Merge branch 'develop' into ck_tile/fmha_receipt_aiter

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_cshuffle_v3.hpp

@@ -138,6 +138,7 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
             if(stream_config.log_level_ > 0)
             {
                 arg.Print();
+                GridwiseGemm::BlockwiseGemmPipe::HotLoopInstList::Print();
             }
 
             if(!GridwiseGemm::CheckValidity(arg))
@@ -733,7 +734,9 @@ struct DeviceGemm_Xdl_CShuffleV3 : public DeviceGemmV2<ALayout,
             << "BlkGemmPipelineVersion: "
             << BlkGemmPipelineVersionToString[BlkGemmPipelineVer] << ", "
             << "BlkGemmPipelinePrefetchStages: "
-            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages;
+            << GridwiseGemm::BlockwiseGemmPipe::PrefetchStages << ", "
+            << "Kpack: "
+            << GridwiseGemm::BlockwiseGemmPipe::AMmaKStride;
         // clang-format on
 
         return str.str();

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle.hpp

@@ -205,7 +205,7 @@ struct DeviceGemmXdlSplitKCShuffle : public DeviceGemmSplitK<ALayout,
 
             const auto b2c_map = DefaultBlock2CTileMap{};
             index_t gdx, gdy, gdz;
-            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
+            ck::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
             const auto K0Padded    = karg.K0Padded;
 
             const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0Padded);

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_splitk_c_shuffle_lds_direct_load.hpp

@@ -183,7 +183,7 @@ struct DeviceGemmXdlSplitKCShuffle_LdsDirectLoad : public DeviceGemmSplitK<ALayo
 
             const auto b2c_map = DefaultBlock2CTileMap{};
             index_t gdx, gdy, gdz;
-            std::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
+            ck::tie(gdx, gdy, gdz) = b2c_map.CalculateGridSize(karg.M, karg.N, karg.k_batch);
             const auto K0Padded    = karg.K0Padded;
 
             const bool has_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0Padded);

include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl_waveletmodel_cshuffle.hpp

@@ -47,8 +47,7 @@ __global__ void
                 e_grid_desc_mblock_mperblock_nblock_nperblock,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,

include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp

@@ -37,8 +37,7 @@ __global__ void
             const BElementwiseOperation b_element_op,
             const CDEElementwiseOperation cde_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -87,8 +87,7 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const ComputePtrOffsetOfN compute_ptr_offset_of_n)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     // offset base pointer for each work-group
     const index_t n_idx = __builtin_amdgcn_readfirstlane(blockIdx.z);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y);

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_multiple_d_xdl_cshuffle.hpp

@@ -60,8 +60,7 @@ __global__ void
             const Block2CTileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
@@ -103,7 +102,7 @@ __global__ void
     compute_ptr_offset_of_batch.GetAPtrOffset(0);
     compute_ptr_offset_of_batch.GetBPtrOffset(0);
     compute_ptr_offset_of_batch.GetCPtrOffset(0);
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 template <index_t NDimSpatial,

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_two_stage_xdl_cshuffle.hpp

@@ -55,8 +55,7 @@ __global__ void
             [[maybe_unused]] const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             [[maybe_unused]] const index_t num_k_per_block)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.z * NumGroupsToMerge);
     const index_t k_idx = __builtin_amdgcn_readfirstlane(blockIdx.y * num_k_per_block);
 
@@ -85,7 +84,7 @@ __global__ void
                                         k_idx);
 #else
     ignore = karg;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 template <typename GridwiseGemm,
@@ -145,7 +144,7 @@ __global__ void
                                              k_idx);
 #else
     ignore = karg;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 template <ck::index_t NDimSpatial,
@@ -1496,10 +1495,13 @@ struct DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle
         // if workspace is not allocated
         if(!arg.p_workspace_)
         {
-            std::cerr << "Warning: Workspace for "
+            if(ck::EnvIsEnabled(CK_ENV(CK_LOGGING)))
+            {
+                std::cout << "Warning: Workspace for "
                              "DeviceGroupedConvBwdWeightTwoStage_Xdl_CShuffle::Argument is not "
                              "allocated, use SetWorkSpacePointer."
                           << std::endl;
+            }
             return false;
         }
         if(!ck::is_xdl_supported())

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp

@@ -9,6 +9,7 @@
 #include <numeric>
 #include <sstream>
 
+#include "ck/library/utility/numeric.hpp"
 #include "ck/utility/common_header.hpp"
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
@@ -98,8 +99,7 @@ __global__ void
             const ComputePtrOffsetOfG compute_ptr_offset_of_groups,
             const ComputePtrOffsetOfN compute_ptr_offset_of_n)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
 
     // offset base pointer for each work-group
     const index_t g_idx = __builtin_amdgcn_readfirstlane(blockIdx.y);
@@ -212,9 +212,13 @@ __global__ void
 }
 
 } // namespace
-
+#ifdef CK_CODE_GEN_RTC
+template <typename T>
+using is_tuple = decltype(ck::declval<T&>().IsTuple());
+#else
 template <typename T>
 using is_tuple = decltype(std::declval<T&>().IsTuple());
+#endif
 
 //
 // @brief      Device Convolution operation.

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle_v3.hpp

@@ -9,6 +9,7 @@
 #include <numeric>
 #include <sstream>
 
+#include "ck/library/utility/numeric.hpp"
 #include "ck/utility/common_header.hpp"
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
@@ -117,7 +118,7 @@ __global__ void
                                         c_grid_desc_mblock_mperblock_nblock_nperblock);
 #else
     ignore = karg;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 template <typename GridwiseGemm,
@@ -183,7 +184,7 @@ __global__ void
                                              c_grid_desc_mblock_mperblock_nblock_nperblock);
 #else
     ignore = karg;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 } // namespace

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp

@@ -155,8 +155,7 @@ __global__ void
             const Block2ETileMap block_2_ctile_map,
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
     const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_large_tensor_cshuffle.hpp

@@ -52,8 +52,7 @@ __global__ void
             const ComputePtrOffset compute_ptr_offset_of_groups,
             const ComputePtrOffset compute_ptr_offset_of_n)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id_x = __builtin_amdgcn_readfirstlane(blockIdx.x);

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_xdl_cshuffle_tile_loop.hpp

@@ -68,8 +68,7 @@ __global__ void
                                            const BElementwiseOperation b_element_op,
                                            const CDEElementwiseOperation cde_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
 
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
     __shared__ uint8_t p_shared[shared_size];
@@ -404,7 +403,7 @@ __global__ void
     ignore = a_element_op;
     ignore = b_element_op;
     ignore = cde_element_op;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 template <typename ALayout,

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp

@@ -43,8 +43,7 @@ __global__ void
             const B1ElementwiseOperation b1_element_op,
             const CElementwiseOperation c_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();
@@ -109,7 +108,7 @@ __global__ void
     ignore = acc_element_op;
     ignore = b1_element_op;
     ignore = c_element_op;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 // Computes C = A * B0 * B1

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp

@@ -38,8 +38,7 @@ __global__ void
                                 const BElementwiseOperation b_element_op,
                                 const CDEElementwiseOperation c_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_fixed_nk.hpp

@@ -50,8 +50,7 @@ __global__ void
                                          const BElementwiseOperation b_element_op,
                                          const CDEElementwiseOperation c_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();

include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp

@@ -40,8 +40,7 @@ __global__ void
                                        const BElementwiseOperation b_element_op,
                                        const CElementwiseOperation c_element_op)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
     __shared__ uint8_t p_shared[shared_size];
 
@@ -80,7 +79,7 @@ __global__ void
     ignore = a_element_op;
     ignore = b_element_op;
     ignore = c_element_op;
-#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
+#endif // end of if (defined(__gfx9__))
 }
 
 template <typename ALayout,

include/ck/tensor_operation/gpu/device/impl/device_image_to_column_impl.hpp

@@ -3,6 +3,7 @@
 
 #pragma once
 
+#include "ck/library/utility/numeric.hpp"
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
 #include "ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp"

include/ck/tensor_operation/gpu/device/impl/device_splitk_contraction_multiple_d_xdl_cshuffle.hpp

@@ -56,8 +56,7 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
             const Block2ETileMap block_2_etile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t num_blocks_per_batch =

include/ck/tensor_operation/gpu/device/tensor_layout.hpp

@@ -430,6 +430,7 @@ struct G_NDHW : public BaseTensorLayout
 
 } // namespace convolution
 
+#ifndef CK_CODE_GEN_RTC
 template <
     typename Layout,
     typename std::enable_if<std::is_base_of<BaseTensorLayout, Layout>::value, bool>::type = false>
@@ -438,6 +439,7 @@ std::ostream& operator<<(std::ostream& os, const Layout&)
     os << Layout::name;
     return os;
 }
+#endif
 
 } // namespace tensor_layout
 } // namespace ck