Update sgemm_wmma_tf32_stage.cu (#79)

DefTruth · Oct 14, 2024 · ba4998d · ba4998d
1 parent cf4f9d7
commit ba4998d
Showing 1 changed file with 21 additions and 0 deletions.
diff --git a/sgemm/sgemm_wmma_tf32_stage.cu b/sgemm/sgemm_wmma_tf32_stage.cu
@@ -283,6 +283,27 @@ void sgemm_wmma_m16n16k8_mma4x2_warp2x4_stage2(
   constexpr int NUM_THREADS= (
     WMMA_TILE_M * WMMA_TILE_N * WARP_SIZE); // 2 * 4 * 32 = 256
 
+  // constexpr int BM = WMMA_M * WMMA_TILE_M * WARP_TILE_M; // 16x4*2=128
+  // constexpr int BN = WMMA_N * WMMA_TILE_N * WARP_TILE_N; // 16x2*4=128
+  // constexpr int BK = WMMA_K; // 8
+  // constexpr int OFFSET=0;
+
+  // int dev_id = 0;
+  // cudaGetDevice(&dev_id);
+  // cudaDeviceProp dev_prop;
+  // cudaGetDeviceProperties(&dev_prop, dev_id);
+  // int smem_max_size = (K_STAGE * BM * (BK+OFFSET) * sizeof(float) + 
+  //                      K_STAGE * BK * (BN+OFFSET) * sizeof(float));
+  // smem_max_size = (smem_max_size < dev_prop.sharedMemPerMultiprocessor ? 
+  //                  smem_max_size : dev_prop.sharedMemPerMultiprocessor);
+
+  // cudaFuncSetAttribute(
+  //   sgemm_wmma_m16n16k8_mma4x2_warp2x4_stages_kernel<
+  //     WMMA_M, WMMA_N, WMMA_K, WMMA_TILE_M, WMMA_TILE_N, 
+  //     WARP_TILE_M, WARP_TILE_N, K_STAGE, 0>, 
+  //   cudaFuncAttributeMaxDynamicSharedMemorySize, 
+  //   smem_max_size);
+
   dim3 block(NUM_THREADS);
   dim3 grid(div_ceil(N, WMMA_N * WMMA_TILE_N * WARP_TILE_N), 
             div_ceil(M, WMMA_M * WMMA_TILE_M * WARP_TILE_M));