Merge branch 'develop' into envloggin

example/01_gemm/gemm_xdl_streamk.cpp

@@ -27,11 +27,15 @@ using DeviceGemmStreamK = ck::tensor_operation::device::DeviceGemmXdlStreamK
 // ######|      Type|      Type|      Type|        Type|        |        |        | Elementwise| Elementwise| Elementwise|   Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
 // ######|          |          |          |            |        |        |        |   Operation|   Operation|   Operation|       |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
 // ######|          |          |          |            |        |        |        |            |            |            |       |      |      |      |   |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
+#if defined(CK_USE_AMD_MFMA_GFX950)
+    < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    256,   256,   128,     4, 16,   32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,      S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
+#else // defined(CK_USE_AMD_MFMA_GFX950)
     < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    256,   128,   128,     4,  8,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,      S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 32, 1, 8>,               8>;
 
           // < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    256,   256,   128,     4,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,         1,           1,           1,               S<1, 32, 1, 8>,              8>;
       //  < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    128,   32,   64,     4,   8,  32,   32,    1,    1,     S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<4, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              8,         1,           1,           1,               S<1, 16, 1, 8>,              8>;
        // < ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout,  AElementOp,  BElementOp,  CElementOp,    128,   32,   128,     4,   8,  32,   32,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,     S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,         1,           1,           1,               S<1, 32, 1, 4>,              8>;
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
 
 
 

example/20_grouped_conv_bwd_weight/CMakeLists.txt

@@ -13,3 +13,9 @@ add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bw
 
 add_example_executable(example_grouped_conv_bwd_weight_dl_fp16 grouped_conv_bwd_weight_dl_fp16.cpp)
 add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_dl_fp16)
+
+add_example_executable(example_grouped_conv_bwd_weight_v3_xdl_bf16 grouped_conv_bwd_weight_v3_xdl_bf16.cpp)
+add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_v3_xdl_bf16)
+
+add_example_executable(example_grouped_conv_bwd_weight_v3_xdl_fp16 grouped_conv_bwd_weight_v3_xdl_fp16.cpp)
+add_example_dependencies(example_grouped_conv_bwd_weight example_grouped_conv_bwd_weight_v3_xdl_fp16)

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_v3_xdl_bf16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle_v3.hpp"
+
+using InDataType = BF16;
+// bf16 kernel use fp32 atomic add to accumulate Weight tensor into global memory
+using WeiDataType = F32;
+using OutDataType = BF16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    // clang-format on
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        32,                   // K0PerBlock
+        8,                    // K1
+        32,                   // MPerXdl
+        32,                   // NPerXdl
+        2,                    // MXdlPerWave
+        2,                    // NXdlPerWave
+        S<4, 16, 1>,          // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<2, 0, 1>,           // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,           // ABlockTransferSrcAccessOrder
+        2,                    // ABlockTransferSrcVectorDim
+        1,                    // ABlockTransferSrcScalarPerVector
+        2,                    // ABlockTransferDstScalarPerVector_K1
+        true,                 // ABlockLdsAddExtraM
+        S<4, 16, 1>,          // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<2, 0, 1>,           // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,           // BBlockTransferSrcAccessOrder
+        2,                    // BBlockTransferSrcVectorDim
+        1,                    // BBlockTransferSrcScalarPerVector
+        2,                    // BBlockTransferDstScalarPerVector_K1
+        true,                 // BBlockLdsAddExtraN
+        1,                    // CShuffleMXdlPerWavePerShuffle
+        1,                    // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
+                                                 // clang-format off
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/20_grouped_conv_bwd_weight/grouped_conv_bwd_weight_v3_xdl_fp16.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "common.hpp"
+
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle_v3.hpp"
+
+using InDataType  = F16;
+using WeiDataType = F16;
+using OutDataType = F16;
+using AccDataType = F32;
+
+using InElementOp  = PassThrough;
+using WeiElementOp = PassThrough;
+using OutElementOp = PassThrough;
+
+template <ck::index_t NDimSpatial>
+using DeviceConvBwdWeightInstance =
+    ck::tensor_operation::device::DeviceGroupedConvBwdWeight_Xdl_CShuffleV3<
+        NDimSpatial,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWC,
+                                      ck::tensor_layout::convolution::GNHWC,
+                                      ck::tensor_layout::convolution::GNDHWC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GKXC,
+                                      ck::tensor_layout::convolution::GKYXC,
+                                      ck::tensor_layout::convolution::GKZYXC>>,
+        ck::tuple_element_t<NDimSpatial - 1,
+                            ck::Tuple<ck::tensor_layout::convolution::GNWK,
+                                      ck::tensor_layout::convolution::GNHWK,
+                                      ck::tensor_layout::convolution::GNDHWK>>,
+        InDataType,           // InDataType
+        WeiDataType,          // WeiDataType
+        OutDataType,          // OutDataType
+        AccDataType,          // AccDataType
+        InElementOp,          // InElementwiseOperation
+        WeiElementOp,         // WeiElementwiseOperation
+        OutElementOp,         // OutElementwiseOperation
+        ConvBwdWeightDefault, // ConvolutionBackwardWeightSpecialization
+        256,                  // BlockSize
+        128,                  // MPerBlock
+        128,                  // NPerBlock
+        32,                   // K0PerBlock
+        8,                    // K1
+        32,                   // MPerXdl
+        32,                   // NPerXdl
+        2,                    // MXdlPerWave
+        2,                    // NXdlPerWave
+        S<4, 16, 1>,          // ABlockTransferThreadClusterLengths_K0_M_K1
+        S<2, 0, 1>,           // ABlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,           // ABlockTransferSrcAccessOrder
+        2,                    // ABlockTransferSrcVectorDim
+        1,                    // ABlockTransferSrcScalarPerVector
+        2,                    // ABlockTransferDstScalarPerVector_K1
+        false,                // ABlockLdsAddExtraM
+        S<4, 16, 1>,          // BBlockTransferThreadClusterLengths_K0_N_K1
+        S<2, 0, 1>,           // BBlockTransferThreadClusterArrangeOrder
+        S<1, 0, 2>,           // BBlockTransferSrcAccessOrder
+        2,                    // BBlockTransferSrcVectorDim
+        1,                    // BBlockTransferSrcScalarPerVector
+        2,                    // BBlockTransferDstScalarPerVector_K1
+        false,                // BBlockLdsAddExtraN
+        1,                    // CShuffleMXdlPerWavePerShuffle
+        1,                    // CShuffleNXdlPerWavePerShuffle
+        S<1, 32, 1, 4>,       // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
+        128 / (sizeof(WeiDataType) * CHAR_BIT)>; // CBlockTransferScalarPerVector_NWaveNPerXdl
+
+template <ck::index_t NDimSpatial>
+using HostConvBwdWeightInstance = ck::tensor_operation::host::ReferenceConvBwdWeight<NDimSpatial,
+                                                                                     InDataType,
+                                                                                     WeiDataType,
+                                                                                     OutDataType,
+                                                                                     InElementOp,
+                                                                                     WeiElementOp,
+                                                                                     OutElementOp>;
+
+#include "run_grouped_conv_bwd_weight_example.inc"
+
+int main(int argc, char* argv[])
+{
+    ExecutionConfig config;
+    ck::utils::conv::ConvParam conv_param = DefaultConvParam;
+
+    if(!parse_cmd_args(argc, argv, config, conv_param))
+    {
+        return 1;
+    }
+
+    switch(conv_param.num_dim_spatial_)
+    {
+    case 1: return !run_grouped_conv_bwd_weight<1>(config, conv_param);
+    case 2: return !run_grouped_conv_bwd_weight<2>(config, conv_param);
+    case 3: return !run_grouped_conv_bwd_weight<3>(config, conv_param);
+    default: break;
+    }
+
+    return 1;
+}

example/ck_tile/03_gemm/run_gemm_example.inc

@@ -107,10 +107,10 @@ int run_gemm_example_with_layouts(int argc,
     ck_tile::index_t stride_B = arg_parser.get_int("stride_b");
     ck_tile::index_t stride_C = arg_parser.get_int("stride_c");
 
-    ck_tile::index_t kbatch = arg_parser.get_int("split_k");
-    int n_warmup            = arg_parser.get_int("warmup");
-    int n_repeat            = arg_parser.get_int("repeat");
-    ck_tile::index_t init_method = arg_parser.get_int("init");  
+    ck_tile::index_t kbatch      = arg_parser.get_int("split_k");
+    int n_warmup                 = arg_parser.get_int("warmup");
+    int n_repeat                 = arg_parser.get_int("repeat");
+    ck_tile::index_t init_method = arg_parser.get_int("init");
 
     stride_A = ck_tile::get_default_stride(M, K, stride_A, is_row_major(a_layout));
     stride_B = ck_tile::get_default_stride(K, N, stride_B, is_row_major(b_layout));
@@ -123,16 +123,23 @@ int run_gemm_example_with_layouts(int argc,
     ck_tile::HostTensor<CDataType> c_m_n_dev_result(
         ck_tile::host_tensor_descriptor(M, N, stride_C, is_row_major(CLayout{})));
 
-    if (init_method == 0) {
+    if(init_method == 0)
+    {
         ck_tile::FillUniformDistribution<ADataType>{-1.f, 1.f}(a_m_k);
         ck_tile::FillUniformDistribution<BDataType>{-1.f, 1.f}(b_k_n);
-    } else if (init_method == 1) {
+    }
+    else if(init_method == 1)
+    {
         ck_tile::FillMonotonicSeq<ADataType>{}(a_m_k);
         ck_tile::FillMonotonicSeq<BDataType>{}(b_k_n);
-    } else if (init_method == 2) {
+    }
+    else if(init_method == 2)
+    {
         ck_tile::FillConstant<ADataType>{static_cast<ADataType>(1)}(a_m_k);
         ck_tile::FillConstant<BDataType>{static_cast<BDataType>(1)}(b_k_n);
-    } else {
+    }
+    else
+    {
         a_m_k.SetZero();
         b_k_n.SetZero();
     }

example/ck_tile/03_gemm/universal_gemm.cpp

@@ -255,20 +255,11 @@ float gemm_calc(const ck_tile::GemmHostArgs& args, const ck_tile::stream_config&
     }
     else
     {
-        // Tail number always Full - #PrefetchStages
-        if(tail_num == ck_tile::TailNumber::Full)
-        {
-            Run(ck_tile::bool_constant<false>{},
-                ck_tile::integral_constant<ck_tile::TailNumber, ck_tile::TailNumber::Full>{});
-        }
-        else
-        {
-            std::ostringstream err;
-            err << "When there's no hot loop, this tail number \"" << tail_num
-                << "\" is not supported! PrefetchStages: " << BaseGemmPipeline::PrefetchStages
-                << "\n File: " << __FILE__ << ":" << __LINE__ << ", in function: " << __func__;
-            throw std::runtime_error(err.str());
-        }
+        std::ostringstream err;
+        err << "Num K loop must be larger than number of prefetech stages."
+            << "\n PrefetchStages: " << BaseGemmPipeline::PrefetchStages << "\n File: " << __FILE__
+            << ":" << __LINE__ << ", in function: " << __func__;
+        throw std::runtime_error(err.str());
     }
 
     return ave_time;

include/ck/tensor_operation/operator_transform/transform_conv_bwd_weight_to_gemm_v2.hpp

 // SPDX-License-Identifier: MIT
-// Copyright (c) 2024, Advanced Micro Devices, Inc. All rights reserved.
+// Copyright (c) 2024-2025, Advanced Micro Devices, Inc. All rights reserved.
 
 #pragma once
 
@@ -34,6 +34,94 @@ struct TransformConvBwdWeightToGemmV2
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
 
+    template <index_t NDim, typename enable_if<NDim == 1, bool>::type = false>
+    constexpr static auto
+    make_out_grid_desc(const index_t N,
+                       const index_t Wo,
+                       const index_t K,
+                       const std::array<index_t, NDimSpatial + 3>& output_strides)
+    {
+        const index_t BatchStride = output_strides[0];
+        const index_t WoStride    = output_strides[3];
+        const auto KStride        = Number<1>{};
+        return make_naive_tensor_descriptor(make_tuple(N * Wo, NumGroupsToMerge, K),
+                                            make_tuple(WoStride, BatchStride, KStride));
+    }
+
+    template <index_t NDim, typename enable_if<NDim == 1, bool>::type = false>
+    constexpr static auto
+    make_in_grid_desc(const index_t N,
+                      const index_t Wi,
+                      const index_t C,
+                      const std::array<index_t, NDimSpatial + 3>& input_strides)
+    {
+        const index_t BatchStride = input_strides[0];
+        const index_t NStride     = input_strides[1];
+        const index_t WiStride    = input_strides[3];
+        const auto CStride        = input_strides[2];
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor(make_tuple(N * Wi, NumGroupsToMerge, C),
+                                                make_tuple(WiStride, BatchStride, CStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(N, Wi, NumGroupsToMerge, C),
+                make_tuple(NStride, WiStride, BatchStride, CStride));
+        }
+    }
+
+    template <index_t NDim, typename enable_if<NDim == 1, bool>::type = false>
+    constexpr static auto
+    make_wei_grid_desc(const index_t K,
+                       const index_t X,
+                       const index_t C,
+                       const std::array<index_t, NDimSpatial + 3>& weights_strides)
+    {
+        const auto CStride     = Number<1>{};
+        const auto KStride     = weights_strides[1];
+        const auto XStride     = weights_strides[3];
+        const auto BatchStride = weights_strides[0];
+        // Add NumGroupsToMerge for Batch+M dimension and, 1 as a placehorder
+        // for Batch+N dimension
+        const auto desc = make_naive_tensor_descriptor(
+            make_tuple(NumGroupsToMerge, K, X, 1, C),
+            make_tuple(BatchStride, KStride, XStride, BatchStride, CStride));
+        // Padd 1 to NumGroupsToMerge
+        const auto padded_desc = transform_tensor_descriptor(
+            desc,
+            make_tuple(make_pass_through_transform(NumGroupsToMerge),
+                       make_pass_through_transform(K),
+                       make_pass_through_transform(X),
+                       make_pad_transform(1, 0, NumGroupsToMerge - 1),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+        // We need only matrices from diagonal. Xor returns 0 for the same
+        // values. So if matrices is not on diagonal then it will be stored in padding.
+        // To avoid use of modulo after xor we assume that NumBatch to merge is power of 2.
+        static_assert(NumGroupsToMerge == 1 || NumGroupsToMerge == 2 || NumGroupsToMerge == 4 ||
+                      NumGroupsToMerge == 8 || NumGroupsToMerge == 16 || NumGroupsToMerge == 32 ||
+                      NumGroupsToMerge == 64);
+        const auto unmerged_padded_desc = transform_tensor_descriptor(
+            padded_desc,
+            make_tuple(make_xor_transform(make_tuple(NumGroupsToMerge, NumGroupsToMerge)),
+                       make_pass_through_transform(K),
+                       make_pass_through_transform(X),
+                       make_pass_through_transform(C)),
+            make_tuple(Sequence<0, 3>{}, Sequence<1>{}, Sequence<2>{}, Sequence<4>{}),
+            make_tuple(Sequence<0, 3>{}, Sequence<1>{}, Sequence<2>{}, Sequence<4>{}));
+        // Merge To M, N
+        return transform_tensor_descriptor(
+            unmerged_padded_desc,
+            make_tuple(make_merge_transform(make_tuple(NumGroupsToMerge, K)),
+                       make_merge_transform(make_tuple(X, NumGroupsToMerge, C))),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3, 4>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+    }
+
     template <index_t NDim, typename enable_if<NDim == 2, bool>::type = false>
     constexpr static auto
     make_out_grid_desc(const index_t N,
@@ -221,6 +309,187 @@ struct TransformConvBwdWeightToGemmV2
             make_tuple(Sequence<0>{}, Sequence<1>{}));
     }
 
+    template <index_t NDim, typename enable_if<NDim == 1, bool>::type = false>
+    static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
+        const index_t N,
+        const index_t K,
+        const index_t C,
+        const std::array<index_t, NDimSpatial>& input_spatial_lengths,
+        const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
+        const std::array<index_t, NDimSpatial>& output_spatial_lengths,
+        const std::array<index_t, NDimSpatial + 3>& input_strides,
+        const std::array<index_t, NDimSpatial + 3>& weights_strides,
+        const std::array<index_t, NDimSpatial + 3>& output_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_strides,
+        const std::array<index_t, NDimSpatial>& conv_filter_dilations,
+        const std::array<index_t, NDimSpatial>& input_left_pads,
+        const std::array<index_t, NDimSpatial>& input_right_pads,
+        const index_t batch_k)
+    {
+        using namespace ck;
+
+        const index_t Wi = input_spatial_lengths[0];
+
+        const index_t Wo = output_spatial_lengths[0];
+
+        const index_t X = filter_spatial_lengths[0];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        const index_t ConvDilationW = conv_filter_dilations[0];
+
+        const index_t InLeftPadW = input_left_pads[0];
+
+        const index_t InRightPadW = input_right_pads[0];
+
+        const index_t GemmKTotal = N * Wo;
+        const index_t GemmM      = K * NumGroupsToMerge;
+        const index_t GemmN      = C * X * NumGroupsToMerge;
+
+        const auto PadGemmM = MPerBlock - GemmM % MPerBlock;
+        const auto PadGemmN = NPerBlock - GemmN % NPerBlock;
+
+        const index_t GemmKBatch = batch_k;
+        const index_t GemmK0 =
+            math::integer_divide_ceil(GemmKTotal, GemmK1Number * K0PerBlock * GemmKBatch) *
+            K0PerBlock;
+        const index_t GemmKPad = GemmKBatch * GemmK0 * GemmK1Number;
+
+        const auto out_grid_desc = make_out_grid_desc<NDim>(N, Wo, K, output_strides);
+        const auto in_grid_desc  = make_in_grid_desc<NDim>(N, Wi, C, input_strides);
+        const auto wei_grid_desc = make_wei_grid_desc<NDim>(K, X, C, weights_strides);
+
+        if constexpr(ConvBackwardWeightSpecialization ==
+                     device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0)
+        {
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(
+                    make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                    make_merge_transform(make_tuple(NumGroupsToMerge, GemmM / NumGroupsToMerge))),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch * GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: input tensor
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_grid_desc,
+                make_tuple(
+                    make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                    make_merge_transform(make_tuple(NumGroupsToMerge, GemmN / NumGroupsToMerge))),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch * GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                              wei_grid_desc);
+        }
+        else
+        {
+            // A: output tensor
+            const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+                out_grid_desc,
+                make_tuple(
+                    make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                    make_merge_transform(make_tuple(NumGroupsToMerge, GemmM / NumGroupsToMerge))),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+                out_gemmkpad_gemmm_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch * GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmM)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // B: input tensor
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(NumGroupsToMerge),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(NumGroupsToMerge),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmktotal_gemmn_grid_desc = transform_tensor_descriptor(
+                in_n_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(X, NumGroupsToMerge, C)),
+                           make_merge_transform(make_tuple(N, Wo))),
+                make_tuple(Sequence<1, 3, 4>{}, Sequence<0, 2>{}),
+                make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+            const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+                in_gemmktotal_gemmn_grid_desc,
+                make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+                in_gemmkpad_gemmn_grid_desc,
+                make_tuple(make_unmerge_transform(make_tuple(GemmKBatch * GemmK0, GemmK1Number)),
+                           make_pass_through_transform(GemmN)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+            // Padd
+            const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch * GemmK0),
+                               make_right_pad_transform(GemmM, PadGemmM),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc =
+                transform_tensor_descriptor(
+                    in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                    make_tuple(make_pass_through_transform(GemmKBatch * GemmK0),
+                               make_right_pad_transform(GemmN, PadGemmN),
+                               make_pass_through_transform(GemmK1Number)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto wei_gemmm_gemmn_pad_grid_desc =
+                transform_tensor_descriptor(wei_grid_desc,
+                                            make_tuple(make_right_pad_transform(GemmM, PadGemmM),
+                                                       make_right_pad_transform(GemmN, PadGemmN)),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_pad_grid_desc,
+                              in_gemmkbatch_gemmk0_gemmn_gemmk1_pad_grid_desc,
+                              wei_gemmm_gemmn_pad_grid_desc);
+        }
+
+    } // function end
+
     template <index_t NDim, typename enable_if<NDim == 2, bool>::type = false>
     static auto MakeABCGridDescriptor_A_K0_M_K1_B_K0_N_K1_C_M_N(
         const index_t N,

include/ck_tile/ops/fmha/block/block_masking.hpp

@@ -310,7 +310,7 @@ struct SimplifiedGenericAttentionMask
 
         const index_t x_per_split = ck_tile::max(1, integer_divide_ceil(x_total, num_splits));
         const index_t split_start = x_per_split * i_split;
-        const index_t split_end   = split_start + x_per_split;
+        const index_t split_end   = ck_tile::min(x_total, split_start + x_per_split);
 
         return ck_tile::make_tuple(ck_tile::max(origin_start, split_start),
                                    ck_tile::min(origin_end, split_end));

include/ck_tile/ops/fmha/kernel/fmha_fwd_splitkv_kernel.hpp

@@ -742,7 +742,7 @@ struct FmhaFwdSplitKVKernel
                 return pad_tensor_view(
                     v_dram_transposed,
                     make_tuple(number<FmhaPipeline::kN1>{}, number<FmhaPipeline::kK1>{}),
-                    sequence<kPadHeadDimV, false>{});
+                    sequence<kPadHeadDimV, kPadSeqLenK>{});
             }
             else
             {

include/ck_tile/ops/fmha/pipeline/block_fmha_fwd_splitkv_pipeline_nwarp_sshuffle_qr_ks_vs.hpp

@@ -343,6 +343,8 @@ struct BlockFmhaFwdSplitKVPipelineNWarpSShuffleQRKSVS
         // moving k_dram_window is an in-page-block operation, so there is
         // no need to invoke k_page_block_navigator.move_tile_window() here.
         move_tile_window(k_dram_window, {0, kK0});
+        // ensure LDS access by Q is done before the over-writting by K
+        block_sync_lds();
         store_tile(k_lds_window, tile_elementwise_in(k_element_func, k_block_tile));
 
         do

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_weight/device_grouped_conv_bwd_weight_v3_xdl_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2025, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_weight_xdl_cshuffle_v3.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using namespace ck::tensor_layout::convolution;
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+#ifdef CK_ENABLE_FP8
+using F8 = ck::f8_t;
+#endif
+
+#ifdef CK_ENABLE_BF8
+using BF8 = ck::bf8_t;
+#endif
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdWeightDefault =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Default;
+
+static constexpr auto ConvBwdWeightFilter1x1Stride1Pad0 =
+    ck::tensor_operation::device::ConvolutionBackwardWeightSpecialization::Filter1x1Stride1Pad0;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f32_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |
+        // generic instance
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F32,     F32,     F32,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    16,    16,     32,   8,   16,   16,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              1,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              1,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_f16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |
+        // generic instance
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              2,              2,      false,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     32,   8,   32,   32,    1,    2,  S<4, 8,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              4,              4,      false,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,   128,     32,   8,   32,   32,    1,    4,  S<4, 4,  1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              8,              8,      false,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     32,   8,   32,   32,    2,    1,  S<4, 16, 1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              4,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,   128,    32,     32,   8,   32,   32,    4,    1,  S<4, 16, 1>, S<2, 0, 1>,  S<1, 0, 2>,                   1,              8,              8,      false,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,       
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    80,     32,   8,   16,   16,    4,    5,  S<4, 16, 1>, S<2, 0, 1>,  S<2, 0, 1>,                   1,              4,              4,      false,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              5,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,    F16,     F16,     F16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,   112,     32,   8,   16,   16,    4,    7,  S<4, 16, 1>, S<2, 0, 1>,  S<2, 0, 1>,                   1,              4,              4,      false,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                1,              7,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>
+    // clang-format on
+    >;
+
+template <ck::index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename ELayout,
+          ConvolutionBackwardWeightSpecialization ConvSpec,
+          BlockGemmPipelineScheduler Scheduler,
+          BlockGemmPipelineVersion PipelineVersion>
+using device_grouped_conv_bwd_weight_v3_xdl_c_shuffle_bf16_instances = std::tuple<
+    // clang-format off
+        //#########################################|     Num| InLayout| WeiLayout| OutLayout| InData| WeiData| OutData| AccData|          In|         Wei|         Out|              ConvBackward| Block|  MPer|  NPer| K0Per| K1| MPer| NPer| MXdl| NXdl|  ABlockTransfer|   ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|  BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds|    CShuffle|    CShuffle|   CBlockTransfer|  CBlockTransfer| BlockGemm| BlockGemm|
+        //#########################################|     Dim|         |          |          |   Type|    Type|    Type|    Type| Elementwise| Elementwise| Elementwise|                    Weight|  Size| Block| Block| Block|   |  XDL|  XDL|  Per|  Per|   ThreadCluster|    ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|   ClusterLengths| ScalarPerVector|  Pipeline|  Pipeline|
+        //#########################################| Spatial|         |          |          |       |        |        |        |   Operation|   Operation|   Operation|            Specialization|      |      |      |      |   |     |     | Wave| Wave| Lengths_K0_M_K1|     ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle| MBlock_MPerBlock|    NWaveNPerXdl| Scheduler|   Version|
+        //#########################################|        |         |          |          |       |        |        |        |            |            |            |                          |      |      |      |      |   |     |     |     |     |                |                 |               |               |               |               |          |                |               |               |              |               |               |          |            |            | NBlock_NPerBlock|                |          |          |
+        // generic instance
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    32,     32,   8,   32,   32,    1,    1,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              2,              2,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              2,              2,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,    64,     32,   8,   32,   32,    1,    2,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              4,              4,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    32,   128,     32,   8,   32,   32,    1,    4,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              8,              8,      false,  S<4, 16,  1>, S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    32,     32,   8,   32,   32,    2,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              4,              4,      false,  S<4, 8,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              4,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,   128,    32,     32,   8,   32,   32,    4,    1,  S<4, 16, 1>,  S<2, 0, 1>,  S<1, 0, 2>,                  1,              8,              8,      false,  S<4, 4,  1>,  S<2, 0, 1>,  S<1, 0, 2>,                1,              8,              8,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,    80,     32,   8,   16,   16,    4,    5,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,               1,              5,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>,
+        DeviceGroupedConvBwdWeight_Xdl_CShuffleV3< NDimSpatial,  ALayout,   BLayout,   ELayout,   BF16,    BF16,    BF16,     F32, PassThrough, PassThrough, PassThrough,                  ConvSpec,    64,    64,   112,     32,   8,   16,   16,    4,    7,  S<4, 16, 1>,  S<2, 0, 1>,  S<2, 0, 1>,                  1,              4,              4,      false,  S<4, 16,  1>,  S<2, 0, 1>,  S<2, 0, 1>,               1,              7,              4,      false,           1,           1,   S<1, 8, 1, 8>,                  2, Scheduler, PipelineVersion>
+    //clang-format on
+    >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_comp_instance.hpp

@@ -61,8 +61,9 @@ using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,  16,  16,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         // Compute friendly
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
@@ -81,7 +82,7 @@ using device_grouped_conv_fwd_xdl_bf16_comp_instances = std::tuple<
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,    64,    64,   8,   8,  32,   32,    2,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,    64,   128,    64,   8,   8,  32,   32,    1,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,    64,    64,    64,   8,   8,  32,   32,    1,    1,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -97,8 +98,9 @@ using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,  16,  16,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    32,   8,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
@@ -113,7 +115,7 @@ using device_grouped_conv_fwd_xdl_f16_comp_instances = std::tuple<
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   256,    32,   8,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   128,    32,   8,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -148,8 +150,9 @@ using device_grouped_conv_fwd_xdl_int8_comp_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,DsLayout,ELayout,int8_t,int8_t,int32_t,   int8_t,    DsLayout,int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,             256,   128,   128,   128,  32,  32,   32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,              16,        1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,             16,             16,         1,            1,           1,              S<1, 64, 1, 4>,              16,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    32,   8,   8,  32,   32,    2,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
@@ -160,7 +163,7 @@ using device_grouped_conv_fwd_xdl_int8_comp_instances = std::tuple<
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   256,    32,   8,   8,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   256,   128,    32,   8,   8,  32,   32,    4,    2,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle_V3<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   int8_t,   int8_t,     int32_t,      int8_t,    DsLayout,   int8_t, PassThrough, PassThrough, PassThrough,       ConvSpec, GemmMNKPadding,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_fwd/device_grouped_conv_fwd_xdl_merged_groups_instance.hpp

@@ -45,13 +45,16 @@ using device_grouped_conv_fwd_xdl_merged_groups_bf16_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|     Type|     Type|  Pipeline|   ToMerge|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|         |         | Scheduler|          |
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |         |         |          |          |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 16>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 32>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         // Instances with NumGroupsPerBatch > 1
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 8>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 16>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,  BF16,  BF16,     F32,     BF16,    DsLayout,  BF16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, BF16, BF16, LoopScheduler::Default, 32>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 
@@ -67,13 +70,17 @@ using device_grouped_conv_fwd_xdl_merged_groups_f16_instances = std::tuple<
         //########################################|    Spatial| Layout| Layout|      Layout| Layout|  Type|  Type|    Type| DataType|    DataType|  Type| Elementwise| Elementwise| Elementwise| Specialization| Specialization| Prefetch|  Size| Block| Block| Block|    |    |  XDL|  XDL|  Per|  Per|   ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar| AddExtraM|   ThreadCluster|  ThreadCluster| SrcAccessOrder|  SrcVectorDim|      SrcScalar|      DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave|         _MBlock_MWaveMPerXdl| ScalarPerVector|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |   Operation|   Operation|   Operation|               |               |    Stage|      |      |      |      |    |    |     |     | Wave| Wave| Lengths_K0_M_K1|   ArrangeOrder|               |               |      PerVector|   PerVector_K1|          | Lengths_K0_N_K1|   ArrangeOrder|               |              |      PerVector|   PerVector_K1|          |  PerShuffle|  PerShuffle|         _NBlock_NWaveNPerXdl|   _NWaveNPerXdl|
         //########################################|           |       |       |            |       |      |      |        |         |            |      |            |            |            |               |               |         |      |      |      |      |    |    |     |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                             |                |
-#if defined(__gfx950__)
-#else
+#if defined(CK_USE_AMD_MFMA_GFX950)
+        // Instances with NumGroupsPerBatch > 1
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 8>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 16>,
+        DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     32,   8, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 32>
+#else // defined(CK_USE_AMD_MFMA_GFX950)
         // Instances with NumGroupsPerBatch > 1
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 8>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 16>,
         DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<NDimSpatial,ALayout,BLayout,    DsLayout,ELayout,   F16,   F16,     F32,      F16,    DsLayout,   F16, PassThrough, PassThrough, PassThrough,                  ConvSpec, GemmMNKPadding,  1,  64,    64,    16,     16,   4, 4,  16,   16,    4,    1,  S< 4, 16,  1>, S<0, 2, 1>,     S<0, 2, 1>,                   1,              4,              4,      1,  S< 4, 16,  1>,   S<1, 0, 2>,     S<1, 0, 2>,             2,              1,              4,      1,           1,           1,   S<1, 16, 1, 4>,                  1, F16, F16, LoopScheduler::Default, 32>
-#endif // defined(__gfx950__)
+#endif // defined(CK_USE_AMD_MFMA_GFX950)
     // clang-format on
     >;
 

library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_backward_weight.hpp

@@ -311,6 +311,11 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_default_pipev1_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f32_pad0_pipev1_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_FP16
@@ -320,6 +325,11 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_default_pipev1_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_gnhwc_gkyxc_gnhwk_f16_pad0_pipev1_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_BF16
@@ -343,6 +353,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_default_pipev5_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f32_pad0_pipev5_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_FP16
@@ -352,6 +371,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_default_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_default_pipev5_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_pad0_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_f16_pad0_pipev5_instances(
+                        op_ptrs);
+
                     add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev1_instances(
                         op_ptrs);
                     add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_f16_pipev2_instances(
@@ -381,6 +410,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_default_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_default_pipev5_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv2d_bwd_weight_xdl_nhwgc_gkyxc_nhwgk_bf16_pad0_pipev5_instances(
+                        op_ptrs);
+
                     add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev1_instances(
                         op_ptrs);
                     add_device_grouped_conv2d_bwd_weight_two_stage_xdl_nhwgc_gkyxc_nhwgk_bf16_pipev2_instances(
@@ -471,6 +510,15 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_default_pipev5_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f32_pad0_pipev2_instances(
+                        op_ptrs);
                 }
 #endif
 #ifdef CK_ENABLE_FP16
@@ -480,6 +528,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_default_pipev5_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pad0_pipev5_instances(
+                        op_ptrs);
+
                     add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev1_instances(
                         op_ptrs);
                     add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_f16_pipev2_instances(
@@ -509,6 +567,16 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
                 {
                     add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instances(
                         op_ptrs);
+
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_default_pipev5_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev2_instances(
+                        op_ptrs);
+                    add_device_grouped_conv3d_bwd_weight_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pad0_pipev5_instances(
+                        op_ptrs);
+
                     add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev1_instances(
                         op_ptrs);
                     add_device_grouped_conv3d_bwd_weight_two_stage_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_pipev2_instances(

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn.hpp

@@ -44,7 +44,9 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn_comp_instances = std::tu
 #if defined(CK_USE_AMD_MFMA_GFX950)
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
         // Compute friendly
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   8,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
@@ -54,7 +56,9 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn_comp_instances = std::tu
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   8,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   224,    64,   8,   8,  16,   16,    8,    7,     S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,                4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   4,   8,  32,   32,    2,    2,     S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,              1,              8,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,
@@ -86,7 +90,9 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_km_nk_mn_mem_instances = std::tup
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   4,   8,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   4,   8,  16,   16,    1,    1,     S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              4,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   4,   8,  16,   16,    1,    1,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   4,   8,  16,   16,    1,    2,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   4,   4,  16,   16,    1,    2,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<16, 8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    64,   4,   8,  16,   16,    1,    4,     S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              4,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Col,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,  PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   2,   8,  16,   16,    1,    4,     S<32, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              2,              2,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn.hpp

@@ -43,13 +43,17 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_comp_instances = std::tu
         //#########################|        |        |        |     |      |      |        |         |            |            |            |              |      |      |      |      |    |    |    |     |     |     |                |               |               |               |               |               |          |                |               |               |              |               |               |          |            |            |                                 |                |                                       |                             |
 #if defined(CK_USE_AMD_MFMA_GFX950)
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<16,16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,          0,          1,           2,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
@@ -80,8 +84,10 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_kn_mn_mem_instances = std::tup
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   8,   4,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   4,   4,  16,   16,    1,    1,     S<16, 4, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,    S<16, 4, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   8,   4,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              4,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   8,   4,  16,   16,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 8, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   8,   4,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
     // clang-format on
     >;

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_bf16_bf16_bf16/device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn.hpp

@@ -44,16 +44,20 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_comp_instances = std::tu
 #if defined(CK_USE_AMD_MFMA_GFX950)
 #endif // defined(CK_USE_AMD_MFMA_GFX950)
         // Compute friendly
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   4,   4,  32,   32,    4,    4,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              4,              4,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   2,   2,  32,   32,    4,    4,     S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,    S<16,16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              2,              2,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         // AGPR Spill
         // DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   8,  16,   16,    8,    8,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,    S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          1,          1,           2,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         // AGPR Spill when use permuted lds layout. so, use padding for these two.
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           2,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   224,    64,   8,   8,  16,   16,    8,    7,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          2,           1,                   S<1, 32, 1, 8>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   8,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,               4,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>
@@ -84,8 +88,10 @@ using device_gemm_xdl_universal_bf16_bf16_bf16_mk_nk_mn_mem_instances = std::tup
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    32,   16,    64,   8,   8,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               2,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,    64,    16,   16,    64,   8,   8,  16,   16,    1,    1,     S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8,  8, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 4>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   32,    64,   8,   8,  16,   16,    1,    1,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,   64,    64,   8,   8,  16,   16,    1,    2,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   128,    16,  128,    64,   8,   8,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 8>,               4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Col,     Row,     BF16,   BF16,  BF16,   F32,     BF16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,    16,  256,    64,   8,   8,  16,   16,    1,    4,     S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 16, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,          1,           1,                   S<1, 16, 1, 16>,              4,  BlkGemmPipeSched, BlockGemmPipelineVersion::v2>
     // clang-format on
     >;

library/src/tensor_operation_instance/gpu/gemm_universal/device_gemm_xdl_universal_f16_f16_f16/device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn.hpp

@@ -47,7 +47,9 @@ using device_gemm_xdl_universal_f16_f16_f16_mk_kn_mn_comp_instances = std::tuple
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v4>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   256,   256,    32,   8,   4,  32,   32,    4,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#if !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   224,   256,    64,   8,   8,  16,   16,    7,    8,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              8,          0,          1,           2,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
+#endif // !defined(CK_USE_AMD_MFMA_GFX950)
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v3>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   128,    64,   8,   4,  32,   32,    2,    2,     S<8, 32, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<16, 16, 1>,    S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Intrawave, BlockGemmPipelineVersion::v5>,
         DeviceGemm_Xdl_CShuffleV3<  Row,     Row,     Row,     F16,   F16,  F16,   F32,     F16,      PassThrough, PassThrough, PassThrough,       GemmSpec,   256,   128,   256,    32,   8,   4,  32,   32,    2,    4,     S<4, 64, 1>,     S<1, 0, 2>,    S<1, 0, 2>,               2,              8,              8,          0,    S<8, 32, 1>,     S<0, 2, 1>,     S<0, 2, 1>,             1,              8,              4,          0,          1,           1,                   S<1, 32, 1, 8>,               8,  BlockGemmPipelineScheduler::Interwave, BlockGemmPipelineVersion::v1>,