Skip to content

Commit

Permalink
Add splitk gemm fp16 @ fp16 with fp8 compute instances (#983)
Browse files Browse the repository at this point in the history 
* Add ComputeType

* Update for compatibility

* Add instances

* Update profiler api
  • Loading branch information
@geyyer
geyyer authored Oct 13, 2023
1 parent 2ce9b56 commit fa753f2
Show file tree
Hide file tree
Showing 10 changed files with 398 additions and 31 deletions.
9 changes: 6 additions & 3 deletions include/ck/tensor_operation/gpu/device/device_gemm_splitk.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,8 @@ template <typename ALayout,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
typename ComputeType = CDataType>
struct DeviceGemmSplitK : public BaseOperator
{
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(const void* p_a,
Expand Down Expand Up @@ -48,7 +49,8 @@ template <typename ALayout,
typename CDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
typename CElementwiseOperation,
typename ComputeType = CDataType>
using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
BLayout,
CLayout,
Expand All @@ -57,7 +59,8 @@ using DeviceGemmSplitKPtr = std::unique_ptr<DeviceGemmSplitK<ALayout,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>>;
CElementwiseOperation,
ComputeType>>;

} // namespace device
} // namespace tensor_operation
Expand Down
3 changes: 2 additions & 1 deletion include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -69,7 +69,8 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
CDataType,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation>
CElementwiseOperation,
ComputeType>
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
Expand Down
56 changes: 52 additions & 4 deletions library/include/ck/library/tensor_operation_instance/gpu/gemm_splitk.hpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -98,14 +98,35 @@ void add_device_gemm_xdl_splitk_f16_f8_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F8, F16, PassThrough, PassThrough, PassThrough>>>&
instances);

void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);

void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);

void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);

void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances);
#endif

template <typename ADataType,
typename BDataType,
typename CDataType,
typename ALayout,
typename BLayout,
typename CLayout>
typename CLayout,
typename ComputeType>
struct DeviceOperationInstanceFactory<
ck::tensor_operation::device::DeviceGemmSplitK<ALayout,
BLayout,
Expand All @@ -115,7 +136,8 @@ struct DeviceOperationInstanceFactory<
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>>
ck::tensor_operation::element_wise::PassThrough,
ComputeType>>
{
using DeviceOp = DeviceGemmSplitK<ALayout,
BLayout,
Expand All @@ -125,7 +147,8 @@ struct DeviceOperationInstanceFactory<
CDataType,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
ck::tensor_operation::element_wise::PassThrough,
ComputeType>;

static auto GetInstances()
{
Expand Down Expand Up @@ -158,7 +181,7 @@ struct DeviceOperationInstanceFactory<
#endif
#ifdef CK_ENABLE_FP16
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t>)
is_same_v<CDataType, half_t> && is_same_v<ComputeType, half_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
Expand Down Expand Up @@ -231,6 +254,31 @@ struct DeviceOperationInstanceFactory<
add_device_gemm_xdl_splitk_f16_f8_f16_km_nk_mn_instances(op_ptrs);
}
}
else if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<CDataType, half_t> && is_same_v<ComputeType, f8_t>)
{
if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_mk_nk_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Col> &&
is_same_v<CLayout, Row>)
{
add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_nk_mn_instances(op_ptrs);
}
}

#endif
return op_ptrs;
}
Expand Down
6 changes: 5 additions & 1 deletion library/src/tensor_operation_instance/gpu/gemm_splitk/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,10 @@ list(APPEND GEMM_SPLITK_INSTANCES device_gemm_xdl_splitk_f32_f32_f32_mk_kn_mn_in
device_gemm_xdl_splitk_f16_fp8_f16_mk_kn_mn_instance.cpp
device_gemm_xdl_splitk_f16_fp8_f16_mk_nk_mn_instance.cpp
device_gemm_xdl_splitk_f16_fp8_f16_km_kn_mn_instance.cpp
device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp)
device_gemm_xdl_splitk_f16_fp8_f16_km_nk_mn_instance.cpp
device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_kn_mn_instance.cpp
device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_mk_nk_mn_instance.cpp
device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_nk_mn_instance.cpp)

add_instance_library(device_gemm_splitk_instance ${GEMM_SPLITK_INSTANCES})
62 changes: 62 additions & 0 deletions ...nstance/gpu/gemm_splitk/device_gemm_xdl_splitk_f16_f16_f16_comp_fp8_km_kn_mn_instance.cpp
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,62 @@
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.

#include <cstdlib>

#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp"

#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"

namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {

using F8 = ck::f8_t;
using F16 = ck::half_t;
using F32 = float;

using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;

template <ck::index_t... Is>
using S = ck::Sequence<Is...>;

using PassThrough = ck::tensor_operation::element_wise::PassThrough;

static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;

// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances = std::tuple<
// clang-format off
//#########################|AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| Compute|
//#########################| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise|Specialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| Type|
//#########################| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| |
//#########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 32, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, F8>,
DeviceGemmXdlSplitKCShuffle< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 1, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 2, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, F8>
// clang-format on
>;

void add_device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances(
std::vector<std::unique_ptr<
DeviceGemmSplitK<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough, F8>>>&
instances)
{
add_device_operation_instances(instances,
device_gemm_xdl_splitk_f16_f16_f16_comp_f8_km_kn_mn_instances{});
}

} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Loading

0 comments on commit fa753f2

Please sign in to comment.