJX`-tId2SIV~ekfaqL<#XH{CjUa2Snkj9{}KL=TK;Y=@o_g72l%$Gul;b^4fMdcjf`;&NKq*gyX)Py zxVDZ(v8>INT2Im`{>Xi^EByB~*ch9`bi0nWs-z{!#BEL}OlhOIy_73*coiu8)!{wa?-oe~CZk+Fgs@4enZqZc{NNQ0DtcBb9`DS!Vf3vQ*>vqt+*% zQLlX6N9|^c*I#)%jb*IwcrcHr-8_g9>E;Kahk-29J(T zqbRnUILt7;X(!kWyR+DUeN`8#ur^<`*5ckO1?P;mt;h=+C0J3U;j@DKNAOtQ$<-;p zEec%c(^;5ngq^5=Vf@Qu4^n$deC2kV3a&BnepP`Ji|P<|lTjB?RL@@;dS#M)>e$%*$%i7jTwqkkk z94VpE!D)Yj%>Y8=Ld~x0Q|8jo8psB!B4+c?2;nmR*DQVrM(o=W|R&+LV2>{)yi;`0R0Y znj?}?ES(xp`G6c+T!w2B7?fy}1_KgwiHlv7PMF(zvx$ux)qU1F3m> z<{DNF_Wf-+eEcjdqFpe9c&5NcB42HU;NkkFsYR*ACpE}IqGSQ+kjz8oqt8n_EbAU{ zYfVc}9)r>ftW_v;L&i=KLGt8TmCb`^#hNxphItw$>jR}sm$u!0xQUn%#jWvMN&Q=o z1Yz;=ZfBx-VXawE&#;ATEevVw!BqGXGOA>;&uyg~&AzF0pLoA=VNB~sE?jm%!;x`ht=wrfM<%SyO?u9W&*!peFUpoYS@*3_k_C4a-&V8mMIRsir?^2(9 zoc=GnKdtW!%S!tT|q{3ck1(S z@c5jp)VolV+nEo(rZ+rFE+%SJhUR@Zc0vJZxB+(=eJ%g!OvjB(Rp#>9H=!*PV+q?G!_?>-b`^Fd;1j3FZ;%XU zv9#rAR#;yE#t(B#&dacII4O4yCH8flWn;--o$xZh!(udRMCt+%UV(u8D4Ge^C<6J> zADpx4ibmg6%roTZbI-gr_SFAA_rXgS)r?O=5xpFjFSpxn*U%m8Pk2b%NN~NM_R0Oc zza&*%3fVCCB7|9x8N3=?+?gUJhvoaNReoulMB4|*_g}eho^qSBmQEh#>LLzjP{@MW zY;Hb3_8kB6vvei}9q2|&tA&sc2U0Z?UfUVo_DH*OyL0-;v2(o_e`-ZcL)4=XC7FM8 zx@Dlw7n4c|cz1XdoKwlxj{zGS7ljBFP?Ee(QnBGBda>Qn+RG0RccY^b$oA6twXI{- zE8RshOqIj~biH_DRJ9H*&OFD>UQSlPBc&-KNkJ3NAx%9$0d2b8`=FH7z#a7-J8yG^`<9%q(txYQh?KeBSTI8(qo5Dg&X)L6h&X4n`n&Yn* zWyi@84FYoBH?VR?6&Kw+VV4%J6h*EaN^~XHu2aXel+5}v2)JT==<)Ut2hFrkQ>WHW zJ02J0%alg37p)@ z>|vj0e5Q%{s%n7D%1^88 zemp6&e77@^d2NOz$WD!L<^Kh>#f-{Z&|ZneYVmtA!$-jyr~iKo;77eZ3Z@M+V+N>; zq<-}Hdl$mU*wpP)0mMyD(!ucCcCyFs>;o2!V_sqDk`5zd<7aL-AG|07l%9f*)m9&G z$?Ka95x48n=-cqk(Fip6P<}bG(${k<(Q{;WF$X%vJ=dT21R?i+^`c%#!OmQN^pn)8 zdV$h>NK5jlcgX5POuz-=OE>8+TF(4FM{#MI&c}w7;j^dJ2<-a=3t($F2}i-I!~1NI znn!yijJ;ym{w(~G^Va?&<}7diR^)XR+sm#=+?rS~QLO**@ZG_O@f@InnOROMDRkmOv+dV!P-?(F;` z=R4w+{tJhy>^Q&IjQHDH&6d^6fy`fP#o%pOUA?x^uWpipy>24<+qK|)!oFPjLps#Tc%G+j==jK-rFm-=j>lIB7C)Abg$x*`?roUIfVikq0RMzf+YOY zv;U!gfF=VJ7GD7{P2SH_sd=`b%PM1yNVH0N;98QYAg+D+BuvdJ0k0v!0xRL$@7a_A z{28L0!!H;D=qcg^GtA$ZrT0~rP_zus!05!4i#=$qy&^4zvys&*q%ZP2+uK>eOgzjr zOD?9Z3H7Bmn^k3jo5IZ@Wx=!S^ptz%6*m`NyzgvKwY${CHa!EqAz6WDArXT`psXUh^CL~I)xA`%XzTax>ImG>E4=+x$ zh+YR*uhYc|^Afj)klc8Ac7nQ|qd6EnWYpL`Uw8iVbc;ATk8RExP#?9WkIwlat+hKD zPvCeU7+9$k(9;FhO{5QyHN(Qr^=zNS7oj#y3p-CDBDU2CR82!3SgNwb1MB7zx}3U{ z)KXPv6f5lFk-xl&nS}c4AbL%4wb#a0RdvnnW!mkXmmaZ+219B223cW4^$o5%#_5;X zzc1 zVW$u>uXNo^9cCJg^h@{!Wt$FAL)H0im^wEE27eM^*I&MNv#>L0dMc;NvvIx8573vY z?d{nB?B;k)PH4w0I~1$p$Ky-?Cdv{0-0Sji6$PNw#ZYRR{=P{ljeh67swR0cww=%p zB(&=;X|D}nZH!>!;0Rf|!(%OZef2Z#BFz59Q=)PaI|ZDa3ZiJ&=W^e22_c7L6`ew! z7t)#F7VvN9Uf}2*(YUqvV>kz{umu?(9ktKLP+ByJQlmNp_rK}#g4Yp|%D2bB^WHbi zA_Umd{lI+4qB<|CS$zmz;>F8&-SxQRxM16@pNsi!JTqMr@N90*u`k7ccgCtn{R(3E z6XR~z=J}5?Cif7eHpX*qqctTIW#83vZ4Dks>yfvQvvX3z?m2ZVdj*`6N%hfTzQyy!PAVDi>-Hb< z?l>!qn4X+-g7?|2Wy}H9|J1<^I0NPoZBe!}Gvuw4)_e#z>>~7mQ#NT$3>3!>#Bd-S zJ{u36ml7_Hl}ZVvQNbC_rA|OA45Y&dkG~oJ7_afZR2kB}2t5GeJG#$a+}X~SpIuu5 zfmjahQwYLBJ2MY~_>S%Ox;MC2+qLKU7C|5luKlO6Y#Z&`AW$yL-m~%JhoHm@AdumH zYrDydEG3+g-?x2#1B6*W3<|%skAZIs%}5C!0L7{8vnWGN2mae{`j)As^WDNiNB&i) zwj$&ti0}CR)UxFn{v=^WzNZ8I*e7V(XN@}P>;DPp^SWc!m0~hCrcz-Tm3*N!z0eV0 zt>JI4?~yc`6>4Lp#6y0hkgd0Ro61-txq6XHYfCMUao5%>!hADf&bI^tKbNmp0sa4P z%6sdTLa)Fkpb|!echCg^u{ufBVW+Tk|i&%&HD~C_#LFLnbh}TF-cqBUT^<8xv3(H4f7?!1zf3r0^-v@8v;_p zAnQ#QuKd+LpS^aJ-vgMpCy5nQm9z&#JhlEDO8hZ1sH;_nxe0E{vGg2_`&eME>%Eot zaL`~2GNhugG*dbWfR=9880+WwG3!S`xj+;H#>~@MY5zK6-(t1x;=49+M`rX-tT_4m z8vP0)VR(J)s((xaPNmc+xfYMDVTaafA^{WX-0 z##*sny5w_enjcC~y2P4Rcf{|Q)wi}Q!?KG5siorAj7Vn+vuqn4xGnw=k$M$oPt-dX z;*oR}EKrdH;^W?**b{zWVmIcyy?D2lChF`x!(h1;UPh#4a^H!RqtZG zsAc|2fj&?7^b1GEk+J7#?>1J&#cEw{NtkVK;Fh_?<;E9B>F*u=EtAr&ZBcyHw0Ju^H71UNuLA`;VB0aLNa@Mhe_CZh z>~BjYvF5V=CHvVu6-YN15n5lkt4pAeUv8C<`pQa&gQJ`^{t};i0T3Va`wV50qKJqLp!tGx&Jx+8kJ z4JGG0-#!H@+h0?gl>v^iGIVe#!g^@NX}76EO4w9r$B&XN&q+_w#R9yl!m}oV2tRhy zK zvB&y@Rk^P*qt3UKANgq+&X^{M#JO8`o-%jXek(B@a?=I?R4nNq*41bK!Dy#!!V77ZG-V5iTRjnU04WBEFk zfMO3K3GElFZha05-3A%tJP}}$qe?1`hH~&h`z13dcR|G-#4uVmBW#_c2^25|!#&?h zR?~o+4Ua+qzDh+NaIYcw&wC9e3!^20c)|UG!-h>-dM`Zo^KLaY8h1Q7fqR`&_&%*v zeiFa@_<=)?v)|C`_dfvs+&IQPHrqbty4qy`x3xH3tc%@=b|Fp;`nlcdZ@3Y8_Ul|u zj0`H}im_(rei0pntAq|tt86gOXR2C@KYXnf8}*odWUQi_gGXakr_A?==jCuOVj&tXa7D^NoxC+>)#7u$e zgeA`W0{e+vJ(c}7@z|;q0Pz8=LZDOZKzr_Y9|!Pds(%a09kJg;ci1k^)pbfSbC-4q zOXnSe&vFY%+y4OGS5+rirn_iQnaWw3&jM5f02!+&aPOI{N3mOFoyIQWm-o8&nGD=> z-sz4Dz~ZYDr+2rBFVK|>9VWYPNJHpQ890N*QQS|iDF_51sf z+>s>8eT91mfaXt$fOMWEkYv~Gw9Mbz8~Za8xPIv=Ig?@$NWJc8X|&nlCj#+vECSs4 z2frvslQ#JmTp1}0FVFpY`M^WAfBEqGR9Xc1ApK;pu#!__j2@KR{BuI53qH3Lx3%_)<5({CNo13huLTmB>20Wv z*rB^X^CAh2B&|(9N$Y09O=kytHbybb&dzFUYs-#!e@Zq`wkWogq2zH3|qNIuwlECNh-==J0bNSV|hb2;ITX+n+0YW^?)#P_N}$+v%p@+ z2|ri~Qi3ZOsW1Q6%byDI&#+yn|P%36TG`Qb7AmN;R*^lBv|Aj^1hgF>Hd`}^yLw@?@5C>C} zq#9c|#PwhO5N{bhIpGkCQN<1uG8A@Fr z4!fkp`XGwAEz}SHS~VzfPMcN4i!#3Uh>7|-JGBw!(H?B_BTdJ2;-DUOEJ95-E*amHWgEg-Wu`dP5#4_}QErr(Pshh(Nsql`tyC7plFbo{1po=$W2!N)h+O zk?PkUW$W;$gl=00j8Z85wx}HBlz79&J!qd*PUKi!V0p!R{wh^SVy$a064>f6lcQpm z?=N~sB2Sru{TGM)zC1fWSL{_Yi@54B^frK3&iihy{~`jMfiUYKRXE~cfvZGDuB&>2 zh#^NyOW0$3$`bY9t%`Tbs$n!rKPu4#8c3ZjK#iDey-H6n9G|bA&NfD2sdFgn8sEXH z2YEr_kU&s43qX#^=gZuq+oBp02}93;VCZQ(^jNh5Dzis_bDA{l0Sn1JIr?HySTXYt*_kpUS{}#?p0W+sw&=$iT} z(XFJ!yZ|wnm=cuNHiYpAP{xoCY?#L3-}97{QSAZUkRxJ)OXW}f-zZkPVjA=yH_KfO zDoQ5SgrA(~Z|>@8SuCPo3)ywM48`U8*+I*!Qf$9Xy?8$K+NL|mDhSU=4GnASLfhW^ zD3eOODu6u+*4M9OO75u%m#;*wH<=WY)&-F(>wr_ql7|pmQ$$Tq&mZRZ5m5Zbh?EIF zaSjIkb~nyp0RZq$qRck$Had&2g`3*J3q{J;x1T;$Z-IU+sgriI2)oa10r9Q1^b9Ne zaxouYHxa{CR$gA*GO>M7m-%ktNicgFMfuG5u6_tJkK#k%cnKX6Xi!ZFm~2ic59Nwr zcpI%8I=eN_NUG7G^GaZbZX3x1t`;z2q^^lzA+`4@2uaEkUw9|RlJVFb7Z(MP`T5=g zAbW8;z?l{~SB&ySS{FW0+Z-~!I=4L5`?y&D$+-W_#+w8SAbDl-Q-q{DUW!IEYz_t0 z3@T{l6m+z<9;-{yqI^Hhap(0XUG_A~_8{4SL$Ys4)uo<)cz2*}2T4h`b4KsQ9P7=k zc}Jz?xz3&rechcgP6r+PpGDjn`jnnLA z%2iZPiM7?#%g0yQy{TXF8{Qd>pXXU-oQsTBl|8eedhA;0=6rywykp@?z)01D(13&N z{@dN_*pt1}^As^g9!0=j8|Jj^TM~N(&RL8f4|h~5u@EY`GllS&L<(^PyKT}-G8fw} z>+XEbV>J;9XL-yG3mWB|l?Rf2a>Sl$v*u0K+I!P2XRY&Gn&ZjrBjriQl?_A{I`wltE(1YU9C+gTPl4Vliq(dvU zDx>SzPj88`H=<_+k`!O0q~sre*q$cQSH1=+;_ro-2!S*nyUEfWxZadfCsU{gisJ)f zQo+kN`=2iZpZiat_dm-4pXmoZz#?q>mz{qiep)5B9sxMJ_-h;+&XOi-O7qz2=0HntQ zoB#;aVugY3#Z~N$=?x(0e3dhOCTx2W6Yz&vASP}|+u#2IZe42ugM&ccK+gVu=u6Yg zpODp)GrbtE@j~N*ATd{3%7=#wlWVva$PXa%W7+lm;lk#77Vh1UIuNaG~kzhl7PSLvlS)O z2$a|1!mO~*n$D-Ip=xXs!0qY}$wzTO&H7|S85kPS`#0^9sQ(^>I{4p!P>nR-C=takA%Yv!-p0(@SGXDMbvw0;Ru1jLgk6GYk^*vfPP$!XOji-n+{uRDD$kqiV z6@f|o1FVB)WwHFlD}!SN0@Bu_fxo8M)e$_|%(#_|e5JB>V;5t9_PMyt-V*YlD!CNKU)Gb=^Pz8D4Wm!?Al8w<^Z)q&{oAv@1`^=TL43+o< z(}&q7))%%G8gKrLp=m3tNx36CI&*wJMbsJ7CEYo~(_?Zx02Z1sKGDvPp_{#P0tZCn zvP7D@eii`DQC-M=2mt-;v(KCeRZt`JX6+vWf<1Y*i-0<3q{RANquTX3hgnuF;A=!q z^afhbIc!%`%&Fn8I{4VFViQv`$YA4&=K(dl(>ZRPeFU_>MNjqERqp%c=*@NO;|hfs zE6^)s3D92Di8e)4-(P*)BSKK< z|F_*1m!OeTkDEu!J*@V28*SkAGmdx3aw%CQ48m<-hl>at2*jz7yK+squ%jsVlFXKazqfAB#4I!3!}(I_sZo%IU&fmGn2=W2b^@-bLeB@%Y`L z>zU3+!M3Vy%A*L)Zy-(LX$pY!jxEB5FDa{OSbvD@}!va#o7A?PA%@Vh3WH@^NlhM25)MLI};~d^f(XZ4uW$ z@8++5?p@EWj`!M#{T=MTBb2G{Bmo>{_?KToq(-s9j$``j=V&MZ)LL?0;x(GJXBsWk zk+Ql)K>7FL2}078;?|mcGABgI8U}Y+8?iJrD0Zfior-aN{9Ld;Q`b750tR2(OEHK| z@QG!=^hg-hVJ7ZGP0q6p3f_deU#`6-tqVB7(m$8z@xo`#*`yQWs64F*Z{w}T$e|bem zRZVIe-xHayHO&gP{iur5+^rjQs&?fk_uq65(z_H({DNLrUhVz5n{Ie`5N&wx-llOS zw}P^Z{VQd^9hY?~@)FnqVG>s>bw&G%y_va+OyJqRe{s$AZ{EFo2jr=d;%8Y;;C`}S zla5T}bm=JQY#@#k*MxEe7upXBGT=TKMs1s!AHTEX$vj_8X}oHB!Y;AtKCPI!G)>I0 zq>G%^t>WidPyfaBSDq~qs!jxM{2jBBZDP6kQH2{mJNH|Q5<>9yFV{jTE@Gpi?ytcX zz!r)8bmTKO4-Htnom75=5v-J(JRa|el?XS{7?W6+#Fnby=H{T-HoYM4^l(9^yl8y~>Y4JSUBdmCUl zWu?9SXqur}991fn%4o-u3P$L%gB;vNd5CyzJgy`5W$we=2}eHqe_f$AlFKo49RCB+ znuX|H8}D9OZEWb?%F1Aa32d+R7o}L++divBRq4Ku43LZ7z*F*nQ z)=?;3oKW+rucek_a+P{nHihG=0)oQD#0pd7{POAFl7s7V{FLyOqmuM8x<0K|ywMd* z*O)u`{w=+jh-<`~)O6OleoQLtn?K|(QE2%zR{lKo-;$09Zw)8s9p86lfsz2zhj$>z zD9|p_?xWP-2FsO_hVItVQ2pt%pW{AfD9;Tch@K(@)W1CjBrH*mrL3?Uo;mv=5cKnc zTKT|H96#pIe97o?l+eiDotwSn=>!r;|73=_t>UwA=pI$&Y8xOGqoZs;P literal 0 HcmV?d00001 diff --git a/example/ck_tile/CMakeLists.txt b/example/ck_tile/CMakeLists.txt index b6a44f76b7..29305405bc 100644 --- a/example/ck_tile/CMakeLists.txt +++ b/example/ck_tile/CMakeLists.txt @@ -14,3 +14,5 @@ add_subdirectory(11_add_rmsnorm2d_rdquant) add_subdirectory(12_smoothquant) add_subdirectory(13_moe_sorting) add_subdirectory(14_moe_smoothquant) +add_subdirectory(15_fused_moe) + diff --git a/include/ck_tile/core.hpp b/include/ck_tile/core.hpp index 3b198502d0..3cf0c2595d 100644 --- a/include/ck_tile/core.hpp +++ b/include/ck_tile/core.hpp @@ -52,6 +52,7 @@ #include "ck_tile/core/tensor/tile_elementwise.hpp" #include "ck_tile/core/tensor/tile_window.hpp" #include "ck_tile/core/tensor/tile_window_linear.hpp" +#include "ck_tile/core/tensor/tile_window_utils.hpp" #include "ck_tile/core/tensor/update_tile.hpp" #include "ck_tile/core/utility/bit_cast.hpp" #include "ck_tile/core/utility/functional.hpp" @@ -62,6 +63,7 @@ #include "ck_tile/core/utility/philox_rand.hpp" #include "ck_tile/core/utility/random.hpp" #include "ck_tile/core/utility/reduce_operator.hpp" +#include "ck_tile/core/utility/static_counter.hpp" #include "ck_tile/core/utility/to_sequence.hpp" #include "ck_tile/core/utility/transpose_vectors.hpp" #include "ck_tile/core/utility/type_traits.hpp" diff --git a/include/ck_tile/core/arch/amd_buffer_addressing.hpp b/include/ck_tile/core/arch/amd_buffer_addressing.hpp index 3feede4d2e..bebf035e9c 100644 --- a/include/ck_tile/core/arch/amd_buffer_addressing.hpp +++ b/include/ck_tile/core/arch/amd_buffer_addressing.hpp @@ -621,6 +621,65 @@ CK_TILE_DEVICE void buffer_load_fence(index_t cnt = 0) asm volatile("s_waitcnt vmcnt(%0)" : : "n"(cnt) : "memory"); } +CK_TILE_DEVICE void lds_load_fence(index_t cnt = 0) +{ + asm volatile("s_waitcnt lgkmcnt(%0)" : : "n"(cnt) : "memory"); +} + +template +struct buffer_atomic_add_if; + +template +struct buffer_atomic_add_if +{ + template + CK_TILE_DEVICE void operator()(const T& value, + int32x4_t res /*buffer resource*/, + index_t v_offset, + index_t /*s_offset*/, + index_t i_offset /*max 0xFFF*/, + index_t flag = 1) + { + static_assert(sizeof(T) == 4); + auto save_exec = __builtin_amdgcn_read_exec(); + using mbuf_t = float; + asm volatile("v_cmpx_le_u32 exec, 1, %4\n" + "global_atomic_pk_add_bf16 %0, %1, %2 offset:%3\n" + "s_mov_b64 exec %5" + : + : "v"(v_offset), + "v"(bit_cast(value)), + "s"(res.xy), + "n"(i_offset), + "v"(flag), + "s"(save_exec) + : "memory"); + } +}; + +template +struct buffer_atomic_add; + +template +struct buffer_atomic_add +{ + template + CK_TILE_DEVICE void operator()(const T& value, + int32x4_t res /*buffer resource*/, + index_t v_offset, + index_t /*s_offset*/, + index_t i_offset /*max 0xFFF*/, + index_t /*flag = 1*/) + { + static_assert(sizeof(T) == 4); + using mbuf_t = float; + asm volatile("global_atomic_pk_add_bf16 %0, %1, %2 offset:%3" + : + : "v"(v_offset), "v"(bit_cast(value)), "s"(res.xy), "n"(i_offset) + : "memory"); + } +}; + namespace impl { // below type indicate the data type used for buffer load inline asm // clang-format off @@ -810,6 +869,11 @@ CK_TILE_DEVICE void buffer_store_fence(index_t cnt = 0) asm volatile("s_waitcnt vmcnt(%0)" : : "n"(cnt) : "memory"); } +CK_TILE_DEVICE auto async_load_fence_raw(index_t cnt = 0) +{ + asm volatile("s_waitcnt vmcnt(%0)" : : "n"(cnt) : "memory"); +} + // buffer load i8 CK_TILE_DEVICE_EXTERN int8_t llvm_amdgcn_raw_buffer_load_i8(int32x4_t srsrc, @@ -2378,6 +2442,45 @@ CK_TILE_DEVICE void amd_buffer_atomic_add(const thread_buffer& src_thread_ #endif } +template +CK_TILE_DEVICE void amd_buffer_atomic_add_raw(const thread_buffer& src_thread_data, + T* p_dst_wave, + const index_t dst_thread_element_offset, + const index_t dst_linear_element_offset, + const bool dst_thread_element_valid, + const index_t dst_element_space_size, + bool_constant = {}) +{ + const int32x4_t dst_wave_buffer_resource = + make_wave_buffer_resource(p_dst_wave, dst_element_space_size * sizeof(T)); + + index_t dst_thread_addr_offset = dst_thread_element_offset * sizeof(T); + index_t dst_linear_addr_offset = dst_linear_element_offset * sizeof(T); + + if constexpr(oob_conditional_check) + { + buffer_atomic_add_if{}(src_thread_data, + dst_wave_buffer_resource, + dst_thread_addr_offset, + 0, + dst_linear_addr_offset, + dst_thread_element_valid); + } + else + { + buffer_atomic_add{}(src_thread_data, + dst_wave_buffer_resource, + dst_thread_addr_offset, + 0, + dst_linear_addr_offset, + 1); + } +} + // buffer_atomic_max requires: // 1) p_dst_wave must point to global memory // 2) p_dst_wave must be a wavewise pointer. diff --git a/include/ck_tile/core/arch/arch.hpp b/include/ck_tile/core/arch/arch.hpp index 65a3a4e2ff..afcf982a63 100644 --- a/include/ck_tile/core/arch/arch.hpp +++ b/include/ck_tile/core/arch/arch.hpp @@ -73,6 +73,24 @@ CK_TILE_DEVICE void block_sync_lds() #endif } +CK_TILE_DEVICE void block_sync_load_raw(index_t cnt = 0) +{ +#ifdef __gfx12__ + asm volatile("s_wait_loadcnt %0 \n" + "s_barrier_signal -1 \n" + "s_barrier_wait -1" + : + : "n"(cnt) + : "memory"); +#else + asm volatile("s_waitcnt vmcnt(%0) \n" + "s_barrier" + : + : "n"(cnt) + : "memory"); +#endif +} + CK_TILE_DEVICE void block_sync_lds_direct_load() { asm volatile("\ diff --git a/include/ck_tile/core/arch/utility.hpp b/include/ck_tile/core/arch/utility.hpp index a88780459b..df0f54c5ed 100644 --- a/include/ck_tile/core/arch/utility.hpp +++ b/include/ck_tile/core/arch/utility.hpp @@ -102,4 +102,28 @@ CK_TILE_DEVICE T warp_shuffle(const T& v_local, uint32_t src_lane) #endif } +template +CK_TILE_DEVICE auto flag_to_exec(const T& v_flag) +{ + static_assert(sizeof(T) == 4); + // per-thread v_flag store into 2x sgpr + uint32x2_t exec_flag; + asm volatile("v_cmp_ge_u32 %[s_exec_flag], %[v_flag], 1" + : [s_exec_flag] "=s"(exec_flag) + : [v_flag] "v"(v_flag)); + return exec_flag; +} + +template +CK_TILE_DEVICE auto cmp_lt_to_exec(const X& x, const Y& y) +{ + static_assert(sizeof(X) == 4 && sizeof(Y) == 4); + // per-thread cmp store into 2x sgpr + uint32x2_t exec_flag; + asm volatile("v_cmp_lt_u32 %[s_exec_flag], %[v_x], %[v_y]" + : [s_exec_flag] "=s"(exec_flag) + : [v_x] "v"(x), [v_y] "v"(y)); + return exec_flag; +} + } // namespace ck_tile diff --git a/include/ck_tile/core/tensor/buffer_view.hpp b/include/ck_tile/core/tensor/buffer_view.hpp index 2cc788d422..7dffa0e555 100644 --- a/include/ck_tile/core/tensor/buffer_view.hpp +++ b/include/ck_tile/core/tensor/buffer_view.hpp @@ -437,34 +437,74 @@ struct buffer_view>::scalar_type, typename vector_traits>::scalar_type>::value, bool>::type = false> - CK_TILE_DEVICE void update(index_t i, index_t linear_offset, bool is_valid_element, const X& x) + CK_TILE_DEVICE void update(index_t i, + index_t linear_offset, + bool is_valid_element, + const X& x, + bool_constant = {}) { if constexpr(Op == memory_operation_enum::set) { - this->template set(i, linear_offset, is_valid_element, x); + this->template set(i, linear_offset, is_valid_element, x); } else if constexpr(Op == memory_operation_enum::atomic_add) { - this->template atomic_add(i, linear_offset, is_valid_element, x); + this->template atomic_add( + i, linear_offset, is_valid_element, x); } else if constexpr(Op == memory_operation_enum::atomic_max) { - this->template atomic_max(i, linear_offset, is_valid_element, x); + this->template atomic_max( + i, linear_offset, is_valid_element, x); } // FIXME: remove memory_operation_enum::add else if constexpr(Op == memory_operation_enum::add) { - auto tmp = this->template get(i, linear_offset, is_valid_element); - this->template set(i, linear_offset, is_valid_element, x + tmp); + auto tmp = + this->template get(i, linear_offset, is_valid_element); + this->template set( + i, linear_offset, is_valid_element, x + tmp); // tmp += x; // this->template set(i, is_valid_element, tmp); } } + // i is offset of T, not X. i should be aligned to X + template >::scalar_type, + typename vector_traits>::scalar_type>::value, + bool>::type = false> + CK_TILE_DEVICE void update_raw(index_t i, + index_t linear_offset, + bool is_valid_element, + const X& x, + bool_constant = {}, + bool_constant = {}) + { + if constexpr(Op == memory_operation_enum::set) + { + this->template set_raw(i, linear_offset, is_valid_element, x); + } + else if constexpr(Op == memory_operation_enum::atomic_add) + { + this->template atomic_add_raw( + i, linear_offset, is_valid_element, x); + } + else if constexpr(Op == memory_operation_enum::atomic_max) + { + // this->template atomic_max_raw(i, linear_offset, is_valid_element, x); + } + } + // i is offset of T, not X. i should be aligned to X template >::scalar_type, typename vector_traits>::scalar_type>::value, @@ -585,6 +626,39 @@ struct buffer_view>::scalar_type, + typename vector_traits>::scalar_type>::value, + bool>::type = false> + CK_TILE_DEVICE void + atomic_add_raw(index_t i, index_t linear_offset, bool is_valid_element, const X& x) + { + // using scalar_t = typename vector_traits>::scalar_type; + + // X contains multiple T + constexpr index_t scalar_per_t_vector = vector_traits>::vector_size; + + constexpr index_t scalar_per_x_vector = vector_traits>::vector_size; + + static_assert(scalar_per_x_vector % scalar_per_t_vector == 0, + "wrong! X should contain multiple T"); + + static_assert(get_address_space() == address_space_enum::global, "only support global mem"); + + constexpr index_t t_per_x = scalar_per_x_vector / scalar_per_t_vector; + + amd_buffer_atomic_add_raw, + t_per_x, + Coherence, + oob_conditional_check, + pre_nop>( + x, p_data_, i, linear_offset, is_valid_element, buffer_size_); + } + + template >::scalar_type, typename vector_traits>::scalar_type>::value, diff --git a/include/ck_tile/core/tensor/load_tile.hpp b/include/ck_tile/core/tensor/load_tile.hpp index f150fc54ca..b280a1725d 100644 --- a/include/ck_tile/core/tensor/load_tile.hpp +++ b/include/ck_tile/core/tensor/load_tile.hpp @@ -22,28 +22,32 @@ template CK_TILE_DEVICE auto load_tile(const tile_window_with_static_distribution& tile_window, + number = {}, bool_constant = {}) { - return tile_window.load(number<-1>{}, bool_constant{}); + return tile_window.load(number{}, bool_constant{}); } template CK_TILE_DEVICE auto load_tile(const tile_window_linear& tile_window, + number = {}, bool_constant = {}) { - return tile_window.load(number<-1>{}, bool_constant{}); + return tile_window.load(number{}, bool_constant{}); } template CK_TILE_DEVICE auto load_tile(DistributedTensor_& dst_tile, const tile_window_with_static_distribution& tile_window, + number = {}, bool_constant = {}) { - return tile_window.load(dst_tile, bool_constant{}); + return tile_window.load(dst_tile, number{}, bool_constant{}); +} + +template +CK_TILE_DEVICE auto load_tile(DistributedTensor_& dst_tile, + const tile_window_linear& tile_window, + number = {}, + bool_constant = {}) +{ + return tile_window.load(dst_tile, number{}, bool_constant{}); } /** @@ -76,6 +100,7 @@ template CK_TILE_DEVICE auto load_tile_raw(T& tile, @@ -83,11 +108,12 @@ CK_TILE_DEVICE auto load_tile_raw(T& tile, WindowLengths_, TileDistribution_, NumCoord>& tile_window, + number = {}, bool_constant = {}, bool_constant = {}) { tile_window.load_raw( - tile, number<-1>{}, bool_constant{}, bool_constant{}); + tile, number{}, bool_constant{}, bool_constant{}); } template CK_TILE_DEVICE auto load_tile_raw(T& tile, @@ -102,11 +129,12 @@ CK_TILE_DEVICE auto load_tile_raw(T& tile, WindowLengths_, TileDistribution_, LinearBottomDims_>& tile_window, + number = {}, bool_constant = {}, bool_constant = {}) { tile_window.load_raw( - tile, number<-1>{}, bool_constant{}, bool_constant{}); + tile, number{}, bool_constant{}, bool_constant{}); } template CK_TILE_DEVICE auto @@ -122,11 +151,14 @@ async_load_tile_raw(LdsTileWindow_&& lds_tile, WindowLengths_, TileDistribution_, NumCoord>& tile_window, + number = {}, bool_constant = {}, bool_constant = {}) { - return tile_window.async_load_raw( - lds_tile, number<-1>{}, bool_constant{}, bool_constant{}); + return tile_window.async_load_raw(lds_tile, + number{}, + bool_constant{}, + bool_constant{}); } template CK_TILE_DEVICE auto async_load_tile_raw(LdsTileWindow_&& lds_tile, @@ -141,11 +174,14 @@ CK_TILE_DEVICE auto async_load_tile_raw(LdsTileWindow_&& lds_tile, WindowLengths_, TileDistribution_, LinearBottomDims_>& tile_window, + number = {}, bool_constant = {}, bool_constant = {}) { - return tile_window.async_load_raw( - lds_tile, number<-1>{}, bool_constant{}, bool_constant{}); + return tile_window.async_load_raw(lds_tile, + number{}, + bool_constant{}, + bool_constant{}); } CK_TILE_DEVICE auto async_load_fence(index_t cnt = 0) diff --git a/include/ck_tile/core/tensor/static_distributed_tensor.hpp b/include/ck_tile/core/tensor/static_distributed_tensor.hpp index 29c20bed00..568d618ec2 100644 --- a/include/ck_tile/core/tensor/static_distributed_tensor.hpp +++ b/include/ck_tile/core/tensor/static_distributed_tensor.hpp @@ -201,4 +201,30 @@ CK_TILE_HOST_DEVICE constexpr auto get_y_unpacks_from_x_unpacks(YLengths, number return unpacks; } +namespace detail { + +// check if 2 static_distributed_tensor has same data type and size of element +// but only difference in distribution +template +struct is_similiar_distributed_tensor +{ + static constexpr bool value = false; +}; + +template +struct is_similiar_distributed_tensor, + static_distributed_tensor> +{ + using Tx = static_distributed_tensor; + using Ty = static_distributed_tensor; + static constexpr bool value = std::is_same_v && + Tx::get_thread_buffer_size() == Ty::get_thread_buffer_size(); +}; + +template +inline constexpr bool is_similiar_distributed_tensor_v = + is_similiar_distributed_tensor::value; + +} // namespace detail + } // namespace ck_tile diff --git a/include/ck_tile/core/tensor/tensor_view.hpp b/include/ck_tile/core/tensor/tensor_view.hpp index 698ce5378d..4c72ed0859 100644 --- a/include/ck_tile/core/tensor/tensor_view.hpp +++ b/include/ck_tile/core/tensor/tensor_view.hpp @@ -333,6 +333,48 @@ struct tensor_view coord.get_offset(), linear_offset, is_valid_element, x); } + // X is vector of DataType. + // "coord" is coordinate of DataType, not X. "coord" should be aligned to X + template >::scalar_type, + typename vector_traits>::scalar_type>, + bool>::type = false> + CK_TILE_HOST_DEVICE constexpr void + update_vectorized_elements_raw(const TensorCoord& coord, + index_t linear_offset, + const X& x, + bool_constant = {}, + bool_constant = {}) + { + buf_.template update_raw( + coord.get_offset(), + linear_offset, + coordinate_has_valid_offset_assuming_top_index_is_valid(desc_, coord), + x); + } + + template >::scalar_type, + typename vector_traits>::scalar_type>, + bool>::type = false> + CK_TILE_HOST_DEVICE constexpr void + update_vectorized_elements_raw(const TensorCoord& coord, + index_t linear_offset, + bool is_valid_element, + const X& x, + bool_constant = {}, + bool_constant = {}) + { + buf_.template update_raw( + coord.get_offset(), linear_offset, is_valid_element, x); + } + CK_TILE_HOST_DEVICE void print() const { printf("tensor_view{"); diff --git a/include/ck_tile/core/tensor/tile_window.hpp b/include/ck_tile/core/tensor/tile_window.hpp index e410246983..caeb038521 100644 --- a/include/ck_tile/core/tensor/tile_window.hpp +++ b/include/ck_tile/core/tensor/tile_window.hpp @@ -292,12 +292,15 @@ struct tile_window_with_static_distribution { constexpr auto tile_dstr = TileDstr{}; auto dst_tensor = make_static_distributed_tensor(tile_dstr); - load(dst_tensor, bool_constant{}); + load(dst_tensor, number{}, bool_constant{}); return dst_tensor; } - template + template CK_TILE_DEVICE auto load(DistributedTensor& dst_tensor, + number = {}, bool_constant = {}) const { using Traits = load_store_traits; @@ -785,6 +788,73 @@ struct tile_window_with_static_distribution }); } + template + CK_TILE_DEVICE void update_raw(const static_distributed_tensor& dstr_tensor, + number = {}, + bool_constant = {}, + bool_constant = {}) const + { + using Traits = load_store_traits; + + using vector_t = typename Traits::vector_t; + using SFC_Ys = typename Traits::SFC_Ys; + + constexpr auto tile_dstr = TileDstr{}; + + // loop over thread tensor space [y0, y1, ...] + static_for<0, NumCoord, 1>{}([&](auto iCoord) { + /// TODO: use structure binding (to be captured later) if compiled in C++20 + auto window_adaptor_thread_coord = pre_computed_coords_[iCoord][I0]; + auto bottom_tensor_thread_coord = pre_computed_coords_[iCoord][I1]; + + static_for<0, NumAccessPerCoord, 1>{}([&](auto iCoordAccess) { + constexpr auto iAccess = number{}; + + // data index [y0, y1, ...] + constexpr auto idx_ys_start = SFC_Ys::get_index(iAccess); + + // read from distributed tensor + vector_t vec_value; + + static_for<0, Traits::ScalarPerVector, 1>{}([&](auto j) { + constexpr auto idx_ys = generate_tuple( + [&](auto jj) { + return jj == Traits::VectorDimY ? (idx_ys_start[jj] + j) + : idx_ys_start[jj]; + }, + number{}); + + constexpr index_t d = + tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys); + + vec_value.template get_as()(j) = + dstr_tensor.get_thread_buffer().template at(); + }); + + // write into bottom tensor + get_bottom_tensor_view().template update_vectorized_elements_raw( + bottom_tensor_thread_coord, + 0, + vec_value, + bool_constant{}, + bool_constant{}); + + // move thread coordinate + if constexpr(iCoordAccess != (NumAccessPerCoord - 1)) + { + constexpr auto idx_diff_ys = SFC_Ys::get_forward_step(iAccess); + + constexpr auto idx_diff_ps_ys = container_concat( + generate_tuple([&](auto) { return number<0>{}; }, number{}), + idx_diff_ys); + + move_window_adaptor_and_bottom_tensor_thread_coordinate( + window_adaptor_thread_coord, bottom_tensor_thread_coord, idx_diff_ps_ys); + } + }); + }); + } + // move thread's botom tensor coordiante // [x0', x1', ... ] ==> [offset] // also move window-origin diff --git a/include/ck_tile/core/tensor/tile_window_linear.hpp b/include/ck_tile/core/tensor/tile_window_linear.hpp index 4b921ec5b9..96a8352c04 100644 --- a/include/ck_tile/core/tensor/tile_window_linear.hpp +++ b/include/ck_tile/core/tensor/tile_window_linear.hpp @@ -432,23 +432,38 @@ struct tile_window_linear CK_TILE_DEVICE static constexpr index_t get_bottom_linear_offset(number) { constexpr auto linear_coord = get_bottom_linear_coordinate(number{}); - // since this is linear offset, we assum bottom X tensor is always linear - constexpr index_t linear_offset = [&]() { - constexpr auto x_idx_ = linear_coord; - constexpr auto x_len_ = TileDstr{}.get_lengths(); - static_assert(x_idx_.size() == x_len_.size()); - constexpr index_t x_dims_ = x_idx_.size(); - index_t cu_stride_ = 1; - index_t cu_offset_ = 0; - static_for<0, x_dims_, 1>{}([&](auto i_) { - auto r_i_ = number{}; - cu_offset_ += x_idx_[r_i_] * cu_stride_; - cu_stride_ *= x_len_[r_i_]; - }); - return cu_offset_; - }(); - - return linear_offset; + constexpr auto is_pure_linear_tensor = + reduce_on_sequence(LinearBottomDims{}, multiplies{}, number<1>{}); + if constexpr(is_pure_linear_tensor) + { + // this case usually is a LDS window, everything is known at compile tile. + // we directly use BottomTensorView transform to compute the offset, in case padding + auto bottom_tensor_coord = + make_tensor_coordinate(BottomTensorView{}.get_tensor_descriptor(), linear_coord); + return bottom_tensor_coord.get_offset(); + } + else + { + // this case usually is a global window, where last dim can be linear + // we hack here, that use the original TileDstr to compute the linear offset + // ... hoping that there is no extra padding between other dims, which make sense + // since that would introduce runtime length (so can't use linear offset) + constexpr index_t linear_offset = [&]() { + constexpr auto x_idx_ = linear_coord; + constexpr auto x_len_ = TileDstr{}.get_lengths(); + static_assert(x_idx_.size() == x_len_.size()); + constexpr index_t x_dims_ = x_idx_.size(); + index_t cu_stride_ = 1; + index_t cu_offset_ = 0; + static_for<0, x_dims_, 1>{}([&](auto i_) { + auto r_i_ = number{}; + cu_offset_ += x_idx_[r_i_] * cu_stride_; + cu_stride_ *= x_len_[r_i_]; + }); + return cu_offset_; + }(); + return linear_offset; + } } CK_TILE_DEVICE constexpr auto get_num_of_access() const { return traits::NumAccess; } @@ -509,6 +524,64 @@ struct tile_window_linear return dst_tensor; } + template + CK_TILE_DEVICE auto load(DstTile& dst_tensor, + number = {}, + bool_constant = {}) const + { + using vector_t = typename traits::vector_t; + using SFC_Ys = typename traits::SFC_Ys; + + constexpr auto tile_dstr = TileDstr{}; + + // auto dst_tensor = make_static_distributed_tensor(tile_dstr); + + auto issue = [&](auto i_access_) { + constexpr auto IAccess = number{}; + + constexpr auto non_linear_id = number{}; + auto bottom_tensor_thread_coord = cached_coords_[non_linear_id]; + auto bottom_tensor_flag = cached_flags_[IAccess]; + + constexpr auto linear_offset = get_bottom_linear_offset(IAccess); + + // read from bottom tensor + const vector_t vec_value = + get_bottom_tensor_view().template get_vectorized_elements( + bottom_tensor_thread_coord, + linear_offset, + bottom_tensor_flag, + bool_constant{}); +#if 1 + // data index [y0, y1, ...] + constexpr auto idx_diff_ys = SFC_Ys::get_index(IAccess); + // write into distributed tensor + static_for<0, traits::ScalarPerVector, 1>{}([&](auto j) { + constexpr auto idx_ys = generate_tuple( + [&](auto jj) { + return jj == traits::VectorDimY ? (idx_diff_ys[jj] + j) : idx_diff_ys[jj]; + }, + number{}); + + constexpr index_t d = tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys); + + dst_tensor.get_thread_buffer().template at() = + vec_value.template get_as()[j]; + }); +#else + constexpr index_t d = tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys_start); + static_assert(d % traits::ScalarPerVector == 0); + + dst_tensor.get_thread_buffer().template get_as()( + number{}) = bit_cast(vec_value); +#endif + }; + + WINDOW_DISPATCH_ISSUE(); + + return dst_tensor; + } + template + CK_TILE_DEVICE void update_raw(const static_distributed_tensor& dstr_tensor, + number = {}, + bool_constant = {}, + bool_constant = {}) const + { + + using vector_t = typename traits::vector_t; + using SFC_Ys = typename traits::SFC_Ys; + + constexpr auto tile_dstr = TileDstr{}; + + // loop over thread tensor space [y0, y1, ...] + auto issue = [&](auto i_access_) { + constexpr auto IAccess = number{}; + constexpr auto non_linear_id = number{}; + auto bottom_tensor_thread_coord = cached_coords_[non_linear_id]; + constexpr auto linear_offset = get_bottom_linear_offset(IAccess); + auto bottom_tensor_flag = cached_flags_[IAccess]; + + // data index [y0, y1, ...] + constexpr auto idx_ys_start = SFC_Ys::get_index(IAccess); + + // read from distributed tensor + vector_t vec_value; + + static_for<0, traits::ScalarPerVector, 1>{}([&](auto j) { + constexpr auto idx_ys = generate_tuple( + [&](auto jj) { + return jj == traits::VectorDimY ? (idx_ys_start[jj] + j) : idx_ys_start[jj]; + }, + number{}); + + constexpr index_t d = tile_dstr.get_ys_to_d_descriptor().calculate_offset(idx_ys); + + vec_value.template get_as()(j) = + dstr_tensor.get_thread_buffer().template at(); + }); + + // write into bottom tensor + get_bottom_tensor_view().template update_vectorized_elements_raw( + bottom_tensor_thread_coord, + linear_offset, + bottom_tensor_flag, + vec_value, + bool_constant{}, + bool_constant{}); + }; + + WINDOW_DISPATCH_ISSUE(); + } + // move thread's botom tensor coordiante // [x0', x1', ... ] ==> [offset] // also move window-origin diff --git a/include/ck_tile/core/tensor/tile_window_utils.hpp b/include/ck_tile/core/tensor/tile_window_utils.hpp new file mode 100644 index 0000000000..71a72329f8 --- /dev/null +++ b/include/ck_tile/core/tensor/tile_window_utils.hpp @@ -0,0 +1,54 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#include "ck_tile/core/arch/arch.hpp" +#include "ck_tile/core/arch/utility.hpp" +#include "ck_tile/core/algorithm/space_filling_curve.hpp" +#include "ck_tile/core/config.hpp" +#include "ck_tile/core/container/array.hpp" +#include "ck_tile/core/container/sequence.hpp" +#include "ck_tile/core/container/tuple.hpp" +#include "ck_tile/core/container/container_helper.hpp" +#include "ck_tile/core/tensor/static_distributed_tensor.hpp" +#include "ck_tile/core/tensor/tensor_adaptor.hpp" +#include "ck_tile/core/tensor/tile_distribution.hpp" +#include "ck_tile/core/utility/functional.hpp" +#include "ck_tile/core/utility/type_traits.hpp" + +#pragma once +namespace ck_tile { + +// input a lds store tile, extract some information from it +// used to set m0 value for gfx9 serious +template +CK_TILE_DEVICE auto get_async_store_smem_info(LdsTileWindow_&& lds_tile) +{ + using LdsTileWindow = remove_cvref_t; + using LdsDataType = typename LdsTileWindow::DataType; + + // issues * warps * lanes + static_assert(LdsTileWindow::get_num_of_dimension() == 3); // TODO: hard coded + + const index_t size_per_buf = + lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset( + make_tuple(number<0>{}, number<0>{}, number<0>{})) * + sizeof(LdsDataType); + + const index_t size_per_wave = + lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset( + make_tuple(number<0>{}, number<1>{}, number<0>{})) * + sizeof(LdsDataType) - + size_per_buf; + + const index_t size_per_issue = + lds_tile.get_bottom_tensor_view().get_tensor_descriptor().calculate_offset( + make_tuple(number<1>{}, number<0>{}, number<0>{})) * + sizeof(LdsDataType) - + size_per_buf; + + const index_t m0_init_value = size_per_buf + size_per_wave * get_warp_id(); + + return make_tuple(m0_init_value, size_per_issue); +} + +} // namespace ck_tile diff --git a/include/ck_tile/core/tensor/update_tile.hpp b/include/ck_tile/core/tensor/update_tile.hpp index fbce7c4083..570abde189 100644 --- a/include/ck_tile/core/tensor/update_tile.hpp +++ b/include/ck_tile/core/tensor/update_tile.hpp @@ -41,15 +41,65 @@ template + typename DataType_, + index_t i_access = -1, + bool oob_conditional_check = true> CK_TILE_DEVICE void update_tile(tile_window_with_static_distribution& tile_window, - const static_distributed_tensor& dstr_tensor) + const static_distributed_tensor& dstr_tensor, + number = {}, + bool_constant = {}) { - tile_window.update(dstr_tensor); + tile_window.update(dstr_tensor, number{}, bool_constant{}); +} + +template +CK_TILE_DEVICE void +update_tile_raw(tile_window_with_static_distribution& tile_window, + const static_distributed_tensor& dstr_tensor, + number = {}, + bool_constant = {}, + bool_constant = {}) +{ + tile_window.update_raw(dstr_tensor, + number{}, + bool_constant{}, + bool_constant{}); +} + +template +CK_TILE_DEVICE auto update_tile_raw( + tile_window_linear& + tile_window, + const static_distributed_tensor& dstr_tensor, + number = {}, + bool_constant = {}, + bool_constant = {}) +{ + tile_window.update_raw(dstr_tensor, + number{}, + bool_constant{}, + bool_constant{}); } } // namespace ck_tile diff --git a/include/ck_tile/core/utility/static_counter.hpp b/include/ck_tile/core/utility/static_counter.hpp new file mode 100644 index 0000000000..84af3dd52f --- /dev/null +++ b/include/ck_tile/core/utility/static_counter.hpp @@ -0,0 +1,116 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core/config.hpp" + +namespace ck_tile { + +template +struct static_counter +{ + public: + template + static constexpr index_t next() + { + return next(0) * Step + Start; + } + + template + static constexpr index_t next() + { + struct Unique + { + }; + return next(0) * Step + Start; + } + + template + static constexpr index_t current() + { + return current(0) * Step + Start; + } + + template + static constexpr index_t current() + { + struct Unique + { + }; + return current(0) * Step + Start; + } + + private: + template + struct slot + { + _Pragma("GCC diagnostic push"); + _Pragma("GCC diagnostic ignored \"-Wundefined-internal\""); + friend constexpr bool slot_allocated(slot); + _Pragma("GCC diagnostic pop"); + }; + + template + struct allocate_slot + { + friend constexpr bool slot_allocated(slot) { return true; } + enum + { + value = I + }; + }; + + // If slot_allocated(slot) has NOT been defined, then SFINAE will keep this function out of + // the overload set... + template ())> + static constexpr index_t next(index_t) + { + return next(0); + } + + // ...And this function will be used, instead, which will define slot_allocated(slot) via + // allocate_slot. + template + static constexpr index_t next(double) + { + return allocate_slot::value; + } + + // If slot_allocated(slot) has NOT been defined, then SFINAE will keep this function out of + // the overload set... + template ())> + static constexpr index_t current(index_t) + { + return current(0); + } + + // ...And this function will be used, instead, which will return the current counter, or assert + // in case next() hasn't been called yet. + template + static constexpr index_t current(double) + { + static_assert(I != 0, "You must invoke next() first"); + + return I - 1; + } +}; + +namespace impl { +template +struct static_counter_uniq_; +} + +#define MAKE_SC() \ + ck_tile::static_counter> {} +#define MAKE_SC_WITH(start_, step_) \ + ck_tile::static_counter, start_, step_> {} +#define NEXT_SC(c_) c_.next<__COUNTER__>() +#define NEXT_SCI(c_, static_i_) c_.next<__COUNTER__ + static_i_>() + +// Usage: +// constexpr auto c = MAKE_SC() +// NEXT_SC(c) // -> constexpr 0 +// NEXT_SC(c) // -> constexpr 1 +// NEXT_SC(c) // -> constexpr 2 +} // namespace ck_tile diff --git a/include/ck_tile/host.hpp b/include/ck_tile/host.hpp index 2e96009ace..2f3a302eea 100644 --- a/include/ck_tile/host.hpp +++ b/include/ck_tile/host.hpp @@ -11,6 +11,7 @@ #include "ck_tile/host/fill.hpp" #include "ck_tile/host/hip_check_error.hpp" #include "ck_tile/host/host_tensor.hpp" +#include "ck_tile/host/joinable_thread.hpp" #include "ck_tile/host/kernel_launch.hpp" #include "ck_tile/host/ranges.hpp" #include "ck_tile/host/reference/reference_batched_dropout.hpp" @@ -20,6 +21,7 @@ #include "ck_tile/host/reference/reference_batched_rotary_position_embedding.hpp" #include "ck_tile/host/reference/reference_batched_softmax.hpp" #include "ck_tile/host/reference/reference_elementwise.hpp" +#include "ck_tile/host/reference/reference_fused_moe.hpp" #include "ck_tile/host/reference/reference_gemm.hpp" #include "ck_tile/host/reference/reference_im2col.hpp" #include "ck_tile/host/reference/reference_layernorm2d_fwd.hpp" diff --git a/include/ck_tile/host/device_memory.hpp b/include/ck_tile/host/device_memory.hpp index 7c8549f74f..13684c0e24 100644 --- a/include/ck_tile/host/device_memory.hpp +++ b/include/ck_tile/host/device_memory.hpp @@ -7,6 +7,7 @@ #include #include #include "ck_tile/host/hip_check_error.hpp" +#include "ck_tile/host/host_tensor.hpp" namespace ck_tile { template @@ -36,6 +37,19 @@ struct DeviceMem mpDeviceBuf = nullptr; } } + template + DeviceMem(const HostTensor& t) : mMemSize(t.get_element_space_size_in_bytes()) + { + if(mMemSize != 0) + { + HIP_CHECK_ERROR(hipMalloc(static_cast(&mpDeviceBuf), mMemSize)); + } + else + { + mpDeviceBuf = nullptr; + } + ToDevice(t.data()); + } void Realloc(std::size_t mem_size) { if(mpDeviceBuf) @@ -92,6 +106,27 @@ struct DeviceMem HIP_CHECK_ERROR(hipMemcpy(p, mpDeviceBuf, cpySize, hipMemcpyDeviceToHost)); } } + + // construct a host tensor with type T + template + HostTensor ToHost(std::size_t cpySize) + { + // TODO: host tensor could be slightly larger than the device tensor + // we just copy all data from GPU buffer + std::size_t host_elements = (cpySize + sizeof(T) - 1) / sizeof(T); + HostTensor h_({host_elements}); + if(mpDeviceBuf) + { + HIP_CHECK_ERROR(hipMemcpy(h_.data(), mpDeviceBuf, cpySize, hipMemcpyDeviceToHost)); + } + return h_; + } + template + HostTensor ToHost() + { + return ToHost(mMemSize); + } + void SetZero() const { if(mpDeviceBuf) diff --git a/include/ck_tile/host/fill.hpp b/include/ck_tile/host/fill.hpp index 335911860a..f24c338755 100644 --- a/include/ck_tile/host/fill.hpp +++ b/include/ck_tile/host/fill.hpp @@ -13,6 +13,7 @@ #include #include "ck_tile/core.hpp" +#include "ck_tile/host/joinable_thread.hpp" namespace ck_tile { @@ -22,13 +23,44 @@ struct FillUniformDistribution float a_{-5.f}; float b_{5.f}; std::optional seed_{11939}; + // ATTENTION: threaded does not guarantee the distribution between thread + bool threaded = false; template void operator()(ForwardIter first, ForwardIter last) const { - std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}()); - std::uniform_real_distribution dis(a_, b_); - std::generate(first, last, [&dis, &gen]() { return ck_tile::type_convert(dis(gen)); }); + if(threaded) + { + uint32_t num_thread = std::thread::hardware_concurrency(); + auto total = static_cast(std::distance(first, last)); + auto work_per_thread = static_cast((total + num_thread - 1) / num_thread); + + std::vector threads(num_thread); + for(std::size_t it = 0; it < num_thread; ++it) + { + std::size_t iw_begin = it * work_per_thread; + std::size_t iw_end = std::min((it + 1) * work_per_thread, total); + auto thread_f = [this, total, iw_begin, iw_end, &first] { + if(iw_begin > total || iw_end > total) + return; + // need to make each thread unique, add an offset to current seed + std::mt19937 gen(seed_.has_value() ? (*seed_ + iw_begin) + : std::random_device{}()); + std::uniform_real_distribution dis(a_, b_); + std::generate(first + iw_begin, first + iw_end, [&dis, &gen]() { + return ck_tile::type_convert(dis(gen)); + }); + }; + threads[it] = joinable_thread(thread_f); + } + } + else + { + std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}()); + std::uniform_real_distribution dis(a_, b_); + std::generate( + first, last, [&dis, &gen]() { return ck_tile::type_convert(dis(gen)); }); + } } template @@ -115,13 +147,44 @@ struct FillNormalDistribution float mean_{0.f}; float variance_{1.f}; std::optional seed_{11939}; + // ATTENTION: threaded does not guarantee the distribution between thread + bool threaded = false; template void operator()(ForwardIter first, ForwardIter last) const { - std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}()); - std::normal_distribution dis(mean_, std::sqrt(variance_)); - std::generate(first, last, [&dis, &gen]() { return ck_tile::type_convert(dis(gen)); }); + if(threaded) + { + uint32_t num_thread = std::thread::hardware_concurrency(); + auto total = static_cast(std::distance(first, last)); + auto work_per_thread = static_cast((total + num_thread - 1) / num_thread); + + std::vector threads(num_thread); + for(std::size_t it = 0; it < num_thread; ++it) + { + std::size_t iw_begin = it * work_per_thread; + std::size_t iw_end = std::min((it + 1) * work_per_thread, total); + auto thread_f = [this, total, iw_begin, iw_end, &first] { + if(iw_begin > total || iw_end > total) + return; + // need to make each thread unique, add an offset to current seed + std::mt19937 gen(seed_.has_value() ? (*seed_ + iw_begin) + : std::random_device{}()); + std::normal_distribution dis(mean_, std::sqrt(variance_)); + std::generate(first + iw_begin, first + iw_end, [&dis, &gen]() { + return ck_tile::type_convert(dis(gen)); + }); + }; + threads[it] = joinable_thread(thread_f); + } + } + else + { + std::mt19937 gen(seed_.has_value() ? *seed_ : std::random_device{}()); + std::normal_distribution dis(mean_, std::sqrt(variance_)); + std::generate( + first, last, [&dis, &gen]() { return ck_tile::type_convert(dis(gen)); }); + } } template @@ -235,6 +298,44 @@ struct FillMonotonicSeq } }; +template +struct FillStepRange +{ + float start_value_{0}; + float end_value_{3}; + float step_{1}; + + template + void operator()(ForwardIter first, ForwardIter last) const + { + std::generate(first, last, [=, n = start_value_]() mutable { + auto tmp = n; + n += step_; + if constexpr(IsAscending) + { + if(n > end_value_) + n = start_value_; + } + else + { + if(n < end_value_) + n = start_value_; + } + + return type_convert(tmp); + }); + } + + template + auto operator()(ForwardRange&& range) const -> std::void_t< + decltype(std::declval()(std::begin(std::forward(range)), + std::end(std::forward(range))))> + { + (*this)(std::begin(std::forward(range)), + std::end(std::forward(range))); + } +}; + template struct FillConstant { diff --git a/include/ck_tile/host/host_tensor.hpp b/include/ck_tile/host/host_tensor.hpp index 5610ba324d..3902cad178 100644 --- a/include/ck_tile/host/host_tensor.hpp +++ b/include/ck_tile/host/host_tensor.hpp @@ -8,12 +8,13 @@ #include #include #include -#include #include #include #include +#include #include "ck_tile/core.hpp" +#include "ck_tile/host/joinable_thread.hpp" #include "ck_tile/host/ranges.hpp" namespace ck_tile { @@ -213,23 +214,6 @@ CK_TILE_HOST HostTensorDescriptor transpose_host_tensor_descriptor_given_new2old return HostTensorDescriptor(new_lengths, new_strides); } -struct joinable_thread : std::thread -{ - template - joinable_thread(Xs&&... xs) : std::thread(std::forward(xs)...) - { - } - - joinable_thread(joinable_thread&&) = default; - joinable_thread& operator=(joinable_thread&&) = default; - - ~joinable_thread() - { - if(this->joinable()) - this->join(); - } -}; - template struct ParallelTensorFunctor { @@ -590,6 +574,107 @@ struct HostTensor size() * FromSize / ToSize}; } + friend std::ostream& operator<<(std::ostream& os, const HostTensor& t) + { + os << t.mDesc; + os << "["; + for(typename Data::size_type idx = 0; idx < t.mData.size(); ++idx) + { + if(0 < idx) + { + os << ", "; + } + if constexpr(std::is_same_v || std::is_same_v) + { + os << type_convert(t.mData[idx]) << " #### "; + } + else + { + os << t.mData[idx]; + } + } + os << "]"; + return os; + } + + // read data from a file, as dtype + // the file could dumped from torch as (targeting tensor is t here) + // numpy.savetxt("f.txt", t.view(-1).numpy()) + // numpy.savetxt("f.txt", t.cpu().view(-1).numpy()) # from cuda to cpu to save + // numpy.savetxt("f.txt", t.cpu().view(-1).numpy(), fmt="%d") # save as int + // will output f.txt, each line is a value + // dtype=float or int, internally will cast to real type + void loadtxt(std::string file_name, std::string dtype = "float") + { + std::ifstream file(file_name); + + if(file.is_open()) + { + std::string line; + + index_t cnt = 0; + while(std::getline(file, line)) + { + if(cnt >= static_cast(mData.size())) + { + throw std::runtime_error(std::string("data read from file:") + file_name + + " is too big"); + } + + if(dtype == "float") + { + mData[cnt] = type_convert(std::stof(line)); + } + else if(dtype == "int" || dtype == "int32") + { + mData[cnt] = type_convert(std::stoi(line)); + } + cnt++; + } + file.close(); + if(cnt < static_cast(mData.size())) + { + std::cerr << "Warning! reading from file:" << file_name + << ", does not match the size of this tensor" << std::endl; + } + } + else + { + // Print an error message to the standard error + // stream if the file cannot be opened. + throw std::runtime_error(std::string("unable to open file:") + file_name); + } + } + + // can save to a txt file and read from torch as: + // torch.from_numpy(np.loadtxt('f.txt', dtype=np.int32/np.float32...)).view([...]).contiguous() + void savetxt(std::string file_name, std::string dtype = "float") + { + std::ofstream file(file_name); + + if(file.is_open()) + { + for(auto& itm : mData) + { + if(dtype == "float") + file << type_convert(itm) << std::endl; + else if(dtype == "int") + file << type_convert(itm) << std::endl; + else + // TODO: we didn't implement operator<< for all custom + // data types, here fall back to float in case compile error + file << type_convert(itm) << std::endl; + } + file.close(); + } + else + { + // Print an error message to the standard error + // stream if the file cannot be opened. + throw std::runtime_error(std::string("unable to open file:") + file_name); + } + } + Descriptor mDesc; Data mData; }; diff --git a/include/ck_tile/host/joinable_thread.hpp b/include/ck_tile/host/joinable_thread.hpp new file mode 100644 index 0000000000..a822f967dc --- /dev/null +++ b/include/ck_tile/host/joinable_thread.hpp @@ -0,0 +1,27 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include + +namespace ck_tile { + +struct joinable_thread : std::thread +{ + template + joinable_thread(Xs&&... xs) : std::thread(std::forward(xs)...) + { + } + + joinable_thread(joinable_thread&&) = default; + joinable_thread& operator=(joinable_thread&&) = default; + + ~joinable_thread() + { + if(this->joinable()) + this->join(); + } +}; +} // namespace ck_tile diff --git a/include/ck_tile/host/reference/reference_fused_moe.hpp b/include/ck_tile/host/reference/reference_fused_moe.hpp new file mode 100644 index 0000000000..bf89f92759 --- /dev/null +++ b/include/ck_tile/host/reference/reference_fused_moe.hpp @@ -0,0 +1,196 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/host/host_tensor.hpp" + +namespace ck_tile { +// [indexing implementation-1] +// using M_a as constexpr block_size to partition all tokens into different slices +// each slice map to one expert, and one expert can have multiple slices +// e.g. num_experts = 6, topk=3, M_a = 4, input_tokens = 5 +// before sort, topk_ids is : [[0, 3, 5], [2, 3, 5], [1, 3, 5], [1, 2, 3], [1, 3, 5]] +// tok-0 tok-1 tok-2 tok-3 tok-4 +// topk_weight is : [[a, b, c], [d, e, f], [g, h, i], [j, k, l], [m, n, o]] (some float +// number) +// +// token_id_per_expert is : [[0], [2, 3, 4], [1, 3], [0, 1, 2, 3, 4], [], [0, 1, 2, 5]] +// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5 +// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]] +// +// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1) +// max_num_tokens_padded : topk * input_tokens + num_experts * M_a - topk (updated) +// * this could be larger than actual, since actual tokens are on GPU +// +// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, +// 0, 1, 2, 5] +// |- exp-0 -|- exp-1 -|- exp-2 -|- exp-3 -|- exp-4 +// -|- exp-5 -| +// sorted_weight_ptr : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *, +// c, f, i, o] +// +// * length is max_num_tokens_padded, actual size is num_tokens_post_padded_ptr +// +// sorted_expert_ids_ptr : [0, 1, 2, 3, 3, 4, 5] +// * length is (max_num_tokens_padded + block_size - 1) / block_size +/// +// num_tokens_post_padded_ptr : [28] +// num_sorted_tiles_ptr : [7] + +template +void reference_fused_moe( + const ck_tile::HostTensor& a_host, // [tokens, hidden_size] + const ck_tile::HostTensor& g_host, // [experts, interme_size_0, hidden_size] + const ck_tile::HostTensor& d_host, // [experts, hidden_size, interme_size_1] + const ck_tile::HostTensor& sa_host, // [tokens, 1], + const ck_tile::HostTensor& sg_host, // [experts, 1, interme_size_0] + const ck_tile::HostTensor& sd_host, // [experts, 1, hidden_size], + const ck_tile::HostTensor& sy_host, // [experts, 1, interme_size_0] + ck_tile::HostTensor& o_host, // [tokens, hidden_size] + const ck_tile::HostTensor& sorted_token_ids_host, // [max_num_tokens_padded] + const ck_tile::HostTensor& sorted_weight_host, // [max_num_tokens_padded] + const ck_tile::HostTensor& + sorted_expert_ids_host, // [(max_num_tokens_padded + block_size - 1) / block_size] + const ck_tile::HostTensor& num_sorted_tiles_host, // [1] + + const ck_tile::HostTensor& + token_ids_host, // [tokens, topk] --> ugly!!! remove in the future + + ck_tile::index_t block_m, + ck_tile::index_t tokens, + ck_tile::index_t experts, + ck_tile::index_t hidden_size, + ck_tile::index_t intermediate_size, // this size is for gate/up + ck_tile::index_t topk, + ck_tile::index_t gate_only) +{ + assert(sorted_token_ids_host.get_num_of_dimension() == 1); + assert(sorted_weight_host.get_num_of_dimension() == 1); + assert(sorted_expert_ids_host.get_num_of_dimension() == 1); + assert(num_sorted_tiles_host.get_element_size() == 1); + ck_tile::index_t num_sorted_tiles = num_sorted_tiles_host.mData[0] / block_m; + ck_tile::index_t intermediate_size_0 = intermediate_size; + ck_tile::index_t intermediate_size_1 = intermediate_size / (gate_only ? 1 : 2); + + // TODO: better remove this in the future, or modify the token_id value + auto get_topk_id = [&](ck_tile::index_t token_id_, ck_tile::index_t expert_id_) { + for(ck_tile::index_t i_ = 0; i_ < topk; i_++) + { + if(token_ids_host(token_id_, i_) == expert_id_) + return i_; + } + throw std::runtime_error("not correct token/expert pair\n"); + return -1; // TODO: not correct!! + }; + + ck_tile::HostTensor out_topk_tokens({tokens, topk, hidden_size}); + + int max_num_tokens_padded = topk * tokens + experts * block_m - topk; + // assert(); + auto f = [&](auto i_flatten) { + ck_tile::index_t i_tile = i_flatten / block_m; + if(i_tile >= num_sorted_tiles) + return; + ck_tile::index_t i_expert = sorted_expert_ids_host.mData[i_tile]; + ck_tile::index_t i_token = sorted_token_ids_host.mData[i_flatten]; + if(i_token >= tokens) + return; + ck_tile::index_t i_topk = get_topk_id(i_token, i_expert); // TODO: ugly + auto weight = sorted_weight_host.mData[i_flatten]; + + ck_tile::HostTensor acc_0({1, intermediate_size_0}); + // first gemm + for(ck_tile::index_t i_n = 0; i_n < intermediate_size_0; i_n++) + { + AccDataType acc = static_cast(0); + for(ck_tile::index_t i_k = 0; i_k < hidden_size; i_k++) + { + acc += type_convert(a_host(i_token, i_k)) * + type_convert(g_host(i_expert, i_n, i_k)); + } + acc_0(0, i_n) = acc; + // printf("ie:%2d, it:%3d, in:%d, %f\n", i_expert, i_token, i_n, acc); + } + + ck_tile::HostTensor y({1, intermediate_size_1}); + if(gate_only) + { + if(intermediate_size_1 != intermediate_size_0) + throw std::runtime_error( + "intermediate_size not correct, 0:" + std::to_string(intermediate_size_0) + + ", 1:" + std::to_string(intermediate_size_1)); + for(ck_tile::index_t i_n = 0; i_n < intermediate_size_1; i_n++) + { + Activation{}(y(0, i_n), acc_0(0, i_n)); + // printf("ie:%2d, it:%3d, in:%d, %f\n", i_expert, i_token, i_n, y(0, i_n)); + } + } + else + { + if(intermediate_size_1 * 2 != intermediate_size_0) + throw std::runtime_error( + "intermediate_size not correct, 0:" + std::to_string(intermediate_size_0) + + ", 1:" + std::to_string(intermediate_size_1)); + for(ck_tile::index_t i_n = 0; i_n < intermediate_size_1; i_n++) + { + AccDataType tmp; + Activation{}(tmp, acc_0(0, i_n)); + y(0, i_n) = tmp * acc_0(0, i_n + intermediate_size_1); // TODO: elementwise mul + } + } + + // second gemm, loop along gemm-n + ck_tile::HostTensor acc_1({1, hidden_size}); + for(ck_tile::index_t i_n = 0; i_n < hidden_size; i_n++) + { + AccDataType acc = static_cast(0); + for(ck_tile::index_t i_k = 0; i_k < intermediate_size_1; i_k++) + { + acc += y(0, i_k) * type_convert(d_host(i_expert, i_n, i_k)); + } + acc_1(0, i_n) = acc * weight; // multiple weight here + } + + for(ck_tile::index_t i_n = 0; i_n < hidden_size; i_n++) + { + out_topk_tokens(i_token, i_topk, i_n) = acc_1(0, i_n); + } + }; + + // make_ParallelTensorFunctor(f, max_num_tokens_padded)(std::thread::hardware_concurrency()); + make_ParallelTensorFunctor(f, max_num_tokens_padded)(1); + + // reduce + auto r = [&](auto i_token) { + for(ck_tile::index_t i_n = 0; i_n < hidden_size; i_n++) + { + AccDataType acc = type_convert(0); + for(ck_tile::index_t i_topk = 0; i_topk < topk; i_topk++) + { + acc += out_topk_tokens(i_token, i_topk, i_n); + } + o_host(i_token, i_n) = type_convert(acc); + } + }; + make_ParallelTensorFunctor(r, tokens)(std::thread::hardware_concurrency()); + + (void)num_sorted_tiles_host; + (void)sa_host; + (void)sg_host; + (void)sd_host; + (void)sy_host; +} +} // namespace ck_tile diff --git a/include/ck_tile/host/reference/reference_permute.hpp b/include/ck_tile/host/reference/reference_permute.hpp index 14ed4f815e..4e0f1a877e 100644 --- a/include/ck_tile/host/reference/reference_permute.hpp +++ b/include/ck_tile/host/reference/reference_permute.hpp @@ -16,7 +16,7 @@ namespace ck_tile { */ template CK_TILE_HOST void -reference_permute(const HostTensor& x, HostTensor& y, std::vector dims) +reference_permute(const HostTensor& x, HostTensor& y, std::vector perm) { const auto x_len = x.mDesc.get_lengths(); const auto y_len = y.mDesc.get_lengths(); @@ -43,7 +43,7 @@ reference_permute(const HostTensor& x, HostTensor& y, std::v std::vector tmp(rank, 0); for(index_t i = 0; i < rank; i++) { - tmp[dims[i]] = y_coord[i]; + tmp[perm[i]] = y_coord[i]; } return tmp; }(); @@ -54,4 +54,23 @@ reference_permute(const HostTensor& x, HostTensor& y, std::v make_ParallelTensorFunctor(f, x_elm)(std::thread::hardware_concurrency()); } + +template +CK_TILE_HOST auto reference_permute(const HostTensor& x, std::vector perm) +{ + auto x_shape = x.get_lengths(); + ck_tile::index_t rank = perm.size(); + std::vector y_shape = [&]() { + std::vector tmp(rank, 0); + for(int i = 0; i < static_cast(rank); i++) + { + tmp[i] = x_shape[perm[i]]; + } + return tmp; + }(); + + HostTensor y(y_shape); + reference_permute(x, y, perm); + return y; +} } // namespace ck_tile diff --git a/include/ck_tile/ops/elementwise/unary_element_wise_operation.hpp b/include/ck_tile/ops/elementwise/unary_element_wise_operation.hpp index 01217e16ce..e24b1ba767 100644 --- a/include/ck_tile/ops/elementwise/unary_element_wise_operation.hpp +++ b/include/ck_tile/ops/elementwise/unary_element_wise_operation.hpp @@ -572,6 +572,105 @@ struct FastGelu } }; +struct FastGeluAsm +{ + template + CK_TILE_HOST void operator()(Y& y, const X& x) const; + + template + CK_TILE_DEVICE void operator()(Y& y, const X& x) const; + + template <> + CK_TILE_HOST void operator()(float& y, const float& x) const + { + // const float u = -2.f * x * (0.035677f * x * x + 0.797885f); + const float c1 = -2.0 * 0.035677f; + const float c2 = -2.0 * 0.797885f; + const float u = x * (c1 * x * x + c2); + const float emu = exp(u); + y = x / (1.f + emu); + } + + // device code, use lower precision "__ocml_exp_f32" and "rcp" + template <> + CK_TILE_DEVICE void operator()(float& y, const float& x) const + { + const uint32_t c1 = 0xbd92220c; // -2.0 * 0.035677f; + const float c2 = -2.0 * 0.797885f; + const uint32_t log2e_ = 0x3fb8aa3b; // log2e_v; + float tmp; + + asm volatile("v_mul_f32 %[v_tmp], %[v_x], %[v_x] ; x*x\n" + "v_fma_f32 %[v_tmp], %[v_tmp], %[s_c1], %[v_c2] ; c1*x*x+c2\n" + "v_mul_f32 %[v_tmp], %[v_tmp], %[v_x] ; x*(c1*x*x+c2)\n" + "v_mul_f32 %[v_tmp], %[v_tmp], %[s_log2e] ; log2e*x*(c1*x*x+c2)\n" + "v_exp_f32 %[v_tmp], %[v_tmp] ; emu = exp2(log2e*x*(c1*x*x+c2))\n" + "s_nop 0 ; hazard for exp\n" + "v_add_f32 %[v_tmp], %[v_tmp], 1.0 ; emu+1.0f\n" + "v_rcp_f32 %[v_tmp], %[v_tmp] ; 1/(emu+1.0f)\n" + "s_nop 0 ; hazard for rcp \n" + "v_mul_f32 %[v_y], %[v_tmp], %[v_x] ; x * 1/(emu+1f)\n" + : [v_y] "=v"(y), [v_tmp] "+v"(tmp) + : [v_x] "v"(x), [s_c1] "s"(c1), [v_c2] "v"(c2), [s_log2e] "s"(log2e_) + :); + } + + template <> + CK_TILE_HOST void operator()(fp32x2_t& y, const fp32x2_t& x) const + { + const float c1 = -2.0 * 0.035677f; + const float c2 = -2.0 * 0.797885f; + const float u0 = x.x * (c1 * x.x * x.x + c2); + const float emu0 = exp(u0); + y.x = x.x / (1.f + emu0); + const float u1 = x.y * (c1 * x.y * x.y + c2); + const float emu1 = exp(u1); + y.y = x.y / (1.f + emu1); + } + + // this is packed verion to remove data hazard for trans + template <> + CK_TILE_DEVICE void operator()(fp32x2_t& y, const fp32x2_t& x) const + { + const uint32_t c1 = 0xbd92220c; // -2.0 * 0.035677f; + float c2 = -2.0 * 0.797885f; + const uint32_t log2e_ = 0x3fb8aa3b; // log2e_v; + float tmp0, tmp1; + float y0 = x.x, y1 = x.y; + + asm volatile( + "v_mul_f32 %[v_tmp0], %[v_y0], %[v_y0] ; x*x\n" + "v_mul_f32 %[v_tmp1], %[v_y1], %[v_y1] ; x*x\n" + "v_fma_f32 %[v_tmp0], %[v_tmp0], %[s_c1], %[v_c2] ; c1*x*x+c2\n" + "v_fma_f32 %[v_tmp1], %[v_tmp1], %[s_c1], %[v_c2] ; c1*x*x+c2\n" + "v_mul_f32 %[v_tmp0], %[v_tmp0], %[v_y0] ; x*(c1*x*x+c2)\n" + "v_mul_f32 %[v_tmp1], %[v_tmp1], %[v_y1] ; x*(c1*x*x+c2)\n" + "v_mul_f32 %[v_tmp0], %[v_tmp0], %[s_log2e] ; log2e*x*(c1*x*x+c2)\n" + "v_mul_f32 %[v_tmp1], %[v_tmp1], %[s_log2e] ; log2e*x*(c1*x*x+c2)\n" + "v_exp_f32 %[v_tmp0], %[v_tmp0] ; emu = exp2(log2e*x*(c1*x*x+c2))\n" + "v_exp_f32 %[v_tmp1], %[v_tmp1] ; emu = exp2(log2e*x*(c1*x*x+c2))\n" + "v_add_f32 %[v_tmp0], %[v_tmp0], 1.0 ; emu+1.0f\n" + "v_add_f32 %[v_tmp1], %[v_tmp1], 1.0 ; emu+1.0f\n" + "v_rcp_f32 %[v_tmp0], %[v_tmp0] ; 1/(emu+1.0f)\n" + "v_rcp_f32 %[v_tmp1], %[v_tmp1] ; 1/(emu+1.0f)\n" + "v_mul_f32 %[v_y0], %[v_tmp0], %[v_y0] ; x * 1/(emu+1f)\n" + "v_mul_f32 %[v_y1], %[v_tmp1], %[v_y1] ; x * 1/(emu+1f)\n" + : [v_y0] "+v"(y0), + [v_y1] "+v"(y1), + [v_c2] "+v"(c2), + // NOTE! it is totally possible that c2/y0/y1 share same register, they are all local + // tmp variables we need to expicitly hint compiler they may read+write, to allow + // allocate different register , the side effect is c2=** may issue for every such + // inline asm block + [v_tmp0] "+v"(tmp0), + [v_tmp1] "+v"(tmp1) + : [s_c1] "s"(c1), [s_log2e] "s"(log2e_) + :); + y.x = y0; + y.y = y1; + } +}; + // https://paperswithcode.com/method/gelu // y = 0.5*x*(1+erf(x/sqrt(2))) struct Gelu diff --git a/include/ck_tile/ops/flatmm.hpp b/include/ck_tile/ops/flatmm.hpp new file mode 100644 index 0000000000..eee80cda4a --- /dev/null +++ b/include/ck_tile/ops/flatmm.hpp @@ -0,0 +1,10 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp" +#include "ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp" +#include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp" +#include "ck_tile/ops/common/generic_2d_block_shape.hpp" +#include "ck_tile/ops/common/tensor_layout.hpp" diff --git a/include/ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp b/include/ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp new file mode 100644 index 0000000000..f5c7caf7df --- /dev/null +++ b/include/ck_tile/ops/flatmm/block/flatmm_32x512x128_1x4x1_16x16x32.hpp @@ -0,0 +1,615 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/ops/gemm/warp/warp_gemm.hpp" +#include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp" + +namespace ck_tile { + +// A async load to LDS, B direct to AGPR +// B matrix preshuffled in br*kr*w +// require 4 wave, occupancy=1c +// agpr useage:256 +// vgpr usage:64(A local) + 64(acc) + 8(os_a) + 8(os_b) = 144 (rem:112) +// +// for this gemm, 4 16x16x16 transposed layout +// input A vpgpr layout +// v0-v15: [ 0:15](gemm_m)x128(gemm_k) +// v16-v31: [16:31](gemm_m)x128(gemm_k) + +// input B vpgpr layout +// v0-v15: [ 0: 15](gemm_n)x128(gemm_k) +// v16-v31: [ 64: 79](gemm_n)x128(gemm_k) +// ...................... +// v111-v127: [448:463](gemm_n)x128(gemm_k) + +// output C vpgpr layout +// v0-v3 : [ 0:15](gemm_m)x[ 0: 15](gemm_n) +// v4-v7 : [16:31](gemm_m)x[ 0: 15](gemm_n) +// v8-v11: [ 0:15](gemm_m)x[64: 79](gemm_n) +// v12-v15: [16:31](gemm_m)x[64: 79](gemm_n) +// ...................... +// v56-v59: [ 0:15](gemm_m)x[448:463](gemm_n) +// v60-v63: [16:31](gemm_m)x[448:463](gemm_n) +struct Flatmm_32x512x128_1x4x1_16x16x32_Base // for f16/bf16 +{ + static constexpr index_t Block_M = 32; + static constexpr index_t Block_N = 512; + static constexpr index_t Block_K = 128; + + static constexpr index_t WarpPerBlock_M = 1; + static constexpr index_t WarpPerBlock_N = 4; + static constexpr index_t WarpPerBlock_K = 1; + + static constexpr index_t NumWarps = 4; + + static constexpr index_t Warp_M = 16; + static constexpr index_t Warp_N = 16; + static constexpr index_t Warp_K = 32; // 16 * SubKPacks + + static constexpr index_t BlockSize = 256; + + static constexpr index_t SubKPacks = 2; // this is used to gurantee every threads can do dwordx4 + + // TODO: note Nr/Kr/W need consider SubKPacks + static constexpr index_t Block_W = Warp_N * Warp_K; // 512 element + static constexpr index_t Block_Nr = Block_N / Warp_N; // 32 element, 4 per wave + static constexpr index_t Block_Kr = Block_K / Warp_K; // 4 + + static constexpr index_t Repeat_M = Block_M / (Warp_M * WarpPerBlock_M); // 2 + static constexpr index_t Repeat_N = Block_N / (Warp_N * WarpPerBlock_N); // 8 + static constexpr index_t Repeat_K = Block_K / (Warp_K * WarpPerBlock_K); // 8/2=4 + + static CK_TILE_DEVICE constexpr auto MakeCBlockDist() + { + constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding< + sequence<>, + tuple, sequence>, + tuple>, + tuple>, + sequence<2, 1>, // !! note here is different + sequence<0, 0>>{}; + + using WG = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; + + constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{}); + constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode); + return c_block_dstr; + } + + static CK_TILE_DEVICE constexpr auto MakeCBlockTile() + { + using CDataType = float; + constexpr auto c_block_dstr = MakeCBlockDist(); + auto c_block_tensor = make_static_distributed_tensor(c_block_dstr); + return c_block_tensor; + } + + CK_TILE_HOST_DEVICE static constexpr auto MakeLdsStoreDesc_A() + { + // A async->LDS + // constexpr index_t Block_M = Problem::BlockShape::Block_M0; + // constexpr index_t Block_K = Problem::BlockShape::Block_K0; + // constexpr index_t BlockSize = Problem::BlockShape::BlockSize; + constexpr index_t warpSize = ck_tile::get_warp_size(); + // constexpr index_t NumWarps = Problem::BlockShape::NumWarps; + + constexpr index_t KPack_ = 8; // GetSmemKPack_A(); // LDS + constexpr index_t KVector = 2; // GetAlignment_A(); // async copy 1 dword + constexpr index_t KPad = KPack_; // pad between warps + + static_assert(Block_K % KVector == 0); + constexpr index_t LanesPerK = Block_K / KVector; // how many thread loading K + if constexpr(LanesPerK >= warpSize) + { + // need multiple waves to load K + static_assert(LanesPerK % warpSize == 0); + constexpr index_t wavesPerK = LanesPerK / warpSize; + if constexpr(wavesPerK > NumWarps) + { + // TODO: need multiple issues along K to load all data + } + else + { + constexpr index_t wavesPerM = NumWarps / wavesPerK; + constexpr index_t NumIssues = Block_M / wavesPerM; + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number{}), // k2 + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number<1>{}), // k2 + number{}, // lds store vector(actually no explicit store) + number<1>{}); + + constexpr auto lds_block_desc_issues_warps_lanes = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple( + make_pass_through_transform(number{}), + make_merge_transform(make_tuple(number{}, number{})), + make_merge_transform(make_tuple(number{}, number{}))), + make_tuple(sequence<0>{}, sequence<1, 2>{}, sequence<3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{})); + + return lds_block_desc_issues_warps_lanes; + } + } + else + { + // lanes within a wave load different M but same K + static_assert(warpSize % LanesPerK == 0); + constexpr index_t LaneGroups = warpSize / LanesPerK; // along m + constexpr index_t NumIssues = Block_M / (LaneGroups * NumWarps); + + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // m2 + number{}, // k0 + number{}), // k1 + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // m2 + number{}, // k0 + number<1>{}), // k1 + number{}, // lds store vector(actually no explicit store) + number<1>{}); + + constexpr auto lds_block_desc_issues_warps_lanes = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple(make_pass_through_transform(number{}), + make_pass_through_transform(number{}), + make_merge_transform(make_tuple( + number{}, number{}, number{}))), + make_tuple(sequence<0>{}, sequence<2>{}, sequence<1, 3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{})); + + return lds_block_desc_issues_warps_lanes; + } + } + + // template + CK_TILE_HOST_DEVICE static constexpr auto MakeLdsLoadDesc_A() + { + // load from LDS to register, every wave has same layout + constexpr index_t KPack_ = 8; // GetSmemKPack_A(); // LDS + constexpr index_t KPad = KPack_; // pad between warps + + constexpr index_t kAMLane = 16; + constexpr index_t kABKLane = 4; + constexpr index_t kABKPerLane = 4; + constexpr index_t kKIter = 2; + static_assert(KPack_ == (kABKPerLane * kKIter)); + + constexpr auto lds_block_desc_0 = + make_naive_tensor_descriptor(make_tuple(number{}, // m0 y + number{}, // m1 p + number{}, // k0 y + number{}, // k1 p + number{}), // k2 y-vector + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number<1>{}), // k2 + number{}, // lds load vector + number<1>{}); + + constexpr auto lds_desc_m_k = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple(make_merge_transform(make_tuple(number{}, number{})), + make_merge_transform( + make_tuple(number{}, number{}, number{}))), + make_tuple(sequence<0, 1>{}, sequence<2, 3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{})); + + return lds_desc_m_k; + } + + static constexpr auto GetGemm_AWarpEnc() + { + constexpr index_t kAMLane = 16; + constexpr index_t kABKLane = 4; + constexpr index_t kABKPerLane = 4; + constexpr index_t kKIter = 2; + + using enc_ = tile_distribution_encoding< + sequence<>, + tuple, sequence>, + tuple>, + tuple>, + sequence<2>, + sequence<1>>; + return enc_{}; + } + + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize() + { + return 32 * (128 + 8) * sizeof(bf16_t); + } +}; + +struct Flatmm_32x512x128_1x4x1_16x16x32_BF16 : public Flatmm_32x512x128_1x4x1_16x16x32_Base +{ + using ADataType = bf16_t; + using BDataType = bf16_t; + + // TODO: need paired with tile_window_linear! + // TODO: need call init_raw() before call this function! + template + CK_TILE_DEVICE auto + operator()(const ARes& res_a, + const ACoords& cached_coords_a, + const BRes& res_b, + const BCoords& cached_coords_b, + CK_TILE_LDS_ADDR void* smem, + index_t k, + index_t tile_offset_a, // for each tile, the offset to move for each unroll + index_t tile_offset_b) // for each tile, the offset to move for each unroll + { + static_assert(ACoords::size() == Block_M * Block_K / BlockSize / 2 /*2x per dword*/); // 8 + static_assert(BCoords::size() == Repeat_N); + + auto a_sst = make_tile_window( + make_tensor_view( + reinterpret_cast(smem), MakeLdsStoreDesc_A()), + MakeLdsStoreDesc_A().get_lengths(), + {0, 0, 0}); + + auto a_sld = [&]() { + constexpr auto a_warp_enc_ = GetGemm_AWarpEnc(); + constexpr auto a_outer_dstr_enc = tile_distribution_encoding< + sequence, + tuple, sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + constexpr auto a_block_dstr_encode = + detail::make_embed_tile_distribution_encoding(a_outer_dstr_enc, a_warp_enc_); + return make_tile_window_linear( + make_tensor_view( + reinterpret_cast(smem), MakeLdsLoadDesc_A()), + MakeLdsLoadDesc_A().get_lengths(), + {0, 0}, + make_static_tile_distribution(a_block_dstr_encode)); + }(); + + const index_t tile_offset_a_bytes = tile_offset_a * sizeof(ADataType); + const index_t tile_offset_b_bytes = tile_offset_b * sizeof(BDataType); + + const auto [m0_init_value, size_per_issue] = get_async_store_smem_info(a_sst); + constexpr auto smem_buf_size = + MakeLdsLoadDesc_A().get_element_space_size() * sizeof(ADataType); + static_assert(a_sld.get_num_of_access() == 8); + constexpr auto sld_os = generate_tuple( + [&](auto i_access) { + return number{}; + }, + number{}); + + index_t loop_cnt = k / Block_K; + + // this is the acc thread buffer + fp32x4_t v_acc[16]{.0f}; + + // B nr->kr +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Winline-asm" + // clang-format off + asm volatile( +#define CK_TILE_FLATMM_UK_MFMA CK_TILE_FLATMM_UK_MFMA_BF16 +#include "uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc" +#undef CK_TILE_FLATMM_UK_MFMA + : [s_loop_cnt]"+s"(loop_cnt), + [v_acc_0]"+v"(v_acc[0]), + [v_acc_1]"+v"(v_acc[1]), + [v_acc_2]"+v"(v_acc[2]), + [v_acc_3]"+v"(v_acc[3]), + [v_acc_4]"+v"(v_acc[4]), + [v_acc_5]"+v"(v_acc[5]), + [v_acc_6]"+v"(v_acc[6]), + [v_acc_7]"+v"(v_acc[7]), + [v_acc_8]"+v"(v_acc[8]), + [v_acc_9]"+v"(v_acc[9]), + [v_acc_10]"+v"(v_acc[10]), + [v_acc_11]"+v"(v_acc[11]), + [v_acc_12]"+v"(v_acc[12]), + [v_acc_13]"+v"(v_acc[13]), + [v_acc_14]"+v"(v_acc[14]), + [v_acc_15]"+v"(v_acc[15]), + [s_mem_]"+r"(smem) + : [s_res_a0]"s"(res_a[0]), + [s_res_a1]"s"(res_a[1]), + [s_res_a2]"s"(res_a[2]), + [s_res_a3]"s"(res_a[3]), + [s_res_b0]"s"(res_b[0]), + [s_res_b1]"s"(res_b[1]), + [s_res_b2]"s"(res_b[2]), + [s_res_b3]"s"(res_b[3]), + [v_os_a0]"v"(static_cast(cached_coords_a[number<0>{}] * sizeof(ADataType))), + [v_os_a1]"v"(static_cast(cached_coords_a[number<1>{}] * sizeof(ADataType))), + [v_os_a2]"v"(static_cast(cached_coords_a[number<2>{}] * sizeof(ADataType))), + [v_os_a3]"v"(static_cast(cached_coords_a[number<3>{}] * sizeof(ADataType))), + [v_os_a4]"v"(static_cast(cached_coords_a[number<4>{}] * sizeof(ADataType))), + [v_os_a5]"v"(static_cast(cached_coords_a[number<5>{}] * sizeof(ADataType))), + [v_os_a6]"v"(static_cast(cached_coords_a[number<6>{}] * sizeof(ADataType))), + [v_os_a7]"v"(static_cast(cached_coords_a[number<7>{}] * sizeof(ADataType))), + + [v_os_b0]"v"(static_cast(cached_coords_b[number<0>{}] * sizeof(BDataType))), + [v_os_b1]"v"(static_cast(cached_coords_b[number<1>{}] * sizeof(BDataType))), + [v_os_b2]"v"(static_cast(cached_coords_b[number<2>{}] * sizeof(BDataType))), + [v_os_b3]"v"(static_cast(cached_coords_b[number<3>{}] * sizeof(BDataType))), + [v_os_b4]"v"(static_cast(cached_coords_b[number<4>{}] * sizeof(BDataType))), + [v_os_b5]"v"(static_cast(cached_coords_b[number<5>{}] * sizeof(BDataType))), + [v_os_b6]"v"(static_cast(cached_coords_b[number<6>{}] * sizeof(BDataType))), + [v_os_b7]"v"(static_cast(cached_coords_b[number<7>{}] * sizeof(BDataType))), + + [v_os_slda]"v"(static_cast(a_sld.cached_coords_[number<0>{}].get_offset() * sizeof(ADataType))), + [s_m0_init]"s"(m0_init_value), + [s_size_per_issue]"s"(size_per_issue), + [smem_sz]"n"(smem_buf_size), //(smem_buf_size), + [sld_os_0]"n"(sld_os[number<0>{}].value), + [sld_os_1]"n"(sld_os[number<1>{}].value), + [sld_os_2]"n"(sld_os[number<2>{}].value), + [sld_os_3]"n"(sld_os[number<3>{}].value), + [sld_os_4]"n"(sld_os[number<4>{}].value), + [sld_os_5]"n"(sld_os[number<5>{}].value), + [sld_os_6]"n"(sld_os[number<6>{}].value), + [sld_os_7]"n"(sld_os[number<7>{}].value), + [s_tile_os_a]"s"(tile_offset_a_bytes), + [s_tile_os_b]"s"(tile_offset_b_bytes) + : "memory", "a0", "a1", "a2", "a3", "a4", "a5", "a6", "a7", "a8", "a9", + "a10", "a11", "a12", "a13", "a14", "a15", "a16", "a17", "a18", "a19", + "a20", "a21", "a22", "a23", "a24", "a25", "a26", "a27", "a28", "a29", + "a30", "a31", "a32", "a33", "a34", "a35", "a36", "a37", "a38", "a39", + "a40", "a41", "a42", "a43", "a44", "a45", "a46", "a47", "a48", "a49", + "a50", "a51", "a52", "a53", "a54", "a55", "a56", "a57", "a58", "a59", + "a60", "a61", "a62", "a63", "a64", "a65", "a66", "a67", "a68", "a69", + "a70", "a71", "a72", "a73", "a74", "a75", "a76", "a77", "a78", "a79", + "a80", "a81", "a82", "a83", "a84", "a85", "a86", "a87", "a88", "a89", + "a90", "a91", "a92", "a93", "a94", "a95", "a96", "a97", "a98", "a99", + "a100", "a101", "a102", "a103", "a104", "a105", "a106", "a107", + "a108", "a109", "a110", "a111", "a112", "a113", "a114", "a115", + "a116", "a117", "a118", "a119", "a120", "a121", "a122", "a123", + "a124", "a125", "a126", "a127", "a128", "a129", "a130", "a131", + "a132", "a133", "a134", "a135", "a136", "a137", "a138", "a139", + "a140", "a141", "a142", "a143", "a144", "a145", "a146", "a147", + "a148", "a149", "a150", "a151", "a152", "a153", "a154", "a155", + "a156", "a157", "a158", "a159", "a160", "a161", "a162", "a163", + "a164", "a165", "a166", "a167", "a168", "a169", "a170", "a171", + "a172", "a173", "a174", "a175", "a176", "a177", "a178", "a179", + "a180", "a181", "a182", "a183", "a184", "a185", "a186", "a187", + "a188", "a189", "a190", "a191", "a192", "a193", "a194", "a195", + "a196", "a197", "a198", "a199", "a200", "a201", "a202", "a203", + "a204", "a205", "a206", "a207", "a208", "a209", "a210", "a211", + "a212", "a213", "a214", "a215", "a216", "a217", "a218", "a219", + "a220", "a221", "a222", "a223", "a224", "a225", "a226", "a227", + "a228", "a229", "a230", "a231", "a232", "a233", "a234", "a235", + "a236", "a237", "a238", "a239", "a240", "a241", "a242", "a243", + "a244", "a245", "a246", "a247", "a248", "a249", "a250", "a251", + "a252", "a253", "a254", "a255", + "s16", "s17", "s18", "s19", "s20", "s21", "s22", "s23", + "s86", // s86 as tmp + "v64", "v65", "v66", "v67", "v68", "v69", + "v70", "v71", "v72", "v73", "v74", "v75", "v76", "v77", "v78", "v79", + "v80", "v81", "v82", "v83", "v84", "v85", "v86", "v87", "v88", "v89", + "v90", "v91", "v92", "v93", "v94", "v95", "v96", "v97", "v98", "v99", + "v100", "v101", "v102", "v103", "v104", "v105", "v106", "v107", + "v108", "v109", "v110", "v111", "v112", "v113", "v114", "v115", + "v116", "v117", "v118", "v119", "v120", "v121", "v122", "v123", + "v124", "v125", "v126", "v127" + ); + // clang-format on +#pragma clang diagnostic pop + + // return local scratch + auto c = MakeCBlockTile(); + for(auto i = 0; i < 16; i++) + { + c.get_thread_buffer()[4 * i + 0] = v_acc[i].x; + c.get_thread_buffer()[4 * i + 1] = v_acc[i].y; + c.get_thread_buffer()[4 * i + 2] = v_acc[i].z; + c.get_thread_buffer()[4 * i + 3] = v_acc[i].w; + } + return c; + } +}; + +struct Flatmm_32x512x128_1x4x1_16x16x32_FP16 : public Flatmm_32x512x128_1x4x1_16x16x32_Base +{ + using ADataType = fp16_t; + using BDataType = fp16_t; + + // TODO: need paired with tile_window_linear! + // TODO: need call init_raw() before call this function! + template + CK_TILE_DEVICE auto + operator()(const ARes& res_a, + const ACoords& cached_coords_a, + const BRes& res_b, + const BCoords& cached_coords_b, + CK_TILE_LDS_ADDR void* smem, + index_t k, + index_t tile_offset_a, // for each tile, the offset to move for each unroll + index_t tile_offset_b) // for each tile, the offset to move for each unroll + { + static_assert(ACoords::size() == Block_M * Block_K / BlockSize / 2 /*2x per dword*/); // 8 + static_assert(BCoords::size() == Repeat_N); + + auto a_sst = make_tile_window( + make_tensor_view( + reinterpret_cast(smem), MakeLdsStoreDesc_A()), + MakeLdsStoreDesc_A().get_lengths(), + {0, 0, 0}); + + auto a_sld = [&]() { + constexpr auto a_warp_enc_ = GetGemm_AWarpEnc(); + constexpr auto a_outer_dstr_enc = tile_distribution_encoding< + sequence, + tuple, sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + constexpr auto a_block_dstr_encode = + detail::make_embed_tile_distribution_encoding(a_outer_dstr_enc, a_warp_enc_); + return make_tile_window_linear( + make_tensor_view( + reinterpret_cast(smem), MakeLdsLoadDesc_A()), + MakeLdsLoadDesc_A().get_lengths(), + {0, 0}, + make_static_tile_distribution(a_block_dstr_encode)); + }(); + + const index_t tile_offset_a_bytes = tile_offset_a * sizeof(ADataType); + const index_t tile_offset_b_bytes = tile_offset_b * sizeof(BDataType); + + const auto [m0_init_value, size_per_issue] = get_async_store_smem_info(a_sst); + constexpr auto smem_buf_size = + MakeLdsLoadDesc_A().get_element_space_size() * sizeof(ADataType); + static_assert(a_sld.get_num_of_access() == 8); + constexpr auto sld_os = generate_tuple( + [&](auto i_access) { + return number{}; + }, + number{}); + + index_t loop_cnt = k / Block_K; + + // this is the acc thread buffer + fp32x4_t v_acc[16]{.0f}; + + // B nr->kr +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Winline-asm" + // clang-format off + asm volatile( +#define CK_TILE_FLATMM_UK_MFMA CK_TILE_FLATMM_UK_MFMA_FP16 +#include "uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc" +#undef CK_TILE_FLATMM_UK_MFMA + : [s_loop_cnt]"+s"(loop_cnt), + [v_acc_0]"+v"(v_acc[0]), + [v_acc_1]"+v"(v_acc[1]), + [v_acc_2]"+v"(v_acc[2]), + [v_acc_3]"+v"(v_acc[3]), + [v_acc_4]"+v"(v_acc[4]), + [v_acc_5]"+v"(v_acc[5]), + [v_acc_6]"+v"(v_acc[6]), + [v_acc_7]"+v"(v_acc[7]), + [v_acc_8]"+v"(v_acc[8]), + [v_acc_9]"+v"(v_acc[9]), + [v_acc_10]"+v"(v_acc[10]), + [v_acc_11]"+v"(v_acc[11]), + [v_acc_12]"+v"(v_acc[12]), + [v_acc_13]"+v"(v_acc[13]), + [v_acc_14]"+v"(v_acc[14]), + [v_acc_15]"+v"(v_acc[15]), + [s_mem_]"+r"(smem) + : [s_res_a0]"s"(res_a[0]), + [s_res_a1]"s"(res_a[1]), + [s_res_a2]"s"(res_a[2]), + [s_res_a3]"s"(res_a[3]), + [s_res_b0]"s"(res_b[0]), + [s_res_b1]"s"(res_b[1]), + [s_res_b2]"s"(res_b[2]), + [s_res_b3]"s"(res_b[3]), + [v_os_a0]"v"(static_cast(cached_coords_a[number<0>{}] * sizeof(ADataType))), + [v_os_a1]"v"(static_cast(cached_coords_a[number<1>{}] * sizeof(ADataType))), + [v_os_a2]"v"(static_cast(cached_coords_a[number<2>{}] * sizeof(ADataType))), + [v_os_a3]"v"(static_cast(cached_coords_a[number<3>{}] * sizeof(ADataType))), + [v_os_a4]"v"(static_cast(cached_coords_a[number<4>{}] * sizeof(ADataType))), + [v_os_a5]"v"(static_cast(cached_coords_a[number<5>{}] * sizeof(ADataType))), + [v_os_a6]"v"(static_cast(cached_coords_a[number<6>{}] * sizeof(ADataType))), + [v_os_a7]"v"(static_cast(cached_coords_a[number<7>{}] * sizeof(ADataType))), + + [v_os_b0]"v"(static_cast(cached_coords_b[number<0>{}] * sizeof(BDataType))), + [v_os_b1]"v"(static_cast(cached_coords_b[number<1>{}] * sizeof(BDataType))), + [v_os_b2]"v"(static_cast(cached_coords_b[number<2>{}] * sizeof(BDataType))), + [v_os_b3]"v"(static_cast(cached_coords_b[number<3>{}] * sizeof(BDataType))), + [v_os_b4]"v"(static_cast(cached_coords_b[number<4>{}] * sizeof(BDataType))), + [v_os_b5]"v"(static_cast(cached_coords_b[number<5>{}] * sizeof(BDataType))), + [v_os_b6]"v"(static_cast(cached_coords_b[number<6>{}] * sizeof(BDataType))), + [v_os_b7]"v"(static_cast(cached_coords_b[number<7>{}] * sizeof(BDataType))), + + [v_os_slda]"v"(static_cast(a_sld.cached_coords_[number<0>{}].get_offset() * sizeof(ADataType))), + [s_m0_init]"s"(m0_init_value), + [s_size_per_issue]"s"(size_per_issue), + [smem_sz]"n"(smem_buf_size), //(smem_buf_size), + [sld_os_0]"n"(sld_os[number<0>{}].value), + [sld_os_1]"n"(sld_os[number<1>{}].value), + [sld_os_2]"n"(sld_os[number<2>{}].value), + [sld_os_3]"n"(sld_os[number<3>{}].value), + [sld_os_4]"n"(sld_os[number<4>{}].value), + [sld_os_5]"n"(sld_os[number<5>{}].value), + [sld_os_6]"n"(sld_os[number<6>{}].value), + [sld_os_7]"n"(sld_os[number<7>{}].value), + [s_tile_os_a]"s"(tile_offset_a_bytes), + [s_tile_os_b]"s"(tile_offset_b_bytes) + : "memory", "a0", "a1", "a2", "a3", "a4", "a5", "a6", "a7", "a8", "a9", + "a10", "a11", "a12", "a13", "a14", "a15", "a16", "a17", "a18", "a19", + "a20", "a21", "a22", "a23", "a24", "a25", "a26", "a27", "a28", "a29", + "a30", "a31", "a32", "a33", "a34", "a35", "a36", "a37", "a38", "a39", + "a40", "a41", "a42", "a43", "a44", "a45", "a46", "a47", "a48", "a49", + "a50", "a51", "a52", "a53", "a54", "a55", "a56", "a57", "a58", "a59", + "a60", "a61", "a62", "a63", "a64", "a65", "a66", "a67", "a68", "a69", + "a70", "a71", "a72", "a73", "a74", "a75", "a76", "a77", "a78", "a79", + "a80", "a81", "a82", "a83", "a84", "a85", "a86", "a87", "a88", "a89", + "a90", "a91", "a92", "a93", "a94", "a95", "a96", "a97", "a98", "a99", + "a100", "a101", "a102", "a103", "a104", "a105", "a106", "a107", + "a108", "a109", "a110", "a111", "a112", "a113", "a114", "a115", + "a116", "a117", "a118", "a119", "a120", "a121", "a122", "a123", + "a124", "a125", "a126", "a127", "a128", "a129", "a130", "a131", + "a132", "a133", "a134", "a135", "a136", "a137", "a138", "a139", + "a140", "a141", "a142", "a143", "a144", "a145", "a146", "a147", + "a148", "a149", "a150", "a151", "a152", "a153", "a154", "a155", + "a156", "a157", "a158", "a159", "a160", "a161", "a162", "a163", + "a164", "a165", "a166", "a167", "a168", "a169", "a170", "a171", + "a172", "a173", "a174", "a175", "a176", "a177", "a178", "a179", + "a180", "a181", "a182", "a183", "a184", "a185", "a186", "a187", + "a188", "a189", "a190", "a191", "a192", "a193", "a194", "a195", + "a196", "a197", "a198", "a199", "a200", "a201", "a202", "a203", + "a204", "a205", "a206", "a207", "a208", "a209", "a210", "a211", + "a212", "a213", "a214", "a215", "a216", "a217", "a218", "a219", + "a220", "a221", "a222", "a223", "a224", "a225", "a226", "a227", + "a228", "a229", "a230", "a231", "a232", "a233", "a234", "a235", + "a236", "a237", "a238", "a239", "a240", "a241", "a242", "a243", + "a244", "a245", "a246", "a247", "a248", "a249", "a250", "a251", + "a252", "a253", "a254", "a255", + "s16", "s17", "s18", "s19", "s20", "s21", "s22", "s23", + "s86", // s86 as tmp + "v64", "v65", "v66", "v67", "v68", "v69", + "v70", "v71", "v72", "v73", "v74", "v75", "v76", "v77", "v78", "v79", + "v80", "v81", "v82", "v83", "v84", "v85", "v86", "v87", "v88", "v89", + "v90", "v91", "v92", "v93", "v94", "v95", "v96", "v97", "v98", "v99", + "v100", "v101", "v102", "v103", "v104", "v105", "v106", "v107", + "v108", "v109", "v110", "v111", "v112", "v113", "v114", "v115", + "v116", "v117", "v118", "v119", "v120", "v121", "v122", "v123", + "v124", "v125", "v126", "v127" + ); + // clang-format on +#pragma clang diagnostic pop + + // return local scratch + auto c = MakeCBlockTile(); + for(auto i = 0; i < 16; i++) + { + c.get_thread_buffer()[4 * i + 0] = v_acc[i].x; + c.get_thread_buffer()[4 * i + 1] = v_acc[i].y; + c.get_thread_buffer()[4 * i + 2] = v_acc[i].z; + c.get_thread_buffer()[4 * i + 3] = v_acc[i].w; + } + return c; + } +}; + +} // namespace ck_tile diff --git a/include/ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp b/include/ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp new file mode 100644 index 0000000000..203c87b9c6 --- /dev/null +++ b/include/ck_tile/ops/flatmm/block/flatmm_sn_32x128x512_1x4x1_16x16x32.hpp @@ -0,0 +1,562 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/ops/gemm/warp/warp_gemm.hpp" +#include "ck_tile/ops/flatmm/block/flatmm_uk_config.hpp" + +namespace ck_tile { + +// "S"tream update output along "N" +// A in smem, B load from global +// require 4 wave, occupancy=1c +struct FlatmmSn_32x128x512_1x4x1_16x16x32_Base +{ + static constexpr index_t Block_M = 32; + static constexpr index_t Block_N = 128; + static constexpr index_t Block_K = 512; + + static constexpr index_t WarpPerBlock_M = 1; + static constexpr index_t WarpPerBlock_N = 4; + static constexpr index_t WarpPerBlock_K = 1; + + static constexpr index_t Warp_M = 16; + static constexpr index_t Warp_N = 16; + static constexpr index_t Warp_K = 32; + + static constexpr index_t BlockSize = 256; + + // static constexpr index_t KPack = 2; // this is used to gurantee every threads can do dwordx4 + + // TODO: note Nr/Kr/W need consider KPack + static constexpr index_t Block_W = Warp_N * Warp_K; // 512 element + static constexpr index_t Block_Nr = Block_N / Warp_N; // 32 element, 4 per wave + static constexpr index_t Block_Kr = Block_K / Warp_K; // 4 + + static constexpr index_t Repeat_M = Block_M / (Warp_M * WarpPerBlock_M); // 2 + static constexpr index_t Repeat_N = Block_N / (Warp_N * WarpPerBlock_N); // 2 + static constexpr index_t Repeat_K = Block_K / (Warp_K * WarpPerBlock_K); // 16 + + static CK_TILE_DEVICE constexpr auto MakeCBlockDist() + { + constexpr auto c_block_outer_dstr_encoding = tile_distribution_encoding< + sequence<>, + tuple, sequence>, + tuple>, + tuple>, + sequence<2, 1>, // !! note here is different + sequence<0, 0>>{}; + + using WG = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; + + constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + c_block_outer_dstr_encoding, typename WG::CWarpDstrEncoding{}); + constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode); + return c_block_dstr; + } + + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize() + { + // y y p p p y + // reg before shfl M0(2)*N0(2)*Nl(4)*Nw(4)*Mw(16)*Nv(4) + // but order is N0*M0*Nv + // in LDS we need store as + // M0(2)* N0(2) * Nl(4) * Nw(4) * (Mw(16)*Nv(4) + 4) + // y y wave-id lid/16 lid%16 v + return 2 * 2 * 4 * 4 * (16 * 4 + 4) * sizeof(bf16_t); + } +}; + +struct FlatmmSn_32x128x512_1x4x1_16x16x32_BF16 : public FlatmmSn_32x128x512_1x4x1_16x16x32_Base +{ + using BDataType = bf16_t; + using ODataType = bf16_t; + + // TODO: need paired with tile_window_linear! + // TODO: need call init_raw() before call this function! + // template + template + CK_TILE_DEVICE auto + operator()(const BRes& res_b, + const BCoords& cached_coords_b, + const ORes& res_o, + const OCoords& cached_coords_o, + const OFlags& o_flags, // this should be in sgpr + CK_TILE_LDS_ADDR void* smem, + index_t n, // loop along n dim + const ScaleTensor& scale_, + index_t tile_offset_b, // stride b is fixed to blockKr * blockW, but still can adjust + index_t tile_offset_o) + { + static_assert(BCoords::size() == 8); // 8 + static_assert(OCoords::size() == 8); + + const index_t tile_stride_b_bytes = tile_offset_b * sizeof(BDataType); + const index_t tile_stride_o_bytes = tile_offset_o * sizeof(ODataType); + + static_assert(ScaleTensor::size() == 2); + float s0 = scale_[number<0>{}]; + float s1 = scale_[number<1>{}]; + + index_t loop_cnt = n / Block_N; + + register float v_c0 asm("v64"); + register float v_c1 asm("v65"); + register float v_c2 asm("v66"); + register float v_c3 asm("v67"); + register float v_c4 asm("v68"); + register float v_c5 asm("v69"); + register float v_c6 asm("v70"); + register float v_c7 asm("v71"); + register float v_c8 asm("v72"); + register float v_c9 asm("v73"); + register float v_c10 asm("v74"); + register float v_c11 asm("v75"); + register float v_c12 asm("v76"); + register float v_c13 asm("v77"); + register float v_c14 asm("v78"); + register float v_c15 asm("v79"); + register float v_c16 asm("v80"); + register float v_c17 asm("v81"); + register float v_c18 asm("v82"); + register float v_c19 asm("v83"); + register float v_c20 asm("v84"); + register float v_c21 asm("v85"); + register float v_c22 asm("v86"); + register float v_c23 asm("v87"); + register float v_c24 asm("v88"); + register float v_c25 asm("v89"); + register float v_c26 asm("v90"); + register float v_c27 asm("v91"); + register float v_c28 asm("v92"); + register float v_c29 asm("v93"); + register float v_c30 asm("v94"); + register float v_c31 asm("v95"); + int32_t nan_hi = 0x7fff0000; + int32_t nan_lo = 0x00007fff; + + // in smem, the layout is M0(2)*K0(128)*M1(16)*K1(4) + // every threads need 8xK in contiguous register + // ... and every wave need the same data + int lane_id = threadIdx.x % 64; + int sld_y_os = (lane_id % 16) * 4 + (lane_id / 16) * 128; + sld_y_os *= 2; + + // y y p p p y + // reg before shfl M0(2)*N0(2)*Nl(4)*Nw(4)*Mw(16)*Nv(4) + // but order is N0*M0*Nv + // in LDS we need store as + // M0(2)* N0(2) * Nl(4) * Nw(4) * (Mw(16)*Nv(4) + 4) + // y y wave-id lid/16 lid%16 v + // sst(v3) = (v0/16*34 + v0%16 * 2 + wid*136) * 4 + int sfl_sst = (threadIdx.x % 16 * 4) + (threadIdx.x / 16) * (64 + 4); + sfl_sst *= 2; + + // from LDS we need load as + // M0(2)* N0(2) * Nl(4) * Nw(4) * (Mw(16) * Nv(4) + 4) + // ( 2 issue) (rem 32-lane) (4 wave*4issue) 2lane*1ussue(pk2) + // sld(v4) = v0/2 *34*4 + v0 % 2 *4 + wid*2 *4 + int sfl_sld = (lane_id % 2) * 2 + (lane_id / 2) * (64 + 4) + (threadIdx.x / 64) * 4; + sfl_sld *= 2; + + // B nr->kr + // clang-format off +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Winline-asm" + asm volatile( +#define CK_TILE_FLATMM_UK_MFMA CK_TILE_FLATMM_UK_MFMA_BF16 +#include "uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc" +#undef CK_TILE_FLATMM_UK_MFMA + :[smem_]"+r"(smem), + [s_loop_cnt]"+s"(loop_cnt), + [c0]"+v" (v_c0), + [c1]"+v" (v_c1), + [c2]"+v" (v_c2), + [c3]"+v" (v_c3), + [c4]"+v" (v_c4), + [c5]"+v" (v_c5), + [c6]"+v" (v_c6), + [c7]"+v" (v_c7), + [c8]"+v" (v_c8), + [c9]"+v" (v_c9), + [c10]"+v"(v_c10), + [c11]"+v"(v_c11), + [c12]"+v"(v_c12), + [c13]"+v"(v_c13), + [c14]"+v"(v_c14), + [c15]"+v"(v_c15), + [c16]"+v"(v_c16), + [c17]"+v"(v_c17), + [c18]"+v"(v_c18), + [c19]"+v"(v_c19), + [c20]"+v"(v_c20), + [c21]"+v"(v_c21), + [c22]"+v"(v_c22), + [c23]"+v"(v_c23), + [c24]"+v"(v_c24), + [c25]"+v"(v_c25), + [c26]"+v"(v_c26), + [c27]"+v"(v_c27), + [c28]"+v"(v_c28), + [c29]"+v"(v_c29), + [c30]"+v"(v_c30), + [c31]"+v"(v_c31) + : + [sld_a_base]"n"(0), + [shfl_base]"n"(0), + [v_sld_y_os]"v"(sld_y_os), + [v_sfl_sld]"v"(sfl_sld), + [v_sfl_sst]"v"(sfl_sst), + [s_res_o0]"s"(res_o[0]), + [s_res_o1]"s"(res_o[1]), + //[s_res_o2]"s"(res_o[2]), + //[s_res_o3]"s"(res_o[3]), + [s_res_b0]"s"(res_b[0]), + [s_res_b1]"s"(res_b[1]), + [s_res_b2]"s"(res_b[2]), + [s_res_b3]"s"(res_b[3]), + [v_os_o0]"v"(static_cast(cached_coords_o[number<0>{}] * sizeof(ODataType))), + [v_os_o1]"v"(static_cast(cached_coords_o[number<1>{}] * sizeof(ODataType))), + [v_os_o2]"v"(static_cast(cached_coords_o[number<2>{}] * sizeof(ODataType))), + [v_os_o3]"v"(static_cast(cached_coords_o[number<3>{}] * sizeof(ODataType))), + [v_os_o4]"v"(static_cast(cached_coords_o[number<4>{}] * sizeof(ODataType))), + [v_os_o5]"v"(static_cast(cached_coords_o[number<5>{}] * sizeof(ODataType))), + [v_os_o6]"v"(static_cast(cached_coords_o[number<6>{}] * sizeof(ODataType))), + [v_os_o7]"v"(static_cast(cached_coords_o[number<7>{}] * sizeof(ODataType))), + [v_os_b0]"v"(static_cast(cached_coords_b[number<0>{}] * sizeof(BDataType))), + [v_os_b1]"v"(static_cast(cached_coords_b[number<1>{}] * sizeof(BDataType))), + [v_os_b2]"v"(static_cast(cached_coords_b[number<2>{}] * sizeof(BDataType))), + [v_os_b3]"v"(static_cast(cached_coords_b[number<3>{}] * sizeof(BDataType))), + [v_os_b4]"v"(static_cast(cached_coords_b[number<4>{}] * sizeof(BDataType))), + [v_os_b5]"v"(static_cast(cached_coords_b[number<5>{}] * sizeof(BDataType))), + [v_os_b6]"v"(static_cast(cached_coords_b[number<6>{}] * sizeof(BDataType))), + [v_os_b7]"v"(static_cast(cached_coords_b[number<7>{}] * sizeof(BDataType))), + + [s_tile_os_o]"s"(tile_stride_o_bytes), + [s_tile_os_b]"s"(tile_stride_b_bytes), + [scale_0]"v"(s0), + [scale_1]"v"(s1), + [v_nan_lo]"v"(nan_lo), + [v_nan_hi]"v"(nan_hi), + [s_execflag_0]"s"(o_flags[number<0>{}]), + [s_execflag_1]"s"(o_flags[number<1>{}]), + [s_execflag_2]"s"(o_flags[number<2>{}]), + [s_execflag_3]"s"(o_flags[number<3>{}]), + [s_execflag_4]"s"(o_flags[number<4>{}]), + [s_execflag_5]"s"(o_flags[number<5>{}]), + [s_execflag_6]"s"(o_flags[number<6>{}]), + [s_execflag_7]"s"(o_flags[number<7>{}]) + : + "memory", "a0", "a1", "a2", "a3", "a4", "a5", "a6", "a7", "a8", "a9", + "a10", "a11", "a12", "a13", "a14", "a15", "a16", "a17", "a18", "a19", + "a20", "a21", "a22", "a23", "a24", "a25", "a26", "a27", "a28", "a29", + "a30", "a31", "a32", "a33", "a34", "a35", "a36", "a37", "a38", "a39", + "a40", "a41", "a42", "a43", "a44", "a45", "a46", "a47", "a48", "a49", + "a50", "a51", "a52", "a53", "a54", "a55", "a56", "a57", "a58", "a59", + "a60", "a61", "a62", "a63", "a64", "a65", "a66", "a67", "a68", "a69", + "a70", "a71", "a72", "a73", "a74", "a75", "a76", "a77", "a78", "a79", + "a80", "a81", "a82", "a83", "a84", "a85", "a86", "a87", "a88", "a89", + "a90", "a91", "a92", "a93", "a94", "a95", "a96", "a97", "a98", "a99", + "a100", "a101", "a102", "a103", "a104", "a105", "a106", "a107", + "a108", "a109", "a110", "a111", "a112", "a113", "a114", "a115", + "a116", "a117", "a118", "a119", "a120", "a121", "a122", "a123", + "a124", "a125", "a126", "a127", "a128", "a129", "a130", "a131", + "a132", "a133", "a134", "a135", "a136", "a137", "a138", "a139", + "a140", "a141", "a142", "a143", "a144", "a145", "a146", "a147", + "a148", "a149", "a150", "a151", "a152", "a153", "a154", "a155", + "a156", "a157", "a158", "a159", "a160", "a161", "a162", "a163", + "a164", "a165", "a166", "a167", "a168", "a169", "a170", "a171", + "a172", "a173", "a174", "a175", "a176", "a177", "a178", "a179", + "a180", "a181", "a182", "a183", "a184", "a185", "a186", "a187", + "a188", "a189", "a190", "a191", "a192", "a193", "a194", "a195", + "a196", "a197", "a198", "a199", "a200", "a201", "a202", "a203", + "a204", "a205", "a206", "a207", "a208", "a209", "a210", "a211", + "a212", "a213", "a214", "a215", "a216", "a217", "a218", "a219", + "a220", "a221", "a222", "a223", "a224", "a225", "a226", "a227", + "a228", "a229", "a230", "a231", "a232", "a233", "a234", "a235", + "a236", "a237", "a238", "a239", "a240", "a241", "a242", "a243", + "a244", "a245", "a246", "a247", "a248", "a249", "a250", "a251", + "a252", "a253", "a254", "a255", + "s8", "s9", "s12", "s13", "s14", "s15", "s38", "s39", "s52", "s86", + "s36", "s37", + "v50", "v54", "v55", + "v64","v65","v66","v67","v68","v69","v70","v71", + "v72","v73","v74","v75","v76","v77","v78","v79", + "v80","v81","v82","v83","v84","v85","v86","v87", + "v88","v89","v90","v91","v92","v93","v94","v95", + "v128", "v129", "v130", "v131", + "v132", "v133", "v134", "v135", "v136", "v137", "v138", "v139", + "v140", "v141", "v142", "v143", "v144", "v145", "v146", "v147", + "v148", "v149", "v150", "v151", "v152", "v153", "v154", "v155", + "v156", "v157", "v158", "v159", "v160", "v161", "v162", "v163", + "v164", "v165", "v166", "v167", "v168", "v169", "v170", "v171", + "v172", "v173", "v174", "v175", "v176", "v177", "v178", "v179", + "v180", "v181", "v182", "v183", "v184", "v185", "v186", "v187", + "v188", "v189", "v190", "v191", "v192", "v193", "v194", "v195", + "v196", "v197", "v198", "v199", "v200", "v201", "v202", "v203", + "v204", "v205", "v206", "v207", "v208", "v209", "v210", "v211", + "v212", "v213", "v214", "v215", "v216", "v217", "v218", "v219", + "v220", "v221", "v222", "v223", "v224", "v225", "v226", "v227", + "v228", "v229", "v230", "v231", "v232", "v233", "v234", "v235", + "v236", "v237", "v238", "v239", "v240", "v241", "v242", "v243", + "v244", "v245", "v246", "v247", "v248", "v249", "v250", "v251", + "v252", "v253", "v254", "v255" + ); +#pragma clang diagnostic pop + // clang-format on + } +}; + +struct FlatmmSn_32x128x512_1x4x1_16x16x32_FP16 : public FlatmmSn_32x128x512_1x4x1_16x16x32_Base +{ + using BDataType = bf16_t; + using ODataType = bf16_t; + + // TODO: need paired with tile_window_linear! + // TODO: need call init_raw() before call this function! + // template + template + CK_TILE_DEVICE auto + operator()(const BRes& res_b, + const BCoords& cached_coords_b, + const ORes& res_o, + const OCoords& cached_coords_o, + const OFlags& o_flags, // this should be in sgpr + CK_TILE_LDS_ADDR void* smem, + index_t n, // loop along n dim + const ScaleTensor& scale_, + index_t tile_offset_b, // stride b is fixed to blockKr * blockW, but still can adjust + index_t tile_offset_o) + { + static_assert(BCoords::size() == 8); // 8 + static_assert(OCoords::size() == 8); + + const index_t tile_stride_b_bytes = tile_offset_b * sizeof(BDataType); + const index_t tile_stride_o_bytes = tile_offset_o * sizeof(ODataType); + + static_assert(ScaleTensor::size() == 2); + float s0 = scale_[number<0>{}]; + float s1 = scale_[number<1>{}]; + + index_t loop_cnt = n / Block_N; + + register float v_c0 asm("v64"); + register float v_c1 asm("v65"); + register float v_c2 asm("v66"); + register float v_c3 asm("v67"); + register float v_c4 asm("v68"); + register float v_c5 asm("v69"); + register float v_c6 asm("v70"); + register float v_c7 asm("v71"); + register float v_c8 asm("v72"); + register float v_c9 asm("v73"); + register float v_c10 asm("v74"); + register float v_c11 asm("v75"); + register float v_c12 asm("v76"); + register float v_c13 asm("v77"); + register float v_c14 asm("v78"); + register float v_c15 asm("v79"); + register float v_c16 asm("v80"); + register float v_c17 asm("v81"); + register float v_c18 asm("v82"); + register float v_c19 asm("v83"); + register float v_c20 asm("v84"); + register float v_c21 asm("v85"); + register float v_c22 asm("v86"); + register float v_c23 asm("v87"); + register float v_c24 asm("v88"); + register float v_c25 asm("v89"); + register float v_c26 asm("v90"); + register float v_c27 asm("v91"); + register float v_c28 asm("v92"); + register float v_c29 asm("v93"); + register float v_c30 asm("v94"); + register float v_c31 asm("v95"); + int32_t nan_hi = 0x7fff0000; + int32_t nan_lo = 0x00007fff; + + // in smem, the layout is M0(2)*K0(128)*M1(16)*K1(4) + // every threads need 8xK in contiguous register + // ... and every wave need the same data + int lane_id = threadIdx.x % 64; + int sld_y_os = (lane_id % 16) * 4 + (lane_id / 16) * 128; + sld_y_os *= 2; + + // y y p p p y + // reg before shfl M0(2)*N0(2)*Nl(4)*Nw(4)*Mw(16)*Nv(4) + // but order is N0*M0*Nv + // in LDS we need store as + // M0(2)* N0(2) * Nl(4) * Nw(4) * (Mw(16)*Nv(4) + 4) + // y y wave-id lid/16 lid%16 v + // sst(v3) = (v0/16*34 + v0%16 * 2 + wid*136) * 4 + int sfl_sst = (threadIdx.x % 16 * 4) + (threadIdx.x / 16) * (64 + 4); + sfl_sst *= 2; + + // from LDS we need load as + // M0(2)* N0(2) * Nl(4) * Nw(4) * (Mw(16) * Nv(4) + 4) + // ( 2 issue) (rem 32-lane) (4 wave*4issue) 2lane*1ussue(pk2) + // sld(v4) = v0/2 *34*4 + v0 % 2 *4 + wid*2 *4 + int sfl_sld = (lane_id % 2) * 2 + (lane_id / 2) * (64 + 4) + (threadIdx.x / 64) * 4; + sfl_sld *= 2; + + // B nr->kr + // clang-format off +#pragma clang diagnostic push +#pragma clang diagnostic ignored "-Winline-asm" + asm volatile( +#define CK_TILE_FLATMM_UK_MFMA CK_TILE_FLATMM_UK_MFMA_FP16 +#include "uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc" +#undef CK_TILE_FLATMM_UK_MFMA + :[smem_]"+r"(smem), + [s_loop_cnt]"+s"(loop_cnt), + [c0]"+v" (v_c0), + [c1]"+v" (v_c1), + [c2]"+v" (v_c2), + [c3]"+v" (v_c3), + [c4]"+v" (v_c4), + [c5]"+v" (v_c5), + [c6]"+v" (v_c6), + [c7]"+v" (v_c7), + [c8]"+v" (v_c8), + [c9]"+v" (v_c9), + [c10]"+v"(v_c10), + [c11]"+v"(v_c11), + [c12]"+v"(v_c12), + [c13]"+v"(v_c13), + [c14]"+v"(v_c14), + [c15]"+v"(v_c15), + [c16]"+v"(v_c16), + [c17]"+v"(v_c17), + [c18]"+v"(v_c18), + [c19]"+v"(v_c19), + [c20]"+v"(v_c20), + [c21]"+v"(v_c21), + [c22]"+v"(v_c22), + [c23]"+v"(v_c23), + [c24]"+v"(v_c24), + [c25]"+v"(v_c25), + [c26]"+v"(v_c26), + [c27]"+v"(v_c27), + [c28]"+v"(v_c28), + [c29]"+v"(v_c29), + [c30]"+v"(v_c30), + [c31]"+v"(v_c31) + : + [sld_a_base]"n"(0), + [shfl_base]"n"(0), + [v_sld_y_os]"v"(sld_y_os), + [v_sfl_sld]"v"(sfl_sld), + [v_sfl_sst]"v"(sfl_sst), + [s_res_o0]"s"(res_o[0]), + [s_res_o1]"s"(res_o[1]), + //[s_res_o2]"s"(res_o[2]), + //[s_res_o3]"s"(res_o[3]), + [s_res_b0]"s"(res_b[0]), + [s_res_b1]"s"(res_b[1]), + [s_res_b2]"s"(res_b[2]), + [s_res_b3]"s"(res_b[3]), + [v_os_o0]"v"(static_cast(cached_coords_o[number<0>{}] * sizeof(ODataType))), + [v_os_o1]"v"(static_cast(cached_coords_o[number<1>{}] * sizeof(ODataType))), + [v_os_o2]"v"(static_cast(cached_coords_o[number<2>{}] * sizeof(ODataType))), + [v_os_o3]"v"(static_cast(cached_coords_o[number<3>{}] * sizeof(ODataType))), + [v_os_o4]"v"(static_cast(cached_coords_o[number<4>{}] * sizeof(ODataType))), + [v_os_o5]"v"(static_cast(cached_coords_o[number<5>{}] * sizeof(ODataType))), + [v_os_o6]"v"(static_cast(cached_coords_o[number<6>{}] * sizeof(ODataType))), + [v_os_o7]"v"(static_cast(cached_coords_o[number<7>{}] * sizeof(ODataType))), + [v_os_b0]"v"(static_cast(cached_coords_b[number<0>{}] * sizeof(BDataType))), + [v_os_b1]"v"(static_cast(cached_coords_b[number<1>{}] * sizeof(BDataType))), + [v_os_b2]"v"(static_cast(cached_coords_b[number<2>{}] * sizeof(BDataType))), + [v_os_b3]"v"(static_cast(cached_coords_b[number<3>{}] * sizeof(BDataType))), + [v_os_b4]"v"(static_cast(cached_coords_b[number<4>{}] * sizeof(BDataType))), + [v_os_b5]"v"(static_cast(cached_coords_b[number<5>{}] * sizeof(BDataType))), + [v_os_b6]"v"(static_cast(cached_coords_b[number<6>{}] * sizeof(BDataType))), + [v_os_b7]"v"(static_cast(cached_coords_b[number<7>{}] * sizeof(BDataType))), + + [s_tile_os_o]"s"(tile_stride_o_bytes), + [s_tile_os_b]"s"(tile_stride_b_bytes), + [scale_0]"v"(s0), + [scale_1]"v"(s1), + [v_nan_lo]"v"(nan_lo), + [v_nan_hi]"v"(nan_hi), + [s_execflag_0]"s"(o_flags[number<0>{}]), + [s_execflag_1]"s"(o_flags[number<1>{}]), + [s_execflag_2]"s"(o_flags[number<2>{}]), + [s_execflag_3]"s"(o_flags[number<3>{}]), + [s_execflag_4]"s"(o_flags[number<4>{}]), + [s_execflag_5]"s"(o_flags[number<5>{}]), + [s_execflag_6]"s"(o_flags[number<6>{}]), + [s_execflag_7]"s"(o_flags[number<7>{}]) + : + "memory", "a0", "a1", "a2", "a3", "a4", "a5", "a6", "a7", "a8", "a9", + "a10", "a11", "a12", "a13", "a14", "a15", "a16", "a17", "a18", "a19", + "a20", "a21", "a22", "a23", "a24", "a25", "a26", "a27", "a28", "a29", + "a30", "a31", "a32", "a33", "a34", "a35", "a36", "a37", "a38", "a39", + "a40", "a41", "a42", "a43", "a44", "a45", "a46", "a47", "a48", "a49", + "a50", "a51", "a52", "a53", "a54", "a55", "a56", "a57", "a58", "a59", + "a60", "a61", "a62", "a63", "a64", "a65", "a66", "a67", "a68", "a69", + "a70", "a71", "a72", "a73", "a74", "a75", "a76", "a77", "a78", "a79", + "a80", "a81", "a82", "a83", "a84", "a85", "a86", "a87", "a88", "a89", + "a90", "a91", "a92", "a93", "a94", "a95", "a96", "a97", "a98", "a99", + "a100", "a101", "a102", "a103", "a104", "a105", "a106", "a107", + "a108", "a109", "a110", "a111", "a112", "a113", "a114", "a115", + "a116", "a117", "a118", "a119", "a120", "a121", "a122", "a123", + "a124", "a125", "a126", "a127", "a128", "a129", "a130", "a131", + "a132", "a133", "a134", "a135", "a136", "a137", "a138", "a139", + "a140", "a141", "a142", "a143", "a144", "a145", "a146", "a147", + "a148", "a149", "a150", "a151", "a152", "a153", "a154", "a155", + "a156", "a157", "a158", "a159", "a160", "a161", "a162", "a163", + "a164", "a165", "a166", "a167", "a168", "a169", "a170", "a171", + "a172", "a173", "a174", "a175", "a176", "a177", "a178", "a179", + "a180", "a181", "a182", "a183", "a184", "a185", "a186", "a187", + "a188", "a189", "a190", "a191", "a192", "a193", "a194", "a195", + "a196", "a197", "a198", "a199", "a200", "a201", "a202", "a203", + "a204", "a205", "a206", "a207", "a208", "a209", "a210", "a211", + "a212", "a213", "a214", "a215", "a216", "a217", "a218", "a219", + "a220", "a221", "a222", "a223", "a224", "a225", "a226", "a227", + "a228", "a229", "a230", "a231", "a232", "a233", "a234", "a235", + "a236", "a237", "a238", "a239", "a240", "a241", "a242", "a243", + "a244", "a245", "a246", "a247", "a248", "a249", "a250", "a251", + "a252", "a253", "a254", "a255", + "s8", "s9", "s12", "s13", "s14", "s15", "s38", "s39", "s52", "s86", + "s36", "s37", + "v50", "v54", "v55", + "v64","v65","v66","v67","v68","v69","v70","v71", + "v72","v73","v74","v75","v76","v77","v78","v79", + "v80","v81","v82","v83","v84","v85","v86","v87", + "v88","v89","v90","v91","v92","v93","v94","v95", + "v128", "v129", "v130", "v131", + "v132", "v133", "v134", "v135", "v136", "v137", "v138", "v139", + "v140", "v141", "v142", "v143", "v144", "v145", "v146", "v147", + "v148", "v149", "v150", "v151", "v152", "v153", "v154", "v155", + "v156", "v157", "v158", "v159", "v160", "v161", "v162", "v163", + "v164", "v165", "v166", "v167", "v168", "v169", "v170", "v171", + "v172", "v173", "v174", "v175", "v176", "v177", "v178", "v179", + "v180", "v181", "v182", "v183", "v184", "v185", "v186", "v187", + "v188", "v189", "v190", "v191", "v192", "v193", "v194", "v195", + "v196", "v197", "v198", "v199", "v200", "v201", "v202", "v203", + "v204", "v205", "v206", "v207", "v208", "v209", "v210", "v211", + "v212", "v213", "v214", "v215", "v216", "v217", "v218", "v219", + "v220", "v221", "v222", "v223", "v224", "v225", "v226", "v227", + "v228", "v229", "v230", "v231", "v232", "v233", "v234", "v235", + "v236", "v237", "v238", "v239", "v240", "v241", "v242", "v243", + "v244", "v245", "v246", "v247", "v248", "v249", "v250", "v251", + "v252", "v253", "v254", "v255" + ); +#pragma clang diagnostic pop + // clang-format on + } +}; + +} // namespace ck_tile diff --git a/include/ck_tile/ops/flatmm/block/flatmm_uk_config.hpp b/include/ck_tile/ops/flatmm/block/flatmm_uk_config.hpp new file mode 100644 index 0000000000..003335c0e7 --- /dev/null +++ b/include/ck_tile/ops/flatmm/block/flatmm_uk_config.hpp @@ -0,0 +1,10 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#define CK_TILE_FLATMM_UK_MFMA_FP16 0 +#define CK_TILE_FLATMM_UK_MFMA_BF16 1 +#define CK_TILE_FLATMM_UK_MFMA_INT8 2 +#define CK_TILE_FLATMM_UK_MFMA_FP8 3 +#define CK_TILE_FLATMM_UK_MFMA_BF8 4 diff --git a/include/ck_tile/ops/flatmm/block/uk/README.md b/include/ck_tile/ops/flatmm/block/uk/README.md new file mode 100644 index 0000000000..84fa132296 --- /dev/null +++ b/include/ck_tile/ops/flatmm/block/uk/README.md @@ -0,0 +1 @@ +the files under this folder should not be included directly! \ No newline at end of file diff --git a/include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc b/include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc new file mode 100644 index 0000000000..8b57611f06 --- /dev/null +++ b/include/ck_tile/ops/flatmm/block/uk/flatmm_sn_uk_gfx9_32x128x512_1x4x1_16x16x16.inc @@ -0,0 +1,613 @@ +#ifndef CK_TILE_FLATMM_UK_MFMA +#define CK_TILE_FLATMM_UK_MFMA CK_TILE_FLATMM_UK_MFMA_BF16 +#endif + +#if CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_BF16 +# define _UK_MFMA_ "v_mfma_f32_16x16x16_bf16" + +# define _UK_PK_CVT_(x0_, x1_, y_) \ + " v_cmp_u_f32 s[36:37], " x0_ ", " x0_ " \n" \ + " v_add3_u32 v50, " x0_ ", %[v_nan_lo], 1 \n" \ + " v_cndmask_b32 v54, v50, %[v_nan_hi], s[36:37] \n" \ + " v_cmp_u_f32 s[36:37], " x1_ ", " x1_ " \n" \ + " v_add3_u32 v50, " x1_ ", %[v_nan_lo], 1 \n" \ + " v_cndmask_b32 v55, v50, %[v_nan_hi], s[36:37] \n" \ + " v_perm_b32 " y_ ", v55, v54, s52 \n" + +# define _UK_ATOMIC_ADD_ "global_atomic_pk_add_bf16" + +#elif CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_FP16 +#define _UK_MFMA_ "v_mfma_f32_16x16x16_f16" + +# define _UK_PK_CVT_(x0_, x1_, y_) \ + " v_cvt_f16_f32 v54, " x0_ " \n" \ + " v_cvt_f16_f32 v55, " x1_ " \n" \ + " v_pack_b32_f16 " y_ ", v54, v55 \n" + +# define _UK_ATOMIC_ADD_ "global_atomic_pk_add_f16" + +#endif + + +";-------------------------------------------------------------\n" +" s_mov_b32 s52, 0x07060302 ; v_perm\n" +" s_mov_b64 s[38:39], exec ; save current exec\n" +" s_mov_b32 s8, %[s_res_o0] \n" +" s_mov_b32 s9, %[s_res_o1] \n" +" s_mov_b32 s12, %[s_res_b0] \n" +" s_mov_b32 s13, %[s_res_b1] \n" +" s_mov_b32 s14, %[s_res_b2] \n" +" s_mov_b32 s15, %[s_res_b3] \n" +" ds_read_b64 v[128:129], %[v_sld_y_os] offset:0 + %[sld_a_base] \n" +" ds_read_b64 v[130:131], %[v_sld_y_os] offset:128 + %[sld_a_base] \n" +" ds_read_b64 v[132:133], %[v_sld_y_os] offset:1024 + %[sld_a_base] \n" +" ds_read_b64 v[134:135], %[v_sld_y_os] offset:1152 + %[sld_a_base] \n" +" ds_read_b64 v[136:137], %[v_sld_y_os] offset:2048 + %[sld_a_base] \n" +" ds_read_b64 v[138:139], %[v_sld_y_os] offset:2176 + %[sld_a_base] \n" +" ds_read_b64 v[140:141], %[v_sld_y_os] offset:3072 + %[sld_a_base] \n" +" ds_read_b64 v[142:143], %[v_sld_y_os] offset:3200 + %[sld_a_base] \n" +" ds_read_b64 v[144:145], %[v_sld_y_os] offset:4096 + %[sld_a_base] \n" +" ds_read_b64 v[146:147], %[v_sld_y_os] offset:4224 + %[sld_a_base] \n" +" ds_read_b64 v[148:149], %[v_sld_y_os] offset:5120 + %[sld_a_base] \n" +" ds_read_b64 v[150:151], %[v_sld_y_os] offset:5248 + %[sld_a_base] \n" +" ds_read_b64 v[152:153], %[v_sld_y_os] offset:6144 + %[sld_a_base] \n" +" ds_read_b64 v[154:155], %[v_sld_y_os] offset:6272 + %[sld_a_base] \n" +" ds_read_b64 v[156:157], %[v_sld_y_os] offset:7168 + %[sld_a_base] \n" +" ds_read_b64 v[158:159], %[v_sld_y_os] offset:7296 + %[sld_a_base] \n" +" ds_read_b64 v[160:161], %[v_sld_y_os] offset:8192 + %[sld_a_base] \n" +" ds_read_b64 v[162:163], %[v_sld_y_os] offset:8320 + %[sld_a_base] \n" +" ds_read_b64 v[164:165], %[v_sld_y_os] offset:9216 + %[sld_a_base] \n" +" ds_read_b64 v[166:167], %[v_sld_y_os] offset:9344 + %[sld_a_base] \n" +" ds_read_b64 v[168:169], %[v_sld_y_os] offset:10240 + %[sld_a_base] \n" +" ds_read_b64 v[170:171], %[v_sld_y_os] offset:10368 + %[sld_a_base] \n" +" ds_read_b64 v[172:173], %[v_sld_y_os] offset:11264 + %[sld_a_base] \n" +" ds_read_b64 v[174:175], %[v_sld_y_os] offset:11392 + %[sld_a_base] \n" +" ds_read_b64 v[176:177], %[v_sld_y_os] offset:12288 + %[sld_a_base] \n" +" ds_read_b64 v[178:179], %[v_sld_y_os] offset:12416 + %[sld_a_base] \n" +" ds_read_b64 v[180:181], %[v_sld_y_os] offset:13312 + %[sld_a_base] \n" +" ds_read_b64 v[182:183], %[v_sld_y_os] offset:13440 + %[sld_a_base] \n" +" ds_read_b64 v[184:185], %[v_sld_y_os] offset:14336 + %[sld_a_base] \n" +" ds_read_b64 v[186:187], %[v_sld_y_os] offset:14464 + %[sld_a_base] \n" +" ds_read_b64 v[188:189], %[v_sld_y_os] offset:15360 + %[sld_a_base] \n" +" ds_read_b64 v[190:191], %[v_sld_y_os] offset:15488 + %[sld_a_base] \n" +" ds_read_b64 v[192:193], %[v_sld_y_os] offset:16384 + %[sld_a_base] \n" +" ds_read_b64 v[194:195], %[v_sld_y_os] offset:16512 + %[sld_a_base] \n" +" ds_read_b64 v[196:197], %[v_sld_y_os] offset:17408 + %[sld_a_base] \n" +" ds_read_b64 v[198:199], %[v_sld_y_os] offset:17536 + %[sld_a_base] \n" +" ds_read_b64 v[200:201], %[v_sld_y_os] offset:18432 + %[sld_a_base] \n" +" ds_read_b64 v[202:203], %[v_sld_y_os] offset:18560 + %[sld_a_base] \n" +" ds_read_b64 v[204:205], %[v_sld_y_os] offset:19456 + %[sld_a_base] \n" +" ds_read_b64 v[206:207], %[v_sld_y_os] offset:19584 + %[sld_a_base] \n" +" ds_read_b64 v[208:209], %[v_sld_y_os] offset:20480 + %[sld_a_base] \n" +" ds_read_b64 v[210:211], %[v_sld_y_os] offset:20608 + %[sld_a_base] \n" +" ds_read_b64 v[212:213], %[v_sld_y_os] offset:21504 + %[sld_a_base] \n" +" ds_read_b64 v[214:215], %[v_sld_y_os] offset:21632 + %[sld_a_base] \n" +" ds_read_b64 v[216:217], %[v_sld_y_os] offset:22528 + %[sld_a_base] \n" +" ds_read_b64 v[218:219], %[v_sld_y_os] offset:22656 + %[sld_a_base] \n" +" ds_read_b64 v[220:221], %[v_sld_y_os] offset:23552 + %[sld_a_base] \n" +" ds_read_b64 v[222:223], %[v_sld_y_os] offset:23680 + %[sld_a_base] \n" +" ds_read_b64 v[224:225], %[v_sld_y_os] offset:24576 + %[sld_a_base] \n" +" ds_read_b64 v[226:227], %[v_sld_y_os] offset:24704 + %[sld_a_base] \n" +" ds_read_b64 v[228:229], %[v_sld_y_os] offset:25600 + %[sld_a_base] \n" +" ds_read_b64 v[230:231], %[v_sld_y_os] offset:25728 + %[sld_a_base] \n" +" ds_read_b64 v[232:233], %[v_sld_y_os] offset:26624 + %[sld_a_base] \n" +" ds_read_b64 v[234:235], %[v_sld_y_os] offset:26752 + %[sld_a_base] \n" +" ds_read_b64 v[236:237], %[v_sld_y_os] offset:27648 + %[sld_a_base] \n" +" ds_read_b64 v[238:239], %[v_sld_y_os] offset:27776 + %[sld_a_base] \n" +" ds_read_b64 v[240:241], %[v_sld_y_os] offset:28672 + %[sld_a_base] \n" +" ds_read_b64 v[242:243], %[v_sld_y_os] offset:28800 + %[sld_a_base] \n" +" ds_read_b64 v[244:245], %[v_sld_y_os] offset:29696 + %[sld_a_base] \n" +" ds_read_b64 v[246:247], %[v_sld_y_os] offset:29824 + %[sld_a_base] \n" +" ds_read_b64 v[248:249], %[v_sld_y_os] offset:30720 + %[sld_a_base] \n" +" ds_read_b64 v[250:251], %[v_sld_y_os] offset:30848 + %[sld_a_base] \n" +" ds_read_b64 v[252:253], %[v_sld_y_os] offset:31744 + %[sld_a_base] \n" +" ds_read_b64 v[254:255], %[v_sld_y_os] offset:31872 + %[sld_a_base] \n" +" s_waitcnt 0 \n" +" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n" +" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n" +" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n" +" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n" +" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n" +" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n" +" s_cselect_b32 s86, %[s_tile_os_b], 0 \n" +" s_add_u32 s12, s86, s12 \n" +" s_addc_u32 s13, 0, s13 \n" +" s_waitcnt 0 \n" +"L_start%=: \n" +" s_waitcnt vmcnt(32) \n" +" s_barrier \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[0:1], v[128:129], 0 \n" +" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[2:3], v[130:131], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[4:5], v[132:133], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[6:7], v[134:135], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[8:9], v[136:137], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[10:11], v[138:139], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[12:13], v[140:141], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[14:15], v[142:143], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[0:1], v[192:193], 0 \n" +" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[2:3], v[194:195], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[4:5], v[196:197], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[6:7], v[198:199], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[8:9], v[200:201], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[10:11], v[202:203], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[12:13], v[204:205], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[14:15], v[206:207], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[16:17], v[128:129], 0 \n" +" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[18:19], v[130:131], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[20:21], v[132:133], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[22:23], v[134:135], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[24:25], v[136:137], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[26:27], v[138:139], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[28:29], v[140:141], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[30:31], v[142:143], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[16:17], v[192:193], 0 \n" +" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[18:19], v[194:195], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[20:21], v[196:197], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[22:23], v[198:199], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[24:25], v[200:201], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[26:27], v[202:203], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[28:29], v[204:205], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[30:31], v[206:207], [%[c12], %[c13], %[c14], %[c15]] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[32:33], v[144:145], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[34:35], v[146:147], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[36:37], v[148:149], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[38:39], v[150:151], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[40:41], v[152:153], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[42:43], v[154:155], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[44:45], v[156:157], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[46:47], v[158:159], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[32:33], v[208:209], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[34:35], v[210:211], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[36:37], v[212:213], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[38:39], v[214:215], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[40:41], v[216:217], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[42:43], v[218:219], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[44:45], v[220:221], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[46:47], v[222:223], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[48:49], v[144:145], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[50:51], v[146:147], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[52:53], v[148:149], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[54:55], v[150:151], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[56:57], v[152:153], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[58:59], v[154:155], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[60:61], v[156:157], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[62:63], v[158:159], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[48:49], v[208:209], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[50:51], v[210:211], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[52:53], v[212:213], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[54:55], v[214:215], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[56:57], v[216:217], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[58:59], v[218:219], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[60:61], v[220:221], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[62:63], v[222:223], [%[c12], %[c13], %[c14], %[c15]] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[64:65], v[160:161], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[66:67], v[162:163], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[68:69], v[164:165], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[70:71], v[166:167], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[72:73], v[168:169], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[74:75], v[170:171], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[76:77], v[172:173], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[78:79], v[174:175], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[64:65], v[224:225], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[66:67], v[226:227], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[68:69], v[228:229], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[70:71], v[230:231], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[72:73], v[232:233], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[74:75], v[234:235], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[76:77], v[236:237], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[78:79], v[238:239], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[80:81], v[160:161], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[82:83], v[162:163], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[84:85], v[164:165], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[86:87], v[166:167], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[88:89], v[168:169], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[90:91], v[170:171], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[92:93], v[172:173], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[94:95], v[174:175], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[80:81], v[224:225], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[82:83], v[226:227], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[84:85], v[228:229], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[86:87], v[230:231], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[88:89], v[232:233], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[90:91], v[234:235], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[92:93], v[236:237], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[94:95], v[238:239], [%[c12], %[c13], %[c14], %[c15]] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[96:97], v[176:177], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[98:99], v[178:179], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[100:101], v[180:181], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[102:103], v[182:183], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[104:105], v[184:185], [%[c0], %[c1], %[c2], %[c3]] \n" +" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[106:107], v[186:187], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[108:109], v[188:189], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c0], %[c1], %[c2], %[c3]], acc[110:111], v[190:191], [%[c0], %[c1], %[c2], %[c3]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[96:97], v[240:241], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[98:99], v[242:243], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[100:101], v[244:245], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[102:103], v[246:247], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[104:105], v[248:249], [%[c4], %[c5], %[c6], %[c7]] \n" +" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[106:107], v[250:251], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[108:109], v[252:253], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c4], %[c5], %[c6], %[c7]], acc[110:111], v[254:255], [%[c4], %[c5], %[c6], %[c7]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[112:113], v[176:177], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[114:115], v[178:179], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[116:117], v[180:181], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[118:119], v[182:183], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[120:121], v[184:185], [%[c8], %[c9], %[c10], %[c11]] \n" +" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[122:123], v[186:187], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[124:125], v[188:189], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c8], %[c9], %[c10], %[c11]], acc[126:127], v[190:191], [%[c8], %[c9], %[c10], %[c11]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[112:113], v[240:241], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[114:115], v[242:243], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[116:117], v[244:245], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[118:119], v[246:247], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[120:121], v[248:249], [%[c12], %[c13], %[c14], %[c15]] \n" +" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[122:123], v[250:251], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[124:125], v[252:253], [%[c12], %[c13], %[c14], %[c15]] \n" +_UK_MFMA_ " [%[c12], %[c13], %[c14], %[c15]], acc[126:127], v[254:255], [%[c12], %[c13], %[c14], %[c15]]\n" +" v_mul_f32 %[c0], %[scale_0], %[c0] \n" +" v_mul_f32 %[c1], %[scale_0], %[c1] \n" +" v_mul_f32 %[c2], %[scale_0], %[c2] \n" +" v_mul_f32 %[c3], %[scale_0], %[c3] \n" +" v_mul_f32 %[c4], %[scale_1], %[c4] \n" +" v_mul_f32 %[c5], %[scale_1], %[c5] \n" +" v_mul_f32 %[c6], %[scale_1], %[c6] \n" +" v_mul_f32 %[c7], %[scale_1], %[c7] \n" +" v_mul_f32 %[c8], %[scale_0], %[c8] \n" +" v_mul_f32 %[c9], %[scale_0], %[c9] \n" +" v_mul_f32 %[c10], %[scale_0], %[c10] \n" +" v_mul_f32 %[c11], %[scale_0], %[c11] \n" +" v_mul_f32 %[c12], %[scale_1], %[c12] \n" +" v_mul_f32 %[c13], %[scale_1], %[c13] \n" +" v_mul_f32 %[c14], %[scale_1], %[c14] \n" +" v_mul_f32 %[c15], %[scale_1], %[c15] \n" +_UK_PK_CVT_("%[c0]", "%[c1]", "%[c0]") +_UK_PK_CVT_("%[c2]", "%[c3]", "%[c1]") +_UK_PK_CVT_("%[c4]", "%[c5]", "%[c2]") +_UK_PK_CVT_("%[c6]", "%[c7]", "%[c3]") +_UK_PK_CVT_("%[c8]", "%[c9]", "%[c4]") +_UK_PK_CVT_("%[c10]", "%[c11]", "%[c5]") +_UK_PK_CVT_("%[c12]", "%[c13]", "%[c6]") +_UK_PK_CVT_("%[c14]", "%[c15]", "%[c7]") +" ;------------------------------ \n" +" ds_write_b64 %[v_sfl_sst], [%[c0],%[c1]] offset:0 + %[shfl_base] \n" +" ds_write_b64 %[v_sfl_sst], [%[c2],%[c3]] offset:4352 + %[shfl_base] \n" +" ds_write_b64 %[v_sfl_sst], [%[c4],%[c5]] offset:2176 + %[shfl_base] \n" +" ds_write_b64 %[v_sfl_sst], [%[c6],%[c7]] offset:6528 + %[shfl_base] \n" +" s_waitcnt lgkmcnt(0) \n" +" s_barrier \n" +" ds_read_b32 %[c0], %[v_sfl_sld] offset:0 + %[shfl_base] \n" +" ds_read_b32 %[c1], %[v_sfl_sld] offset:32 + %[shfl_base] \n" +" ds_read_b32 %[c2], %[v_sfl_sld] offset:64 + %[shfl_base] \n" +" ds_read_b32 %[c3], %[v_sfl_sld] offset:96 + %[shfl_base] \n" +" ds_read_b32 %[c4], %[v_sfl_sld] offset:4352 + %[shfl_base] \n" +" ds_read_b32 %[c5], %[v_sfl_sld] offset:4384 + %[shfl_base] \n" +" ds_read_b32 %[c6], %[v_sfl_sld] offset:4416 + %[shfl_base] \n" +" ds_read_b32 %[c7], %[v_sfl_sld] offset:4448 + %[shfl_base] \n" +" s_waitcnt lgkmcnt(0) \n" +" s_mov_b64 exec, %[s_execflag_0] \n" +_UK_ATOMIC_ADD_ " %[v_os_o0], %[c0], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_1] \n" +_UK_ATOMIC_ADD_ " %[v_os_o1], %[c1], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_2] \n" +_UK_ATOMIC_ADD_ " %[v_os_o2], %[c2], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_3] \n" +_UK_ATOMIC_ADD_ " %[v_os_o3], %[c3], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_4] \n" +_UK_ATOMIC_ADD_ " %[v_os_o4], %[c4], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_5] \n" +_UK_ATOMIC_ADD_ " %[v_os_o5], %[c5], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_6] \n" +_UK_ATOMIC_ADD_ " %[v_os_o6], %[c6], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_7] \n" +_UK_ATOMIC_ADD_ " %[v_os_o7], %[c7], s[8:9] \n" +" s_mov_b64 exec, s[38:39] \n" +" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 ; k-- \n" +" s_cmp_gt_i32 %[s_loop_cnt] 0 \n" +" s_cbranch_scc0 L_end%= \n" +" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n" +" s_cselect_b32 s86, %[s_tile_os_b], 0 \n" +" s_add_u32 s12, s86, s12 \n" +" s_addc_u32 s13, 0, s13 \n" +" s_add_u32 s8, %[s_tile_os_o], s8 \n" +" s_addc_u32 s9, 0, s9 \n" +" s_waitcnt vmcnt(32) \n" +" s_barrier \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[128:129], v[128:129], 0 \n" +" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[130:131], v[130:131], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[132:133], v[132:133], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[134:135], v[134:135], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[136:137], v[136:137], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[138:139], v[138:139], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[140:141], v[140:141], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[142:143], v[142:143], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[128:129], v[192:193], 0 \n" +" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[130:131], v[194:195], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[132:133], v[196:197], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[134:135], v[198:199], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[136:137], v[200:201], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[138:139], v[202:203], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[140:141], v[204:205], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[142:143], v[206:207], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[144:145], v[128:129], 0 \n" +" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[146:147], v[130:131], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[148:149], v[132:133], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[150:151], v[134:135], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[152:153], v[136:137], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[154:155], v[138:139], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[156:157], v[140:141], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[158:159], v[142:143], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[144:145], v[192:193], 0 \n" +" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[146:147], v[194:195], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[148:149], v[196:197], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[150:151], v[198:199], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[152:153], v[200:201], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[154:155], v[202:203], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[156:157], v[204:205], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[158:159], v[206:207], [%[c28],%[c29],%[c30],%[c31]] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[160:161], v[144:145], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[162:163], v[146:147], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[164:165], v[148:149], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[166:167], v[150:151], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[168:169], v[152:153], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[170:171], v[154:155], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[172:173], v[156:157], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[174:175], v[158:159], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[160:161], v[208:209], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[162:163], v[210:211], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[164:165], v[212:213], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[166:167], v[214:215], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[168:169], v[216:217], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[170:171], v[218:219], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[172:173], v[220:221], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[174:175], v[222:223], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[176:177], v[144:145], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[178:179], v[146:147], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[180:181], v[148:149], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[182:183], v[150:151], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[184:185], v[152:153], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[186:187], v[154:155], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[188:189], v[156:157], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[190:191], v[158:159], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[176:177], v[208:209], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[178:179], v[210:211], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[180:181], v[212:213], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[182:183], v[214:215], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[184:185], v[216:217], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[186:187], v[218:219], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[188:189], v[220:221], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[190:191], v[222:223], [%[c28],%[c29],%[c30],%[c31]] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[192:193], v[160:161], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[194:195], v[162:163], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[196:197], v[164:165], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[198:199], v[166:167], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[200:201], v[168:169], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[202:203], v[170:171], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[204:205], v[172:173], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[206:207], v[174:175], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[192:193], v[224:225], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[194:195], v[226:227], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[196:197], v[228:229], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[198:199], v[230:231], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[200:201], v[232:233], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[202:203], v[234:235], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[204:205], v[236:237], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[206:207], v[238:239], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[208:209], v[160:161], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[210:211], v[162:163], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[212:213], v[164:165], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[214:215], v[166:167], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[216:217], v[168:169], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[218:219], v[170:171], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[220:221], v[172:173], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[222:223], v[174:175], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[208:209], v[224:225], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[210:211], v[226:227], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[212:213], v[228:229], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[214:215], v[230:231], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[216:217], v[232:233], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[218:219], v[234:235], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[220:221], v[236:237], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[222:223], v[238:239], [%[c28],%[c29],%[c30],%[c31]] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[224:225], v[176:177], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[226:227], v[178:179], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[228:229], v[180:181], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[230:231], v[182:183], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[232:233], v[184:185], [%[c16],%[c17],%[c18],%[c19]] \n" +" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[234:235], v[186:187], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[236:237], v[188:189], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c16],%[c17],%[c18],%[c19]], acc[238:239], v[190:191], [%[c16],%[c17],%[c18],%[c19]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[224:225], v[240:241], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[226:227], v[242:243], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[228:229], v[244:245], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[230:231], v[246:247], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[232:233], v[248:249], [%[c20],%[c21],%[c22],%[c23]] \n" +" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[234:235], v[250:251], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[236:237], v[252:253], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c20],%[c21],%[c22],%[c23]], acc[238:239], v[254:255], [%[c20],%[c21],%[c22],%[c23]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[240:241], v[176:177], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[12:15], 0 offen \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[242:243], v[178:179], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[244:245], v[180:181], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[246:247], v[182:183], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[248:249], v[184:185], [%[c24],%[c25],%[c26],%[c27]] \n" +" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[12:15], 0 offen offset:1024 \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[250:251], v[186:187], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[252:253], v[188:189], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c24],%[c25],%[c26],%[c27]], acc[254:255], v[190:191], [%[c24],%[c25],%[c26],%[c27]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[240:241], v[240:241], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[12:15], 0 offen offset:2048 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[242:243], v[242:243], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[244:245], v[244:245], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[246:247], v[246:247], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[248:249], v[248:249], [%[c28],%[c29],%[c30],%[c31]] \n" +" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[12:15], 0 offen offset:3072 \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[250:251], v[250:251], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[252:253], v[252:253], [%[c28],%[c29],%[c30],%[c31]] \n" +_UK_MFMA_ " [%[c28],%[c29],%[c30],%[c31]], acc[254:255], v[254:255], [%[c28],%[c29],%[c30],%[c31]]\n" +" v_mul_f32 %[c16], %[scale_0], %[c16] \n" +" v_mul_f32 %[c17], %[scale_0], %[c17] \n" +" v_mul_f32 %[c18], %[scale_0], %[c18] \n" +" v_mul_f32 %[c19], %[scale_0], %[c19] \n" +" v_mul_f32 %[c20], %[scale_1], %[c20] \n" +" v_mul_f32 %[c21], %[scale_1], %[c21] \n" +" v_mul_f32 %[c22], %[scale_1], %[c22] \n" +" v_mul_f32 %[c23], %[scale_1], %[c23] \n" +" v_mul_f32 %[c24], %[scale_0], %[c24] \n" +" v_mul_f32 %[c25], %[scale_0], %[c25] \n" +" v_mul_f32 %[c26], %[scale_0], %[c26] \n" +" v_mul_f32 %[c27], %[scale_0], %[c27] \n" +" v_mul_f32 %[c28], %[scale_1], %[c28] \n" +" v_mul_f32 %[c29], %[scale_1], %[c29] \n" +" v_mul_f32 %[c30], %[scale_1], %[c30] \n" +" v_mul_f32 %[c31], %[scale_1], %[c31] \n" + +_UK_PK_CVT_("%[c16]", "%[c17]", "%[c16]") +_UK_PK_CVT_("%[c18]", "%[c19]", "%[c17]") +_UK_PK_CVT_("%[c20]", "%[c21]", "%[c18]") +_UK_PK_CVT_("%[c22]", "%[c23]", "%[c19]") +_UK_PK_CVT_("%[c24]", "%[c25]", "%[c20]") +_UK_PK_CVT_("%[c26]", "%[c27]", "%[c21]") +_UK_PK_CVT_("%[c28]", "%[c29]", "%[c22]") +_UK_PK_CVT_("%[c30]", "%[c31]", "%[c23]") + +" ;------------------------------ \n" +" ds_write_b64 %[v_sfl_sst], [%[c16],%[c17]] offset:0 + %[shfl_base] \n" +" ds_write_b64 %[v_sfl_sst], [%[c18],%[c19]] offset:4352 + %[shfl_base] \n" +" ds_write_b64 %[v_sfl_sst], [%[c20],%[c21]] offset:2176 + %[shfl_base] \n" +" ds_write_b64 %[v_sfl_sst], [%[c22],%[c23]] offset:6528 + %[shfl_base] \n" +" s_waitcnt lgkmcnt(0) \n" +" s_barrier \n" +" ds_read_b32 %[c16], %[v_sfl_sld] offset:0 + %[shfl_base] \n" +" ds_read_b32 %[c17], %[v_sfl_sld] offset:32 + %[shfl_base] \n" +" ds_read_b32 %[c18], %[v_sfl_sld] offset:64 + %[shfl_base] \n" +" ds_read_b32 %[c19], %[v_sfl_sld] offset:96 + %[shfl_base] \n" +" ds_read_b32 %[c20], %[v_sfl_sld] offset:4352 + %[shfl_base] \n" +" ds_read_b32 %[c21], %[v_sfl_sld] offset:4384 + %[shfl_base] \n" +" ds_read_b32 %[c22], %[v_sfl_sld] offset:4416 + %[shfl_base] \n" +" ds_read_b32 %[c23], %[v_sfl_sld] offset:4448 + %[shfl_base] \n" +" s_waitcnt lgkmcnt(0) \n" +" s_mov_b64 exec, %[s_execflag_0] \n" +_UK_ATOMIC_ADD_ " %[v_os_o0], %[c16], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_1] \n" +_UK_ATOMIC_ADD_ " %[v_os_o1], %[c17], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_2] \n" +_UK_ATOMIC_ADD_ " %[v_os_o2], %[c18], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_3] \n" +_UK_ATOMIC_ADD_ " %[v_os_o3], %[c19], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_4] \n" +_UK_ATOMIC_ADD_ " %[v_os_o4], %[c20], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_5] \n" +_UK_ATOMIC_ADD_ " %[v_os_o5], %[c21], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_6] \n" +_UK_ATOMIC_ADD_ " %[v_os_o6], %[c22], s[8:9] \n" +" s_mov_b64 exec, %[s_execflag_7] \n" +_UK_ATOMIC_ADD_ " %[v_os_o7], %[c23], s[8:9] \n" +" s_mov_b64 exec, s[38:39] \n" +" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 ; k-- \n" +" s_cmp_gt_i32 %[s_loop_cnt] 0 \n" +" s_cbranch_scc0 L_end%= \n" +" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n" +" s_cselect_b32 s86, %[s_tile_os_b], 0 \n" +" s_add_u32 s12, s86, s12 \n" +" s_addc_u32 s13, 0, s13 \n" +" s_add_u32 s8, %[s_tile_os_o], s8 \n" +" s_addc_u32 s9, 0, s9 \n" +" s_branch L_start%= \n" +"L_end%=: \n" + +#undef _UK_MFMA_ +#undef _UK_PK_CVT_ +#undef _UK_ATOMIC_ADD_ diff --git a/include/ck_tile/ops/flatmm/block/uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc b/include/ck_tile/ops/flatmm/block/uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc new file mode 100644 index 0000000000..a34a21d39f --- /dev/null +++ b/include/ck_tile/ops/flatmm/block/uk/flatmm_uk_gfx9_32x512x128_1x1x1_16x16x16.inc @@ -0,0 +1,516 @@ +#ifndef CK_TILE_FLATMM_UK_MFMA +#define CK_TILE_FLATMM_UK_MFMA CK_TILE_FLATMM_UK_MFMA_BF16 +#endif + +#if CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_BF16 +#define _UK_MFMA_ "v_mfma_f32_16x16x16_bf16" +#elif CK_TILE_FLATMM_UK_MFMA == CK_TILE_FLATMM_UK_MFMA_FP16 +#define _UK_MFMA_ "v_mfma_f32_16x16x16_f16" +#endif + +"s_mov_b32 s16, %[s_res_a0] \n" +"s_mov_b32 s17, %[s_res_a1] \n" +"s_mov_b32 s18, %[s_res_a2] \n" +"s_mov_b32 s19, %[s_res_a3] \n" +"s_mov_b32 s20, %[s_res_b0] \n" +"s_mov_b32 s21, %[s_res_b1] \n" +"s_mov_b32 s22, %[s_res_b2] \n" +"s_mov_b32 s23, %[s_res_b3] \n" +// "s_nop 4\n" +"; -- prefetch A0\n" +"s_add_u32 m0, 0, %[s_m0_init] \n" +"buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[smem_sz], %[s_m0_init] \n" +"s_cmp_gt_i32 %[s_loop_cnt] 1 ; move a with cond \n" +"s_cselect_b32 s86, %[s_tile_os_a], 0 ; move a with cond \n" +"s_add_u32 s16, s86, s16 ; move a with cond \n" +"s_addc_u32 s17, 0, s17 ; move a with cond \n" +"; -- prefetch A1\n" +"buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n" +"s_add_u32 m0, %[s_size_per_issue], m0 \n" +"buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n" +"s_add_u32 m0, 0, %[s_m0_init] \n" +"s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n" +"s_cselect_b32 s86, %[s_tile_os_a], 0 ; move a with cond \n" +"s_add_u32 s16, s86, s16 ; move a with cond \n" +"s_addc_u32 s17, 0, s17 ; move a with cond \n" +"; -- prefetch B0\n" +"buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[20:23], 0 offen offset:3072 \n" +"buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[20:23], 0 offen \n" +"buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[20:23], 0 offen offset:1024 \n" +"buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[20:23], 0 offen offset:2048 \n" +"buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[20:23], 0 offen offset:3072 \n" +"s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n" +"s_cselect_b32 s86, %[s_tile_os_b], 0 ; move b with cond \n" +"s_add_u32 s20, s86, s20 ; move b with cond \n" +"s_addc_u32 s21, 0, s21 ; move b with cond \n" +"s_waitcnt vmcnt(40) \n" +"s_barrier \n" +"ds_read_b128 v[64:67], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_0]\n" // 1024: N stride, 64 K stride +"ds_read_b128 v[68:71], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_1]\n" +"ds_read_b128 v[72:75], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_2]\n" +"ds_read_b128 v[76:79], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_3]\n" +"ds_read_b128 v[80:83], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_4]\n" +"ds_read_b128 v[84:87], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_5]\n" +"ds_read_b128 v[88:91], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_6]\n" +"ds_read_b128 v[92:95], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_7]\n" +"L_start%=: \n" +" s_waitcnt vmcnt(24) & lgkmcnt(0) \n" +" s_barrier \n" +_UK_MFMA_ " %[v_acc_0], acc[0:1], v[64:65], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[2:3], v[66:67], %[v_acc_0] \n" +" buffer_load_dwordx4 acc[128:131], %[v_os_b0], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_0], acc[4:5], v[68:69], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[6:7], v[70:71], %[v_acc_0] \n" +" buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_0], acc[8:9], v[72:73], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[10:11], v[74:75], %[v_acc_0] \n" +" buffer_load_dwordx4 acc[132:135], %[v_os_b0], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_0], acc[12:13], v[76:77], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[14:15], v[78:79], %[v_acc_0] \n" +" buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_1], acc[0:1], v[80:81], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[2:3], v[82:83], %[v_acc_1] \n" +" buffer_load_dwordx4 acc[136:139], %[v_os_b0], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_1], acc[4:5], v[84:85], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[6:7], v[86:87], %[v_acc_1] \n" +" buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_1], acc[8:9], v[88:89], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[10:11], v[90:91], %[v_acc_1] \n" +" buffer_load_dwordx4 acc[140:143], %[v_os_b0], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_1], acc[12:13], v[92:93], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[14:15], v[94:95], %[v_acc_1] \n" +" buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_2], acc[16:17], v[64:65], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[18:19], v[66:67], %[v_acc_2] \n" +" buffer_load_dwordx4 acc[144:147], %[v_os_b1], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_2], acc[20:21], v[68:69], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[22:23], v[70:71], %[v_acc_2] \n" +" buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_2], acc[24:25], v[72:73], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[26:27], v[74:75], %[v_acc_2] \n" +" buffer_load_dwordx4 acc[148:151], %[v_os_b1], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_2], acc[28:29], v[76:77], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[30:31], v[78:79], %[v_acc_2] \n" +" buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_3], acc[16:17], v[80:81], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[18:19], v[82:83], %[v_acc_3] \n" +" buffer_load_dwordx4 acc[152:155], %[v_os_b1], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_3], acc[20:21], v[84:85], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[22:23], v[86:87], %[v_acc_3] \n" +" buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_3], acc[24:25], v[88:89], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[26:27], v[90:91], %[v_acc_3] \n" +" buffer_load_dwordx4 acc[156:159], %[v_os_b1], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_3], acc[28:29], v[92:93], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[30:31], v[94:95], %[v_acc_3] \n" +" buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[smem_sz], %[s_m0_init] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " %[v_acc_4], acc[32:33], v[64:65], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[34:35], v[66:67], %[v_acc_4] \n" +" buffer_load_dwordx4 acc[160:163], %[v_os_b2], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_4], acc[36:37], v[68:69], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[38:39], v[70:71], %[v_acc_4] \n" +" ds_read_b128 v[96:99], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_0] \n" +_UK_MFMA_ " %[v_acc_4], acc[40:41], v[72:73], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[42:43], v[74:75], %[v_acc_4] \n" +" buffer_load_dwordx4 acc[164:167], %[v_os_b2], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_4], acc[44:45], v[76:77], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[46:47], v[78:79], %[v_acc_4] \n" +" ds_read_b128 v[100:103], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_1] \n" +_UK_MFMA_ " %[v_acc_5], acc[32:33], v[80:81], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[34:35], v[82:83], %[v_acc_5] \n" +" buffer_load_dwordx4 acc[168:171], %[v_os_b2], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_5], acc[36:37], v[84:85], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[38:39], v[86:87], %[v_acc_5] \n" +" ds_read_b128 v[104:107], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_2] \n" +_UK_MFMA_ " %[v_acc_5], acc[40:41], v[88:89], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[42:43], v[90:91], %[v_acc_5] \n" +" buffer_load_dwordx4 acc[172:175], %[v_os_b2], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_5], acc[44:45], v[92:93], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[46:47], v[94:95], %[v_acc_5] \n" +" ds_read_b128 v[108:111], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_3] \n" +_UK_MFMA_ " %[v_acc_6], acc[48:49], v[64:65], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[50:51], v[66:67], %[v_acc_6] \n" +" buffer_load_dwordx4 acc[176:179], %[v_os_b3], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_6], acc[52:53], v[68:69], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[54:55], v[70:71], %[v_acc_6] \n" +" ds_read_b128 v[112:115], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_4] \n" +_UK_MFMA_ " %[v_acc_6], acc[56:57], v[72:73], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[58:59], v[74:75], %[v_acc_6] \n" +" buffer_load_dwordx4 acc[180:183], %[v_os_b3], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_6], acc[60:61], v[76:77], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[62:63], v[78:79], %[v_acc_6] \n" +" ds_read_b128 v[116:119], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_5] \n" +_UK_MFMA_ " %[v_acc_7], acc[48:49], v[80:81], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[50:51], v[82:83], %[v_acc_7] \n" +" buffer_load_dwordx4 acc[184:187], %[v_os_b3], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_7], acc[52:53], v[84:85], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[54:55], v[86:87], %[v_acc_7] \n" +" ds_read_b128 v[120:123], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_6] \n" +_UK_MFMA_ " %[v_acc_7], acc[56:57], v[88:89], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[58:59], v[90:91], %[v_acc_7] \n" +" buffer_load_dwordx4 acc[188:191], %[v_os_b3], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_7], acc[60:61], v[92:93], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[62:63], v[94:95], %[v_acc_7] \n" +" ds_read_b128 v[124:127], %[v_os_slda], offset:1*%[smem_sz] + %[sld_os_7] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " %[v_acc_8], acc[64:65], v[64:65], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[66:67], v[66:67], %[v_acc_8] \n" +" buffer_load_dwordx4 acc[192:195], %[v_os_b4], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_8], acc[68:69], v[68:69], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[70:71], v[70:71], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[72:73], v[72:73], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[74:75], v[74:75], %[v_acc_8] \n" +" buffer_load_dwordx4 acc[196:199], %[v_os_b4], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_8], acc[76:77], v[76:77], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[78:79], v[78:79], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_9], acc[64:65], v[80:81], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[66:67], v[82:83], %[v_acc_9] \n" +" buffer_load_dwordx4 acc[200:203], %[v_os_b4], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_9], acc[68:69], v[84:85], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[70:71], v[86:87], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[72:73], v[88:89], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[74:75], v[90:91], %[v_acc_9] \n" +" buffer_load_dwordx4 acc[204:207], %[v_os_b4], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_9], acc[76:77], v[92:93], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[78:79], v[94:95], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_10], acc[80:81], v[64:65], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[82:83], v[66:67], %[v_acc_10] \n" +" buffer_load_dwordx4 acc[208:211], %[v_os_b5], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_10], acc[84:85], v[68:69], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[86:87], v[70:71], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[88:89], v[72:73], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[90:91], v[74:75], %[v_acc_10] \n" +" buffer_load_dwordx4 acc[212:215], %[v_os_b5], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_10], acc[92:93], v[76:77], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[94:95], v[78:79], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_11], acc[80:81], v[80:81], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[82:83], v[82:83], %[v_acc_11] \n" +" buffer_load_dwordx4 acc[216:219], %[v_os_b5], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_11], acc[84:85], v[84:85], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[86:87], v[86:87], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[88:89], v[88:89], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[90:91], v[90:91], %[v_acc_11] \n" +" buffer_load_dwordx4 acc[220:223], %[v_os_b5], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_11], acc[92:93], v[92:93], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[94:95], v[94:95], %[v_acc_11] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " %[v_acc_12], acc[96:97], v[64:65], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[98:99], v[66:67], %[v_acc_12] \n" +" buffer_load_dwordx4 acc[224:227], %[v_os_b6], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_12], acc[100:101], v[68:69], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[102:103], v[70:71], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[104:105], v[72:73], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[106:107], v[74:75], %[v_acc_12] \n" +" buffer_load_dwordx4 acc[228:231], %[v_os_b6], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_12], acc[108:109], v[76:77], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[110:111], v[78:79], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_13], acc[96:97], v[80:81], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[98:99], v[82:83], %[v_acc_13] \n" +" buffer_load_dwordx4 acc[232:235], %[v_os_b6], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_13], acc[100:101], v[84:85], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[102:103], v[86:87], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[104:105], v[88:89], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[106:107], v[90:91], %[v_acc_13] \n" +" buffer_load_dwordx4 acc[236:239], %[v_os_b6], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_13], acc[108:109], v[92:93], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[110:111], v[94:95], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_14], acc[112:113], v[64:65], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[114:115], v[66:67], %[v_acc_14] \n" +" buffer_load_dwordx4 acc[240:243], %[v_os_b7], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_14], acc[116:117], v[68:69], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[118:119], v[70:71], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[120:121], v[72:73], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[122:123], v[74:75], %[v_acc_14] \n" +" buffer_load_dwordx4 acc[244:247], %[v_os_b7], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_14], acc[124:125], v[76:77], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[126:127], v[78:79], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_15], acc[112:113], v[80:81], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[114:115], v[82:83], %[v_acc_15] \n" +" buffer_load_dwordx4 acc[248:251], %[v_os_b7], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_15], acc[116:117], v[84:85], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[118:119], v[86:87], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[120:121], v[88:89], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[122:123], v[90:91], %[v_acc_15] \n" +" buffer_load_dwordx4 acc[252:255], %[v_os_b7], s[20:23], 0 offen offset:3072\n" +_UK_MFMA_ " %[v_acc_15], acc[124:125], v[92:93], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[126:127], v[94:95], %[v_acc_15] \n" +" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 \n" +" s_cmp_gt_i32 %[s_loop_cnt] 0 \n" +" s_cbranch_scc0 L_end%= \n" +" s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n" +" s_cselect_b32 s86, %[s_tile_os_a], 0 \n" +" s_add_u32 s16, s86, s16 \n" +" s_addc_u32 s17, 0, s17 \n" +" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n" +" s_cselect_b32 s86, %[s_tile_os_b], 0 \n" +" s_add_u32 s20, s86, s20 \n" +" s_addc_u32 s21, 0, s21 \n" +" ;------------------------------------------ \n" +" s_waitcnt vmcnt(24) & lgkmcnt(0) \n" +" s_barrier \n" +_UK_MFMA_ " %[v_acc_0], acc[128:129], v[96:97], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[130:131], v[98:99], %[v_acc_0] \n" +" buffer_load_dwordx4 acc[0:3], %[v_os_b0], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_0], acc[132:133], v[100:101], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[134:135], v[102:103], %[v_acc_0] \n" +" buffer_load_dword %[v_os_a0], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_0], acc[136:137], v[104:105], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[138:139], v[106:107], %[v_acc_0] \n" +" buffer_load_dwordx4 acc[4:7], %[v_os_b0], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_0], acc[140:141], v[108:109], %[v_acc_0] \n" +_UK_MFMA_ " %[v_acc_0], acc[142:143], v[110:111], %[v_acc_0] \n" +" buffer_load_dword %[v_os_a1], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_1], acc[128:129], v[112:113], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[130:131], v[114:115], %[v_acc_1] \n" +" buffer_load_dwordx4 acc[8:11], %[v_os_b0], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_1], acc[132:133], v[116:117], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[134:135], v[118:119], %[v_acc_1] \n" +" buffer_load_dword %[v_os_a2], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_1], acc[136:137], v[120:121], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[138:139], v[122:123], %[v_acc_1] \n" +" buffer_load_dwordx4 acc[12:15], %[v_os_b0], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_1], acc[140:141], v[124:125], %[v_acc_1] \n" +_UK_MFMA_ " %[v_acc_1], acc[142:143], v[126:127], %[v_acc_1] \n" +" buffer_load_dword %[v_os_a3], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_2], acc[144:145], v[96:97], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[146:147], v[98:99], %[v_acc_2] \n" +" buffer_load_dwordx4 acc[16:19], %[v_os_b1], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_2], acc[148:149], v[100:101], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[150:151], v[102:103], %[v_acc_2] \n" +" buffer_load_dword %[v_os_a4], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_2], acc[152:153], v[104:105], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[154:155], v[106:107], %[v_acc_2] \n" +" buffer_load_dwordx4 acc[20:23], %[v_os_b1], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_2], acc[156:157], v[108:109], %[v_acc_2] \n" +_UK_MFMA_ " %[v_acc_2], acc[158:159], v[110:111], %[v_acc_2] \n" +" buffer_load_dword %[v_os_a5], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_3], acc[144:145], v[112:113], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[146:147], v[114:115], %[v_acc_3] \n" +" buffer_load_dwordx4 acc[24:27], %[v_os_b1], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_3], acc[148:149], v[116:117], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[150:151], v[118:119], %[v_acc_3] \n" +" buffer_load_dword %[v_os_a6], s[16:19], 0 offen lds \n" +" s_add_u32 m0, %[s_size_per_issue], m0 \n" +_UK_MFMA_ " %[v_acc_3], acc[152:153], v[120:121], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[154:155], v[122:123], %[v_acc_3] \n" +" buffer_load_dwordx4 acc[28:31], %[v_os_b1], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_3], acc[156:157], v[124:125], %[v_acc_3] \n" +_UK_MFMA_ " %[v_acc_3], acc[158:159], v[126:127], %[v_acc_3] \n" +" buffer_load_dword %[v_os_a7], s[16:19], 0 offen lds \n" +" s_add_u32 m0, 0, %[s_m0_init] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " %[v_acc_4], acc[160:161], v[96:97], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[162:163], v[98:99], %[v_acc_4] \n" +" buffer_load_dwordx4 acc[32:35], %[v_os_b2], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_4], acc[164:165], v[100:101], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[166:167], v[102:103], %[v_acc_4] \n" +" ds_read_b128 v[64:67], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_0] \n" +_UK_MFMA_ " %[v_acc_4], acc[168:169], v[104:105], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[170:171], v[106:107], %[v_acc_4] \n" +" buffer_load_dwordx4 acc[36:39], %[v_os_b2], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_4], acc[172:173], v[108:109], %[v_acc_4] \n" +_UK_MFMA_ " %[v_acc_4], acc[174:175], v[110:111], %[v_acc_4] \n" +" ds_read_b128 v[68:71], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_1] \n" +_UK_MFMA_ " %[v_acc_5], acc[160:161], v[112:113], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[162:163], v[114:115], %[v_acc_5] \n" +" buffer_load_dwordx4 acc[40:43], %[v_os_b2], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_5], acc[164:165], v[116:117], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[166:167], v[118:119], %[v_acc_5] \n" +" ds_read_b128 v[72:75], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_2] \n" +_UK_MFMA_ " %[v_acc_5], acc[168:169], v[120:121], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[170:171], v[122:123], %[v_acc_5] \n" +" buffer_load_dwordx4 acc[44:47], %[v_os_b2], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_5], acc[172:173], v[124:125], %[v_acc_5] \n" +_UK_MFMA_ " %[v_acc_5], acc[174:175], v[126:127], %[v_acc_5] \n" +" ds_read_b128 v[76:79], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_3] \n" +_UK_MFMA_ " %[v_acc_6], acc[176:177], v[96:97], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[178:179], v[98:99], %[v_acc_6] \n" +" buffer_load_dwordx4 acc[48:51], %[v_os_b3], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_6], acc[180:181], v[100:101], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[182:183], v[102:103], %[v_acc_6] \n" +" ds_read_b128 v[80:83], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_4] \n" +_UK_MFMA_ " %[v_acc_6], acc[184:185], v[104:105], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[186:187], v[106:107], %[v_acc_6] \n" +" buffer_load_dwordx4 acc[52:55], %[v_os_b3], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_6], acc[188:189], v[108:109], %[v_acc_6] \n" +_UK_MFMA_ " %[v_acc_6], acc[190:191], v[110:111], %[v_acc_6] \n" +" ds_read_b128 v[84:87], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_5] \n" +_UK_MFMA_ " %[v_acc_7], acc[176:177], v[112:113], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[178:179], v[114:115], %[v_acc_7] \n" +" buffer_load_dwordx4 acc[56:59], %[v_os_b3], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_7], acc[180:181], v[116:117], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[182:183], v[118:119], %[v_acc_7] \n" +" ds_read_b128 v[88:91], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_6] \n" +_UK_MFMA_ " %[v_acc_7], acc[184:185], v[120:121], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[186:187], v[122:123], %[v_acc_7] \n" +" buffer_load_dwordx4 acc[60:63], %[v_os_b3], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_7], acc[188:189], v[124:125], %[v_acc_7] \n" +_UK_MFMA_ " %[v_acc_7], acc[190:191], v[126:127], %[v_acc_7] \n" +" ds_read_b128 v[92:95], %[v_os_slda] offset:0*%[smem_sz] + %[sld_os_7] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " %[v_acc_8], acc[192:193], v[96:97], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[194:195], v[98:99], %[v_acc_8] \n" +" buffer_load_dwordx4 acc[64:67], %[v_os_b4], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_8], acc[196:197], v[100:101], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[198:199], v[102:103], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[200:201], v[104:105], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[202:203], v[106:107], %[v_acc_8] \n" +" buffer_load_dwordx4 acc[68:71], %[v_os_b4], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_8], acc[204:205], v[108:109], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_8], acc[206:207], v[110:111], %[v_acc_8] \n" +_UK_MFMA_ " %[v_acc_9], acc[192:193], v[112:113], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[194:195], v[114:115], %[v_acc_9] \n" +" buffer_load_dwordx4 acc[72:75], %[v_os_b4], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_9], acc[196:197], v[116:117], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[198:199], v[118:119], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[200:201], v[120:121], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[202:203], v[122:123], %[v_acc_9] \n" +" buffer_load_dwordx4 acc[76:79], %[v_os_b4], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_9], acc[204:205], v[124:125], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_9], acc[206:207], v[126:127], %[v_acc_9] \n" +_UK_MFMA_ " %[v_acc_10], acc[208:209], v[96:97], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[210:211], v[98:99], %[v_acc_10] \n" +" buffer_load_dwordx4 acc[80:83], %[v_os_b5], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_10], acc[212:213], v[100:101], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[214:215], v[102:103], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[216:217], v[104:105], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[218:219], v[106:107], %[v_acc_10] \n" +" buffer_load_dwordx4 acc[84:87], %[v_os_b5], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_10], acc[220:221], v[108:109], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_10], acc[222:223], v[110:111], %[v_acc_10] \n" +_UK_MFMA_ " %[v_acc_11], acc[208:209], v[112:113], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[210:211], v[114:115], %[v_acc_11] \n" +" buffer_load_dwordx4 acc[88:91], %[v_os_b5], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_11], acc[212:213], v[116:117], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[214:215], v[118:119], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[216:217], v[120:121], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[218:219], v[122:123], %[v_acc_11] \n" +" buffer_load_dwordx4 acc[92:95], %[v_os_b5], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_11], acc[220:221], v[124:125], %[v_acc_11] \n" +_UK_MFMA_ " %[v_acc_11], acc[222:223], v[126:127], %[v_acc_11] \n" +" s_waitcnt vmcnt(32) \n" +_UK_MFMA_ " %[v_acc_12], acc[224:225], v[96:97], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[226:227], v[98:99], %[v_acc_12] \n" +" buffer_load_dwordx4 acc[96:99], %[v_os_b6], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_12], acc[228:229], v[100:101], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[230:231], v[102:103], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[232:233], v[104:105], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[234:235], v[106:107], %[v_acc_12] \n" +" buffer_load_dwordx4 acc[100:103], %[v_os_b6], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_12], acc[236:237], v[108:109], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_12], acc[238:239], v[110:111], %[v_acc_12] \n" +_UK_MFMA_ " %[v_acc_13], acc[224:225], v[112:113], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[226:227], v[114:115], %[v_acc_13] \n" +" buffer_load_dwordx4 acc[104:107], %[v_os_b6], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_13], acc[228:229], v[116:117], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[230:231], v[118:119], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[232:233], v[120:121], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[234:235], v[122:123], %[v_acc_13] \n" +" buffer_load_dwordx4 acc[108:111], %[v_os_b6], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_13], acc[236:237], v[124:125], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_13], acc[238:239], v[126:127], %[v_acc_13] \n" +_UK_MFMA_ " %[v_acc_14], acc[240:241], v[96:97], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[242:243], v[98:99], %[v_acc_14] \n" +" buffer_load_dwordx4 acc[112:115], %[v_os_b7], s[20:23], 0 offen \n" +_UK_MFMA_ " %[v_acc_14], acc[244:245], v[100:101], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[246:247], v[102:103], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[248:249], v[104:105], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[250:251], v[106:107], %[v_acc_14] \n" +" buffer_load_dwordx4 acc[116:119], %[v_os_b7], s[20:23], 0 offen offset:1024 \n" +_UK_MFMA_ " %[v_acc_14], acc[252:253], v[108:109], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_14], acc[254:255], v[110:111], %[v_acc_14] \n" +_UK_MFMA_ " %[v_acc_15], acc[240:241], v[112:113], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[242:243], v[114:115], %[v_acc_15] \n" +" buffer_load_dwordx4 acc[120:123], %[v_os_b7], s[20:23], 0 offen offset:2048 \n" +_UK_MFMA_ " %[v_acc_15], acc[244:245], v[116:117], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[246:247], v[118:119], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[248:249], v[120:121], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[250:251], v[122:123], %[v_acc_15] \n" +" buffer_load_dwordx4 acc[124:127], %[v_os_b7], s[20:23], 0 offen offset:3072 \n" +_UK_MFMA_ " %[v_acc_15], acc[252:253], v[124:125], %[v_acc_15] \n" +_UK_MFMA_ " %[v_acc_15], acc[254:255], v[126:127], %[v_acc_15] \n" +" s_sub_i32 %[s_loop_cnt], %[s_loop_cnt], 1 \n" +" s_cmp_gt_i32 %[s_loop_cnt] 0 \n" +" s_cbranch_scc0 L_end%= \n" +" s_cmp_gt_i32 %[s_loop_cnt] 2 ; move a with cond \n" +" s_cselect_b32 s86, %[s_tile_os_a], 0 \n" +" s_add_u32 s16, s86, s16 \n" +" s_addc_u32 s17, 0, s17 \n" +" s_cmp_gt_i32 %[s_loop_cnt] 1 ; move b with cond \n" +" s_cselect_b32 s86, %[s_tile_os_b], 0 \n" +" s_add_u32 s20, s86, s20 \n" +" s_addc_u32 s21, 0, s21 \n" +" s_branch L_start%= \n" +"L_end%=: \n" +" s_nop 2 \n" + +#undef _UK_MFMA_ diff --git a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp index 10bb01168f..173887513e 100644 --- a/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp +++ b/include/ck_tile/ops/fmha/pipeline/block_fmha_pipeline_qr_ks_vs_async.hpp @@ -331,7 +331,8 @@ struct BlockFmhaPipelineQRKSVSAsync Policy::template MakeVDramTileDistribution()); // prefetch K tile - async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, k_oob_ck, k_pre_np); + async_load_tile_raw( + k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, number<-1>{}, k_oob_ck, k_pre_np); move_tile_window(k_dram_window, {0, kK0}); __builtin_amdgcn_sched_barrier(0); @@ -355,6 +356,7 @@ struct BlockFmhaPipelineQRKSVSAsync static_for<0, k0_loops - 1, 1>{}([&](auto i_k0) { async_load_tile_raw(k_lds_store(number{})>{}), k_dram_window, + number<-1>{}, k_oob_ck, k_pre_np); if constexpr(i_k0 < k0_loops - 1) @@ -386,7 +388,7 @@ struct BlockFmhaPipelineQRKSVSAsync __builtin_amdgcn_s_barrier(); const auto bias_tile = load_tile(bias_dram_window); // load bias tile - auto v_buf = load_tile(v_dram_window, bool_constant{}); + auto v_buf = load_tile(v_dram_window, number<-1>{}, bool_constant{}); __builtin_amdgcn_sched_barrier(0); { // tail gemm_0(s_acc, @@ -514,7 +516,8 @@ struct BlockFmhaPipelineQRKSVSAsync move_tile_window( v_dram_window, {0, kK1}); // will have scratch if move this right after load_tile(v_dram)... - v_buf = load_tile(v_dram_window, bool_constant{}); // load next v_buf + v_buf = load_tile( + v_dram_window, number<-1>{}, bool_constant{}); // load next v_buf } __builtin_amdgcn_sched_barrier(0); @@ -618,7 +621,8 @@ struct BlockFmhaPipelineQRKSVSAsync static_for<0, k1_loops - 1, 1>{}([&](auto i_k1) { if constexpr(i_k1 != 0 && i_k1 < k1_loops - 1) { - v_buf = load_tile(v_dram_window, bool_constant{}); // load next v_buf + v_buf = load_tile( + v_dram_window, number<-1>{}, bool_constant{}); // load next v_buf } block_sync_lds(); gemm_1(o_acc, @@ -665,8 +669,11 @@ struct BlockFmhaPipelineQRKSVSAsync if constexpr(k1_loops >= 2 && LdsSeq.at(number<0>{}) == LdsSeq.at(number{})) __builtin_amdgcn_s_barrier(); - async_load_tile_raw( - k_lds_store(LdsSeq.at(number<0>{})), k_dram_window, k_oob_ck, k_pre_np); + async_load_tile_raw(k_lds_store(LdsSeq.at(number<0>{})), + k_dram_window, + number<-1>{}, + k_oob_ck, + k_pre_np); move_tile_window(k_dram_window, {0, kK0}); } // tail diff --git a/include/ck_tile/ops/fused_moe.hpp b/include/ck_tile/ops/fused_moe.hpp index b74607f061..d23af0af8d 100644 --- a/include/ck_tile/ops/fused_moe.hpp +++ b/include/ck_tile/ops/fused_moe.hpp @@ -3,7 +3,15 @@ #pragma once +#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp" +#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp" +#include "ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp" #include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp" #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp" #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp" #include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp" diff --git a/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp b/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp new file mode 100644 index 0000000000..2d25d44f3c --- /dev/null +++ b/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_kernel.hpp @@ -0,0 +1,421 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/ops/common.hpp" +#include "ck_tile/ops/elementwise.hpp" +#include +#include + +// clang-format off +// [indexing implementation-1] +// using M_a as constexpr block_size to partition all tokens into different slices +// each slice map to one expert, and one expert can have multiple slices +// e.g. num_experts = 6, topk=3, M_a = 4, input_tokens = 5 +// before sort, topk_ids is : [[0, 3, 5], [2, 3, 5], [1, 3, 5], [1, 2, 3], [1, 3, 5]] +// tok-0 tok-1 tok-2 tok-3 tok-4 +// topk_weight is : [[a, b, c], [d, e, f], [g, h, i], [j, k, l], [m, n, o]] (some float number) +// +// token_id_per_expert is : [[0], [2, 3, 4], [1, 3], [0, 1, 2, 3, 4], [], [0, 1, 2, 5]] +// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5 +// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]] +// +// max_num_tokens_padded : topk * input_tokens + num_experts * (M_a - 1) +// * this could be larger than actual, since actual tokens are on GPU +// +// sorted_token_ids_ptr : [0, 6, 6, 6, 2, 3, 4, 6, 1, 3, 6, 6, 0, 1, 2, 3, 4, 6, 6, 6, 6, 6, 6, 6, 0, 1, 2, 5] +// |- exp-0 -|- exp-1 -|- exp-2 -|- exp-3 -|- exp-4 -|- exp-5 -| +// sorted_weight_ptr : [a, *, *, *, g, j, m, *, d, k, *, *, b, e, h, l, n, *, *, *, *, *, *, *, c, f, i, o] +// +// * length is max_num_tokens_padded, actual size is num_tokens_post_padded_ptr +// +// * Note on token_id_per_expert/sorted_token_ids_ptr data: +// currently we do not have topk information from the data of token_id_per_expert/sorted_token_ids_ptr. +// In some cases(like smooth-quant), we need topk information to indexing into tokens quant from +// different expert smooth quant. So we modify the number stored inside token_id_per_expert/sorted_token_ids_ptr +// +// 32bit 0........23 24.....31 bit +// (data) -> (token_id | topk_id) +// low 24 bit is for token id, top 8 bit is for topk id +// +// the input after smooth-quant is [token, topk, hidden_dim], originally it is [token, hidden_dim] +// the input scale for token is [topk, token, 1], the smooth-quant scale for first gemm is [expert, interm_dim] +// +// sorted_expert_ids_ptr : [0, 1, 2, 3, 3, 4, 5] +// * length is (max_num_tokens_padded + block_size - 1) / block_size +// +// num_tokens_post_padded_ptr : [28] +// num_sorted_tiles_ptr : [7] +// +// * different from vLLM +// 1) token_id stored in sorted_token_ids_ptr is actual token_id, not token_id*top_K expanded id +// 2）need sorted_weight_ptr +// 3) use num_sorted_tiles_ptr, already divided by M_a +// +// * below used for indexing +// 1) sorted_token_ids_ptr [max_num_tokens_padded] +// 2) sorted_weight_ptr +// 3) sorted_expert_ids_ptr +// 4）num_tokens_post_padded_ptr/num_sorted_tiles_ptr (select one) +// +// max_num_tokens_padded: opk_ids.numel() + num_experts * (block_size - 1) +// +// [indexing implementation-2] +// before sort, topk_ids is : [[0, 3, 5], [2, 3, 5], [1, 3, 5], [1, 2, 3], [1, 3, 5]] +// tok-0 tok-1 tok-2 tok-3 tok-4 +// topk_weight is : [[a, b, c], [d, e, f], [g, h, i], [j, k, l], [m, n, o]] (some float number) +// +// we generate original rol/col id as +// topk_rc_ids : [[0, 5, A], [1, 6, B], [2, 7, C], [3, 8, D], [4, 9, E]] +// let x be one element of above, we can get: +// tpok_row_id(token_id) = x % num_tokens(5) +// tpok_col_id(expert_Id) = x / num_tokens +// topk_row_id/col_id can be used to access original topk_ids/topk_weight +// +// token_id_per_expert is : [[0], [2, 3, 4], [1, 3], [0, 1, 2, 3, 4], [], [0, 1, 5, 5]] +// (only for reference) exp-0 exp-1 exp-2 exp-3 exp-4 exp-5 +// weight_id_per_expert is: [[a], [g, j, m], [d, k], [b, e, h, l, n], [], [c, f, i, o]] +// +// we can get permuted_rc_ids: +// [[0], [2, 3, 4], [1, 8], [5, 6, 7, D, 9], [], [A, B, C, E]] +// +// +// clang-format on +// +namespace ck_tile { + +// m: num_tokens (or token*input-batch) +// k: intermediate_size +// n: intermediate_size used between 2 FC (TP slice this) +// e: num expert +// if doing pre-shuffle +// nr : n / Block_Nr +// kr : k / Block_Kr +// w : fattened 1d wave buffer +struct FusedMoeGemmHostArgs +{ + const void* a_ptr; // [m, k], input token + const void* a_scale_ptr; // [m, 1], token scale + const void* g_ptr; // [e, n, k]/[e, 2*n, k], pre-shuffle([e, nr, kr, w]) + const void* d_ptr; // [e, n, k], pre-shuffle([e, nr, kr, w]) + const void* g_scale_ptr; // [e, 1, n], gate(up) scale + const void* d_scale_ptr; // [e, 1, k], down scale + const void* y_smooth_scale_ptr; // [e, 1, n], smooth-quant-scale for 2nd gemm input + void* o_ptr; // [m, k], output token + + const void* sorted_token_ids_ptr; // [max_num_tokens_padded] + const void* sorted_weight_ptr; // [max_num_tokens_padded] + const void* sorted_expert_ids_ptr; // [(max_num_tokens_padded + block_size - 1) / block_size] + const void* num_sorted_tiles_ptr; // [1] + + index_t hidden_size; // k + index_t intermediate_size; // n / TP, for Gate. if Gate+Up, Down need divide by 2 + index_t num_tokens; // input number of tokens for current iteration + index_t num_experts; // number of groups + index_t topk; // need this? + + index_t stride_token; // for input/output, stride for each row, should >= hidden_size +}; + +// This is scatter/gather b2b group-gemm +template +struct FusedMoeGemmKernel +{ + using Partitioner = remove_cvref_t; + using Pipeline = remove_cvref_t; + using Epilogue = remove_cvref_t; // TODO: not used + // static constexpr index_t kBlockPerCu = Pipeline::kBlockPerCu; + // static_assert(kBlockPerCu > 0); + + using BlockShape = typename Pipeline::BlockShape; // this is FusedMoeGemmShape + static constexpr index_t BlockSize_ = BlockShape::BlockSize; + + using ADataType = typename Pipeline::Problem::ADataType; + using GDataType = typename Pipeline::Problem::GDataType; + using DDataType = typename Pipeline::Problem::DDataType; + using AccDataType = typename Pipeline::Problem::AccDataType; + using ODataType = typename Pipeline::Problem::ODataType; + using AScaleDataType = typename Pipeline::Problem::AScaleDataType; + using GScaleDataType = typename Pipeline::Problem::GScaleDataType; + using DScaleDataType = typename Pipeline::Problem::DScaleDataType; + using YSmoothScaleDataType = typename Pipeline::Problem::YSmoothScaleDataType; + using TopkWeightDataType = typename Pipeline::Problem::TopkWeightDataType; + using IndexDataType = typename Pipeline::Problem::IndexDataType; + using YDataType = typename Pipeline::Problem::YDataType; + + using Traits = typename Pipeline::Problem::Traits; + static constexpr bool UseUK = true; + + static constexpr bool IsGateOnly = Traits::IsGateOnly; + static constexpr bool UseSmoothQuant = Traits::UseSmoothQuant; + static constexpr bool PadHiddenSize = Traits::PadHiddenSize; + static constexpr bool PadIntermediateSize = Traits::PadIntermediateSize; + + // clang-format off + template struct t2s; + template <> struct t2s { static constexpr const char * name = "fp32"; }; + template <> struct t2s { static constexpr const char * name = "fp16"; }; + template <> struct t2s { static constexpr const char * name = "bf16"; }; + template <> struct t2s { static constexpr const char * name = "fp8"; }; + template <> struct t2s { static constexpr const char * name = "bf8"; }; + template <> struct t2s { static constexpr const char * name = "int8"; }; + // clang-format on + + CK_TILE_HOST static std::string GetName() + { +#define _SS_ std::string +#define _TS_ std::to_string + // clang-format off + using S_ = BlockShape; + + auto prec_str = [&] () { + std::string base_str = _SS_(t2s::name); + if (!std::is_same_v) { + base_str += _SS_("_") + _SS_(t2s::name); + } + return base_str; + }(); + + return _SS_("fused_moe_") + _SS_(prec_str) + "_" + + _TS_(S_::Block_M0) + "x" + _TS_(S_::Block_N0) + "x" + _TS_(S_::Block_K0) + "x" + _TS_(S_::Block_N1) + "_" + + _TS_(S_::WarpPerBlock_M0) + "x" + _TS_(S_::WarpPerBlock_N0) + "x" + _TS_(S_::WarpPerBlock_K0) + "_" + + _TS_(S_::Warp_M0) + "x" + _TS_(S_::Warp_N0) + "x" + _TS_(S_::Warp_K0) + "_" + _SS_(Pipeline::name); +#undef _SS_ +#undef _TS_ + // clang-format on + } + + struct FusedMoeGemmKargs + { + const void* a_ptr; // [m, k], input token + const void* a_scale_ptr; // [m, 1], token scale + const void* g_ptr; // [e, n, k]/[e, 2*n, k], pre-shuffle([e, nr, kr, w]) + const void* d_ptr; // [e, n, k], pre-shuffle([e, nr, kr, w]) + const void* g_scale_ptr; // [e, 1, n], gate(up) scale + const void* d_scale_ptr; // [e, 1, k], down scale + const void* y_smooth_scale_ptr; // [e, 1, n], smooth-quant-scale for 2nd gemm input + void* o_ptr; // [m, k], output token + + const void* sorted_token_ids_ptr; + const void* sorted_weight_ptr; + const void* sorted_expert_ids_ptr; + const void* num_sorted_tiles_ptr; + + index_t hidden_size; // k + index_t intermediate_size; // n / TP, for Gate. if Gate+Up, Down need divide by 2 + index_t num_tokens; // input number of tokens for current iteration + index_t num_experts; // number of groups + index_t topk; // need this? + + index_t stride_token; // for input/output, stride for each row, should >= hidden_size + }; + + // TODO: switch karg based on + using Kargs = FusedMoeGemmKargs; + using Hargs = FusedMoeGemmHostArgs; + + CK_TILE_HOST static constexpr Kargs MakeKargs(const Hargs& hargs) + { + // TODO: hargs/kargs not guranteed to be the same + return bit_cast(hargs); + } + + CK_TILE_HOST static constexpr auto GridSize(const Hargs& hargs) + { + constexpr index_t block_m = BlockShape::Block_M0; + int max_num_tokens_padded = + hargs.topk * hargs.num_tokens + hargs.num_experts * block_m - hargs.topk; + // printf("xxx max_num_tokens_padded:%d\n", max_num_tokens_padded); + return Partitioner::GridSize(max_num_tokens_padded, hargs.intermediate_size); + } + + CK_TILE_HOST static constexpr auto BlockSize() { return dim3(BlockSize_); } + + CK_TILE_HOST_DEVICE static constexpr index_t GetSmemSize() { return Pipeline::GetSmemSize(); } + + CK_TILE_DEVICE void operator()(Kargs kargs) const + { + if constexpr(UseUK) + { + __shared__ CK_TILE_LDS_ADDR ADataType smem[GetSmemSize()]; + IndexDataType num_sorted_tiles = __builtin_amdgcn_readfirstlane( + *reinterpret_cast(kargs.num_sorted_tiles_ptr)); + + num_sorted_tiles = num_sorted_tiles / BlockShape::Block_M0; + + const auto [sorted_tile_id, intermediate_tile_id] = + Partitioner{}(num_sorted_tiles, kargs.intermediate_size); + // if(threadIdx.x == 0) + // printf("bid:%d,%d, num_sorted_tiles:%d, sorted_tile_id:%d(%d), + // intermediate_tile_id:%d\n", static_cast(blockIdx.x), + // static_cast(blockIdx.y), num_sorted_tiles, sorted_tile_id, sorted_tile_id >= + // num_sorted_tiles? 1 : 0, intermediate_tile_id); + if(sorted_tile_id >= num_sorted_tiles) + return; + + Pipeline{}(kargs, smem, sorted_tile_id, intermediate_tile_id); + } + else + { + // allocate LDS + // __shared__ char smem_ptr[GetSmemSize()]; + IndexDataType num_sorted_tiles = __builtin_amdgcn_readfirstlane( + *reinterpret_cast(kargs.num_sorted_tiles_ptr)); + constexpr index_t hidden_radio_0 = IsGateOnly ? 1 : 2; + + index_t nr_0 = kargs.intermediate_size / BlockShape::Block_Nr0; + index_t kr_0 = kargs.hidden_size / BlockShape::Block_Kr0; + index_t nr_1 = kargs.hidden_size / BlockShape::Block_Nr1; // should be same as kr_0 + index_t kr_1 = + kargs.intermediate_size / BlockShape::Block_Kr1; // should be same as nr_0 + + index_t expert_stride_0 = kargs.intermediate_size * hidden_radio_0 * kargs.hidden_size; + index_t expert_stride_1 = kargs.intermediate_size * kargs.hidden_size; + + __shared__ CK_TILE_LDS_ADDR ADataType smem[GetSmemSize()]; + + // note this is in unit of tile, need multiple tile size to get the index + const auto [sorted_tile_id, intermediate_tile_id] = + Partitioner{}(num_sorted_tiles, kargs.intermediate_size); + if(sorted_tile_id >= num_sorted_tiles) + return; + + const IndexDataType expert_id = + __builtin_amdgcn_readfirstlane(reinterpret_cast( + kargs.sorted_expert_ids_ptr)[sorted_tile_id]); + + // index along intermediate_size + // index_t hidden_idx = __builtin_amdgcn_readfirstlane(intermediate_tile_id * + // BlockShape::Block_N0); + index_t interm_idx_nr = + __builtin_amdgcn_readfirstlane(intermediate_tile_id * BlockShape::Block_Nr0); + + const auto a_coord = Pipeline::GetACoord(); // 2d thread offset, [i_row, i_col] + const auto sorted_token_id = + a_coord[number<0>{}] + sorted_tile_id * BlockShape::Block_M0; + + index_t token_id = + reinterpret_cast(kargs.sorted_token_ids_ptr)[sorted_token_id]; + auto topk_weight = reinterpret_cast( + kargs.sorted_weight_ptr)[sorted_token_id]; + + const auto a_window = [&]() { + // A is already pre-padded in previous kernel + const ADataType* a_ptr = reinterpret_cast(kargs.a_ptr); + const auto a_view_ = make_naive_tensor_view( + a_ptr, + make_tuple(kargs.num_tokens, kargs.hidden_size), + make_tuple(kargs.stride_token, 1), + number{}, + number<1>{}); + + // gather is here use indexing transform + const auto a_gather_view_ = transform_tensor_view( + a_view_, + make_tuple(make_indexing_transform(kargs.num_tokens, token_id), + make_pass_through_transform(kargs.hidden_size)), + make_tuple(sequence<0>{}, sequence<1>{}), + make_tuple(sequence<0>{}, sequence<1>{})); + + const auto a_window_ = make_tile_window( + a_gather_view_, + make_tuple(number{}, number{}), + {0, 0}); + return a_window_; + }(); + + // TODO: gtile using NSub to have less register pressure + const auto g_window = [&]() { + const GDataType* g_ptr = reinterpret_cast(kargs.g_ptr) + + static_cast(expert_id) * expert_stride_0 + + interm_idx_nr * kr_0 * BlockShape::Block_W0; + const auto g_view_ = make_naive_tensor_view( + g_ptr, + make_tuple(nr_0, kr_0, number{}), + make_tuple(kr_0 * BlockShape::Block_W0, number{}, 1), + number{}, + number<1>{}); + const auto g_view_1_ = + pad_tensor_view(g_view_, + make_tuple(number{}, + number{}, + number{}), + sequence{}); + + const auto g_window_ = make_tile_window(g_view_1_, + make_tuple(number{}, + number{}, + number{}), + {0, 0, 0}); + return g_window_; + }(); + + const auto d_window = [&]() { + const DDataType* d_ptr = reinterpret_cast(kargs.d_ptr) + + static_cast(expert_id) * expert_stride_1 + + interm_idx_nr * BlockShape::Block_W1; + // note interm_idx_nr is along the gemm-k dim of 2nd gemm + + const auto d_view_ = make_naive_tensor_view( + d_ptr, + make_tuple(nr_1, kr_1, BlockShape::Block_W1), + make_tuple(kr_1 * BlockShape::Block_W1, BlockShape::Block_W1, 1), + number{}, + number<1>{}); + const auto d_view_1_ = + pad_tensor_view(d_view_, + make_tuple(number{}, + number{}, + number{}), + sequence{}); + + const auto d_window_ = make_tile_window(d_view_1_, + make_tuple(number{}, + number{}, + number{}), + {0, 0, 0}); + return d_window_; + }(); + + auto o_window = [&]() { + ODataType* o_ptr = reinterpret_cast(kargs.o_ptr); + auto o_view_ = make_naive_tensor_view( + o_ptr, + make_tuple(kargs.num_tokens, kargs.hidden_size), + make_tuple(kargs.stride_token, 1), + number{}, + number<1>{}); + + // gather is here + auto o_scatter_view_ = transform_tensor_view( + o_view_, + make_tuple(make_indexing_transform(kargs.num_tokens, token_id), + make_pass_through_transform(kargs.hidden_size)), + make_tuple(sequence<0>{}, sequence<1>{}), + make_tuple(sequence<0>{}, sequence<1>{})); + + auto o_window_ = make_tile_window( + o_scatter_view_, + make_tuple(number{}, number{}), + {0, 0}); + return o_window_; + }(); + + // do compute yeah + Pipeline{}(a_window, + g_window, + d_window, + o_window, + topk_weight, + smem, + kargs.hidden_size, + kargs.intermediate_size, + kargs.stride_token); + } + } +}; + +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp b/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp new file mode 100644 index 0000000000..4f3f8bb7d3 --- /dev/null +++ b/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_shape.hpp @@ -0,0 +1,125 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" + +namespace ck_tile { + +/* +tensors: +1. act (A): input feature map +2. gate (G): B matrix for first gemm, output will do activation(Silu) +3. up (U): B matrix for first gemm +4. down (D): B matrix for second gemm + N1 + / \ + +----------+ | + | Down | | + x----------x | + hidden hidden K1 | | | + N0 N0 x----------x | + | +------x-----x------+------x-----x------+ | | | + dim | | Gate | | | Up | | | | | | + contiguous | | | | | | | | | | | + | | | | | | | | | | | + v +------x-----x------+------x-----x------+ +----------+ V + K0 | | | | | contiguous + / \ v v v v | + +---------+ +------x-----x------+------x-----x------+ | +M0 | A | | | | | | | | | + +---------+ +------x-----x------+------x-----x------+ | + ----------> | | | + contiguous | V V + | x-----x +----------+ + +------------> M1 | Y | ---------> | Out(O) | + ACT x-----x +----------+ + K1 = N0 dim + +* Note: Act could be Gelu/Silu/... +* Note: some model does not have Up +*/ +template +struct FusedMoeGemmShape +{ + using BlockTile_0 = remove_cvref_t; + using WarpPerBlock_0 = remove_cvref_t; + using WarpTile_0 = remove_cvref_t; + using BlockTile_1 = remove_cvref_t; + using WarpPerBlock_1 = remove_cvref_t; + using WarpTile_1 = remove_cvref_t; + + static constexpr index_t NumWarps = + reduce_on_sequence(WarpPerBlock_0{}, multiplies{}, number<1>{}); + + // TODO: we don't support half warps aound to 1 warp here + static_assert(NumWarps == reduce_on_sequence(WarpPerBlock_1{}, multiplies{}, number<1>{})); + + static constexpr index_t Block_M0 = BlockTile_0::at(number<0>{}); + static constexpr index_t Block_N0 = BlockTile_0::at(number<1>{}); + static constexpr index_t Block_K0 = BlockTile_0::at(number<2>{}); + static constexpr index_t WarpPerBlock_M0 = WarpPerBlock_0::at(number<0>{}); + static constexpr index_t WarpPerBlock_N0 = WarpPerBlock_0::at(number<1>{}); + static constexpr index_t WarpPerBlock_K0 = WarpPerBlock_0::at(number<2>{}); + static constexpr index_t Warp_M0 = WarpTile_0::at(number<0>{}); + static constexpr index_t Warp_N0 = WarpTile_0::at(number<1>{}); + static constexpr index_t Warp_K0 = WarpTile_0::at(number<2>{}); + + static constexpr index_t ThreadPerBlock_M0 = Warp_M0 * WarpPerBlock_M0; + static constexpr index_t ThreadPerBlock_N0 = Warp_N0 * WarpPerBlock_N0; + static constexpr index_t ThreadPerBlock_K0 = Warp_K0 * WarpPerBlock_K0; + static_assert(Block_M0 % ThreadPerBlock_M0 == 0); + static_assert(Block_N0 % ThreadPerBlock_N0 == 0); + static_assert(Block_K0 % ThreadPerBlock_K0 == 0); + static constexpr index_t Repeat_M0 = Block_M0 / ThreadPerBlock_M0; + static constexpr index_t Repeat_N0 = Block_N0 / ThreadPerBlock_N0; + static constexpr index_t Repeat_K0 = Block_K0 / ThreadPerBlock_K0; + + static constexpr index_t Block_M1 = BlockTile_1::at(number<0>{}); + static constexpr index_t Block_N1 = BlockTile_1::at(number<1>{}); + static constexpr index_t Block_K1 = BlockTile_1::at(number<2>{}); + static constexpr index_t WarpPerBlock_M1 = WarpPerBlock_1::at(number<0>{}); + static constexpr index_t WarpPerBlock_N1 = WarpPerBlock_1::at(number<1>{}); + static constexpr index_t WarpPerBlock_K1 = WarpPerBlock_1::at(number<2>{}); + static constexpr index_t Warp_M1 = WarpTile_1::at(number<0>{}); + static constexpr index_t Warp_N1 = WarpTile_1::at(number<1>{}); + static constexpr index_t Warp_K1 = WarpTile_1::at(number<2>{}); + + static constexpr index_t ThreadPerBlock_M1 = Warp_M1 * WarpPerBlock_M1; + static constexpr index_t ThreadPerBlock_N1 = Warp_N1 * WarpPerBlock_N1; + static constexpr index_t ThreadPerBlock_K1 = Warp_K1 * WarpPerBlock_K1; + static_assert(Block_M1 % ThreadPerBlock_M1 == 0); + static_assert(Block_N1 % ThreadPerBlock_N1 == 0); + static_assert(Block_K1 % ThreadPerBlock_K1 == 0); + static constexpr index_t Repeat_M1 = Block_M1 / ThreadPerBlock_M1; + static constexpr index_t Repeat_N1 = Block_N1 / ThreadPerBlock_N1; + static constexpr index_t Repeat_K1 = Block_K1 / ThreadPerBlock_K1; + + static constexpr index_t BlockSize = warpSize * NumWarps; + + // some assert + static_assert(Block_M0 == Block_M1); + static_assert(Block_N0 == Block_K1 || (Block_N0 / 2) == Block_K1); // Gate Only or Gate+Up + + // pre-shuffle tile size compute (assume only for B matrix) + // we flatten the each wave tile to a 1d linear tensor(at model loading time) + // e.g. originally we have Block_N*Block_K tile size, after pre-shuffle + // we can have Block_Nr*Block_Kr*Block_W, where Block_W is Warp_N*Warp_K, + // and Block_Nr=Block_N/Warp_N, Block_Kr=Block_K/Warp_K + static constexpr index_t Block_W0 = Warp_N0 * Warp_K0; + static constexpr index_t Block_Nr0 = Block_N0 / Warp_N0; + static constexpr index_t Block_Kr0 = Block_K0 / Warp_K0; + static constexpr index_t Block_W1 = Warp_N1 * Warp_K1; + static constexpr index_t Block_Nr1 = Block_N1 / Warp_N1; + static constexpr index_t Block_Kr1 = Block_K1 / Warp_K1; + + static_assert(Block_W0 == Block_W1); + // static_assert(Block_Nr0 == Block_Kr1); +}; +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp b/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp new file mode 100644 index 0000000000..381edb650d --- /dev/null +++ b/include/ck_tile/ops/fused_moe/kernel/fused_moegemm_tile_partitioner.hpp @@ -0,0 +1,33 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +namespace ck_tile { + +template +struct FusedMoeGemmTilePartitioner_Linear +{ + // FusedMoeGemmShape + using BlockShape = ck_tile::remove_cvref_t; + + static constexpr const char* name = "lin"; + + CK_TILE_DEVICE auto operator()(ck_tile::index_t /*num_sorted_tiles*/, + ck_tile::index_t /*intermediate_size*/) + { + index_t i_n = blockIdx.x; + index_t i_m = blockIdx.y; + + return ck_tile::make_tuple(i_m, i_n); + } + + CK_TILE_HOST static constexpr auto GridSize(index_t max_tokens, index_t intermediate_size) + { + // TODO: this may need tuning + index_t ms = ck_tile::integer_divide_ceil(max_tokens, BlockShape::Block_M0); + index_t ns = ck_tile::integer_divide_ceil(intermediate_size, BlockShape::Block_N0); + return dim3(ns, ms, 1); + } +}; +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp new file mode 100644 index 0000000000..e9577e2304 --- /dev/null +++ b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_ex.hpp @@ -0,0 +1,651 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/ops/common/tensor_layout.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp" + +namespace ck_tile { + +/* +This pipeline deal with a gemm(actually 2 gemm) with one very small(token), one very big(weight) +we need to design the pipeline such that all waves along gemm-N dim (gemm-m only 1 wave) + + <----- gemm-N ------> + +----+----+----+----+ + | w0 | w1 | w2 | w3 | gemm-m + +----+----+----+----+ +*/ +template +struct FusedMoeGemmPipeline_FlatmmEx +{ + using Problem = remove_cvref_t; + using Policy = remove_cvref_t; + + using BlockShape = typename Problem::BlockShape; // this is FusedMoeGemmShape + + using ADataType = typename Problem::ADataType; + using GDataType = typename Problem::GDataType; + using DDataType = typename Problem::DDataType; + using AccDataType = typename Problem::AccDataType; + using ODataType = typename Problem::ODataType; + using AScaleDataType = typename Problem::AScaleDataType; + using GScaleDataType = typename Problem::GScaleDataType; + using DScaleDataType = typename Problem::DScaleDataType; + using YSmoothScaleDataType = typename Problem::YSmoothScaleDataType; + using TopkWeightDataType = typename Problem::TopkWeightDataType; + using IndexDataType = typename Problem::IndexDataType; + using YDataType = typename Problem::YDataType; + + using Traits = typename Problem::Traits; + + static constexpr bool IsGateOnly = Traits::IsGateOnly; + static constexpr bool UseSmoothQuant = Traits::UseSmoothQuant; + static constexpr bool PadHiddenSize = Traits::PadHiddenSize; + static constexpr bool PadIntermediateSize = Traits::PadIntermediateSize; + + static constexpr index_t kAlignmentA = Policy::template GetAlignment_A(); + static constexpr index_t kAlignmentG = Policy::template GetAlignment_G(); + static constexpr index_t kAlignmentD = Policy::template GetAlignment_D(); + static constexpr index_t kAlignmentO = Policy::template GetAlignment_O(); + + static constexpr index_t SLD_A = static_cast(FusedMoeGemmPipelineSequencerEnum::SLD_A); + static constexpr index_t GLD_A = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_A); + static constexpr index_t GLD_B = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_B); + static constexpr index_t GST_O = static_cast(FusedMoeGemmPipelineSequencerEnum::GST_O); + + static constexpr index_t kBlockPerCu = []() { + if constexpr(Problem::kBlockPerCu != -1) + return Problem::kBlockPerCu; + else + { + // minimize occupancy + return 2; + } + }(); + + static constexpr const char* name = "fused_moe_flatmm"; + + // TODO: there are multiple buffers + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize_A() + { + return Policy::template GetSmemSize_A(); + } + + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize() + { + return Policy::template GetSmemSize(); + } + + // this is the thread-offset along row/col + CK_TILE_HOST_DEVICE static auto GetACoord() + { + constexpr auto a_dist = Policy::template MakeGlobalTileDistribution_A(); + const auto a_coord = a_dist.calculate_index(); + return a_coord; + } + + // this is the thread-offset along row/col + CK_TILE_HOST_DEVICE static auto GetOCoord() + { + constexpr auto o_dist = Policy::template MakeOGlobalTileDistribution(); + const auto o_coord = o_dist.calculate_index(); + return o_coord; + } + + template + CK_TILE_DEVICE auto operator()(const AWindow& a_window_, + const GWindow& g_window_, + const DWindow& d_window_, + OWindow& o_window_, + TopkWeightDataType /*topk_weight*/, + CK_TILE_LDS_ADDR void* smem, + index_t hidden_size, + index_t intermediate_size) + { + _Pragma("clang diagnostic push") _Pragma("clang diagnostic ignored \"-Wc++20-extensions\""); + constexpr auto NEG1 = number<-1>{}; + constexpr auto I0 = number<0>{}; + constexpr auto I1 = number<1>{}; + constexpr auto TRUE = bool_constant{}; + constexpr auto FALSE = bool_constant{}; + + CK_TILE_LDS_ADDR ADataType* smem_0 = reinterpret_cast(smem); + CK_TILE_LDS_ADDR ADataType* smem_1 = reinterpret_cast( + reinterpret_cast(smem) + + Policy::template GetSmemSize_A()); + + auto g_view = g_window_.get_bottom_tensor_view(); + + auto u_view = [&]() { + if constexpr(IsGateOnly) + { + return g_view; + } + else + { + index_t nr_0 = intermediate_size / BlockShape::Block_Nr0; + index_t kr_0 = hidden_size / BlockShape::Block_Kr0; + + const GDataType* g_ptr = + g_window_.get_bottom_tensor_view().get_buffer_view().p_data_; + const GDataType* u_ptr = g_ptr + (nr_0 / 2) * kr_0 * number{}; + + const auto u_view_ = make_naive_tensor_view( + u_ptr, + make_tuple(nr_0, kr_0, number{}), + make_tuple(kr_0 * BlockShape::Block_W0, number{}, 1), + number{}, + number<1>{}); + const auto u_view_1_ = + pad_tensor_view(u_view_, + make_tuple(number{}, + number{}, + number{}), + sequence{}); + return u_view_1_; + } + }(); + + auto a_win = make_tile_window_linear( + a_window_, Policy::template MakeGlobalTileDistribution_A()); + auto g_win = + make_tile_window_linear(g_window_, + Policy::template MakeGlobalTileDistribution_G(), + sequence<0, 1, 1>{}); + auto d_win = + make_tile_window_linear(d_window_, + Policy::template MakeGlobalTileDistribution_D(), + sequence<0, 1, 1>{}); + auto o_win = make_tile_window_linear( + o_window_, Policy::template MakeGlobalTileDistribution_O()); + + using g_thread_type = decltype(load_tile(g_win)); + using d_thread_type = decltype(load_tile(d_win)); + + using WarpGemm0 = decltype(Policy::template GetWarpGemm0()); + using WarpGemm1 = decltype(Policy::template GetWarpGemm1()); + auto warp_gemm_0 = WarpGemm0{}; + auto warp_gemm_1 = WarpGemm1{}; + + // issues_warps_lanes + auto a_sst_win0 = + make_tile_window(make_tensor_view( + smem_0, Policy::template MakeLdsStoreDesc_A()), + Policy::template MakeLdsStoreDesc_A().get_lengths(), + {0, 0, 0}); + + auto a_sst_win1 = + make_tile_window(make_tensor_view( + smem_1, Policy::template MakeLdsStoreDesc_A()), + Policy::template MakeLdsStoreDesc_A().get_lengths(), + {0, 0, 0}); + // m*k + auto a_sld_win0 = [&]() { + using WG = WarpGemm0; + constexpr auto a_outer_dstr_enc = tile_distribution_encoding< + sequence<>, + tuple, + sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + a_outer_dstr_enc, typename WG::AWarpDstrEncoding{}); + return make_tile_window_linear( + make_tensor_view( + smem_0, Policy::template MakeLdsLoadDesc_A()), + Policy::template MakeLdsLoadDesc_A().get_lengths(), + {0, 0}, + make_static_tile_distribution(a_block_dstr_encode)); + }(); + + // m*k + auto a_sld_win1 = [&]() { + using WG = WarpGemm0; + constexpr auto a_outer_dstr_enc = tile_distribution_encoding< + sequence<>, + tuple, + sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + constexpr auto a_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + a_outer_dstr_enc, typename WG::AWarpDstrEncoding{}); + return make_tile_window_linear( + make_tensor_view( + smem_1, Policy::template MakeLdsLoadDesc_A()), + Policy::template MakeLdsLoadDesc_A().get_lengths(), + {0, 0}, + make_static_tile_distribution(a_block_dstr_encode)); + }(); + + auto bridge_sst_win = [&]() { + return make_tile_window( + make_tensor_view( + reinterpret_cast(smem), + Policy::template MakeBridgeLdsStoreDesc()), + Policy::template MakeBridgeLdsStoreDesc().get_lengths(), + {0, 0}); + }(); + + auto bridge_sld_win = [&]() { + return make_tile_window_linear( + make_tensor_view( + reinterpret_cast(smem), + Policy::template MakeBridgeLdsLoadDesc()), + Policy::template MakeBridgeLdsLoadDesc().get_lengths(), + {0, 0}, + Policy::template MakeYTileDistribution()); + }(); + + // also OK with C array, 2 register buffer + statically_indexed_array gs; + + constexpr auto issues_a = number{}; + constexpr auto issues_g = number{}; + // constexpr auto issues_d = number{}; + // constexpr auto issues_o = number{}; + constexpr auto issues_gemm0 = + number{}; + constexpr auto issues_gemm1 = + number{}; + // constexpr auto issues_sld_a = number{}; + + const index_t num_blocks_k0 = + (hidden_size + BlockShape::Block_K0 - 1) / BlockShape::Block_K0; + const index_t num_blocks_n1 = + (hidden_size + BlockShape::Block_N1 - 1) / BlockShape::Block_N1; + + using a_thread_type = decltype(load_tile(a_sld_win0)); + statically_indexed_array as; + + auto gld_a = [&]>( + auto& a_store_, auto i_access, PreNop = {}) + { + async_load_tile_raw(a_store_, a_win, i_access, PreNop{}); + }; + auto move_a = [&]() { + move_tile_window(a_win, {number<0>{}, number{}}); + }; + auto sld_a = [&](auto& a_, auto& win_, auto i_access) { + load_tile_raw(a_, win_, i_access); + }; + + auto gld_g = [&]>( + auto& g_, auto i_access, PreNop = {}) + { + if constexpr(IsGateOnly) + { + // TODO: hack! + if constexpr(i_access.value == 0) + { + g_win.bottom_tensor_view_ = g_view; + } + else if constexpr(i_access.value == issues_g / 2) + { + g_win.bottom_tensor_view_ = u_view; + } + } + load_tile_raw(g_, g_win, i_access, FALSE, PreNop{}); + }; + auto move_g = [&]() { + move_tile_window(g_win, {number<0>{}, number{}, number<0>{}}); + }; + statically_indexed_array ds; + + auto gld_d = [&]>( + auto& d_, auto i_access, PreNop = {}) + { + load_tile_raw(d_, d_win, i_access, FALSE, PreNop{}); + }; + auto move_d = [&]() { + // d move along gemm-n + move_tile_window(d_win, {number{}, number<0>{}}); + }; + + auto atomic_add_o = [&]>( + auto& o_, auto i_access, PreNop = {}) + { + update_tile_raw(o_win, o_, i_access, TRUE, PreNop{}); + }; + + auto acc_0 = Policy::template MakeCBlockTile_Gemm0(); + auto acc_1s = generate_tuple( + [&](auto) { return Policy::template MakeCBlockTile_Gemm1(); }, number<2>{}); + + // clang-format off + auto gemm_0 = [&]> + (auto& t_c, auto& t_a, auto& t_b, auto i_access, PostNop = {}) { + using WarpGemm = remove_cvref_t; + + constexpr auto repeat_sub = WarpGemm::get_num_of_access(); + constexpr auto repeat_m = BlockShape::Repeat_M0; + // constexpr auto repeat_n = BlockShape::Repeat_N0; + constexpr auto repeat_k = BlockShape::Repeat_K0; + // loop order n->m->k + constexpr auto i_sub = i_access % repeat_sub; + constexpr auto i_k = (i_access / repeat_sub) % repeat_k; + constexpr auto i_m = (i_access / (repeat_sub * repeat_k )) % repeat_m; + constexpr auto i_n = (i_access / (repeat_sub * repeat_k )) / repeat_m; + + using AWarpTensor = typename WarpGemm::AWarpTensor; + using BWarpTensor = typename WarpGemm::BWarpTensor; + using CWarpTensor = typename WarpGemm::CWarpTensor; + using AWarpDstr = typename WarpGemm::AWarpDstr; + using BWarpDstr = typename WarpGemm::BWarpDstr; + using CWarpDstr = typename WarpGemm::CWarpDstr; + + constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t{}; + constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t{}; + constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t{}; + + constexpr auto a_warp_y_lengths = to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths()); + constexpr auto b_warp_y_lengths = to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths()); + constexpr auto c_warp_y_lengths = to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths()); + + AWarpTensor w_a; + w_a.get_thread_buffer() = t_a.get_y_sliced_thread_data( + merge_sequences(sequence{}, a_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, a_warp_y_lengths)); + + BWarpTensor w_b; + w_b.get_thread_buffer() = t_b.get_y_sliced_thread_data( + merge_sequences(sequence{}, b_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, b_warp_y_lengths)); + + CWarpTensor w_c; + w_c.get_thread_buffer() = t_c.get_y_sliced_thread_data( + merge_sequences(sequence{}, c_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, c_warp_y_lengths)); + + warp_gemm_0(w_c, w_a, w_b, number{}, PostNop{}); + + t_c.set_y_sliced_thread_data( + merge_sequences(sequence{}, c_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, c_warp_y_lengths), + w_c.get_thread_buffer()); + }; + // clang-format on + + // clang-format off + auto gemm_1 = [&]> + (auto& t_c, auto& t_a, auto& t_b, auto i_access, PostNop = {}) { + using WarpGemm = remove_cvref_t; + + constexpr auto repeat_sub = WarpGemm::get_num_of_access(); + constexpr auto repeat_m = BlockShape::Repeat_M0; + // constexpr auto repeat_n = BlockShape::Repeat_N0; + constexpr auto repeat_k = BlockShape::Repeat_K0; + // loop order n->m->k + constexpr auto i_sub = i_access % repeat_sub; + constexpr auto i_k = (i_access / repeat_sub) % repeat_k; + constexpr auto i_m = (i_access / (repeat_sub * repeat_k )) % repeat_m; + constexpr auto i_n = (i_access / (repeat_sub * repeat_k )) / repeat_m; + + using AWarpTensor = typename WarpGemm::AWarpTensor; + using BWarpTensor = typename WarpGemm::BWarpTensor; + using CWarpTensor = typename WarpGemm::CWarpTensor; + using AWarpDstr = typename WarpGemm::AWarpDstr; + using BWarpDstr = typename WarpGemm::BWarpDstr; + using CWarpDstr = typename WarpGemm::CWarpDstr; + + constexpr auto a_warp_y_index_zeros = uniform_sequence_gen_t{}; + constexpr auto b_warp_y_index_zeros = uniform_sequence_gen_t{}; + constexpr auto c_warp_y_index_zeros = uniform_sequence_gen_t{}; + + constexpr auto a_warp_y_lengths = to_sequence(AWarpDstr{}.get_ys_to_d_descriptor().get_lengths()); + constexpr auto b_warp_y_lengths = to_sequence(BWarpDstr{}.get_ys_to_d_descriptor().get_lengths()); + constexpr auto c_warp_y_lengths = to_sequence(CWarpDstr{}.get_ys_to_d_descriptor().get_lengths()); + + AWarpTensor w_a; + w_a.get_thread_buffer() = t_a.get_y_sliced_thread_data( + merge_sequences(sequence{}, a_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, a_warp_y_lengths)); + + BWarpTensor w_b; + w_b.get_thread_buffer() = t_b.get_y_sliced_thread_data( + merge_sequences(sequence{}, b_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, b_warp_y_lengths)); + + CWarpTensor w_c; + w_c.get_thread_buffer() = t_c.get_y_sliced_thread_data( + merge_sequences(sequence{}, c_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, c_warp_y_lengths)); + + warp_gemm_1(w_c, w_a, w_b, number{}, PostNop{}); + + t_c.set_y_sliced_thread_data( + merge_sequences(sequence{}, c_warp_y_index_zeros), + merge_sequences(sequence<1, 1>{}, c_warp_y_lengths), + w_c.get_thread_buffer()); + }; + // clang-format on + _Pragma("clang diagnostic pop"); + + // this gemm pipeline is designed with assumption that issues of buffer-load/ds_read can + // be hide under mfma. In other words, issues of mfma is >= memory this is true if we + // pre-shuffle B matrix, and A matrix is relatively small we prefer use multiple mfma + // paired with 1 buffer-load B matrix, to get max throughput of buffer_load. and by + // preshuffle, we always pack to dwordx4 load, and this will already extend to multiple + // mfma but that is already consumed inside warpgemm-impl. So indeed how many extra + // mfma(that can reuse the B matrix) only affected by M repeat. + auto pipeline_gemm0 = [&]() { + constexpr index_t total_loops = issues_gemm0; + constexpr auto sr = Policy::template GetSequencer_0(); + static_assert(sr.size() == total_loops); + + constexpr auto c_sld_a_0 = MAKE_SC(); + constexpr auto c_gld_a_0 = MAKE_SC(); + constexpr auto c_gld_b_0 = MAKE_SC(); + // compute buffer 1 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_0(acc_0, as[I0], gs[I0], i_issue); + constexpr index_t slot = sr.at(i_issue); + + if constexpr(slot & SLD_A) + sld_a(as[I1], a_sld_win1, number{}); + if constexpr(slot & GLD_A) + gld_a(a_sst_win0, number{}); + if constexpr(slot & GLD_B) + gld_g(gs[I0], number{}); + }); + move_g(); + move_a(); + block_sync_load_raw(issues_a + issues_g); + lds_load_fence(); + + constexpr auto c_sld_a_1 = MAKE_SC(); + constexpr auto c_gld_a_1 = MAKE_SC(); + constexpr auto c_gld_b_1 = MAKE_SC(); + + // compute buffer 1 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_0(acc_0, as[I1], gs[I1], i_issue); + constexpr index_t slot = sr.at(i_issue); + + if constexpr(slot & SLD_A) + sld_a(as[I0], a_sld_win0, number{}); + if constexpr(slot & GLD_A) + gld_a(a_sst_win1, number{}); + if constexpr(slot & GLD_B) + gld_g(gs[I1], number{}); + }); + move_g(); + move_a(); + block_sync_load_raw(issues_a + issues_g); + lds_load_fence(); + }; + + auto pipeline_gemm0_tail = [&]() { + constexpr index_t total_loops = issues_gemm0; + constexpr auto sr = Policy::template GetSequencer_0(); + static_assert(sr.size() == total_loops); + + constexpr auto c_gld_b_0 = MAKE_SC(); + + // compute buffer 0 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_0(acc_0, as[I0], gs[I0], i_issue); + constexpr index_t slot = sr.at(i_issue); + + if constexpr(slot & GLD_B) + gld_g(gs[I1], number{}); + }); + + block_sync_load_raw(issues_g); + sld_a(as[I1], a_sld_win1, NEG1); + + // compute buffer 1 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + constexpr auto last_nop = [&]() { + if constexpr(i_issue == (total_loops - 1)) + return TRUE; + else + return FALSE; + }(); + gemm_0(acc_0, as[I1], gs[I1], i_issue, last_nop); // last gemm has nop + }); + }; + + auto y = Policy::template MakeYBlockTile(); + + auto pipeline_bridge = [&]() { + // cast to Y data + auto y_pre = cast_tile(acc_0); + store_tile(bridge_sst_win, y_pre); + clear_tile(acc_1s(I0)); + // wave_barrier(); + load_tile(y, bridge_sld_win); + clear_tile(acc_1s(I1)); + }; + + // note, gemm-1 start from idx-1 to N-2 (0, 1, 2....N-1) + auto pipeline_gemm1 = [&]() { + constexpr index_t total_loops = issues_gemm1; + constexpr auto sr = Policy::template GetSequencer_1(); + static_assert(sr.size() == total_loops); + + constexpr auto c_gld_b_0 = MAKE_SC(); + constexpr auto c_gst_o_0 = MAKE_SC(); + constexpr auto c_gld_b_1 = MAKE_SC(); + constexpr auto c_gst_o_1 = MAKE_SC(); + + // compute buffer 0 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_1(acc_1s[I1], y, ds[I1], i_issue); + constexpr index_t slot = sr.at(i_issue); + if constexpr(slot & GLD_B) + gld_d(ds[I0], number{}); + + if constexpr(slot & GST_O) + { + auto out = cast_tile(acc_1s[I0]); + atomic_add_o(out, number{}); + } + }); + move_d(); + // move_o(); + + // compute buffer 1 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_1(acc_1s[I0], y, ds[I0], i_issue); + constexpr index_t slot = sr.at(i_issue); + if constexpr(slot & GLD_B) + gld_d(ds[I1], number{}); + + if constexpr(slot & GST_O) + { + auto out = cast_tile(acc_1s[I1]); + atomic_add_o(out, number{}); + } + }); + move_d(); + }; + + auto pipeline_gemm1_head = [&]() { + constexpr index_t total_loops = issues_gemm1; + constexpr auto sr = Policy::template GetSequencer_1(); + static_assert(sr.size() == total_loops); + + constexpr auto c_gld_b_0 = MAKE_SC(); + + // compute buffer 0 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_1(acc_1s[I0], y, ds[I0], i_issue); + constexpr index_t slot = sr.at(i_issue); + if constexpr(slot & GLD_B) + gld_d(ds[I1], number{}); + }); + move_d(); + }; + auto pipeline_gemm1_tail = [&]() { + constexpr index_t total_loops = issues_gemm1; + constexpr auto sr = Policy::template GetSequencer_1(); + static_assert(sr.size() == total_loops); + + constexpr auto c_gst_o_0 = MAKE_SC(); + + // compute buffer 1 + static_for<0, total_loops, 1>{}([&](auto i_issue) { + gemm_1(acc_1s[I1], y, ds[I1], i_issue); + + constexpr index_t slot = sr.at(i_issue); + if constexpr(slot & GST_O) + { + auto out = cast_tile(acc_1s[I0]); + atomic_add_o(out, number{}); + } + }); + { + auto out = cast_tile(acc_1s[I1]); + atomic_add_o(out, NEG1); + } + }; + + // start of pipeline + // clang-format off + gld_a(a_sst_win0, NEG1, TRUE); + gld_g(gs[I0], NEG1, TRUE); + move_a(); + move_g(); + clear_tile(acc_0); + + // preload for next round + gld_a(a_sst_win1, NEG1); + gld_g(gs[I1], NEG1); + + // make sure a,g loaded + block_sync_load_raw(issues_a + issues_g); + lds_load_fence(); + + // we manually unroll double buffer inside hot loop + const index_t iters_0 = (num_blocks_k0 - 2) / 2; + index_t i_0 = 0; // (void)i_0; (void)iters_0; (void)pipeline_gemm0; + while(i_0++ < iters_0) + { + pipeline_gemm0(); + } + pipeline_gemm0_tail(); + + pipeline_bridge(); + + const index_t iters_1 = (num_blocks_n1 - 2) / 2; + index_t i_1 = 0; // (void) i_1; (void)iters_1; (void)pipeline_gemm1; + pipeline_gemm1_head(); + while(i_1++ < iters_1) + { + pipeline_gemm1(); + } + pipeline_gemm1_tail(); + // clang-format on + } +}; + +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp new file mode 100644 index 0000000000..fea30f0297 --- /dev/null +++ b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp @@ -0,0 +1,831 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp" +#include "ck_tile/ops/flatmm.hpp" +#include "ck_tile/ops/gemm/warp/warp_gemm.hpp" +#include "ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp" + +namespace ck_tile { + +struct FusedMoeGemmPipelineFlatmmPolicy +{ + CK_TILE_HOST_DEVICE static constexpr index_t GetAsyncCopyDwords() + { + // TODO: always 1 dword + return 1; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetAlignment_A() + { + // using async + constexpr index_t copy_bytes = 4 * GetAsyncCopyDwords(); + constexpr index_t data_bytes = sizeof(typename Problem::ADataType); + static_assert(copy_bytes % data_bytes == 0); + return copy_bytes / data_bytes; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetAlignment_G() + { + constexpr index_t copy_bytes = [&]() { return 16; }(); + constexpr index_t data_bytes = sizeof(typename Problem::GDataType); + static_assert(copy_bytes % data_bytes == 0); + return copy_bytes / data_bytes; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetAlignment_D() + { + constexpr index_t copy_bytes = [&]() { return 16; }(); + constexpr index_t data_bytes = sizeof(typename Problem::DDataType); + static_assert(copy_bytes % data_bytes == 0); + return copy_bytes / data_bytes; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetAlignment_O() + { + if constexpr(Problem::Traits::OAtomic == 1) + { + // pack fp16/bf16 atomic + static_assert(sizeof(typename Problem::ODataType) == 2); + return 2; + } + else if constexpr(Problem::Traits::OAtomic == 2) + { + // fp32 atomic + return 1; + } + else + { + return 16 / sizeof(typename Problem::ODataType); + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetSmemKPack() + { + // TODO: this is for 3d layout + return 16 / sizeof(remove_cvref_t); + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetSmemKPack_A() + { + return GetSmemKPack(); + } + + // used for bridge LDS shuffle + template + CK_TILE_HOST_DEVICE static constexpr auto GetSmemKPack_Y() + { + // TODO: this should match mfma layout + return 16 / sizeof(typename Problem::YDataType); + } + + template + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize_A() + { + constexpr auto a_sld_desc = MakeLdsLoadDesc_A(); + constexpr auto a_sst_desc = MakeLdsStoreDesc_A(); + static_assert(a_sld_desc.get_element_space_size() == a_sst_desc.get_element_space_size()); + return a_sld_desc.get_element_space_size(); + } + + template + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize_Bridge() + { + constexpr auto bridge_sld_desc = MakeBridgeLdsLoadDesc(); + constexpr auto bridge_sst_desc = MakeBridgeLdsStoreDesc(); + static_assert(bridge_sld_desc.get_element_space_size() == + bridge_sst_desc.get_element_space_size()); + return bridge_sld_desc.get_element_space_size(); + } + + template + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize() + { + constexpr index_t a_lds = GetSmemSize_A(); + constexpr index_t bridge_lds = GetSmemSize_Bridge(); + return max(a_lds, bridge_lds); + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_SimpleMxK() + { + constexpr index_t K_vec = Alignment; + constexpr index_t K_rem = KPerBlock / K_vec; + + if constexpr(get_warp_size() < K_rem) + { + static_assert(K_rem % get_warp_size() == 0); + constexpr index_t K_lan = get_warp_size(); // lane within same wave is along gemm-k + constexpr index_t K_wav = K_rem / get_warp_size(); + static_assert(K_wav <= NumWarps, "not not support thread has repeat along K yet"); + constexpr index_t M_wav = NumWarps / K_wav; + static_assert(MPerBlock % M_wav == 0, "this tile size is too small please check"); + constexpr index_t M_rep = MPerBlock / M_wav; + + return make_static_tile_distribution( + tile_distribution_encoding< + sequence<1>, + tuple, sequence>, + tuple, sequence<2>>, + tuple, sequence<1>>, + sequence<1, 2>, + sequence<0, 2>>{}); + } + else + { + constexpr index_t K_lan = K_rem; + constexpr index_t M_lan = get_warp_size() / K_lan; + constexpr index_t M_wav = NumWarps; + static_assert(MPerBlock % (M_lan * M_wav) == 0, + "this tile size is too small please check"); + constexpr index_t M_rep = MPerBlock / (M_lan * M_wav); + return make_static_tile_distribution( + tile_distribution_encoding< + sequence<1>, + tuple, sequence>, + tuple, sequence<1, 2>>, + tuple, sequence<2, 0>>, + sequence<1, 2>, + sequence<0, 1>>{}); + } + } + + // optimized version for async, not same as simple MXK dist(pay attention!!) + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_SimpleMxK_Async() + { + constexpr index_t K_vec = Alignment; + constexpr index_t K_rem = KPerBlock / K_vec; + + if constexpr(get_warp_size() <= K_rem) + { + static_assert(K_rem % get_warp_size() == 0); + constexpr index_t K_lan = get_warp_size(); // lane within same wave is along gemm-k + constexpr index_t K_wav = K_rem / get_warp_size(); + static_assert(K_wav <= NumWarps, "do not support thread has repeat along K yet"); + constexpr index_t M_wav = NumWarps / K_wav; + static_assert(MPerBlock % M_wav == 0, "this tile size is too small please check"); + constexpr index_t M_rep = MPerBlock / M_wav; + // NOTE: no swap, but hard to avoid LDS bank conflict + return make_static_tile_distribution( + tile_distribution_encoding< + sequence<1>, + tuple, sequence>, + tuple, sequence<2>>, + tuple, sequence<1>>, + sequence<1, 2>, + sequence<0, 2>>{}); + } + else + { + constexpr index_t K_lan = K_rem; + constexpr index_t M_lan = get_warp_size() / K_lan; + constexpr index_t M_wav = NumWarps; + static_assert(MPerBlock % (M_lan * M_wav) == 0, + "this tile size is too small please check"); + constexpr index_t M_rep = MPerBlock / (M_lan * M_wav); + // NOTE: swapped for LDS load bank conflict free + return make_static_tile_distribution( + tile_distribution_encoding< + sequence<1>, + // Note M_wave(num waves) is the fastest dim, different from sipmle 2d + // distribution + tuple, sequence>, + tuple, sequence<1, 2>>, + tuple, sequence<1, 0>>, + sequence<1, 2>, + sequence<0, 1>>{}); + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_Nr_Kr_W() + { + return make_static_tile_distribution( + tile_distribution_encoding, + tuple, + sequence, + sequence>, + tuple, sequence<3>>, + tuple, sequence<0>>, + sequence<1, 2, 3>, + sequence<0, 0, 1>>{}); + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_A() + { + constexpr index_t Block_M_ = Problem::BlockShape::Block_M0; + constexpr index_t Block_K_ = Problem::BlockShape::Block_K0; + constexpr index_t NumWarps_ = Problem::BlockShape::NumWarps; + constexpr index_t Alignment_ = GetAlignment_A(); + return MakeGlobalTileDistribution_SimpleMxK_Async(); + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_G() + { + constexpr auto PermuteEnum = Problem::Traits::PermuteEnum; + // constexpr index_t hidden_radio_0 = Problem::Traits::IsGateOnly ? 1 : 2; + using S_ = typename Problem::BlockShape; + if constexpr(PermuteEnum == FusedMoeGemmWeightPermuteEnum::b_nr_kr_waveflatten) + { + // number{}.rrr(); + // number{}.eee(); + return MakeGlobalTileDistribution_Nr_Kr_W()>(); + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_D() + { + constexpr auto PermuteEnum = Problem::Traits::PermuteEnum; + using S_ = typename Problem::BlockShape; + if constexpr(PermuteEnum == FusedMoeGemmWeightPermuteEnum::b_nr_kr_waveflatten) + { + return MakeGlobalTileDistribution_Nr_Kr_W()>(); + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeGlobalTileDistribution_O() + { + using S_ = remove_cvref_t; + using WarpGemm = remove_cvref_t())>; + // using CDataType = typename WarpGemm::CDataType; + + constexpr auto c_block_outer_dstr_encoding = + tile_distribution_encoding, + tuple, + sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + + constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{}); + constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode); + return c_block_dstr; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeLdsStoreDesc_A() + { + // A async->LDS + constexpr index_t Block_M = Problem::BlockShape::Block_M0; + constexpr index_t Block_K = Problem::BlockShape::Block_K0; + // constexpr index_t BlockSize = Problem::BlockShape::BlockSize; + constexpr index_t warpSize = ck_tile::get_warp_size(); + constexpr index_t NumWarps = Problem::BlockShape::NumWarps; + + constexpr index_t KPack = GetSmemKPack_A(); // LDS + constexpr index_t KVector = GetAlignment_A(); // async copy 1 dword + constexpr index_t KPad = KPack; // pad between warps + + static_assert(Block_K % KVector == 0); + constexpr index_t LanesPerK = Block_K / KVector; // how many thread loading K + if constexpr(LanesPerK >= warpSize) + { + // need multiple waves to load K + static_assert(LanesPerK % warpSize == 0); + constexpr index_t wavesPerK = LanesPerK / warpSize; + if constexpr(wavesPerK > NumWarps) + { + // TODO: need multiple issues along K to load all data + } + else + { + constexpr index_t wavesPerM = NumWarps / wavesPerK; + constexpr index_t NumIssues = Block_M / wavesPerM; + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number{}), // k2 + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number<1>{}), // k2 + number{}, // lds store vector(actually no explicit store) + number<1>{}); + + constexpr auto lds_block_desc_issues_warps_lanes = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple( + make_pass_through_transform(number{}), + make_merge_transform(make_tuple(number{}, number{})), + make_merge_transform(make_tuple(number{}, number{}))), + make_tuple(sequence<0>{}, sequence<1, 2>{}, sequence<3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{})); + + return lds_block_desc_issues_warps_lanes; + } + } + else + { + // lanes within a wave load different M but same K + static_assert(warpSize % LanesPerK == 0); + constexpr index_t LaneGroups = warpSize / LanesPerK; // along m + constexpr index_t NumIssues = Block_M / (LaneGroups * NumWarps); + + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // m2 + number{}, // k0 + number{}), // k1 + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // m2 + number{}, // k0 + number<1>{}), // k1 + number{}, // lds store vector(actually no explicit store) + number<1>{}); + + constexpr auto lds_block_desc_issues_warps_lanes = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple(make_pass_through_transform(number{}), + make_pass_through_transform(number{}), + make_merge_transform(make_tuple( + number{}, number{}, number{}))), + make_tuple(sequence<0>{}, sequence<2>{}, sequence<1, 3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{}, sequence<2>{})); + + return lds_block_desc_issues_warps_lanes; + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeLdsLoadDesc_A() + { + // A async->LDS + // Note that, this descriptor is only to construct the layout inside LDS + // in real Gemm pipeline, ds_read may not follow this pattern + // (may follow that in tile_distribution) + // below code is almost the same as SmemStore dist, with difference: + // 1). modify the GuaranteedLastDimensionVectorLength of naive tensor desc + // 2). return discriptor is in NxK 2d layout + constexpr index_t Block_M = Problem::BlockShape::Block_M0; + constexpr index_t Block_K = Problem::BlockShape::Block_K0; + // constexpr index_t BlockSize = Problem::BlockShape::BlockSize; + constexpr index_t warpSize = ck_tile::get_warp_size(); + constexpr index_t NumWarps = Problem::BlockShape::NumWarps; + + constexpr index_t KPack = GetSmemKPack_A(); // LDS + constexpr index_t KVector = GetAlignment_A(); // async copy 1 dword + constexpr index_t KPad = KPack; // pad between warps + + static_assert(Block_K % KVector == 0); + constexpr index_t LanesPerK = Block_K / KVector; // how many thread loading K + if constexpr(LanesPerK >= warpSize) + { + // need multiple waves to load K + static_assert(LanesPerK % warpSize == 0); + constexpr index_t wavesPerK = LanesPerK / warpSize; + if constexpr(wavesPerK >= NumWarps) + { + // TODO: need multiple issues along K to load all data + } + else + { + constexpr index_t wavesPerM = NumWarps / wavesPerK; + constexpr index_t NumIssues = Block_M / wavesPerM; + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number{}), // k2 + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // k0 + number{}, // k1 + number<1>{}), // k2 + number{}, // lds load vector + number<1>{}); + + constexpr auto lds_desc_m_k = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple( + make_merge_transform(make_tuple(number{}, number{})), + make_merge_transform(make_tuple( + number{}, number{}, number{}))), + make_tuple(sequence<0, 1>{}, sequence<2, 3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{})); + + return lds_desc_m_k; + } + } + else + { + // lanes within a wave load different M but same K + static_assert(warpSize % LanesPerK == 0); + constexpr index_t LaneGroups = warpSize / LanesPerK; // along m + constexpr index_t NumIssues = Block_M / (LaneGroups * NumWarps); + + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // m2 + number{}, // k0 + number{}), // k1 + make_tuple(number{}, // m0 + number{}, // m1 + number{}, // m2 + number{}, // k0 + number<1>{}), // k1 + number{}, // lds load vector + number<1>{}); + + constexpr auto lds_desc_m_k = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple( + make_merge_transform( + make_tuple(number{}, number{}, number{})), + make_merge_transform(make_tuple(number{}, number{}))), + make_tuple(sequence<0, 1, 2>{}, sequence<3, 4>{}), + make_tuple(sequence<0>{}, sequence<1>{})); + + return lds_desc_m_k; + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeBridgeLdsLoadDesc() + { + constexpr index_t Block_M = Problem::BlockShape::Block_M0; + constexpr index_t Block_N = Problem::BlockShape::Block_N0; + + constexpr index_t KVector = GetSmemKPack_Y(); // async copy 1 dword + constexpr index_t KPad = 0; // pad between warps + + constexpr auto desc = + make_naive_tensor_descriptor(make_tuple(number{}, number{}), + make_tuple(number{}, number<1>{}), + number{}, + number<1>{}); + return desc; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeBridgeLdsStoreDesc() + { + constexpr index_t Block_M = Problem::BlockShape::Block_M0; + constexpr index_t Block_N = Problem::BlockShape::Block_N0; + + constexpr index_t KVector = GetSmemKPack_Y(); // async copy 1 dword + constexpr index_t KPad = 0; // KVector; // pad between warps + + constexpr auto desc = + make_naive_tensor_descriptor(make_tuple(number{}, number{}), + make_tuple(number{}, number<1>{}), + number{}, + number<1>{}); + return desc; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeBridgeLdsStoreForUKDesc() + { + constexpr index_t WarpPerBlock_N = Problem::BlockShape::WarpPerBlock_N0; + constexpr index_t Repeat_N = Problem::BlockShape::Repeat_N0; + constexpr index_t Repeat_M = Problem::BlockShape::Repeat_M0; + + constexpr index_t kAMLane = 16; + constexpr index_t kABKLane = 4; + constexpr index_t kABKPerLane = 4; + + constexpr index_t KPack = kABKPerLane; + + constexpr auto lds_block_desc_0 = make_naive_tensor_descriptor( + make_tuple(number{}, // m + number{}, // n + number{}, // n + number{}, // n + number{}, // m + number{}), // n + make_tuple(number{}, // m + number{}, // n + number{}, // n + number{}, // n + number{}, // m + number<1>{}), // n + number{}, // lds store vector(actually no explicit store) + number<1>{}); + + constexpr auto desc = transform_tensor_descriptor( + lds_block_desc_0, + make_tuple(make_merge_transform(make_tuple(number{}, number{})), + make_merge_transform(make_tuple(number{}, + number{}, + number{}, + number{}))), + make_tuple(sequence<0, 4>{}, sequence<1, 2, 3, 5>{}), + make_tuple(sequence<0>{}, sequence<1>{})); + + return desc; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemm0() + { + using S_ = typename Problem::BlockShape; + // A is vgpr, B is agpr. But since we transposed, so also need swap this + // TODO: this is ugly + constexpr auto wg_ctrl = WGAttrCtlEnum::Raw_avv; + // TODO: ugly + if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M0 == 32 && S_::Warp_N0 == 32 && S_::Warp_K0 == 16) + { + return WarpGemmImpl, + 2>>{}; + } + else if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M0 == 32 && S_::Warp_N0 == 32 && S_::Warp_K0 == 32) + { + return WarpGemmImpl, + 2>>{}; + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetSequencer_0() + { + // this function return seq<...> used to identify gld/sld/valu... inside mfma sequence + // the purpose is to hide thoes instructions under mfma + // every value inside seq<...> is a mask, indicating a specific operation + using S_ = typename Problem::BlockShape; + constexpr index_t SLD_A = static_cast(FusedMoeGemmPipelineSequencerEnum::SLD_A); + constexpr index_t GLD_A = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_A); + constexpr index_t GLD_B = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_B); + if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M0 == 32 && S_::Warp_N0 == 32 && S_::Warp_K0 == 16 && + S_::Block_M0 == 32 && S_::Block_N0 == 512 && S_::Block_K0 == 128 && + S_::Block_N1 == 128) + { + // Total 64 instructions, 32 buffer-load-dwordx4 gld_b, 8x buffer-load-dwordx1-async + // gld_a 8x ds_read_b128 sld_a total 64 slot :) + // clang-format off + constexpr auto seq_all = + // 0 1 2 3 4 5 6 7 + sequence{}; // 7 + return seq_all; + // clang-format on + } + else if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M0 == 32 && S_::Warp_N0 == 32 && S_::Warp_K0 == 16 && + S_::Block_M0 == 32 && S_::Block_N0 == 256 && S_::Block_K0 == 128 && + S_::Block_N1 == 128) + { + // Total 32 instructions, 16 buffer-load-dwordx4 gld_b, 8x buffer-load-dwordx1-async + // gld_a 8x ds_read_b128 sld_a total 64 slot :) + // clang-format off + constexpr auto seq_all = + // 0 1 2 3 4 5 6 7 + sequence{}; // 3 + return seq_all; + // clang-format on + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetSequencer_1() + { + // this function return seq<...> used to identify gld/sld/valu... inside mfma sequence + // the purpose is to hide thoes instructions under mfma + // every value inside seq<...> is a mask, indicating a specific operation + using S_ = typename Problem::BlockShape; + constexpr index_t GLD_B = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_B); + constexpr index_t GST_O = static_cast(FusedMoeGemmPipelineSequencerEnum::GST_O); + if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M1 == 32 && S_::Warp_N1 == 32 && S_::Warp_K1 == 16 && + S_::Block_M0 == 32 && S_::Block_N0 == 512 && S_::Block_K0 == 128 && + S_::Block_N1 == 128) + { + // Total 64 instructions, 32 buffer-load-dwordx4 gld_b, 8x buffer-load-dwordx1-async + // gld_a 8x ds_read_b128 sld_a total 64 slot :) + // clang-format off + constexpr auto seq_all = + // 0 1 2 3 4 5 6 7 + sequence{}; // 7 + return seq_all; + // clang-format on + } + else if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M1 == 32 && S_::Warp_N1 == 32 && S_::Warp_K1 == 16 && + S_::Block_M0 == 32 && S_::Block_N0 == 256 && S_::Block_K0 == 128 && + S_::Block_N1 == 128) + { + // Total 64 instructions, 32 buffer-load-dwordx4 gld_b, 8x buffer-load-dwordx1-async + // gld_a 8x ds_read_b128 sld_a total 64 slot :) + // clang-format off + constexpr auto seq_all = + // 0 1 2 3 4 5 6 7 + sequence{}; // 3 + return seq_all; + // clang-format on + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetWarpGemm1() + { + using S_ = typename Problem::BlockShape; + constexpr auto wg_ctrl = WGAttrCtlEnum::Raw_avv; + // TODO: ugly + if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M0 == 32 && S_::Warp_N0 == 32 && S_::Warp_K0 == 16) + { + return WarpGemmImpl, + 2>>{}; + } + else if constexpr(std::is_same_v && + std::is_same_v && + S_::Warp_M0 == 32 && S_::Warp_N0 == 32 && S_::Warp_K0 == 32) + { + return WarpGemmImpl, + 2>>{}; + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeCBlockTile_Gemm0() + { + using S_ = remove_cvref_t; + using WarpGemm = remove_cvref_t())>; + using CDataType = typename WarpGemm::CDataType; + + constexpr auto c_block_outer_dstr_encoding = + tile_distribution_encoding, + tuple, + sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + + constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{}); + constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode); + auto c_block_tensor = make_static_distributed_tensor(c_block_dstr); + return c_block_tensor; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeCBlockTile_Gemm1() + { + using S_ = remove_cvref_t; + using WarpGemm = remove_cvref_t())>; + using CDataType = typename WarpGemm::CDataType; + + constexpr auto c_block_outer_dstr_encoding = + tile_distribution_encoding, + tuple, + sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + + constexpr auto c_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + c_block_outer_dstr_encoding, typename WarpGemm::CWarpDstrEncoding{}); + constexpr auto c_block_dstr = make_static_tile_distribution(c_block_dstr_encode); + auto c_block_tensor = make_static_distributed_tensor(c_block_dstr); + return c_block_tensor; + } + + // this is used as A matrix for 2nd gemm + template + CK_TILE_HOST_DEVICE static constexpr auto MakeYTileDistribution() + { + using S_ = remove_cvref_t; + using WarpGemm = remove_cvref_t())>; + + // TODO: all waves a along different N, but same M + constexpr auto y_outer_dstr_enc = + tile_distribution_encoding, + tuple, sequence>, + tuple>, + tuple>, + sequence<1, 2>, + sequence<0, 0>>{}; + + constexpr auto y_block_dstr_encode = detail::make_embed_tile_distribution_encoding( + y_outer_dstr_enc, typename WarpGemm::AWarpDstrEncoding{}); + constexpr auto y_block_dstr = make_static_tile_distribution(y_block_dstr_encode); + return y_block_dstr; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto MakeYBlockTile() + { + constexpr auto y_block_dstr = MakeYTileDistribution(); + auto y_block_tensor = + make_static_distributed_tensor(y_block_dstr); + return y_block_tensor; + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetUK_0() + { + using S_ = typename Problem::BlockShape; + if constexpr(std::is_same_v && + std::is_same_v && + S_::Block_M0 == 32 && S_::Block_N0 == 512 && S_::Block_K0 == 128 && + S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32) + { + return Flatmm_32x512x128_1x4x1_16x16x32_BF16{}; + } + else if constexpr(std::is_same_v && + std::is_same_v && + S_::Block_M0 == 32 && S_::Block_N0 == 512 && S_::Block_K0 == 128 && + S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32) + { + return Flatmm_32x512x128_1x4x1_16x16x32_FP16{}; + } + } + + template + CK_TILE_HOST_DEVICE static constexpr auto GetUK_1() + { + using S_ = typename Problem::BlockShape; + if constexpr(std::is_same_v && + std::is_same_v && + std::is_same_v && + S_::Block_M1 == 32 && S_::Block_N1 == 128 && S_::Block_K1 == 512 && + S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32) + { + return FlatmmSn_32x128x512_1x4x1_16x16x32_BF16{}; + } + else if constexpr(std::is_same_v && + std::is_same_v && + std::is_same_v && + S_::Block_M1 == 32 && S_::Block_N1 == 128 && S_::Block_K1 == 512 && + S_::Warp_M0 == 16 && S_::Warp_N0 == 16 && S_::Warp_K0 == 32) + { + return FlatmmSn_32x128x512_1x4x1_16x16x32_FP16{}; + } + } +}; +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp new file mode 100644 index 0000000000..a6f71eafac --- /dev/null +++ b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_uk.hpp @@ -0,0 +1,354 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" +#include "ck_tile/ops/common/tensor_layout.hpp" +#include "ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_flatmm_policy.hpp" + +namespace ck_tile { + +/* +This pipeline deal with a gemm(actually 2 gemm) with one very small(token), one very big(weight) +we need to design the pipeline such that all waves along gemm-N dim (gemm-m only 1 wave) + + <----- gemm-N ------> + +----+----+----+----+ + | w0 | w1 | w2 | w3 | gemm-m + +----+----+----+----+ +*/ +template +struct FusedMoeGemmPipeline_FlatmmUk +{ + using Problem = remove_cvref_t; + using Policy = remove_cvref_t; + + using BlockShape = typename Problem::BlockShape; // this is FusedMoeGemmShape + + using ADataType = typename Problem::ADataType; + using GDataType = typename Problem::GDataType; + using DDataType = typename Problem::DDataType; + using AccDataType = typename Problem::AccDataType; + using ODataType = typename Problem::ODataType; + using AScaleDataType = typename Problem::AScaleDataType; + using GScaleDataType = typename Problem::GScaleDataType; + using DScaleDataType = typename Problem::DScaleDataType; + using YSmoothScaleDataType = typename Problem::YSmoothScaleDataType; + using TopkWeightDataType = typename Problem::TopkWeightDataType; + using IndexDataType = typename Problem::IndexDataType; + using YDataType = typename Problem::YDataType; + + using Traits = typename Problem::Traits; + + static constexpr bool IsGateOnly = Traits::IsGateOnly; + static constexpr bool UseSmoothQuant = Traits::UseSmoothQuant; + static constexpr bool PadHiddenSize = Traits::PadHiddenSize; + static constexpr bool PadIntermediateSize = Traits::PadIntermediateSize; + + static constexpr index_t kAlignmentA = Policy::template GetAlignment_A(); + static constexpr index_t kAlignmentG = Policy::template GetAlignment_G(); + static constexpr index_t kAlignmentD = Policy::template GetAlignment_D(); + static constexpr index_t kAlignmentO = Policy::template GetAlignment_O(); + + static constexpr index_t SLD_A = static_cast(FusedMoeGemmPipelineSequencerEnum::SLD_A); + static constexpr index_t GLD_A = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_A); + static constexpr index_t GLD_B = static_cast(FusedMoeGemmPipelineSequencerEnum::GLD_B); + static constexpr index_t GST_O = static_cast(FusedMoeGemmPipelineSequencerEnum::GST_O); + + static constexpr index_t kBlockPerCu = []() { + if constexpr(Problem::kBlockPerCu != -1) + return Problem::kBlockPerCu; + else + { + // minimize occupancy + return 2; + } + }(); + + static constexpr const char* name = "flatmm_uk"; + + CK_TILE_HOST_DEVICE static constexpr ck_tile::index_t GetSmemSize() + { + constexpr index_t smem_0 = Policy::template GetUK_0().GetSmemSize(); + constexpr index_t smem_1 = Policy::template GetUK_1().GetSmemSize(); + constexpr index_t smem_bridge = + BlockShape::Block_M0 * BlockShape::Block_N0 * sizeof(YDataType); + return max(smem_0, max(smem_1, smem_bridge)); + } + + // this is the thread-offset along row/col + CK_TILE_HOST_DEVICE static auto GetACoord() + { + constexpr auto a_dist = Policy::template MakeGlobalTileDistribution_A(); + const auto a_coord = a_dist.calculate_index(); + return a_coord; + } + + // this is the thread-offset along row/col + CK_TILE_HOST_DEVICE static auto GetOCoord() + { + constexpr auto o_dist = Policy::template MakeOGlobalTileDistribution(); + const auto o_coord = o_dist.calculate_index(); + return o_coord; + } + + CK_TILE_DEVICE constexpr auto GetNumRowCoords_A() + { + constexpr index_t KLans = BlockShape::Block_K0 / kAlignmentA; + constexpr index_t MLans = BlockShape::BlockSize / KLans; + constexpr index_t MRepeat = BlockShape::Block_M0 / MLans; + + return MRepeat; + } + + // TODO: properlly support scatter/gather + CK_TILE_DEVICE auto GetRowCoords_A(index_t base_offset) + { + constexpr index_t KLans = BlockShape::Block_K0 / kAlignmentA; + constexpr index_t MLans = BlockShape::BlockSize / KLans; + constexpr index_t MRepeat = BlockShape::Block_M0 / MLans; + + auto base_coord = threadIdx.x / KLans + base_offset; + + array coords; + static_for<0, MRepeat, 1>{}([&](auto i) { coords.at(i) = base_coord + i * MLans; }); + + return coords; + } + + template + CK_TILE_DEVICE auto GetRowID(const ROW_COORDS coords, const IndexDataType* sorted_token_ids_ptr) + { + constexpr index_t n_size = coords.size(); + + array row_ids; + static_for<0, n_size, 1>{}([&](auto i) { + row_ids.at(i) = sorted_token_ids_ptr[coords[i]]; // base_coord + i * MLans; + }); + + return row_ids; + } + + template + CK_TILE_DEVICE auto GetWeightScale(const ROW_COORDS coords, + const TopkWeightDataType* sorted_weight_ptr) + { + constexpr index_t n_size = coords.size(); + + array w; + static_for<0, n_size, 1>{}([&](auto i) { + w.at(i) = sorted_weight_ptr[coords[i]]; // base_coord + i * MLans; + }); + + return w; + } + + // TODO: this row id is before shuffle atomic, need use acc distribution + CK_TILE_DEVICE auto GetRowCoords_O(index_t base_offset) + { + constexpr index_t MLanes = BlockShape::Warp_M1; + constexpr index_t Repeat_M = BlockShape::Repeat_M1; + + auto base_coord = threadIdx.x % MLanes + base_offset; + + array coords; + static_for<0, Repeat_M, 1>{}([&](auto i) { coords.at(i) = base_coord + i * MLanes; }); + + return coords; + } + + template + CK_TILE_DEVICE auto operator()(const Karg& kargs, + CK_TILE_LDS_ADDR void* smem, + index_t sorted_tile_id, + index_t intermediate_tile_id) + { + constexpr index_t hidden_radio_0 = IsGateOnly ? 1 : 2; + ck_tile::index_t shared_intermediate_size_0 = kargs.intermediate_size; + ck_tile::index_t shared_intermediate_size_1 = kargs.intermediate_size / hidden_radio_0; + + index_t nr_0 = shared_intermediate_size_0 / BlockShape::Warp_N0; // divide N in W + index_t kr_0 = kargs.hidden_size / BlockShape::Warp_K0; // divide K in W + index_t nr_1 = kargs.hidden_size / BlockShape::Warp_N1; + index_t kr_1 = shared_intermediate_size_1 / BlockShape::Warp_K1; + + const IndexDataType expert_id = __builtin_amdgcn_readfirstlane( + reinterpret_cast(kargs.sorted_expert_ids_ptr)[sorted_tile_id]); + index_t expert_stride_0 = shared_intermediate_size_0 * kargs.hidden_size; + index_t expert_stride_1 = shared_intermediate_size_1 * kargs.hidden_size; + + // nr*kr*w + index_t interm_idx_nr0 = __builtin_amdgcn_readfirstlane( + intermediate_tile_id * + BlockShape::Block_Nr0); // intermediate_tile_id * Block_N / (N in W) + + index_t interm_idx_kr1 = __builtin_amdgcn_readfirstlane( + intermediate_tile_id * + BlockShape::Block_Kr1); // intermediate_tile_id * Block_N / (N in W) + + auto row_coords_a = GetRowCoords_A(sorted_tile_id * BlockShape::Block_M0); + auto row_ids_a = GetRowID( + row_coords_a, reinterpret_cast(kargs.sorted_token_ids_ptr)); + auto a_coords = generate_tuple( + [&](auto i) { + return row_ids_a[i] * kargs.stride_token + + threadIdx.x % (BlockShape::Block_K0 / kAlignmentA) * kAlignmentA; + }, + number{}); + auto a_res = + make_wave_buffer_resource(reinterpret_cast(kargs.a_ptr), + kargs.num_tokens * kargs.stride_token * sizeof(ADataType)); + + auto g_win = [&]() { + const GDataType* g_ptr = reinterpret_cast(kargs.g_ptr) + + static_cast(expert_id) * expert_stride_0 + + interm_idx_nr0 * kr_0 * BlockShape::Block_W0; + auto g_view_ = make_naive_tensor_view( + g_ptr, + make_tuple(nr_0, kr_0, number{}), + make_tuple(kr_0 * BlockShape::Block_W0, number{}, 1), + number{}, + number<1>{}); + + auto g_window_ = make_tile_window_linear_raw( + g_view_, + make_tuple(number{}, + number{}, + number{}), + {0, 0, 0}, + Policy::template MakeGlobalTileDistribution_G(), + sequence<0, 1, 1>{}); + return g_window_; + }(); + + auto g_res = g_win.get_bottom_tensor_view().get_buffer_view().cached_buf_res_; + auto g_coords = generate_tuple([&](auto i) { return g_win.cached_coords_[i].get_offset(); }, + number{}); + + const auto d_win = [&]() { + const DDataType* d_ptr = reinterpret_cast(kargs.d_ptr) + + static_cast(expert_id) * expert_stride_1 + + interm_idx_kr1 * BlockShape::Block_W1; + // note interm_idx_nr0 is along the gemm-k dim of 2nd gemm + + const auto d_view_ = make_naive_tensor_view( + d_ptr, + make_tuple(nr_1, kr_1, BlockShape::Block_W1), + make_tuple(kr_1 * BlockShape::Block_W1, BlockShape::Block_W1, 1), + number{}, + number<1>{}); + + const auto d_window_ = make_tile_window_linear_raw( + d_view_, + make_tuple(number{}, + number{}, + number{}), + {0, 0, 0}, + Policy::template MakeGlobalTileDistribution_D(), + sequence<0, 1, 1>{}); + return d_window_; + }(); + auto d_res = d_win.get_bottom_tensor_view().get_buffer_view().cached_buf_res_; + + // TODO: load D order is N0.K0...127, N64.K0...127, N0.K128...255, N64.K128...255 + // block-k=512, block-n=128 + // wg |<----- W_ ----->| + // Nr(2)*Nw(4)* Kr *Kr0(4)*Kr1(4) * [Kl(4)*Nl(16)*Kv(8)]->one issue + // y p y y p p y + // 1 2 0(imm) + auto d_coords = [&]() { + constexpr index_t Nr_ = 2; + constexpr index_t Nw_ = 4; + constexpr index_t Kr0_ = 4; + constexpr index_t Kr1_ = 4; + constexpr index_t Kl_ = 4; + constexpr index_t Nl_ = 16; + constexpr index_t Kv_ = 8; + constexpr index_t W_ = Kl_ * Nl_ * Kv_; + constexpr index_t num_offsets_ = Nr_ * Kr0_; + index_t base_os_ = (threadIdx.x % 64) * Kv_ + (threadIdx.x / 64) * + shared_intermediate_size_1 * + Nl_; // Kr0_ * Kr1_ * W_; + return generate_tuple( + [&](auto i) { + constexpr auto i_nr_ = number{}; + constexpr auto i_kr0_ = number{}; + + return i_nr_ * shared_intermediate_size_1 * Nw_ * Nl_ + i_kr0_ * Kr1_ * W_ + + base_os_; + }, + number{}); + }(); + + auto o_coords = generate_tuple( + [&](auto i) { + return row_ids_a[i] * kargs.stride_token + + threadIdx.x % (BlockShape::Block_N1 / kAlignmentO) * kAlignmentO; + }, + number{}); + + auto o_flags = + generate_tuple([&](auto i) { return cmp_lt_to_exec(row_ids_a[i], kargs.num_tokens); }, + number{}); + + auto bridge_sst_win = [&]() { + constexpr auto desc_ = Policy::template MakeBridgeLdsStoreForUKDesc(); + constexpr auto dist_ = Policy::template GetUK_0().MakeCBlockDist(); + return make_tile_window_linear(make_tensor_view( + reinterpret_cast(smem), desc_), + desc_.get_lengths(), + {0, 0}, + dist_); + }(); + auto o_res = + make_wave_buffer_resource(reinterpret_cast(kargs.o_ptr), + kargs.num_tokens * kargs.stride_token * sizeof(ODataType)); + + auto row_coords_o = GetRowCoords_O(sorted_tile_id * BlockShape::Block_M0); + auto w_scale = GetWeightScale( + row_coords_o, reinterpret_cast(kargs.sorted_weight_ptr)); + + auto uk_0 = Policy::template GetUK_0(); + auto acc_0 = uk_0(a_res, + a_coords, + g_res, + g_coords, + smem, + kargs.hidden_size, + BlockShape::Block_K0, // tile offset for B matrix each unroll + BlockShape::Block_Kr0 * + BlockShape::Block_W0); // tile offset for B matrix each unroll + + sweep_tile( + acc_0, + [&](auto idx0, auto idx1) { + fp32x2_t v_{acc_0(idx0), acc_0(idx1)}; + typename Problem::GateActivation{}(v_, v_); + acc_0(idx0) = v_.x; + acc_0(idx1) = v_.y; + }, + sequence<1, 2>{}); + + auto y_pre = cast_tile(acc_0); + + block_sync_lds(); + + store_tile(bridge_sst_win, y_pre); + block_sync_lds(); + + auto uk_1 = Policy::template GetUK_1(); + uk_1(d_res, + d_coords, + o_res, + o_coords, + o_flags, + smem, + kargs.hidden_size, // total n number + w_scale, + BlockShape::Block_Nr1 * kr_1 * BlockShape::Block_W1, // along N + BlockShape::Block_N1); // along N + } +}; + +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp new file mode 100644 index 0000000000..6089c2558f --- /dev/null +++ b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_pipeline_problem.hpp @@ -0,0 +1,46 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" + +namespace ck_tile { + +// TODO: alow 2 gemm have different type +template +struct FusedMoeGemmPipelineProblem +{ + using ADataType = remove_cvref_t; + using GDataType = remove_cvref_t; + using DDataType = remove_cvref_t; + using AccDataType = remove_cvref_t; + using ODataType = remove_cvref_t; + using AScaleDataType = remove_cvref_t; + using GScaleDataType = remove_cvref_t; + using DScaleDataType = remove_cvref_t; + using YSmoothScaleDataType = remove_cvref_t; + using TopkWeightDataType = remove_cvref_t; + using IndexDataType = remove_cvref_t; + + // the input for next gemm should have same time as + using YDataType = ADataType; + + using GateActivation = remove_cvref_t; + using BlockShape = remove_cvref_t; + using Traits = remove_cvref_t; +}; +} // namespace ck_tile diff --git a/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp new file mode 100644 index 0000000000..d7127b098c --- /dev/null +++ b/include/ck_tile/ops/fused_moe/pipeline/fused_moegemm_traits.hpp @@ -0,0 +1,48 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include "ck_tile/core.hpp" + +namespace ck_tile { + +enum class FusedMoeGemmWeightPermuteEnum +{ + // permute_b_n0_k0_n1_k1_n2_k2 = 0, // 0,1,4,2,5,3,6 + // permute_b_n0_n1_k0_k1_n2_k2 = 1, // 0,1,2,4,5,3,6 + no_permute = 0, + b_nr_kr_kw_nw_kv = 1, // 0,1,3,4,2,5 + b_nr_kr_waveflatten = b_nr_kr_kw_nw_kv, +}; + +template +struct FusedMoeGemmTraits +{ + // Gate+Up or Gate only + static constexpr bool IsGateOnly = IsGateOnly_; + static constexpr bool UseSmoothQuant = UseSmoothQuant_; + static constexpr index_t OAtomic = OAtomic_; + static constexpr FusedMoeGemmWeightPermuteEnum PermuteEnum = PermuteEnum_; + static constexpr bool PadHiddenSize = PadHiddenSize_; + static constexpr bool PadIntermediateSize = PadIntermediateSize_; +}; + +// Note: this need to be a bit mask +enum class FusedMoeGemmPipelineSequencerEnum +{ + SLD_A = 1 << 0, // shared load a + SLD_B = 1 << 1, + GLD_A = 1 << 2, // global load a + GLD_B = 1 << 3, + SST_A = 1 << 4, // shared store a + SST_B = 1 << 5, + GST_O = 1 << 6, // global store out +}; +} // namespace ck_tile diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp index 7ca4a697a7..89ea82c5bd 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm.hpp @@ -10,114 +10,134 @@ namespace ck_tile { // fp16 -using WarpGemmMfmaF16F16F32M32N32K8 = - WarpGemmImpl>; -using WarpGemmMfmaF16F16F32M16N16K16 = - WarpGemmImpl>; +using WarpGemmMfmaF16F16F32M32N32K8 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; -using WarpGemmMfmaF16F16F32M32N32K16 = - WarpGemmImpl>; +using WarpGemmMfmaF16F16F32M16N16K16 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; -using WarpGemmMfmaF16F16F32M16N16K32 = - WarpGemmImpl>; +using WarpGemmMfmaF16F16F32M32N32K16 = WarpGemmImpl, + 2>>; -using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl< - WarpGemmAtrributeMfmaIterateK_SwizzleA>; +using WarpGemmMfmaF16F16F32M16N16K32 = WarpGemmImpl, + 2>>; -using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl< - WarpGemmAtrributeMfmaIterateK_SwizzleA>; +using WarpGemmMfmaF16F16F32M32N32K8SwizzleA = WarpGemmImpl, + 1>>; -using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfmaF16F16F32M32N32K16SwizzleA = WarpGemmImpl, + 2>>; -using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution = + WarpGemmImpl>>; + +using WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution = + WarpGemmImpl>>; using WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution = WarpGemmImpl, 2>>; using WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution = WarpGemmImpl, 2>>; using WarpGemmMfmaF16F16F32M32N32K16SwizzleBTransposedCDistribution = WarpGemmImpl, 2>>; // bf16 -using WarpGemmMfmaBf16Bf16F32M32N32K8 = - WarpGemmImpl>; -using WarpGemmMfmaBf16Bf16F32M16N16K16 = - WarpGemmImpl>; +using WarpGemmMfmaBf16Bf16F32M32N32K8 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; + +using WarpGemmMfmaBf16Bf16F32M16N16K16 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; -using WarpGemmMfmaBf16Bf16F32M32N32K16 = - WarpGemmImpl>; +using WarpGemmMfmaBf16Bf16F32M32N32K16 = WarpGemmImpl, + 2>>; -using WarpGemmMfmaBf16Bf16F32M16N16K32 = - WarpGemmImpl>; +using WarpGemmMfmaBf16Bf16F32M16N16K32 = WarpGemmImpl, + 2>>; -using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl< - WarpGemmAtrributeMfmaIterateK_SwizzleA>; +using WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA = WarpGemmImpl, + 1>>; -using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA = WarpGemmImpl< - WarpGemmAtrributeMfmaIterateK_SwizzleA>; +using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA = + WarpGemmImpl, + 2>>; -using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution = + WarpGemmImpl>>; -using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution = + WarpGemmImpl>>; using WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution = WarpGemmImpl, 2>>; using WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution = WarpGemmImpl, 2>>; using WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleBTransposedCDistribution = WarpGemmImpl, 2>>; // fp8 -using WarpGemmMfma_f32_32x32x16_fp8_fp8 = - WarpGemmImpl>; -using WarpGemmMfma_f32_32x32x16_fp8_bf8 = - WarpGemmImpl>; +using WarpGemmMfma_f32_32x32x16_fp8_fp8 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; + +using WarpGemmMfma_f32_32x32x16_fp8_bf8 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; -using WarpGemmMfma_f32_32x32x16_bf8_fp8 = - WarpGemmImpl>; +using WarpGemmMfma_f32_32x32x16_bf8_fp8 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; -using WarpGemmMfma_f32_32x32x16_bf8_bf8 = - WarpGemmImpl>; +using WarpGemmMfma_f32_32x32x16_bf8_bf8 = WarpGemmImpl< + WarpGemmAtrributeMfma>>; -using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed = + WarpGemmImpl>>; -using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed = + WarpGemmImpl>>; -using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed = + WarpGemmImpl>>; -using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed = WarpGemmImpl< - WarpGemmAtrributeMfmaTransposedCDistribution>; +using WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed = + WarpGemmImpl>>; template using WarpGemmMfmaFp8Fp8F32M32N32K16SwizzleBTransposedCDistribution = WarpGemmImpl, + WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base, 2, swizzle_factor>>; diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp index d80e5198e6..0a8d2dfbe3 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma.hpp @@ -25,6 +25,8 @@ struct WarpGemmAtrributeMfma static constexpr index_t kN = Impl::kN; static constexpr index_t kK = Impl::kK; + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple, sequence>, @@ -51,10 +53,13 @@ struct WarpGemmAtrributeMfma sequence<0, 2>>; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { - Impl{}(c_vec, a_vec, b_vec); + Impl{}(c_vec, a_vec, b_vec, bool_constant{}); } // c_vec = a_vec * b_vec @@ -85,6 +90,8 @@ struct WarpGemmAtrributeMfmaIterateK static constexpr index_t kN = Impl::kN; static constexpr index_t kK = Impl::kK * kKIter; + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple, sequence>, @@ -111,8 +118,11 @@ struct WarpGemmAtrributeMfmaIterateK sequence<0, 2>>; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { using buf_a = thread_buffer; using buf_b = thread_buffer; @@ -122,10 +132,33 @@ struct WarpGemmAtrributeMfmaIterateK reinterpret_cast(a_vec) .template get_as()[iKIter], reinterpret_cast(b_vec) - .template get_as()[iKIter]); + .template get_as()[iKIter], + bool_constant{}); }); } + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + number, + bool_constant = {}) const + { + using buf_a = thread_buffer; + using buf_b = thread_buffer; + + static_assert(iKIter < kKIter); + + // static_for<0, kKIter, 1>{}([&](auto iKIter) { + Impl{}(c_vec, + reinterpret_cast(a_vec) + .template get_as()[iKIter], + reinterpret_cast(b_vec) + .template get_as()[iKIter], + bool_constant{}); + //}); + } + // c_vec = a_vec * b_vec CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const { @@ -168,6 +201,8 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution static constexpr index_t kN = Impl::kM; static constexpr index_t kK = Impl::kK; + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple, sequence>, @@ -194,11 +229,14 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution sequence<0, 2>>; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { // swap A and B - Impl{}(c_vec, b_vec, a_vec); + Impl{}(c_vec, b_vec, a_vec, bool_constant{}); } // c_vec = a_vec * b_vec @@ -226,6 +264,8 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB static constexpr index_t kN = Impl::kM; static constexpr index_t kK = Impl::kK; + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return 1; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple, sequence>, @@ -255,12 +295,15 @@ struct WarpGemmAtrributeMfmaTransposedCDistribution_SwizzleB sequence<2, 2>, sequence<0, 2>>; + template // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { // swap A and B - Impl{}(c_vec, b_vec, a_vec); + Impl{}(c_vec, b_vec, a_vec, bool_constant{}); } // c_vec = a_vec * b_vec @@ -291,6 +334,8 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution static constexpr index_t kN = Impl::kM; static constexpr index_t kK = Impl::kK * kKIter; + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple, sequence>, @@ -316,9 +361,12 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution sequence<2, 2>, sequence<0, 2>>; + template // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { using buf_a = thread_buffer; using buf_b = thread_buffer; @@ -328,10 +376,34 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution reinterpret_cast(b_vec) .template get_as()[iKIter], reinterpret_cast(a_vec) - .template get_as()[iKIter]); + .template get_as()[iKIter], + bool_constant{}); }); } + template + // c_vec += a_vec * b_vec + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + number, + bool_constant = {}) const + { + using buf_a = thread_buffer; + using buf_b = thread_buffer; + + static_assert(iKIter < kKIter); + // swap A and B, value and type + // static_for<0, kKIter, 1>{}([&](auto iKIter) { + Impl{}(c_vec, + reinterpret_cast(b_vec) + .template get_as()[iKIter], + reinterpret_cast(a_vec) + .template get_as()[iKIter], + bool_constant{}); + //}); + } + // c_vec = a_vec * b_vec CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const { @@ -377,6 +449,8 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB static constexpr index_t kK = Impl::kK * kKIter; static constexpr index_t SFactor = SFactor_; // group how many CM1 together + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple, sequence>, @@ -429,8 +503,11 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB sequence<0, 2>>; #endif // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { using buf_a = thread_buffer; using buf_b = thread_buffer; @@ -440,10 +517,33 @@ struct WarpGemmAtrributeMfmaIterateKAndTransposedCDistribution_SwizzleB reinterpret_cast(b_vec) .template get_as()[iKIter], reinterpret_cast(a_vec) - .template get_as()[iKIter]); + .template get_as()[iKIter], + bool_constant{}); }); } + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + number, + bool_constant = {}) const + { + using buf_a = thread_buffer; + using buf_b = thread_buffer; + + static_assert(iKIter < kKIter); + // swap A and B, value and type + // static_for<0, kKIter, 1>{}([&](auto iKIter) { + Impl{}(c_vec, + reinterpret_cast(b_vec) + .template get_as()[iKIter], + reinterpret_cast(a_vec) + .template get_as()[iKIter], + bool_constant{}); + //}); + } + // c_vec = a_vec * b_vec CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const { @@ -488,6 +588,8 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA static constexpr index_t kK = Impl::kK * kKIter; static constexpr index_t SFactor = SFactor_; // group how many CM1 together + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { return kKIter; } + using AWarpDstrEncoding = tile_distribution_encoding< sequence<>, tuple>; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { using buf_a = thread_buffer; using buf_b = thread_buffer; @@ -529,10 +634,33 @@ struct WarpGemmAtrributeMfmaIterateK_SwizzleA reinterpret_cast(a_vec) .template get_as()[iKIter], reinterpret_cast(b_vec) - .template get_as()[iKIter]); + .template get_as()[iKIter], + bool_constant{}); }); } + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + number, + bool_constant = {}) const + { + using buf_a = thread_buffer; + using buf_b = thread_buffer; + + static_assert(iKIter < kKIter); + + // static_for<0, kKIter, 1>{}([&](auto iKIter) { + Impl{}(c_vec, + reinterpret_cast(a_vec) + .template get_as()[iKIter], + reinterpret_cast(b_vec) + .template get_as()[iKIter], + bool_constant{}); + //}); + } + // c_vec = a_vec * b_vec CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const { diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp index bb59a72982..0aba1f5355 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm_attribute_mfma_impl.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -7,12 +7,68 @@ namespace ck_tile { +// TODO: refactor warp-gemm +// currently there is a discrepency for vav/vva if we need transpose C/D +// e.g. if we want A:agpr, B:vgpr, we have to use vva in WGAttrEnum +// because we swap the A/B pointer in _impl code (but not known this info here) +enum class WGAttrCtlEnum +{ + Default_ = 0, + Raw_vvv = 1, // c-vgpr, a-vgpr, b-vgpr + Raw_vaa = 2, // c-vgpr, a-agpr, b-agpr + Raw_vav = 3, // c-vgpr, a-agpr, b-vgpr + Raw_vva = 4, // c-vgpr, a-vgpr, b-agpr + Raw_avv = 5, // c-agpr, a-vgpr, b-vgpr + // raw_a_a_a = 3, // c-agpr, a-agpr, b-agpr +}; + +#define DISPATCH_MFMA_(mfma_, dmod_, amod_, bmod_, cmod_) \ + if constexpr(post_nop_) \ + { \ + asm volatile(mfma_ " %0, %1, %2, %3 ; yyy\n" \ + "s_nop 3" \ + : dmod_(c_vec) \ + : amod_(a_vec), bmod_(b_vec), cmod_(c_vec) \ + :); \ + } \ + else \ + { \ + asm volatile(mfma_ " %0, %1, %2, %3\n" \ + : dmod_(c_vec) \ + : amod_(a_vec), bmod_(b_vec), cmod_(c_vec) \ + :); \ + } + +#define DISPATCH_MFMA_CTRL_(mfma_, ctrl_) \ + if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vvv) \ + { \ + DISPATCH_MFMA_(mfma_, "+v", "v", "v", "v") \ + } \ + else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vaa) \ + { \ + DISPATCH_MFMA_(mfma_, "+v", "a", "a", "v") \ + } \ + else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vav) \ + { \ + DISPATCH_MFMA_(mfma_, "+v", "a", "v", "v") \ + } \ + else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_vva) \ + { \ + DISPATCH_MFMA_(mfma_, "+v", "v", "a", "v") \ + } \ + else if constexpr(ctrl_ == WGAttrCtlEnum::Raw_avv) \ + { \ + DISPATCH_MFMA_(mfma_, "+a", "v", "v", "a") \ + } + // FP16 +template struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8 { - using ADataType = fp16_t; - using BDataType = fp16_t; - using CDataType = float; + static constexpr WGAttrCtlEnum Ctrl = Ctrl_; + using ADataType = fp16_t; + using BDataType = fp16_t; + using CDataType = float; using AVecType = ext_vector_t; using BVecType = ext_vector_t; @@ -33,16 +89,23 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8 static constexpr index_t kCM1PerLane = 4; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { + DISPATCH_MFMA_CTRL_("v_mfma_f32_32x32x8f16", Ctrl) + else + { #if defined(__gfx9__) - c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0); + c_vec = __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, c_vec, 0, 0, 0); #else - ignore = c_vec; - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = c_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; #endif + } } // c_vec = a_vec * b_vec @@ -52,18 +115,20 @@ struct WarpGemmAttributeMfmaImplF16F16F32M32N32K8 return bit_cast( __builtin_amdgcn_mfma_f32_32x32x8f16(a_vec, b_vec, fp32x16_t{0.f}, 0, 0, 0)); #else - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; return CVecType{0.f}; #endif } }; +template struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16 { - using ADataType = fp16_t; - using BDataType = fp16_t; - using CDataType = float; + static constexpr WGAttrCtlEnum Ctrl = Ctrl_; + using ADataType = fp16_t; + using BDataType = fp16_t; + using CDataType = float; using AVecType = ext_vector_t; using BVecType = ext_vector_t; @@ -84,16 +149,23 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16 static constexpr index_t kCM1PerLane = 4; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { + DISPATCH_MFMA_CTRL_("v_mfma_f32_16x16x16f16", Ctrl) + else + { #if defined(__gfx9__) - c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0); + c_vec = __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, c_vec, 0, 0, 0); #else - ignore = c_vec; - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = c_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; #endif + } } // c_vec = a_vec * b_vec @@ -103,19 +175,21 @@ struct WarpGemmAttributeMfmaImplF16F16F32M16N16K16 return bit_cast( __builtin_amdgcn_mfma_f32_16x16x16f16(a_vec, b_vec, fp32x4_t{0.f}, 0, 0, 0)); #else - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; return CVecType{0.f}; #endif } }; // Bf16 +template struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8 { - using ADataType = bf16_t; - using BDataType = bf16_t; - using CDataType = float; + static constexpr WGAttrCtlEnum Ctrl = Ctrl_; + using ADataType = bf16_t; + using BDataType = bf16_t; + using CDataType = float; using AVecType = ext_vector_t; using BVecType = ext_vector_t; @@ -136,28 +210,35 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8 static constexpr index_t kCM1PerLane = 4; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { + DISPATCH_MFMA_CTRL_("v_mfma_f32_32x32x8bf16_1k", Ctrl) + else + { #if defined(__gfx90a__) || defined(__gfx94__) - c_vec = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0); + c_vec = __builtin_amdgcn_mfma_f32_32x32x8bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0); #elif defined(__gfx908__) - static_for<0, 2, 1>{}([&](auto k) { - c_vec = __builtin_amdgcn_mfma_f32_32x32x4bf16( - reinterpret_cast&>(a_vec) - .template get_as>()[number{}], - reinterpret_cast&>(b_vec) - .template get_as>()[number{}], - c_vec, - 0, - 0, - 0); - }); + static_for<0, 2, 1>{}([&](auto k) { + c_vec = __builtin_amdgcn_mfma_f32_32x32x4bf16( + reinterpret_cast&>(a_vec) + .template get_as>()[number{}], + reinterpret_cast&>(b_vec) + .template get_as>()[number{}], + c_vec, + 0, + 0, + 0); + }); #else - ignore = c_vec; - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = c_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; #endif + } } // c_vec = a_vec * b_vec @@ -181,18 +262,20 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M32N32K8 }); return c_vec; #else - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; return CVecType{0.f}; #endif } }; +template struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16 { - using ADataType = bf16_t; - using BDataType = bf16_t; - using CDataType = float; + static constexpr WGAttrCtlEnum Ctrl = Ctrl_; + using ADataType = bf16_t; + using BDataType = bf16_t; + using CDataType = float; using AVecType = ext_vector_t; using BVecType = ext_vector_t; @@ -213,28 +296,34 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16 static constexpr index_t kCM1PerLane = 4; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { + DISPATCH_MFMA_CTRL_("v_mfma_f32_16x16x16bf16_1k", Ctrl) + { #if defined(__gfx90a__) || defined(__gfx94__) - c_vec = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0); + c_vec = __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a_vec, b_vec, c_vec, 0, 0, 0); #elif defined(__gfx908__) - static_for<0, 2, 1>{}([&](auto k) { - c_vec = __builtin_amdgcn_mfma_f32_16x16x8bf16( - reinterpret_cast&>(a_vec) - .template get_as>()[number{}], - reinterpret_cast&>(b_vec) - .template get_as>()[number{}], - c_vec, - 0, - 0, - 0); - }); + static_for<0, 2, 1>{}([&](auto k) { + c_vec = __builtin_amdgcn_mfma_f32_16x16x8bf16( + reinterpret_cast&>(a_vec) + .template get_as>()[number{}], + reinterpret_cast&>(b_vec) + .template get_as>()[number{}], + c_vec, + 0, + 0, + 0); + }); #else - ignore = c_vec; - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = c_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; #endif + } } // c_vec = a_vec * b_vec @@ -258,20 +347,21 @@ struct WarpGemmAttributeMfmaImplBf16Bf16F32M16N16K16 }); return c_vec; #else - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; return CVecType{0.f}; #endif } }; // FP8 -template +template struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base { - using ADataType = AType_; - using BDataType = BType_; - using CDataType = float; + static constexpr WGAttrCtlEnum Ctrl = Ctrl_; + using ADataType = AType_; + using BDataType = BType_; + using CDataType = float; using AVecType = ext_vector_t; using BVecType = ext_vector_t; @@ -292,38 +382,120 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base static constexpr index_t kCM1PerLane = 4; // c_vec += a_vec * b_vec - CK_TILE_DEVICE void - operator()(CVecType& c_vec, const AVecType& a_vec, const BVecType& b_vec) const + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const { + if constexpr(Ctrl == WGAttrCtlEnum::Raw_vvv) + { + if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "v", "v", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "v", "v", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "v", "v", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "v", "v", "v") + } + } + else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vaa) + { + if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "a", "a", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "a", "a", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "a", "a", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "a", "a", "v") + } + } + else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vav) + { + if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "a", "v", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "a", "v", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "a", "v", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "a", "v", "v") + } + } + else if constexpr(Ctrl == WGAttrCtlEnum::Raw_vva) + { + if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_fp8", "+v", "v", "a", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_fp8_bf8", "+v", "v", "a", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_fp8", "+v", "v", "a", "v") + } + else if constexpr(std::is_same_v && std::is_same_v) + { + DISPATCH_MFMA_("mfma_f32_32x32x16_bf8_bf8", "+v", "v", "a", "v") + } + } + else + { #if defined(__gfx94__) - if constexpr(std::is_same_v && std::is_same_v) - c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8( - bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); - else if constexpr(std::is_same_v && std::is_same_v) - c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8( - bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); - else if constexpr(std::is_same_v && std::is_same_v) - c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8( - bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); - else if constexpr(std::is_same_v && std::is_same_v) - c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8( - bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); + if constexpr(std::is_same_v && std::is_same_v) + c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_fp8( + bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); + else if constexpr(std::is_same_v && std::is_same_v) + c_vec = __builtin_amdgcn_mfma_f32_32x32x16_fp8_bf8( + bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); + else if constexpr(std::is_same_v && std::is_same_v) + c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_fp8( + bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); + else if constexpr(std::is_same_v && std::is_same_v) + c_vec = __builtin_amdgcn_mfma_f32_32x32x16_bf8_bf8( + bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); #elif defined(__gfx908__) || defined(__gfx90a__) - static_for<0, 8, 1>{}([&](auto k) { - float a_f32 = - type_convert(reinterpret_cast&>(a_vec) - .template get_as()[number{}]); - float b_f32 = - type_convert(reinterpret_cast&>(b_vec) - .template get_as()[number{}]); - - c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0); - }); + static_for<0, 8, 1>{}([&](auto k) { + float a_f32 = + type_convert(reinterpret_cast&>(a_vec) + .template get_as()[number{}]); + float b_f32 = + type_convert(reinterpret_cast&>(b_vec) + .template get_as()[number{}]); + + c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0); + }); #else - ignore = c_vec; - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = c_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; #endif + } } // c_vec = a_vec * b_vec @@ -356,20 +528,97 @@ struct WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base }); return c_vec; #else - ignore = a_vec; - ignore = b_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; return CVecType{0.f}; #endif } }; +template using WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_fp8 = - WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + +template using WarpGemmAttributeMfmaImpl_f32_32x32x16_fp8_bf8 = - WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + +template using WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_fp8 = - WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + +template using WarpGemmAttributeMfmaImpl_f32_32x32x16_bf8_bf8 = - WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + WarpGemmAttributeMfmaImpl_f32_32x32x16_f8_base; + +// int8 +template +struct WarpGemmAttributeMfmaImpl_i32_32x32x16_i8 +{ + static constexpr WGAttrCtlEnum Ctrl = Ctrl_; + using ADataType = int8_t; + using BDataType = int8_t; + using CDataType = int32_t; + + using AVecType = ext_vector_t; + using BVecType = ext_vector_t; + using CVecType = ext_vector_t; + + static constexpr index_t kM = 32; + static constexpr index_t kN = 32; + static constexpr index_t kK = 16; + + static constexpr index_t kAMLane = 32; + static constexpr index_t kBNLane = 32; + static constexpr index_t kABKLane = 2; + static constexpr index_t kABKPerLane = 8; + + static constexpr index_t kCMLane = 2; + static constexpr index_t kCNLane = 32; + static constexpr index_t kCM0PerLane = 4; + static constexpr index_t kCM1PerLane = 4; + + // c_vec += a_vec * b_vec + template + CK_TILE_DEVICE void operator()(CVecType& c_vec, + const AVecType& a_vec, + const BVecType& b_vec, + bool_constant = {}) const + { + DISPATCH_MFMA_CTRL_("v_mfma_i32_32x32x16_i8", Ctrl) + else + { +#if defined(__gfx94__) + c_vec = __builtin_amdgcn_mfma_i32_32x32x8i8( + bit_cast(a_vec), bit_cast(b_vec), c_vec, 0, 0, 0); +#elif defined(__gfx908__) || defined(__gfx90a__) + static_for<0, 8, 1>{}([&](auto k) { + float a_f32 = + type_convert(reinterpret_cast&>(a_vec) + .template get_as()[number{}]); + float b_f32 = + type_convert(reinterpret_cast&>(b_vec) + .template get_as()[number{}]); + + c_vec = __builtin_amdgcn_mfma_f32_32x32x2f32(a_f32, b_f32, c_vec, 0, 0, 0); + }); +#else + ck_tile::ignore = c_vec; + ck_tile::ignore = a_vec; + ck_tile::ignore = b_vec; +#endif + } + } + + // c_vec = a_vec * b_vec + CK_TILE_DEVICE CVecType operator()(const AVecType& a_vec, const BVecType& b_vec) const + { + CVecType c_vec{0}; + operator()(c_vec, a_vec, b_vec); + return c_vec; + } +}; + +#undef DISPATCH_MFMA_ } // namespace ck_tile diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp index 4183d9cb95..99cd5d787e 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm_dispatcher.hpp @@ -1,5 +1,5 @@ // SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. +// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved. #pragma once @@ -21,40 +21,40 @@ struct WarpGemmMfmaDispatcher; // clang-format off // fp16 -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K8; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K16; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K16; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K32; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K8; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K8TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K16; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K16TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K16; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K16TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K32; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M16N16K32TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K8SwizzleA; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K8SwizzleA; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaF16F16F32M32N32K16SwizzleA; }; // bf16 -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K16TransposedCDistribution; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M16N16K32TransposedCDistribution; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K8SwizzleA; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfmaBf16Bf16F32M32N32K16SwizzleA; }; // fp8 -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8; }; -template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_fp8_CTransposed; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_fp8_bf8_CTransposed; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_fp8_CTransposed; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8; }; +template<> struct WarpGemmMfmaDispatcher { using Type = WarpGemmMfma_f32_32x32x16_bf8_bf8_CTransposed; }; // clang-format on } // namespace impl diff --git a/include/ck_tile/ops/gemm/warp/warp_gemm_impl.hpp b/include/ck_tile/ops/gemm/warp/warp_gemm_impl.hpp index eb9dbf127d..182d023a00 100644 --- a/include/ck_tile/ops/gemm/warp/warp_gemm_impl.hpp +++ b/include/ck_tile/ops/gemm/warp/warp_gemm_impl.hpp @@ -31,11 +31,21 @@ struct WarpGemmImpl using BWarpTensor = static_distributed_tensor; using CWarpTensor = static_distributed_tensor; - CK_TILE_DEVICE void operator()(CWarpTensor& c, const AWarpTensor& a, const BWarpTensor& b) const + CK_TILE_HOST_DEVICE static constexpr auto get_num_of_access() { - using AVec = ext_vector_t; - using BVec = ext_vector_t; - using CVec = ext_vector_t; + return WarpGemmAttribute_::get_num_of_access(); + } + + template + CK_TILE_DEVICE void + operator()(CTensor& c, const ATensor& a, const BTensor& b, bool_constant = {}) const + { + static_assert(detail::is_similiar_distributed_tensor_v && + detail::is_similiar_distributed_tensor_v && + detail::is_similiar_distributed_tensor_v); + using AVec = ext_vector_t; + using BVec = ext_vector_t; + using CVec = ext_vector_t; constexpr auto I0 = number<0>{}; @@ -44,18 +54,49 @@ struct WarpGemmImpl auto c_vec = c.get_thread_buffer().template get_as()[I0]; // c_vec += a_vec * b_vec - WarpGemmAttribute{}(c_vec, a_vec, b_vec); + WarpGemmAttribute{}(c_vec, a_vec, b_vec, bool_constant{}); c.get_thread_buffer().template set_as(I0, c_vec); } - CK_TILE_DEVICE auto operator()(const AWarpTensor& a, const BWarpTensor& b) const + template + CK_TILE_DEVICE void operator()(CTensor& c, + const ATensor& a, + const BTensor& b, + number, + bool_constant = {}) const { - CWarpTensor c; + using AVec = ext_vector_t; + using BVec = ext_vector_t; + using CVec = ext_vector_t; + + constexpr auto I0 = number<0>{}; - using AVec = ext_vector_t; - using BVec = ext_vector_t; - using CVec = ext_vector_t; + const auto a_vec = a.get_thread_buffer().template get_as()[I0]; + const auto b_vec = b.get_thread_buffer().template get_as()[I0]; + auto c_vec = c.get_thread_buffer().template get_as()[I0]; + + // c_vec += a_vec * b_vec + WarpGemmAttribute{}(c_vec, a_vec, b_vec, number{}, bool_constant{}); + + c.get_thread_buffer().template set_as(I0, c_vec); + } + + template + CK_TILE_DEVICE auto operator()(const ATensor& a, const BTensor& b) const + { + using CTensor = CWarpTensor; + static_assert(detail::is_similiar_distributed_tensor_v && + detail::is_similiar_distributed_tensor_v); + CTensor c; + + using AVec = ext_vector_t; + using BVec = ext_vector_t; + using CVec = ext_vector_t; constexpr auto I0 = number<0>{}; diff --git a/include/ck_tile/ops/moe_sorting.hpp b/include/ck_tile/ops/moe_sorting.hpp deleted file mode 100644 index b74607f061..0000000000 --- a/include/ck_tile/ops/moe_sorting.hpp +++ /dev/null @@ -1,11 +0,0 @@ -// SPDX-License-Identifier: MIT -// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved. - -#pragma once - -#include "ck_tile/ops/fused_moe/kernel/moe_sorting_kernel.hpp" -#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_pipeline.hpp" -#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_policy.hpp" -#include "ck_tile/ops/fused_moe/pipeline/moe_sorting_problem.hpp" -#include "ck_tile/ops/common/generic_2d_block_shape.hpp" -#include "ck_tile/ops/common/tensor_layout.hpp"