lite

build/Linux/Release/amdgpu/onnxruntime/core/providers/rocm/tensor/tile_impl.cu

@@ -10,7 +10,7 @@ namespace rocm {
 
 #ifdef USE_ROCM
 constexpr int num_elements_per_thread = 2;
-constexpr int num_threads_per_block = 512;
+constexpr int num_threads_per_block = 256;
 #else
 constexpr int num_elements_per_thread = GridDim::maxElementsPerThread;
 constexpr int num_threads_per_block = GridDim::maxThreadsPerBlock;

build/Linux/Release/amdgpu/onnxruntime/core/providers/rocm/tensor/transpose_impl.cu

@@ -133,6 +133,7 @@ bool CanDoTranspose4DParallelizeMultipleElementsPerThreadInInnermostDim(const hi
                                                                         const gsl::span<const int64_t>& input_dims,
                                                                         const gsl::span<const size_t>& permutations,
                                                                         dim3& grid_size, dim3& block_size) {
+  //printf("maxThreadsPerBlock:%d \n",prop.maxThreadsPerBlock);
   if (rank == 4 &&
       // the permutations is not on the last dimension.
       permutations[3] == 3) {
@@ -142,7 +143,9 @@ bool CanDoTranspose4DParallelizeMultipleElementsPerThreadInInnermostDim(const hi
     // dims[2]: block.y + grid.x
     // dims[1]: grid.y
     // dims[0]: grid.z
-    if (input_dims[3] / num_elements_per_thread <= prop.maxThreadsPerBlock &&
+    
+    const int maxThreadsPerBlock = prop.maxThreadsPerBlock;
+    if (input_dims[3] / num_elements_per_thread <= maxThreadsPerBlock &&
         (input_dims[3] % num_elements_per_thread) == 0 &&
         input_dims[1] <= prop.maxGridSize[1] &&
         input_dims[0] <= prop.maxGridSize[2]) {
@@ -150,7 +153,7 @@ bool CanDoTranspose4DParallelizeMultipleElementsPerThreadInInnermostDim(const hi
       // 1. block_size_x * block_size_y <= prop.maxThreadsPerBlock
       // 2. block_size_y * num_block_ext >= input_dims[2]
       int64_t block_size_x = input_dims[3] / num_elements_per_thread;
-      int64_t max_block_size_y = prop.maxThreadsPerBlock / block_size_x;
+      int64_t max_block_size_y = maxThreadsPerBlock / block_size_x;
       int64_t block_size_y = min(input_dims[2], max_block_size_y);
       int64_t num_block_ext = CeilDiv(input_dims[2], block_size_y);
 
@@ -255,14 +258,15 @@ bool CanDoTranspose4DParallelizeOneElementPerThread(const hipDeviceProp_t& prop,
     // dims[2]: block.y + grid.x
     // dims[1]: grid.y
     // dims[0]: grid.z
-    if (input_dims[3] <= prop.maxThreadsPerBlock &&
+    const int maxThreadsPerBlock = prop.maxThreadsPerBlock;
+    if (input_dims[3] <= maxThreadsPerBlock &&
         input_dims[1] <= prop.maxGridSize[1] &&
         input_dims[0] <= prop.maxGridSize[2]) {
       // There are 2 constrains when luanching the kernels
       // 1. block_size_x * block_size_y <= prop.maxThreadsPerBlock
       // 2. block_size_y * num_block_ext >= input_dims[2]
       int64_t block_size_x = input_dims[3];
-      int64_t max_block_size_y = prop.maxThreadsPerBlock / block_size_x;
+      int64_t max_block_size_y = maxThreadsPerBlock / block_size_x;
       int64_t block_size_y = std::min(input_dims[2], max_block_size_y);
       int64_t num_block_ext = CeilDiv(input_dims[2], block_size_y);