use_c2c_tiles to params.hjson, remove legacy gemm_baseline

pulp-platform · Feb 9, 2024 · e37ba95 · e37ba95
1 parent c056eac
commit e37ba95
Show file tree

Hide file tree

Showing 7 changed files with 40 additions and 192 deletions.
diff --git a/sw/blas/gemm/data/datagen.py b/sw/blas/gemm/data/datagen.py
@@ -122,6 +122,7 @@ def emit_header(**kwargs):
     # gemmImpl
     data_str += ["// -- gemmImpl"]
     data_str += [f"#define USE_METHOD {gemmImpl['method']}"]
+    data_str += [f"#define USE_C2C_TILES {int(gemmImpl['use_c2c_tiles'])}"]
     data_str += [f"#define L1_M {gemmImpl['L1_M']}"]
     data_str += [f"#define L1_N {gemmImpl['L1_N']}"]
     data_str += [f"#define L1_K {gemmImpl['L1_K']}"]

diff --git a/sw/blas/gemm/data/params.hjson b/sw/blas/gemm/data/params.hjson
@@ -21,11 +21,12 @@
     }
 
     gemmImpl: {
-        method: "baseline",
-        L1_M: 8,
-        L1_N: 8,
-        L1_K: 8,
-        ta_tile: false,
+        method: "2dpipe",
+        use_c2c_tiles: true,
+        L1_M: 16,
+        L1_N: 16,
+        L1_K: 16,
+        ta_tile: true,
         tb_tile: false,
         tc_tile: false, // not implemented
         expand: 0,

diff --git a/sw/blas/gemm/src/dma_xfer_test.h b/sw/blas/gemm/src/dma_xfer_test.h
@@ -1,5 +1,11 @@
 #include "gemm_decls.h"
 
+/**
+ * Test DMA effective bandwidth.
+ * Transfers an array from HBM to TCDM, 
+ * then rotates between cluster TCDMs with C2C communication, 
+ * then stores the result back to HBM.
+*/
 void dma_xfer_test(const double* A, const uint32_t N, const bool bench) {
     if (!snrt_is_dm_core()) return;
 

diff --git a/sw/blas/gemm/src/gemm_baseline.h b/sw/blas/gemm/src/gemm_baseline.h
diff --git a/sw/blas/gemm/src/gemm_kernel_legacy.h b/sw/blas/gemm/src/gemm_kernel_legacy.h
@@ -23,6 +23,33 @@ typedef __fp16 v4f16 __attribute__((vector_size(8)));
 typedef char v8f8 __attribute__((vector_size(8)));
 #endif
 
+/**
+ * \brief Each cluster performs a GEMM for A, B, C inside each TCDM
+ */
+void gemm_cluster_kernel_baseline(double alpha, double beta, uint32_t M, uint32_t N,
+                                  uint32_t K, double* const A, double* const B,
+                                  double* const C, int lda, int ldb, int ldc) {
+    uint32_t p[3], P[3];
+    ocrt_thread_idx(p);
+    ocrt_compute_thread_num(P);
+
+    for (uint32_t i = p[0]; i < M; i += P[0]) {
+        for (uint32_t j = 0; j < N; j++) {
+            uint32_t cIdx = i * ldc + j;  // C[i][j]
+            register double c0 = beta * C[cIdx];
+
+            for (uint32_t k = 0; k < K; k++) {
+                uint32_t aIdx = i * lda + k;  // A[i][k]
+                uint32_t bIdx = k * ldb + j;  // B[k][j]
+
+                c0 += A[aIdx] * B[bIdx];
+            }
+            C[cIdx] = c0;
+        }
+    }
+    snrt_fpu_fence();
+}
+
 void gemm_fp64_baseline(uint32_t M, uint32_t N, uint32_t K, double* A,
                         uint32_t ldA, uint32_t ta, double* B, uint32_t ldB,
                         uint32_t tb, double* C, uint32_t ldC, double ALPHA) {

diff --git a/sw/blas/gemm/src/gemm_tiling_2dpipe_tpl.h b/sw/blas/gemm/src/gemm_tiling_2dpipe_tpl.h
@@ -10,8 +10,6 @@
 
 void SNBLAS_GEMM_TILING(2dpipe, FLOAT_T, IS_DM_CORE) (const SnblasGemmInfo info, const SNBLAS_GEMM_ARGS(FLOAT_T) args, const SnblasGemmImpl impl) {
 
-#define USE_C2C_TILES true
-
     /**
      * Problem is double buffered in L1. The buffer that is used is toggled at
      * each iteration. The DMA cores are one index step ahead so they load the

diff --git a/sw/blas/gemm/src/main.c b/sw/blas/gemm/src/main.c
@@ -21,8 +21,6 @@
 NAMED_DUMP(uint32_t, err, 0x7)
 NAMED_DUMP(uint32_t, bench_iter, 0x7)
 
-#define BIST
-#include "data.h"
 
 int main() {
     const bool setup_ssr = true;