Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel...

Merge branch 'develop' of github.com:ROCmSoftwarePlatform/composable_kernel into barkocot/grouped-conv-weight-fp16-c1-k1

include/ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp

@@ -4,7 +4,8 @@
 #pragma once
 
 #include "ck/utility/common_header.hpp"
-#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
+#include "ck/utility/loop_scheduler.hpp"
+#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
 
 namespace ck {
 

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_streamk.hpp

@@ -37,7 +37,8 @@ __global__ void
                                    index_t StrideC,
                                    typename GridwiseGemm::Block2CTileMap block_mapping)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
 
     __shared__ uint8_t p_shared[shared_size];

library/include/ck/library/reference_tensor_operation/cpu/reference_pool_fwd.hpp

@@ -256,10 +256,12 @@ struct ReferencePoolingFwd : public device::BaseOperator
 
                     for(ck::index_t y = 0; y < arg.window_spatial_lengths_[0]; ++y)
                     {
-                        ck::index_t hi = ho * arg.window_strides_[0] + y - arg.in_left_pads_[0];
+                        ck::index_t hi = ho * arg.window_strides_[0] +
+                                         y * arg.window_dilations_[0] - arg.in_left_pads_[0];
                         for(ck::index_t x = 0; x < arg.window_spatial_lengths_[1]; ++x)
                         {
-                            ck::index_t wi = wo * arg.window_strides_[1] + x - arg.in_left_pads_[1];
+                            ck::index_t wi = wo * arg.window_strides_[1] +
+                                             x * arg.window_dilations_[1] - arg.in_left_pads_[1];
                             if(hi >= 0 &&
                                hi < static_cast<ck::index_t>(arg.in_.mDesc.GetLengths()[2]) &&
                                wi >= 0 &&