Merge pull request #295 from ROCm/gfx950

Merge gfx950 changes into develop branch.

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

@@ -50,9 +50,8 @@ __global__ void
             const CGridDesc_M0_M10_M11_N0_N10_N11 e_grid_desc_m0_m10_m11_n0_n10_n11,
             const Block2CTileMap block_2_ctile_map)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||         \
-    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
-    defined(__gfx12__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx9__) || \
+    defined(__gfx103__) || defined(__gfx11__) || defined(__gfx12__))
 
     constexpr index_t shared_block_size =
         GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);

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

@@ -63,8 +63,7 @@ __global__ void
             const Block2ETileMap block_2_etile_map,
             index_t NRaw)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx94__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     __shared__ char p_shared[GridwiseGemmWelford::GetSharedMemoryNumberOfByte()];
 
     GridwiseGemmWelford::template Run<HasMainKBlockLoop>(

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

@@ -60,8 +60,7 @@ __global__ void
             const RsGridDescriptor_MBlock_MPerBlock rs_grid_desc_mblock_mperblock,
             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_gemm_multiple_d_xdl_cshuffle.hpp

@@ -52,8 +52,7 @@ __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_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,

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

@@ -98,8 +98,7 @@ __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);

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

@@ -117,7 +117,7 @@ __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 +183,7 @@ __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_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/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp

@@ -607,6 +607,7 @@ struct GridwiseBatchedGemmGemm_Xdl_CShuffle
         // with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
         // cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
         // therefore we may just as well assign Gemm1KPack = group_size
+
         constexpr index_t Gemm1KPack =
             MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;