Merge branch 'develop' into lwpck-1815

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

@@ -133,8 +133,13 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
     static constexpr auto NWaves = NPerBlock / (NRepeat * NPerWmma);
     static constexpr auto WmmaK  = K1 == 16 ? 32 : 16;
 
-    static constexpr auto AEnableLds_auto = NWaves == 1 ? false : true;
-    static constexpr auto BEnableLds_auto = MWaves == 1 ? false : true;
+    static constexpr auto MaxVectorLoadA = K1 * sizeof(ADataType) == 16 ? true : false;
+    static constexpr auto MaxVectorLoadB = K1 * sizeof(BDataType) == 16 ? true : false;
+
+    static constexpr auto AEnableLds_auto =
+        (NWaves == 1 && (MaxVectorLoadA || MRepeat == 1)) ? false : true;
+    static constexpr auto BEnableLds_auto =
+        (MWaves == 1 && (MaxVectorLoadB || NRepeat == 1)) ? false : true;
 
     // If true, LDS is used unconditionally
     static constexpr auto AEnableLds_manu = false;
@@ -829,7 +834,7 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {
@@ -869,13 +874,17 @@ struct DeviceBatchedContractionMultipleD_Wmma_CShuffle
         }
         else
         {
-            if(!(arg.a_kz_stride_ == 1 &&
-                 arg.a_grid_desc_.GetLength(I2) % ABlockTransferSrcScalarPerVector == 0))
+            if(!(arg.a_kz_stride_ == 1))
+            {
+                index_t LastK =
+                    AEnableLds ? arg.a_grid_desc_.GetLength(I2) : arg.a_grid_desc_.GetLength(I6);
+                if(LastK % ABlockTransferSrcScalarPerVector == 0)
                 {
                     printf("DeviceOp: Vector Access A-k check failure\n");
                     return false;
                 }
             }
+        }
 
         // vector memory access of B: could be on N or BK1 dimension
         if constexpr(BBlockTransferSrcVectorDim == 1)

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

@@ -71,7 +71,8 @@ __global__ void
             const Block2CTileMap block_2_ctile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||         \
-    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__))
+    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
+    defined(__gfx12__))
 
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -648,7 +649,7 @@ struct DeviceBatchedGemmMultipleD_Dl : public DeviceBatchedGemmMultiD<ALayout,
     static bool IsSupportedArgument(const Argument& arg)
     {
         if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
-           ck::is_gfx103_supported() || ck::is_gfx11_supported())
+           ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             bool pass = true;
             pass      = pass && arg.K_ % K1 == 0;

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

@@ -56,7 +56,7 @@ __global__ void
                                                        bool input_permute,
                                                        bool output_permute)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
 
     // clang-format off
 // ***************************************************
@@ -159,6 +159,7 @@ __global__ void
     ignore = O;
     ignore = G0;
     ignore = G1;
+    ignore = alpha;
     ignore = input_permute;
     ignore = output_permute;
 #endif // end of if (defined(__gfx11__))
@@ -187,7 +188,7 @@ __global__ void
                                            index_t head_size,
                                            float alpha)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
 
     // clang-format off
 // ***************************************************
@@ -321,7 +322,7 @@ __global__ void
                                             index_t head_size,
                                             float alpha)
 {
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__))
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx11__) || defined(__gfx12__))
 
     // clang-format off
 // ***************************************************
@@ -858,7 +859,7 @@ struct DeviceBatchedGemmSoftmaxGemmPermute_Wmma_CShuffle
 
     static bool IsSupportedArgument(const RawArg& arg)
     {
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<Acc0DataType, float> || is_same_v<Acc0DataType, int32_t>))
             {

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

@@ -592,9 +592,7 @@ struct DeviceContractionMultipleD_Xdl_CShuffle
             return false;
         }
 
-        if(ck::get_device_name() != "gfx90a" && ck::get_device_name() != "gfx940" &&
-           ck::get_device_name() != "gfx941" && ck::get_device_name() != "gfx942" &&
-           std::is_same<ADataType, double>::value)
+        if(!ck::is_lds_direct_load_supported() && std::is_same<ADataType, double>::value)
         {
             return false;
         }

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

@@ -1393,7 +1393,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Dl
     {
         // check device
         if(!(ck::get_device_name() == "gfx906" || ck::is_gfx103_supported() ||
-             ck::is_gfx11_supported()))
+             ck::is_gfx11_supported() || ck::is_gfx12_supported()))
         {
             return false;
         }

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

@@ -509,7 +509,7 @@ struct DeviceFpAintBGemm_Wmma_CShuffle : public DeviceGemm_dequantB<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, ck::half_t> ||
                            is_same_v<AccDataType, int32_t>))

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

@@ -536,7 +536,7 @@ struct DeviceGemmDl : public DeviceGemm<ALayout,
         }
 
         if(ck::get_device_name() == "gfx906" || ck::is_gfx103_supported() ||
-           ck::is_gfx11_supported())
+           ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             return GridwiseGemm::CheckValidity(
                 arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);

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

@@ -51,7 +51,8 @@ __global__ void
             const Block2CTileMap block_2_ctile_map)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||         \
-    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__))
+    defined(__gfx90a__) || defined(__gfx94__) || defined(__gfx103__) || defined(__gfx11__) || \
+    defined(__gfx12__))
 
     constexpr index_t shared_block_size =
         GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType);
@@ -552,7 +553,7 @@ struct DeviceGemmMultipleD_Dl : public DeviceGemmMultipleD<ALayout,
     static bool IsSupportedArgument(const Argument& arg)
     {
         if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
-           ck::is_gfx103_supported() || ck::is_gfx11_supported())
+           ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             return GridwiseGemm::CheckValidity(
                 arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.e_grid_desc_m_n_);

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

@@ -515,7 +515,7 @@ struct DeviceGemmMultipleD_Wmma_CShuffle : public DeviceGemmMultipleD<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -87,11 +87,18 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
     static constexpr auto MWaves         = MPerBlock / (MRepeat * MPerWmma);
     static constexpr auto NWaves         = NPerBlock / (NRepeat * NPerWmma);
     static constexpr auto WmmaK          = K1 == 16 ? 32 : 16;
+    static constexpr auto MaxVectorLoadA = K1 * sizeof(ADataType) == 16 ? true : false;
+    static constexpr auto MaxVectorLoadB = K1 * sizeof(BDataType) == 16 ? true : false;
 
-    static constexpr auto AEnableLds_auto =
-        (NWaves == 1 && is_same<tensor_layout::gemm::RowMajor, ALayout>::value) ? false : true;
+    static constexpr auto AEnableLds_auto = (NWaves == 1 && (MaxVectorLoadA || MRepeat == 1) &&
+                                             is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
+                                                ? false
+                                                : true;
     static constexpr auto BEnableLds_auto =
-        (MWaves == 1 && is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value) ? false : true;
+        (MWaves == 1 && (MaxVectorLoadB || NRepeat == 1) &&
+         is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
+            ? false
+            : true;
 
     // If true, LDS is used unconditionally
     static constexpr auto AEnableLds_manu = false;
@@ -443,7 +450,7 @@ struct DeviceGemmWmma_CShuffle : public DeviceGemm<ALayout,
 
     static bool IsSupportedArgument(const Argument& arg)
     {
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, ck::half_t> ||
                            is_same_v<AccDataType, int32_t>))

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

@@ -629,7 +629,7 @@ struct DeviceGroupedConvBwdDataMultipleD_Wmma_CShuffle
     static bool IsSupportedArgument(const Argument& arg)
     {
         // check device
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -49,7 +49,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) ||         \
-    defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__))
+    defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__) || \
+    defined(__gfx12__))
     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_bwd_weight_two_stage_xdl_cshuffle.hpp

@@ -47,12 +47,12 @@ __global__ void
 #endif
         kernel_grouped_conv_bwd_weight_xdl_cshuffle_v3(
             typename GridwiseGemm::Argument karg,
-            const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
-            const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
-            const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            [[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
+            [[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
+            [[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
                 c_grid_desc_mblock_mperblock_nblock_nperblock,
-            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
-            const index_t num_k_per_block)
+            [[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__))
@@ -103,12 +103,12 @@ __global__ void
 #endif
         kernel_grouped_conv_bwd_weight_xdl_cshuffle_v3_2lds(
             typename GridwiseGemm::Argument karg,
-            const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
-            const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
-            const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            [[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
+            [[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
+            [[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
                 c_grid_desc_mblock_mperblock_nblock_nperblock,
-            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch,
-            const index_t num_k_per_block)
+            [[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(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))

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

@@ -692,7 +692,7 @@ struct DeviceGroupedConvBwdWeight_Wmma_CShuffle
     static bool IsSupportedArgument(const Argument& arg)
     {
         // check device
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -91,7 +91,8 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) ||         \
-    defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__))
+    defined(__gfx90a__) || defined(__gfx908__) || defined(__gfx94__) || defined(__gfx11__) || \
+    defined(__gfx12__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -667,7 +668,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
 
         // check device
         if(!(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
-             ck::is_gfx103_supported() || ck::is_gfx11_supported()))
+             ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported()))
         {
             return false;
         }

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

@@ -107,7 +107,7 @@ __global__ void
             const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx103__) || \
-    defined(__gfx11__))
+    defined(__gfx11__) || defined(__gfx12__))
     // offset base pointer for each work-group
     const index_t num_blocks_per_batch =
         __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
@@ -603,7 +603,7 @@ struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd<NDimS
 
         // check device
         if(!(ck::get_device_name() == "gfx906" || ck::is_gfx103_supported() ||
-             ck::is_gfx11_supported()))
+             ck::is_gfx11_supported() || ck::is_gfx12_supported()))
         {
             return false;
         }

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

@@ -69,14 +69,15 @@ __global__ void
 #if CK_USE_LAUNCH_BOUNDS
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, MinimumOccupancy)
 #endif
-        kernel_grouped_conv_fwd_xdl_cshuffle_v3(typename GridwiseGemm::Argument karg,
-                                                const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
-                                                const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
-                                                const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+        kernel_grouped_conv_fwd_xdl_cshuffle_v3(
+            typename GridwiseGemm::Argument karg,
+            [[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
+            [[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
+            [[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
                 c_grid_desc_mblock_mperblock_nblock_nperblock,
-                                                const ComputePtrOffset compute_ptr_offset_of_groups,
-                                                const ComputePtrOffset compute_ptr_offset_of_n,
-                                                const index_t groups_count)
+            [[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_groups,
+            [[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_n,
+            [[maybe_unused]] const index_t groups_count)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     // offset base pointer for each work-group
@@ -132,13 +133,13 @@ __global__ void
 #endif
         kernel_grouped_conv_fwd_xdl_cshuffle_v3_2lds(
             typename GridwiseGemm::Argument karg,
-            const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
-            const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
-            const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            [[maybe_unused]] const AGridDesc_AK0_M_K1 a_grid_desc_ak0_m_ak1,
+            [[maybe_unused]] const BGridDesc_BK0_N_K1 b_grid_desc_bk0_n_bk1,
+            [[maybe_unused]] const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
                 c_grid_desc_mblock_mperblock_nblock_nperblock,
-            const ComputePtrOffset compute_ptr_offset_of_groups,
-            const ComputePtrOffset compute_ptr_offset_of_n,
-            const index_t groups_count)
+            [[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_groups,
+            [[maybe_unused]] const ComputePtrOffset compute_ptr_offset_of_n,
+            [[maybe_unused]] const index_t groups_count)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx9__))
     // offset base pointer for each work-group

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

@@ -582,7 +582,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
         namespace ctc = tensor_layout::convolution;
 
         // check device
-        if(ck::is_gfx11_supported())
+        if(ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, int32_t>))
             {

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

@@ -40,7 +40,8 @@ __global__ void
                                           const CDEElementwiseOperation cde_element_op)
 {
 #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||         \
-    defined(__gfx90a__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx94__))
+    defined(__gfx90a__) || defined(__gfx103__) || defined(__gfx11__) || defined(__gfx94__) || \
+    defined(__gfx12__))
     __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
 
     const index_t block_id = get_block_1d_id();
@@ -673,7 +674,7 @@ struct DeviceGroupedGemmMultipleD_Dl : public DeviceGroupedGemm<ALayout,
         }
 
         if(ck::get_device_name() == "gfx906" || ck::is_xdl_supported() ||
-           ck::is_gfx103_supported() || ck::is_gfx11_supported())
+           ck::is_gfx103_supported() || ck::is_gfx11_supported() || ck::is_gfx12_supported())
         {
             for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
             {