Merge remote-tracking branch 'upstream/develop' into merge_upstream_1129

also fix regression

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

@@ -14,6 +14,7 @@
 #include "ck/tensor_operation/gpu/device/device_grouped_conv_bwd_weight.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_backward_weight_specialization.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 
@@ -21,34 +22,9 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-namespace {
-
-struct ComputePtrOffsetOfStridedBatch
-{
-    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideA_);
-    }
-
-    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideB_);
-    }
-
-    __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideC_);
-    }
-
-    index_t BatchStrideA_;
-    index_t BatchStrideB_;
-    index_t BatchStrideC_;
-};
-
-} // namespace
-
 template <typename GridwiseGemm,
-          typename FloatAB,
+          typename FloatA,
+          typename FloatB,
           typename FloatC,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
@@ -64,8 +40,8 @@ __global__ void
     __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
 #endif
         kernel_batched_gemm_xdlops_bwd_weight(
-            const FloatAB* __restrict__ p_a_grid,
-            const FloatAB* __restrict__ p_b_grid,
+            const FloatA* __restrict__ p_a_grid,
+            const FloatB* __restrict__ p_b_grid,
             FloatC* __restrict__ p_c_grid,
             const AElementwiseOperation a_element_op,
             const BElementwiseOperation b_element_op,
@@ -91,7 +67,7 @@ __global__ void
     const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane(
         static_cast<long_index_t>(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx)));
 
-    __shared__ FloatAB p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB)];
+    __shared__ FloatA p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatA)];
 
     GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
                                                   p_b_grid + b_batch_offset,
@@ -163,7 +139,9 @@ template <ck::index_t NDimSpatial,
           index_t CShuffleMXdlPerWavePerShuffle,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
-          index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
+          index_t CBlockTransferScalarPerVector_NWaveNPerXdl,
+          typename ComputeTypeA = InDataType,
+          typename ComputeTypeB = ComputeTypeA>
 struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
     : public DeviceGroupedConvBwdWeight<NDimSpatial,
                                         InLayout,
@@ -174,7 +152,9 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                                         OutDataType,
                                         InElementwiseOperation,
                                         WeiElementwiseOperation,
-                                        OutElementwiseOperation>
+                                        OutElementwiseOperation,
+                                        ComputeTypeA,
+                                        ComputeTypeB>
 {
     using DeviceOp = DeviceGroupedConvBwdWeight_Xdl_CShuffle;
 
@@ -1045,7 +1025,8 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
 
     using GridwiseGemm = GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight<
         BlockSize,
-        ADataType, // TODO: distinguish A/B datatype
+        ADataType,
+        BDataType,
         AccDataType,
         CDataType,
         InMemoryDataOperationEnum::AtomicAdd,
@@ -1090,7 +1071,11 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         CBlockTransferScalarPerVector_NWaveNPerXdl,
         CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         true,
-        true>;
+        true,
+        1,
+        PipelineVersion::v1,
+        ComputeTypeA,
+        ComputeTypeB>;
 
     // Argument
     using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
@@ -1212,13 +1197,13 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
         Block2CTileMap block_2_ctile_map_;
 
         // for computing batch offset
-        ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_;
+        ComputePtrOffsetOfStridedBatch<> compute_ptr_offset_of_batch_;
 
         index_t M01_;
         index_t N01_;
 
-        InElementwiseOperation a_element_op_;
-        OutElementwiseOperation b_element_op_;
+        OutElementwiseOperation a_element_op_;
+        InElementwiseOperation b_element_op_;
         WeiElementwiseOperation c_element_op_;
 
         // for checking IsSupportedArgument()
@@ -1281,7 +1266,8 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
 
                 const auto kernel = kernel_batched_gemm_xdlops_bwd_weight<
                     GridwiseGemm,
-                    ADataType, // TODO: distiguish A/B datatype
+                    ADataType,
+                    BDataType,
                     CDataType,
                     OutElementwiseOperation,
                     InElementwiseOperation,
@@ -1290,7 +1276,7 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
                     remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
                     remove_reference_t<DeviceOp::CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock>,
                     remove_reference_t<DeviceOp::Block2CTileMap>,
-                    ComputePtrOffsetOfStridedBatch,
+                    ComputePtrOffsetOfStridedBatch<>,
                     has_main_loop>;
 
                 return launch_and_time_kernel(stream_config,
@@ -1337,6 +1323,10 @@ struct DeviceGroupedConvBwdWeight_Xdl_CShuffle
 
     static bool IsSupportedArgument(const Argument& arg)
     {
+        if(!ck::is_xdl_supported())
+        {
+            return false;
+        }
         if constexpr(NDimSpatial == 1)
         {
             if constexpr(!is_GNWK_GKXC_GNWC)

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

@@ -15,10 +15,11 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
 #include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 #include "ck/host_utility/io.hpp"
@@ -29,51 +30,6 @@ namespace device {
 
 namespace {
 
-template <index_t NumDTensor>
-struct ComputePtrOffsetOfStridedBatch
-{
-    ComputePtrOffsetOfStridedBatch() = default;
-
-    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
-                                   index_t BatchStrideB,
-                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
-                                   index_t BatchStrideE)
-        : BatchStrideA_(BatchStrideA),
-          BatchStrideB_(BatchStrideB),
-          BatchStrideDs_(BatchStrideDs),
-          BatchStrideE_(BatchStrideE)
-    {
-    }
-
-    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideA_);
-    }
-
-    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideB_);
-    }
-
-    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
-    {
-        Array<long_index_t, NumDTensor> ds_offset;
-        static_for<0, NumDTensor, 1>{}(
-            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
-        return ds_offset;
-    }
-
-    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideE_);
-    }
-
-    index_t BatchStrideA_;
-    index_t BatchStrideB_;
-    Array<ck::index_t, NumDTensor> BatchStrideDs_;
-    index_t BatchStrideE_;
-};
-
 /*
  * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
  *
@@ -260,18 +216,18 @@ template <index_t NDimSpatial,
           index_t CThreadTransferSrcDstVectorDim,
           index_t CThreadTransferDstScalarPerVector>
 struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
-    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
-                                           ALayout,
-                                           BLayout,
-                                           DsLayout,
-                                           ELayout,
-                                           ADataType,
-                                           BDataType,
-                                           DsDataType,
-                                           EDataType,
-                                           AElementwiseOperation,
-                                           BElementwiseOperation,
-                                           CDEElementwiseOperation>
+    : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
+                                             ALayout,
+                                             BLayout,
+                                             DsLayout,
+                                             ELayout,
+                                             ADataType,
+                                             BDataType,
+                                             DsDataType,
+                                             EDataType,
+                                             AElementwiseOperation,
+                                             BElementwiseOperation,
+                                             CDEElementwiseOperation>
 {
     using DeviceOp = DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK;
 
@@ -581,7 +537,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
         DefaultBlock2CTileMap block_2_ctile_map_;
 
         // for computing batch offset
-        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+        ComputePtrOffsetOfStridedBatch<I1, I1, NumDTensor> compute_ptr_offset_of_batch_;
 
         // element-wise op
         AElementwiseOperation a_element_op_;
@@ -645,7 +601,7 @@ struct DeviceGroupedConvFwdDlMultipleD_NHWC_KYXC_NHWK
                     DeviceOp::DsGridDesc_M0_M10_M11_N0_N10_N11,
                     DeviceOp::CGridDesc_M0_M10_M11_N0_N10_N11,
                     DefaultBlock2CTileMap,
-                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    ComputePtrOffsetOfStridedBatch<I1, I1, NumDTensor>,
                     has_main_loop,
                     has_double_loop>;
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <functional>
+#include <iostream>
+#include <iterator>
+#include <numeric>
+#include <sstream>
+
+#include "ck/utility/common_header.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
+#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
+#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+#include "ck/host_utility/io.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+namespace {
+
+/*
+ * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
+ *
+ * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
+ * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
+ * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
+ * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
+ * limitations.
+ *
+ * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
+ * returns the 2D index of the tile that it computes. \see
+ * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
+ *
+ * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
+ * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
+ * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
+ * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
+ * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
+ * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
+ *
+ * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
+ * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
+ * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
+ *
+ */
+template <typename GridwiseGemm,
+          typename AsPointer, // tuples if multi AB, pointers if no
+          typename BsPointer,
+          typename DsPointer,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          typename AGridDesc_AK0_M_AK1,
+          typename BGridDesc_BK0_N_BK1,
+          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+          typename Block2ETileMap,
+          typename ComputePtrOffsetOfBatch,
+          bool HasMainKBlockLoop,
+          bool isMultiA,
+          bool isMultiB>
+__global__ void
+#if CK_USE_LAUNCH_BOUNDS
+    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
+#endif
+        kernel_grouped_conv_fwd_multiple_abd_xdl_cshuffle(
+            AsPointer p_as_grid,
+            BsPointer p_bs_grid,
+            DsPointer p_ds_grid,
+            EDataType* __restrict__ p_e_grid,
+            const AElementwiseOperation a_element_op,
+            const BElementwiseOperation b_element_op,
+            const CDEElementwiseOperation cde_element_op,
+            const index_t batch_count,
+            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
+            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
+            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+                e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            const Block2ETileMap block_2_ctile_map,
+            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
+{
+#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
+    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
+    // offset base pointer for each work-group
+    const index_t num_blocks_per_batch =
+        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
+    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
+
+    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
+        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
+    const auto& ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
+
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    DsPointer p_ds_grid_grp;
+
+    static constexpr index_t NumDTensor =
+        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
+
+    static_for<0, NumDTensor, 1>{}(
+        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
+
+    if constexpr(isMultiA || isMultiB)
+    {
+        AsPointer p_as_grid_grp;
+        BsPointer p_bs_grid_grp;
+
+        const auto& as_batch_offset = compute_ptr_offset_of_batch.GetAsPtrOffset(g_idx);
+
+        static constexpr index_t NumATensor = AGridDesc_AK0_M_AK1::Size();
+        static_for<0, NumATensor, 1>{}(
+            [&](auto i) { p_as_grid_grp(i) = p_as_grid[i] + as_batch_offset[i]; });
+
+        const auto& bs_batch_offset = compute_ptr_offset_of_batch.GetBsPtrOffset(g_idx);
+
+        static constexpr index_t NumBTensor = BGridDesc_BK0_N_BK1::Size();
+        static_for<0, NumBTensor, 1>{}(
+            [&](auto i) { p_bs_grid_grp(i) = p_bs_grid[i] + bs_batch_offset[i]; });
+
+        GridwiseGemm::template Run<HasMainKBlockLoop>(
+            p_as_grid_grp,
+            p_bs_grid_grp,
+            p_ds_grid_grp,
+            p_e_grid + e_batch_offset,
+            p_shared,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            a_grid_desc_k0_m_k1,
+            b_grid_desc_k0_n_k1,
+            ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            block_2_ctile_map);
+    }
+    else
+    {
+        const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
+        const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
+            static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
+
+        GridwiseGemm::template Run<HasMainKBlockLoop>(
+            p_as_grid + a_batch_offset,
+            p_bs_grid + b_batch_offset,
+            p_ds_grid_grp,
+            p_e_grid + e_batch_offset,
+            p_shared,
+            a_element_op,
+            b_element_op,
+            cde_element_op,
+            a_grid_desc_k0_m_k1,
+            b_grid_desc_k0_n_k1,
+            ds_grid_desc_mblock_mperblock_nblock_nperblock,
+            e_grid_desc_mblock_mperblock_nblock_nperblock_,
+            block_2_ctile_map);
+    }
+#else
+    ignore = p_as_grid;
+    ignore = p_bs_grid;
+    ignore = p_ds_grid;
+    ignore = p_e_grid;
+    ignore = batch_count;
+    ignore = a_grid_desc_k0_m_k1;
+    ignore = b_grid_desc_k0_n_k1;
+    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
+    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
+    ignore = a_element_op;
+    ignore = b_element_op;
+    ignore = cde_element_op;
+    ignore = compute_ptr_offset_of_batch;
+    ignore = block_2_ctile_map;
+#endif
+}
+
+} // namespace
+
+template <typename T>
+using is_tuple = decltype(std::declval<T&>().IsTuple());
+
+//
+// @brief      Device Convolution operation.
+//
+// Supports:
+//  @li         Forward convolution with up to 3 spatial dimentions
+//  @li         Input tensor in GNWC data format
+//  @li         Weight tensor in GKXC data format
+//  @li         Output tensor in GNWK data format
+//
+// 1D:
+// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C]
+// 2D:
+// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C]
+// 3D:
+// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C]
+//
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename AccDataType,
+          typename CShuffleDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation,
+          ConvolutionForwardSpecialization ConvForwardSpecialization,
+          GemmSpecialization GemmSpec,
+          index_t NumGemmKPrefetchStage,
+          index_t BlockSize,
+          index_t MPerBlock,
+          index_t NPerBlock,
+          index_t KPerBlock,
+          index_t AK1,
+          index_t BK1,
+          index_t MPerXDL,
+          index_t NPerXDL,
+          index_t MXdlPerWave,
+          index_t NXdlPerWave,
+          typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
+          typename ABlockTransferThreadClusterArrangeOrder,
+          typename ABlockTransferSrcAccessOrder,
+          index_t ABlockTransferSrcVectorDim,
+          index_t ABlockTransferSrcScalarPerVector,
+          index_t ABlockTransferDstScalarPerVector_AK1,
+          index_t ABlockLdsExtraM,
+          typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
+          typename BBlockTransferThreadClusterArrangeOrder,
+          typename BBlockTransferSrcAccessOrder,
+          index_t BBlockTransferSrcVectorDim,
+          index_t BBlockTransferSrcScalarPerVector,
+          index_t BBlockTransferDstScalarPerVector_BK1,
+          index_t BBlockLdsExtraN,
+          index_t CShuffleMXdlPerWavePerShuffle,
+          index_t CShuffleNXdlPerWavePerShuffle,
+          typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+          index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          typename ComputeDataType =
+              decltype(UnpackDataType<is_detected<is_tuple, ADataType>::value,
+                                      Number<0>,
+                                      ADataType>()), // ComputeType is InputType by default (first
+                                                     // in tuple for MultiAB), unpack if tuple was
+                                                     // passed
+          LoopScheduler LoopSched = make_default_loop_scheduler()>
+struct DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle
+    : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
+                                             ALayout,
+                                             BLayout,
+                                             DsLayout,
+                                             ELayout,
+                                             ADataType,
+                                             BDataType,
+                                             DsDataType,
+                                             EDataType,
+                                             AElementwiseOperation,
+                                             BElementwiseOperation,
+                                             CDEElementwiseOperation,
+                                             ComputeDataType>
+{
+    using DeviceOp = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle;
+
+    static constexpr bool isMultiA = is_detected<is_tuple, ADataType>::value;
+    static constexpr bool isMultiB = is_detected<is_tuple, BDataType>::value;
+
+    static constexpr index_t NumATensor = GetNumABTensors<isMultiA, ADataType>();
+    static constexpr index_t NumBTensor = GetNumABTensors<isMultiB, BDataType>();
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    static constexpr auto conv_to_gemm_transformer =
+        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
+
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay>
+    static auto
+    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 auto in_gemmmraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
+                                                                        a_g_n_c_wis_strides,
+                                                                        b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides,
+                                                                        e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides,
+                                                                        conv_filter_strides,
+                                                                        conv_filter_dilations,
+                                                                        input_left_pads,
+                                                                        input_right_pads);
+
+        const auto in_gemmm_gemmk_desc =
+            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
+
+        return in_gemmm_gemmk_desc;
+    }
+
+    template <typename BLay>
+    static auto
+    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
+    {
+        const auto wei_gemmnraw_gemmkraw_desc =
+            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
+                                                                        b_g_k_c_xs_strides);
+
+        const auto wei_gemmn_gemmk_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
+
+        return wei_gemmn_gemmk_desc;
+    }
+
+    template <typename ELay>
+    static auto
+    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
+    {
+        const auto out_gemmmraw_gemmnraw_desc =
+            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
+                                                                        e_g_n_k_wos_strides);
+
+        const auto out_gemmm_gemmn_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
+
+        return out_gemmm_gemmn_desc;
+    }
+
+    static auto MakeDsGridDescriptor_M_N(
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
+                                                                  ds_g_n_k_wos_strides[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // desc for problem definition
+    using AGridDesc_M_K  = remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(
+        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
+    using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
+    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
+
+    // If we are using multiAB and one of the template datatype parameters is not a tuple, convert
+    // it to it
+    using GemmADataType = std::conditional_t<!isMultiA && isMultiB, Tuple<ADataType>, ADataType>;
+    using GemmBDataType = std::conditional_t<!isMultiB && isMultiA, Tuple<BDataType>, BDataType>;
+
+#define GridwiseGemmTemplateParameters                                                          \
+    GemmADataType, GemmBDataType, ComputeDataType, AccDataType, CShuffleDataType, DsDataType,   \
+        EDataType, AElementwiseOperation, BElementwiseOperation, CDEElementwiseOperation,       \
+        InMemoryDataOperationEnum::Set, NumGemmKPrefetchStage, BlockSize, MPerBlock, NPerBlock, \
+        KPerBlock, AK1, BK1, MPerXDL, NPerXDL, MXdlPerWave, NXdlPerWave,                        \
+        ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterArrangeOrder,  \
+        ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim,                               \
+        ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_AK1, false,          \
+        ABlockLdsExtraM, BBlockTransferThreadClusterLengths_BK0_N_BK1,                          \
+        BBlockTransferThreadClusterArrangeOrder, BBlockTransferSrcAccessOrder,                  \
+        BBlockTransferSrcVectorDim, BBlockTransferSrcScalarPerVector,                           \
+        BBlockTransferDstScalarPerVector_BK1, false, BBlockLdsExtraN,                           \
+        CShuffleMXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle,                           \
+        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,                       \
+        CDEBlockTransferScalarPerVector_NPerBlock, LoopSched
+    // Use appropriate gridwise gemm
+    using GridwiseGemm =
+        std::conditional_t<isMultiA || isMultiB,
+                           GridwiseGemmMultipleABD_xdl_cshuffle<GridwiseGemmTemplateParameters>,
+                           GridwiseGemmMultipleD_xdl_cshuffle<GridwiseGemmTemplateParameters>>;
+
+    // If ADataTypes or BDataTypes is tuple, user has to pass std::array with pointers.
+    using APointers =
+        std::conditional_t<isMultiA, std::array<const void*, NumATensor>&, const void*>;
+    using BPointers =
+        std::conditional_t<isMultiB, std::array<const void*, NumBTensor>&, const void*>;
+    // Use Tuple for the both cases for GridPointer to initialize it in Argument constructor (not
+    // in initializer list what is required for single const pointer).
+    using AGridPointer = remove_cvref_t<
+        decltype(GetAGridPointer < isMultiA || isMultiB, GridwiseGemm, ADataType > ())>;
+    using BGridPointer = remove_cvref_t<
+        decltype(GetBGridPointer < isMultiA || isMultiB, GridwiseGemm, BDataType > ())>;
+
+    // desc for blockwise copy
+    using AGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
+            AGridDesc_M_K{}))>;
+    using BGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
+            BGridDesc_N_K{}))>;
+    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
+        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            DsGridDesc_M_N{}))>;
+    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
+        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+            EGridDesc_M_N{}))>;
+
+    // block-to-e-tile map
+    using Block2ETileMap =
+        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
+
+    // Argument
+    struct Argument : public BaseArgument
+    {
+        Argument(APointers p_as,
+                 BPointers p_bs,
+                 const std::array<const void*, NumDTensor>& p_ds,
+                 void* p_e,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_lengths,
+                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
+                     ds_g_n_k_wos_strides,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 AElementwiseOperation& a_element_op,
+                 const BElementwiseOperation& b_element_op,
+                 const CDEElementwiseOperation& cde_element_op)
+            : p_as_grid_{},
+              p_bs_grid_{},
+              p_ds_grid_{},
+              p_e_grid_{static_cast<EDataType*>(p_e)},
+              num_group_{a_g_n_c_wis_lengths[0]},
+              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
+                                                                          a_g_n_c_wis_strides,
+                                                                          b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides,
+                                                                          e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides,
+                                                                          conv_filter_strides,
+                                                                          conv_filter_dilations,
+                                                                          input_left_pads,
+                                                                          input_right_pads)},
+              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
+                                                                          b_g_k_c_xs_strides)},
+              ds_grid_desc_m_n_{},
+              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
+                                                                          e_g_n_k_wos_strides)},
+              a_grid_desc_ak0_m_ak1_{
+                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
+              b_grid_desc_bk0_n_bk1_{
+                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
+              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
+              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
+              compute_ptr_offset_of_batch_{},
+              a_element_op_{a_element_op},
+              b_element_op_{b_element_op},
+              cde_element_op_{cde_element_op},
+              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
+              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
+              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
+              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
+              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
+              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
+              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
+              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
+              conv_filter_strides_{conv_filter_strides},
+              conv_filter_dilations_{conv_filter_dilations},
+              input_left_pads_{input_left_pads},
+              input_right_pads_{input_right_pads}
+        {
+            // A/B/E Batch Stride
+            if constexpr(isMultiA || isMultiB)
+            {
+                static_for<0, NumATensor, 1>{}([&](auto i) {
+                    // Init compute_ptr_offset_of_batch_ for multiple AB
+                    compute_ptr_offset_of_batch_.BatchStrideA_(i) = a_g_n_c_wis_strides[0];
+
+                    // Use GemmADataType/GemmBDataType to iterate over tuple (even if passed data
+                    // type is not tuple)
+                    using DataType = remove_cvref_t<tuple_element_t<i.value, GemmADataType>>;
+                    // It is possible that one of the AB is a pointer and one is a tuple.
+                    // Then also use multiAB but we have to cast single pointer instead of tuple of
+                    // pointer.
+                    if constexpr(isMultiA)
+                    {
+                        // p_as is tuple
+                        p_as_grid_(i) = static_cast<const DataType*>(p_as[i.value]);
+                    }
+                    else
+                    {
+                        // if MultiB and not MultiA then p_as is single pointer
+                        p_as_grid_(i) = static_cast<const DataType*>(p_as);
+                    }
+                });
+                static_for<0, NumBTensor, 1>{}([&](auto i) {
+                    // Init compute_ptr_offset_of_batch_ for multiple AB
+                    compute_ptr_offset_of_batch_.BatchStrideB_(i) = b_g_k_c_xs_strides[0];
+
+                    using DataType = remove_cvref_t<tuple_element_t<i.value, GemmBDataType>>;
+                    // It is possible that one of the AB is a pointer and one is a tuple.
+                    // Then also use multiAB but we have to cast single pointer instead of tuple of
+                    // pointer.
+                    if constexpr(isMultiB)
+                    {
+                        // p_bs is tuple
+                        p_bs_grid_(i) = static_cast<const DataType*>(p_bs[i.value]);
+                    }
+                    else
+                    {
+                        // if MultiA and not MultiB then p_bs is single pointer
+                        p_bs_grid_(i) = static_cast<const DataType*>(p_bs);
+                    }
+                });
+            }
+            else
+            {
+                compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
+                compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
+
+                // p_as and p_bs are pointers
+                p_as_grid_(I0) = static_cast<const ADataType*>(p_as);
+                p_bs_grid_(I0) = static_cast<const BDataType*>(p_bs);
+            }
+
+            // populate pointer, batch stride, desc for Ds
+            static_for<0, NumDTensor, 1>{}([&](auto i) {
+                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
+
+                // D pointer
+                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
+
+                // D batch stride
+                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
+
+                // D desc
+                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
+            });
+            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
+
+            // populate desc for Ds/E
+            if constexpr(isMultiA || isMultiB)
+            {
+                const auto as_grid_desc_ak0_m_ak1 =
+                    generate_tuple([&](auto) { return a_grid_desc_m_k_; }, Number<NumATensor>{});
+                const auto bs_grid_desc_bk0_n_bk1 =
+                    generate_tuple([&](auto) { return b_grid_desc_n_k_; }, Number<NumBTensor>{});
+
+                if(GridwiseGemm::CheckValidity(as_grid_desc_ak0_m_ak1,
+                                               bs_grid_desc_bk0_n_bk1,
+                                               ds_grid_desc_m_n_,
+                                               e_grid_desc_m_n_,
+                                               block_2_etile_map_))
+                {
+                    e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            e_grid_desc_m_n_);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            ds_grid_desc_m_n_);
+                }
+            }
+            else
+            {
+                if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
+                                               b_grid_desc_n_k_,
+                                               ds_grid_desc_m_n_,
+                                               e_grid_desc_m_n_,
+                                               block_2_etile_map_))
+                {
+                    e_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            e_grid_desc_m_n_);
+
+                    ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
+                        GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            ds_grid_desc_m_n_);
+                }
+            }
+        }
+
+        void Print() const
+        {
+            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
+            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
+            static_for<0, NumDTensor, 1>{}(
+                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
+            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
+        }
+
+        //  private:
+        // pointers (tuple if multi AB, pointer if no)
+        AGridPointer p_as_grid_;
+        BGridPointer p_bs_grid_;
+        typename GridwiseGemm::DsGridPointer p_ds_grid_;
+        EDataType* p_e_grid_;
+
+        // tensor descriptors for problem definiton
+        index_t num_group_;
+        AGridDesc_M_K a_grid_desc_m_k_;
+        BGridDesc_N_K b_grid_desc_n_k_;
+        DsGridDesc_M_N ds_grid_desc_m_n_;
+        EGridDesc_M_N e_grid_desc_m_n_;
+
+        // tensor descriptors for block/thread-wise copy
+        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
+        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
+        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        // for computing batch offset
+        ComputePtrOffsetOfStridedBatch<NumATensor, NumBTensor, NumDTensor>
+            compute_ptr_offset_of_batch_;
+
+        // element-wise op
+        AElementwiseOperation a_element_op_;
+        BElementwiseOperation b_element_op_;
+        CDEElementwiseOperation cde_element_op_;
+
+        // for checking IsSupportedArgument()
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
+        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
+        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
+        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
+        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_strides_;
+        std::array<index_t, NDimSpatial> conv_filter_dilations_;
+        std::array<index_t, NDimSpatial> input_left_pads_;
+        std::array<index_t, NDimSpatial> input_right_pads_;
+    };
+
+    // Invoker
+    struct Invoker : public BaseInvoker
+    {
+        using Argument = DeviceOp::Argument;
+
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            if(stream_config.log_level_ > 0)
+            {
+                arg.Print();
+            }
+
+            const index_t grid_size =
+                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
+
+            const auto K =
+                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
+
+            auto launch_kernel = [&](auto has_main_k_block_loop) {
+                constexpr bool has_main_loop = has_main_k_block_loop.value;
+
+                if constexpr(isMultiA || isMultiB)
+                {
+                    // Generate tuples with grid descriptors for each A and B
+                    const auto as_grid_desc_ak0_m_ak1 = generate_tuple(
+                        [&](auto) { return arg.a_grid_desc_ak0_m_ak1_; }, Number<NumATensor>{});
+                    const auto bs_grid_desc_bk0_n_bk1 = generate_tuple(
+                        [&](auto) { return arg.b_grid_desc_bk0_n_bk1_; }, Number<NumBTensor>{});
+
+                    const auto kernel = kernel_grouped_conv_fwd_multiple_abd_xdl_cshuffle<
+                        GridwiseGemm,
+                        AGridPointer,
+                        BGridPointer,
+                        typename GridwiseGemm::DsGridPointer,
+                        EDataType,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CDEElementwiseOperation,
+                        decltype(as_grid_desc_ak0_m_ak1),
+                        decltype(bs_grid_desc_bk0_n_bk1),
+                        DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        Block2ETileMap,
+                        ComputePtrOffsetOfStridedBatch<NumATensor, NumBTensor, NumDTensor>,
+                        has_main_loop,
+                        isMultiA,
+                        isMultiB>;
+
+                    return launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_as_grid_,
+                        arg.p_bs_grid_,
+                        arg.p_ds_grid_,
+                        arg.p_e_grid_,
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.cde_element_op_,
+                        arg.a_g_n_c_wis_lengths_[0], // Group count
+                        as_grid_desc_ak0_m_ak1,
+                        bs_grid_desc_bk0_n_bk1,
+                        arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                        arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                        arg.block_2_etile_map_,
+                        arg.compute_ptr_offset_of_batch_);
+                }
+                else
+                {
+                    const auto kernel = kernel_grouped_conv_fwd_multiple_abd_xdl_cshuffle<
+                        GridwiseGemm,
+                        const ADataType*,
+                        const BDataType*,
+                        typename GridwiseGemm::DsGridPointer,
+                        EDataType,
+                        AElementwiseOperation,
+                        BElementwiseOperation,
+                        CDEElementwiseOperation,
+                        DeviceOp::AGridDesc_AK0_M_AK1,
+                        DeviceOp::BGridDesc_BK0_N_BK1,
+                        DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
+                        Block2ETileMap,
+                        ComputePtrOffsetOfStridedBatch<NumATensor, NumBTensor, NumDTensor>,
+                        has_main_loop,
+                        isMultiA,
+                        isMultiB>;
+
+                    return launch_and_time_kernel(
+                        stream_config,
+                        kernel,
+                        dim3(grid_size),
+                        dim3(BlockSize),
+                        0,
+                        arg.p_as_grid_.At(I0), // Pass just A descriptor instead of tuple
+                        arg.p_bs_grid_.At(I0), // Pass just B descriptor instead of tuple
+                        arg.p_ds_grid_,
+                        arg.p_e_grid_,
+                        arg.a_element_op_,
+                        arg.b_element_op_,
+                        arg.cde_element_op_,
+                        arg.a_g_n_c_wis_lengths_[0], // Group count
+                        arg.a_grid_desc_ak0_m_ak1_,
+                        arg.b_grid_desc_bk0_n_bk1_,
+                        arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
+                        arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
+                        arg.block_2_etile_map_,
+                        arg.compute_ptr_offset_of_batch_);
+                }
+            };
+
+            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
+            {
+                return launch_kernel(integral_constant<bool, true>{});
+            }
+            else
+            {
+                return launch_kernel(integral_constant<bool, false>{});
+            }
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    static bool IsSupportedArgument(const Argument& arg)
+    {
+        namespace ctc = tensor_layout::convolution;
+
+        // check device
+        if(get_device_name() == "gfx908")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t>))
+            {
+                return false;
+            }
+        }
+        else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940" ||
+                get_device_name() == "gfx941" || get_device_name() == "gfx942")
+        {
+            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
+                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check ConvolutionForwardSpecialization
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            // check if it's 1x1, stride=1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t ConvStride = arg.conv_filter_strides_[i];
+                const index_t LeftPad    = arg.input_left_pads_[i];
+                const index_t RightPad   = arg.input_right_pads_[i];
+
+                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // check if it's 1x1 conv
+            for(index_t i = 0; i < NDimSpatial; ++i)
+            {
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
+                const index_t LeftPad  = arg.input_left_pads_[i];
+                const index_t RightPad = arg.input_right_pads_[i];
+
+                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
+                {
+                    return false;
+                }
+            }
+        }
+
+        // check vector access of A
+        // FIXME: layout
+        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
+                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
+                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
+                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
+                     is_same_v<ALayout, ctc::NDHWGC>)
+        {
+            const index_t C = arg.a_g_n_c_wis_lengths_[2];
+
+            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check vector access of B
+        // FIXME: layout
+        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
+                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
+                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
+                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
+                     is_same_v<BLayout, ctc::KZYXGC>)
+
+        {
+            const index_t C = arg.b_g_k_c_xs_lengths_[2];
+
+            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        //  check vector access of Ds
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+            // FIXME: layout
+            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
+                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
+                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
+                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
+                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
+                         is_same_v<DLayout, ctc::G_K>)
+            {
+                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
+
+                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+                {
+                    valid = false;
+                }
+            }
+            else
+            {
+                valid = false;
+            }
+        });
+
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check vector access of E
+        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
+                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
+                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
+                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
+                     is_same_v<ELayout, ctc::NDHWGK>)
+        {
+            const index_t K = arg.e_g_n_k_wos_lengths_[2];
+
+            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
+            {
+                return false;
+            }
+        }
+        else
+        {
+            return false;
+        }
+
+        // check Gridwise GEMM
+        if constexpr(isMultiA || isMultiB)
+        {
+            // Genarate tuples with the same descriptors
+            const auto as_grid_desc_ak0_m_ak1 =
+                generate_tuple([&](auto) { return arg.a_grid_desc_m_k_; }, Number<NumATensor>{});
+            const auto bs_grid_desc_bk0_n_bk1 =
+                generate_tuple([&](auto) { return arg.b_grid_desc_n_k_; }, Number<NumBTensor>{});
+            return GridwiseGemm::CheckValidity(as_grid_desc_ak0_m_ak1,
+                                               bs_grid_desc_bk0_n_bk1,
+                                               arg.ds_grid_desc_m_n_,
+                                               arg.e_grid_desc_m_n_,
+                                               arg.block_2_etile_map_);
+        }
+        else
+        {
+            return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
+                                               arg.b_grid_desc_n_k_,
+                                               arg.ds_grid_desc_m_n_,
+                                               arg.e_grid_desc_m_n_,
+                                               arg.block_2_etile_map_);
+        }
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
+    }
+
+    static auto MakeArgument(
+        APointers p_as,
+        BPointers p_bs,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op)
+    {
+        return Argument{p_as,
+                        p_bs,
+                        p_ds,
+                        p_e,
+                        a_g_n_c_wis_lengths,
+                        a_g_n_c_wis_strides,
+                        b_g_k_c_xs_lengths,
+                        b_g_k_c_xs_strides,
+                        ds_g_n_k_wos_lengths,
+                        ds_g_n_k_wos_strides,
+                        e_g_n_k_wos_lengths,
+                        e_g_n_k_wos_strides,
+                        conv_filter_strides,
+                        conv_filter_dilations,
+                        input_left_pads,
+                        input_right_pads,
+                        a_element_op,
+                        b_element_op,
+                        cde_element_op};
+    }
+
+    static auto MakeInvoker() { return Invoker{}; }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(
+        APointers p_a,
+        BPointers p_b,
+        const std::array<const void*, NumDTensor>& p_ds,
+        void* p_e,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
+        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
+        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
+        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
+        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 AElementwiseOperation& a_element_op,
+        const BElementwiseOperation& b_element_op,
+        const CDEElementwiseOperation& cde_element_op) override
+    {
+        return std::make_unique<Argument>(p_a,
+                                          p_b,
+                                          p_ds,
+                                          p_e,
+                                          a_g_n_c_wis_lengths,
+                                          a_g_n_c_wis_strides,
+                                          b_g_k_c_xs_lengths,
+                                          b_g_k_c_xs_strides,
+                                          ds_g_n_k_wos_lengths,
+                                          ds_g_n_k_wos_strides,
+                                          e_g_n_k_wos_lengths,
+                                          e_g_n_k_wos_strides,
+                                          conv_filter_strides,
+                                          conv_filter_dilations,
+                                          input_left_pads,
+                                          input_right_pads,
+                                          a_element_op,
+                                          b_element_op,
+                                          cde_element_op);
+    }
+
+    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>(Invoker{});
+    }
+
+    std::string GetTypeString() const override
+    {
+        auto str = std::stringstream();
+
+        // clang-format off
+        str << "DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle"
+            << "<"
+            << BlockSize << ", "
+            << MPerBlock << ", "
+            << NPerBlock << ", "
+            << KPerBlock << ", "
+            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
+            << MPerXDL << ", "
+            << NPerXDL << ", "
+            << MXdlPerWave << ", "
+            << NXdlPerWave << ", "
+            << ABlockTransferSrcScalarPerVector << ", "
+            << BBlockTransferSrcScalarPerVector << ", "
+            << CDEBlockTransferScalarPerVector_NPerBlock << ", "
+            << CShuffleMXdlPerWavePerShuffle << ", "
+            << CShuffleNXdlPerWavePerShuffle
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -834,7 +834,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
             // check if it's 1x1, stride=1 conv
             for(index_t i = 0; i < NDimSpatial; ++i)
             {
-                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
                 const index_t ConvStride = arg.conv_filter_strides_[i];
                 const index_t LeftPad    = arg.input_left_pads_[i];
                 const index_t RightPad   = arg.input_right_pads_[i];
@@ -851,7 +851,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
             // check if it's 1x1 conv
             for(index_t i = 0; i < NDimSpatial; ++i)
             {
-                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
                 const index_t LeftPad  = arg.input_left_pads_[i];
                 const index_t RightPad = arg.input_right_pads_[i];
 
@@ -1090,7 +1090,7 @@ struct DeviceGroupedConvFwdMultipleDMultipleR_Xdl_CShuffle
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
+        str << "DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle"
             << "<"
             << BlockSize << ", "
             << MPerBlock << ", "

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

@@ -15,10 +15,11 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
 #include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_wmma_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
 #include "ck/host_utility/device_prop.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 #include "ck/host_utility/io.hpp"
@@ -27,55 +28,6 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 
-namespace {
-
-template <index_t NumDTensor>
-struct ComputePtrOffsetOfStridedBatch
-{
-    ComputePtrOffsetOfStridedBatch() = default;
-
-    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
-                                   index_t BatchStrideB,
-                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
-                                   index_t BatchStrideE)
-        : BatchStrideA_(BatchStrideA),
-          BatchStrideB_(BatchStrideB),
-          BatchStrideDs_(BatchStrideDs),
-          BatchStrideE_(BatchStrideE)
-    {
-    }
-
-    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideA_);
-    }
-
-    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideB_);
-    }
-
-    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
-    {
-        Array<long_index_t, NumDTensor> ds_offset;
-        static_for<0, NumDTensor, 1>{}(
-            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
-        return ds_offset;
-    }
-
-    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideE_);
-    }
-
-    index_t BatchStrideA_;
-    index_t BatchStrideB_;
-    Array<ck::index_t, NumDTensor> BatchStrideDs_;
-    index_t BatchStrideE_;
-};
-
-} // namespace
-
 //
 // @brief      Device Convolution operation.
 //
@@ -140,18 +92,18 @@ template <index_t NDimSpatial,
           LoopScheduler LoopSched         = make_default_loop_scheduler(),
           ck::PipelineVersion PipelineVer = ck::PipelineVersion::v1>
 struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
-    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
-                                           ALayout,
-                                           BLayout,
-                                           DsLayout,
-                                           ELayout,
-                                           ADataType,
-                                           BDataType,
-                                           DsDataType,
-                                           EDataType,
-                                           AElementwiseOperation,
-                                           BElementwiseOperation,
-                                           CDEElementwiseOperation>
+    : public DeviceGroupedConvFwdMultipleABD<NDimSpatial,
+                                             ALayout,
+                                             BLayout,
+                                             DsLayout,
+                                             ELayout,
+                                             ADataType,
+                                             BDataType,
+                                             DsDataType,
+                                             EDataType,
+                                             AElementwiseOperation,
+                                             BElementwiseOperation,
+                                             CDEElementwiseOperation>
 {
     using DeviceOp = DeviceGroupedConvFwdMultipleD_Wmma_CShuffle;
 
@@ -476,7 +428,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
         typename GridwiseOp::DefaultBlock2CTileMap block_2_etile_map_;
 
         // for computing batch offset
-        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
+        ComputePtrOffsetOfStridedBatch<I1, I1, NumDTensor> compute_ptr_offset_of_batch_;
 
         // element-wise op
         AElementwiseOperation a_element_op_;
@@ -519,7 +471,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
             auto launch_kernel = [&](auto has_main_k_block_loop) {
                 constexpr bool has_main_loop = has_main_k_block_loop.value;
 
-                const auto kernel = kernel_grouped_conv_fwd_multiple_d_wmma_cshuffle<
+                const auto kernel = kernel_grouped_conv_multiple_d_wmma_cshuffle<
                     GridwiseOp,
                     ADataType,
                     BDataType,
@@ -533,7 +485,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
                     typename GridwiseOp::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
                     typename GridwiseOp::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
                     remove_reference_t<typename GridwiseOp::DefaultBlock2CTileMap>,
-                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
+                    ComputePtrOffsetOfStridedBatch<I1, I1, NumDTensor>,
                     has_main_loop>;
 
                 return launch_and_time_kernel(stream_config,
@@ -599,7 +551,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
             // check if it's 1x1, stride=1 conv
             for(index_t i = 0; i < NDimSpatial; ++i)
             {
-                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 3];
                 const index_t ConvStride = arg.conv_filter_strides_[i];
                 const index_t LeftPad    = arg.input_left_pads_[i];
                 const index_t RightPad   = arg.input_right_pads_[i];
@@ -616,7 +568,7 @@ struct DeviceGroupedConvFwdMultipleD_Wmma_CShuffle
             // check if it's 1x1 conv
             for(index_t i = 0; i < NDimSpatial; ++i)
             {
-                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
+                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 3];
                 const index_t LeftPad  = arg.input_left_pads_[i];
                 const index_t RightPad = arg.input_right_pads_[i];
 

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

@@ -3,200 +3,20 @@
 
 #pragma once
 
-#include <functional>
-#include <iostream>
-#include <iterator>
-#include <numeric>
-#include <sstream>
-
-#include "ck/utility/common_header.hpp"
-#include "ck/tensor_description/tensor_descriptor.hpp"
-#include "ck/tensor_description/tensor_descriptor_helper.hpp"
-#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_abd_xdl_cshuffle.hpp"
+#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_abd.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
-#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
-#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
-#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp"
-#include "ck/host_utility/device_prop.hpp"
-#include "ck/host_utility/kernel_launch.hpp"
-#include "ck/host_utility/io.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
-namespace {
-
-template <index_t NumDTensor>
-struct ComputePtrOffsetOfStridedBatch
-{
-    ComputePtrOffsetOfStridedBatch() = default;
-
-    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
-                                   index_t BatchStrideB,
-                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
-                                   index_t BatchStrideE)
-        : BatchStrideA_(BatchStrideA),
-          BatchStrideB_(BatchStrideB),
-          BatchStrideDs_(BatchStrideDs),
-          BatchStrideE_(BatchStrideE)
-    {
-    }
-
-    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideA_);
-    }
-
-    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideB_);
-    }
-
-    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
-    {
-        Array<long_index_t, NumDTensor> ds_offset;
-        static_for<0, NumDTensor, 1>{}(
-            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
-        return ds_offset;
-    }
-
-    __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
-    {
-        return g_idx * static_cast<long_index_t>(BatchStrideE_);
-    }
-
-    index_t BatchStrideA_;
-    index_t BatchStrideB_;
-    Array<ck::index_t, NumDTensor> BatchStrideDs_;
-    index_t BatchStrideE_;
-};
-
-/*
- * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM.
- *
- * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix
- * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly
- * strided batched, but we can easily extend to other layouts. The returned offset can be either \p
- * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB
- * limitations.
- *
- * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and
- * returns the 2D index of the tile that it computes. \see
- * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run().
- *
- * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
- * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
- * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
- * impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
- * \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
- * computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
- *
- * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
- * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
- * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion).
- *
- */
-template <typename GridwiseGemm,
-          typename ABDataType,
-          typename DsPointer,
-          typename EDataType,
-          typename AElementwiseOperation,
-          typename BElementwiseOperation,
-          typename CDEElementwiseOperation,
-          typename AGridDesc_AK0_M_AK1,
-          typename BGridDesc_BK0_N_BK1,
-          typename DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
-          typename EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
-          typename Block2ETileMap,
-          typename ComputePtrOffsetOfBatch,
-          bool HasMainKBlockLoop>
-__global__ void
-#if CK_USE_LAUNCH_BOUNDS
-    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
-#endif
-        kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle(
-            const ABDataType* __restrict__ p_a_grid,
-            const ABDataType* __restrict__ p_b_grid,
-            DsPointer p_ds_grid,
-            EDataType* __restrict__ p_e_grid,
-            const AElementwiseOperation a_element_op,
-            const BElementwiseOperation b_element_op,
-            const CDEElementwiseOperation cde_element_op,
-            const index_t batch_count,
-            const AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1,
-            const BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1,
-            const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
-                ds_grid_desc_mblock_mperblock_nblock_nperblock,
-            const EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
-                e_grid_desc_mblock_mperblock_nblock_nperblock_,
-            const Block2ETileMap block_2_ctile_map,
-            const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch)
-{
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
-    defined(__gfx940__) || defined(__gfx941__) || defined(__gfx942__))
-    // offset base pointer for each work-group
-    const index_t num_blocks_per_batch =
-        __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count);
-    const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch);
-
-    const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane(
-        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx)));
-    const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane(
-        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx)));
-    const long_index_t e_batch_offset = __builtin_amdgcn_readfirstlane(
-        static_cast<long_index_t>(compute_ptr_offset_of_batch.GetEPtrOffset(g_idx)));
-
-    const auto ds_batch_offset = compute_ptr_offset_of_batch.GetDsPtrOffset(g_idx);
-
-    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
-
-    DsPointer p_ds_grid_grp;
-
-    static constexpr index_t NumDTensor =
-        DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock::Size();
-
-    static_for<0, NumDTensor, 1>{}(
-        [&](auto i) { p_ds_grid_grp(i) = p_ds_grid[i] + ds_batch_offset[i]; });
-
-    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid + a_batch_offset,
-                                                  p_b_grid + b_batch_offset,
-                                                  p_ds_grid_grp,
-                                                  p_e_grid + e_batch_offset,
-                                                  p_shared,
-                                                  a_element_op,
-                                                  b_element_op,
-                                                  cde_element_op,
-                                                  a_grid_desc_k0_m_k1,
-                                                  b_grid_desc_k0_n_k1,
-                                                  ds_grid_desc_mblock_mperblock_nblock_nperblock,
-                                                  e_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                                  block_2_ctile_map);
-#else
-    ignore = p_a_grid;
-    ignore = p_b_grid;
-    ignore = p_ds_grid;
-    ignore = p_e_grid;
-    ignore = batch_count;
-    ignore = a_grid_desc_k0_m_k1;
-    ignore = b_grid_desc_k0_n_k1;
-    ignore = ds_grid_desc_mblock_mperblock_nblock_nperblock;
-    ignore = e_grid_desc_mblock_mperblock_nblock_nperblock_;
-    ignore = a_element_op;
-    ignore = b_element_op;
-    ignore = cde_element_op;
-    ignore = compute_ptr_offset_of_batch;
-    ignore = block_2_ctile_map;
-#endif
-}
-
-} // namespace
-
 //
 // @brief      Device Convolution operation.
-//
+// @note This structure is deprecated (left for backwards compatibility). Please use
+//       DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle.
 // Supports:
 //  @li         Forward convolution with up to 3 spatial dimentions
 //  @li         Input tensor in GNWC data format
@@ -255,715 +75,61 @@ template <index_t NDimSpatial,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          typename ComputeDataType =
+              decltype(UnpackDataType<is_detected<is_tuple, ADataType>::value,
+                                      Number<0>,
+                                      ADataType>()), // ComputeType is InputType by default (first
+                                                     // in tuple for MultiAB), unpack if tuple was
+                                                     // passed
           LoopScheduler LoopSched = make_default_loop_scheduler()>
-struct DeviceGroupedConvFwdMultipleD_Xdl_CShuffle
-    : public DeviceGroupedConvFwdMultipleD<NDimSpatial,
-                                           ALayout,
-                                           BLayout,
-                                           DsLayout,
-                                           ELayout,
-                                           ADataType,
-                                           BDataType,
-                                           DsDataType,
-                                           EDataType,
-                                           AElementwiseOperation,
-                                           BElementwiseOperation,
-                                           CDEElementwiseOperation>
-{
-    using DeviceOp = DeviceGroupedConvFwdMultipleD_Xdl_CShuffle;
-
-    static constexpr index_t NumDTensor = DsDataType::Size();
-
-    static constexpr auto I0 = Number<0>{};
-    static constexpr auto I1 = Number<1>{};
-    static constexpr auto I2 = Number<2>{};
-    static constexpr auto I3 = Number<3>{};
-
-    static constexpr auto conv_to_gemm_transformer =
-        TransformConvFwdToGemm<NDimSpatial, ConvForwardSpecialization>{};
-
-    static constexpr auto matrix_padder =
-        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
-
-    template <typename ALay>
-    static auto
-    MakeAGridDescriptor_M_K(const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
-                            const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
-                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
-                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
-                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
-                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 auto in_gemmmraw_gemmkraw_desc =
-            conv_to_gemm_transformer.template MakeADescriptor_M_K<ALay>(a_g_n_c_wis_lengths,
-                                                                        a_g_n_c_wis_strides,
-                                                                        b_g_k_c_xs_lengths,
-                                                                        b_g_k_c_xs_strides,
-                                                                        e_g_n_k_wos_lengths,
-                                                                        e_g_n_k_wos_strides,
-                                                                        conv_filter_strides,
-                                                                        conv_filter_dilations,
-                                                                        input_left_pads,
-                                                                        input_right_pads);
-
-        const auto in_gemmm_gemmk_desc =
-            matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc);
-
-        return in_gemmm_gemmk_desc;
-    }
-
-    template <typename BLay>
-    static auto
-    MakeBGridDescriptor_N_K(const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
-                            const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides)
-    {
-        const auto wei_gemmnraw_gemmkraw_desc =
-            conv_to_gemm_transformer.template MakeBDescriptor_N_K<BLay>(b_g_k_c_xs_lengths,
-                                                                        b_g_k_c_xs_strides);
-
-        const auto wei_gemmn_gemmk_desc =
-            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc);
-
-        return wei_gemmn_gemmk_desc;
-    }
-
-    template <typename ELay>
-    static auto
-    MakeEGridDescriptor_M_N(const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
-                            const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_strides)
-    {
-        const auto out_gemmmraw_gemmnraw_desc =
-            conv_to_gemm_transformer.template MakeCDescriptor_M_N<ELay>(e_g_n_k_wos_lengths,
-                                                                        e_g_n_k_wos_strides);
-
-        const auto out_gemmm_gemmn_desc =
-            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc);
-
-        return out_gemmm_gemmn_desc;
-    }
-
-    static auto MakeDsGridDescriptor_M_N(
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides)
-    {
-        return generate_tuple(
-            [&](auto i) {
-                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
-
-                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(ds_g_n_k_wos_lengths[i],
-                                                                  ds_g_n_k_wos_strides[i]);
-            },
-            Number<NumDTensor>{});
-    }
-
-    // desc for problem definition
-    using AGridDesc_M_K  = remove_cvref_t<decltype(MakeAGridDescriptor_M_K<ALayout>(
-        {}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>;
-    using BGridDesc_N_K  = remove_cvref_t<decltype(MakeBGridDescriptor_N_K<BLayout>({}, {}))>;
-    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}))>;
-    using EGridDesc_M_N  = remove_cvref_t<decltype(MakeEGridDescriptor_M_N<ELayout>({}, {}))>;
-
-    using ComputeDataType = ADataType;
-
-    // GridwiseGemm
-    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
-        ADataType, // TODO: distinguish A/B datatype
-        BDataType,
-        ComputeDataType,
-        AccDataType,
-        CShuffleDataType,
-        DsDataType,
-        EDataType,
-        AElementwiseOperation,
-        BElementwiseOperation,
-        CDEElementwiseOperation,
-        InMemoryDataOperationEnum::Set,
-        NumGemmKPrefetchStage,
-        BlockSize,
-        MPerBlock,
-        NPerBlock,
-        KPerBlock,
-        AK1,
-        BK1,
-        MPerXDL,
-        NPerXDL,
-        MXdlPerWave,
-        NXdlPerWave,
-        ABlockTransferThreadClusterLengths_AK0_M_AK1,
-        ABlockTransferThreadClusterArrangeOrder,
-        ABlockTransferSrcAccessOrder,
-        ABlockTransferSrcVectorDim,
-        ABlockTransferSrcScalarPerVector,
-        ABlockTransferDstScalarPerVector_AK1,
-        false,
-        ABlockLdsExtraM,
-        BBlockTransferThreadClusterLengths_BK0_N_BK1,
-        BBlockTransferThreadClusterArrangeOrder,
-        BBlockTransferSrcAccessOrder,
-        BBlockTransferSrcVectorDim,
-        BBlockTransferSrcScalarPerVector,
-        BBlockTransferDstScalarPerVector_BK1,
-        false,
-        BBlockLdsExtraN,
-        CShuffleMXdlPerWavePerShuffle,
-        CShuffleNXdlPerWavePerShuffle,
-        CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
-        CDEBlockTransferScalarPerVector_NPerBlock,
-        LoopSched>;
-
-    // desc for blockwise copy
-    using AGridDesc_AK0_M_AK1 =
-        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(
-            AGridDesc_M_K{}))>;
-    using BGridDesc_BK0_N_BK1 =
-        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(
-            BGridDesc_N_K{}))>;
-    using DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<
-        decltype(GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-            DsGridDesc_M_N{}))>;
-    using EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
-        remove_cvref_t<decltype(GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-            EGridDesc_M_N{}))>;
-
-    // block-to-e-tile map
-    using Block2ETileMap =
-        remove_cvref_t<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
-
-    // Argument
-    struct Argument : public BaseArgument
-    {
-        Argument(const void* p_a,
-                 const void* p_b,
-                 const std::array<const void*, NumDTensor>& p_ds,
-                 void* p_e,
-                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
-                 const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
-                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
-                 const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
-                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
-                     ds_g_n_k_wos_lengths,
-                 const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>&
-                     ds_g_n_k_wos_strides,
-                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
-                 const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 AElementwiseOperation& a_element_op,
-                 const BElementwiseOperation& b_element_op,
-                 const CDEElementwiseOperation& cde_element_op)
-            : p_a_grid_{static_cast<const ADataType*>(p_a)},
-              p_b_grid_{static_cast<const BDataType*>(p_b)},
-              p_ds_grid_{},
-              p_e_grid_{static_cast<EDataType*>(p_e)},
-              num_group_{a_g_n_c_wis_lengths[0]},
-              a_grid_desc_m_k_{DeviceOp::MakeAGridDescriptor_M_K<ALayout>(a_g_n_c_wis_lengths,
-                                                                          a_g_n_c_wis_strides,
-                                                                          b_g_k_c_xs_lengths,
-                                                                          b_g_k_c_xs_strides,
-                                                                          e_g_n_k_wos_lengths,
-                                                                          e_g_n_k_wos_strides,
-                                                                          conv_filter_strides,
-                                                                          conv_filter_dilations,
-                                                                          input_left_pads,
-                                                                          input_right_pads)},
-              b_grid_desc_n_k_{DeviceOp::MakeBGridDescriptor_N_K<BLayout>(b_g_k_c_xs_lengths,
-                                                                          b_g_k_c_xs_strides)},
-              ds_grid_desc_m_n_{},
-              e_grid_desc_m_n_{DeviceOp::MakeEGridDescriptor_M_N<ELayout>(e_g_n_k_wos_lengths,
-                                                                          e_g_n_k_wos_strides)},
-              a_grid_desc_ak0_m_ak1_{
-                  GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k_)},
-              b_grid_desc_bk0_n_bk1_{
-                  GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
-              ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
-              e_grid_desc_mblock_mperblock_nblock_nperblock_{},
-              block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)},
-              compute_ptr_offset_of_batch_{},
-              a_element_op_{a_element_op},
-              b_element_op_{b_element_op},
-              cde_element_op_{cde_element_op},
-              a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths},
-              a_g_n_c_wis_strides_{a_g_n_c_wis_strides},
-              b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths},
-              b_g_k_c_xs_strides_{b_g_k_c_xs_strides},
-              ds_g_n_k_wos_lengths_{ds_g_n_k_wos_lengths},
-              ds_g_n_k_wos_strides_{ds_g_n_k_wos_strides},
-              e_g_n_k_wos_lengths_{e_g_n_k_wos_lengths},
-              e_g_n_k_wos_strides_{e_g_n_k_wos_strides},
-              conv_filter_strides_{conv_filter_strides},
-              conv_filter_dilations_{conv_filter_dilations},
-              input_left_pads_{input_left_pads},
-              input_right_pads_{input_right_pads}
-        {
-            // A/B/E Batch Stride
-            compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0];
-            compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0];
-            compute_ptr_offset_of_batch_.BatchStrideE_ = e_g_n_k_wos_strides[0];
-
-            // populate pointer, batch stride, desc for Ds
-            static_for<0, NumDTensor, 1>{}([&](auto i) {
-                using DLayout   = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
-                using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
-
-                // D pointer
-                p_ds_grid_(i) = static_cast<const DDataType*>(p_ds[i]);
-
-                // D batch stride
-                compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_k_wos_strides[i][0];
-
-                // D desc
-                ds_grid_desc_m_n_(i) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
-                    ds_g_n_k_wos_lengths[i], ds_g_n_k_wos_strides[i]);
-            });
-
-            // populate desc for Ds/E
-            if(GridwiseGemm::CheckValidity(a_grid_desc_m_k_,
-                                           b_grid_desc_n_k_,
-                                           ds_grid_desc_m_n_,
-                                           e_grid_desc_m_n_,
-                                           block_2_etile_map_))
-            {
-                e_grid_desc_mblock_mperblock_nblock_nperblock_ =
-                    GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                        e_grid_desc_m_n_);
-
-                ds_grid_desc_mblock_mperblock_nblock_nperblock_ =
-                    GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                        ds_grid_desc_m_n_);
-            }
-        }
-
-        void Print() const
-        {
-            std::cout << "A[M, K]: " << a_grid_desc_m_k_ << std::endl;
-            std::cout << "B[N, K]: " << b_grid_desc_n_k_ << std::endl;
-            static_for<0, NumDTensor, 1>{}(
-                [&](auto i) { std::cout << "Ds[M, N]: " << ds_grid_desc_m_n_[i] << std::endl; });
-            std::cout << "E[M, N]: " << e_grid_desc_m_n_ << std::endl;
-        }
-
-        //  private:
-        // pointers
-        const ADataType* p_a_grid_;
-        const BDataType* p_b_grid_;
-        typename GridwiseGemm::DsGridPointer p_ds_grid_;
-        EDataType* p_e_grid_;
-
-        // tensor descriptors for problem definiton
-        index_t num_group_;
-        AGridDesc_M_K a_grid_desc_m_k_;
-        BGridDesc_N_K b_grid_desc_n_k_;
-        DsGridDesc_M_N ds_grid_desc_m_n_;
-        EGridDesc_M_N e_grid_desc_m_n_;
-
-        // tensor descriptors for block/thread-wise copy
-        AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
-        BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
-        DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
-            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
-        EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock e_grid_desc_mblock_mperblock_nblock_nperblock_;
-
-        // block-to-e-tile map
-        Block2ETileMap block_2_etile_map_;
-
-        // for computing batch offset
-        ComputePtrOffsetOfStridedBatch<NumDTensor> compute_ptr_offset_of_batch_;
-
-        // element-wise op
-        AElementwiseOperation a_element_op_;
-        BElementwiseOperation b_element_op_;
-        CDEElementwiseOperation cde_element_op_;
-
-        // for checking IsSupportedArgument()
-        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_lengths_;
-        std::array<index_t, NDimSpatial + 3> a_g_n_c_wis_strides_;
-        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_lengths_;
-        std::array<index_t, NDimSpatial + 3> b_g_k_c_xs_strides_;
-        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_lengths_;
-        std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor> ds_g_n_k_wos_strides_;
-        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_lengths_;
-        std::array<index_t, NDimSpatial + 3> e_g_n_k_wos_strides_;
-        std::array<index_t, NDimSpatial> conv_filter_strides_;
-        std::array<index_t, NDimSpatial> conv_filter_dilations_;
-        std::array<index_t, NDimSpatial> input_left_pads_;
-        std::array<index_t, NDimSpatial> input_right_pads_;
-    };
-
-    // Invoker
-    struct Invoker : public BaseInvoker
-    {
-        using Argument = DeviceOp::Argument;
-
-        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
-        {
-            if(stream_config.log_level_ > 0)
-            {
-                arg.Print();
-            }
-
-            if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
-                                            arg.b_grid_desc_n_k_,
-                                            arg.ds_grid_desc_m_n_,
-                                            arg.e_grid_desc_m_n_,
-                                            arg.block_2_etile_map_))
-            {
-                throw std::runtime_error(
-                    "wrong! GridwiseGemmMultipleD_xdl_cshuffle has invalid setting");
-            }
-
-            const index_t grid_size =
-                arg.block_2_etile_map_.CalculateGridSize(arg.e_grid_desc_m_n_) * arg.num_group_;
-
-            const auto K =
-                arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) * arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
-
-            auto launch_kernel = [&](auto has_main_k_block_loop) {
-                constexpr bool has_main_loop = has_main_k_block_loop.value;
-
-                const auto kernel = kernel_grouped_conv_fwd_multiple_d_xdl_cshuffle<
-                    GridwiseGemm,
-                    ADataType, // TODO: distiguish A/B datatype
-                    typename GridwiseGemm::DsGridPointer,
-                    EDataType,
-                    AElementwiseOperation,
-                    BElementwiseOperation,
-                    CDEElementwiseOperation,
-                    DeviceOp::AGridDesc_AK0_M_AK1,
-                    DeviceOp::BGridDesc_BK0_N_BK1,
-                    DeviceOp::DsGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
-                    DeviceOp::EGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
-                    Block2ETileMap,
-                    ComputePtrOffsetOfStridedBatch<NumDTensor>,
-                    has_main_loop>;
-
-                return launch_and_time_kernel(stream_config,
-                                              kernel,
-                                              dim3(grid_size),
-                                              dim3(BlockSize),
-                                              0,
-                                              arg.p_a_grid_,
-                                              arg.p_b_grid_,
-                                              arg.p_ds_grid_,
-                                              arg.p_e_grid_,
-                                              arg.a_element_op_,
-                                              arg.b_element_op_,
-                                              arg.cde_element_op_,
-                                              arg.a_g_n_c_wis_lengths_[0], // Group count
-                                              arg.a_grid_desc_ak0_m_ak1_,
-                                              arg.b_grid_desc_bk0_n_bk1_,
-                                              arg.ds_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                              arg.e_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                              arg.block_2_etile_map_,
-                                              arg.compute_ptr_offset_of_batch_);
-            };
-
-            if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
-            {
-                return launch_kernel(integral_constant<bool, true>{});
-            }
-            else
-            {
-                return launch_kernel(integral_constant<bool, false>{});
-            }
-        }
-
-        float Run(const BaseArgument* p_arg,
-                  const StreamConfig& stream_config = StreamConfig{}) override
-        {
-            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
-        }
-    };
-
-    static bool IsSupportedArgument(const Argument& arg)
-    {
-        namespace ctc = tensor_layout::convolution;
-
-        // check device
-        if(get_device_name() == "gfx908")
-        {
-            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
-                           is_same_v<AccDataType, int32_t>))
-            {
-                return false;
-            }
-        }
-        else if(get_device_name() == "gfx90a" || get_device_name() == "gfx940" ||
-                get_device_name() == "gfx941" || get_device_name() == "gfx942")
-        {
-            if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
-                           is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
-            {
-                return false;
-            }
-        }
-        else
-        {
-            return false;
-        }
-
-        // check ConvolutionForwardSpecialization
-        if constexpr(ConvForwardSpecialization ==
-                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
-        {
-            // check if it's 1x1, stride=1 conv
-            for(index_t i = 0; i < NDimSpatial; ++i)
-            {
-                const index_t X          = arg.b_g_k_c_xs_lengths_[i + 2];
-                const index_t ConvStride = arg.conv_filter_strides_[i];
-                const index_t LeftPad    = arg.input_left_pads_[i];
-                const index_t RightPad   = arg.input_right_pads_[i];
-
-                if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0))
-                {
-                    return false;
-                }
-            }
-        }
-        else if constexpr(ConvForwardSpecialization ==
-                          ConvolutionForwardSpecialization::Filter1x1Pad0)
-        {
-            // check if it's 1x1 conv
-            for(index_t i = 0; i < NDimSpatial; ++i)
-            {
-                const index_t X        = arg.b_g_k_c_xs_lengths_[i + 2];
-                const index_t LeftPad  = arg.input_left_pads_[i];
-                const index_t RightPad = arg.input_right_pads_[i];
-
-                if(!(X == 1 && LeftPad == 0 && RightPad == 0))
-                {
-                    return false;
-                }
-            }
-        }
-
-        // check vector access of A
-        // FIXME: layout
-        if constexpr(is_same_v<ALayout, ctc::G_NW_C> || is_same_v<ALayout, ctc::G_NHW_C> ||
-                     is_same_v<ALayout, ctc::G_NDHW_C> || is_same_v<ALayout, ctc::GNWC> ||
-                     is_same_v<ALayout, ctc::GNHWC> || is_same_v<ALayout, ctc::GNDHWC> ||
-                     is_same_v<ALayout, ctc::NWGC> || is_same_v<ALayout, ctc::NHWGC> ||
-                     is_same_v<ALayout, ctc::NDHWGC>)
-        {
-            const index_t C = arg.a_g_n_c_wis_lengths_[2];
-
-            if(!(ABlockTransferSrcVectorDim == 2 && C % ABlockTransferSrcScalarPerVector == 0))
-            {
-                return false;
-            }
-        }
-        else
-        {
-            return false;
-        }
-
-        // check vector access of B
-        // FIXME: layout
-        if constexpr(is_same_v<BLayout, ctc::G_K_X_C> || is_same_v<BLayout, ctc::G_K_YX_C> ||
-                     is_same_v<BLayout, ctc::G_K_ZYX_C> || is_same_v<BLayout, ctc::GKXC> ||
-                     is_same_v<BLayout, ctc::GKYXC> || is_same_v<BLayout, ctc::GKZYXC> ||
-                     is_same_v<BLayout, ctc::KXGC> || is_same_v<BLayout, ctc::KYXGC> ||
-                     is_same_v<BLayout, ctc::KZYXGC>)
-
-        {
-            const index_t C = arg.b_g_k_c_xs_lengths_[2];
-
-            if(!(BBlockTransferSrcVectorDim == 2 && C % BBlockTransferSrcScalarPerVector == 0))
-            {
-                return false;
-            }
-        }
-        else
-        {
-            return false;
-        }
-
-        //  check vector access of Ds
-        bool valid = true;
-
-        static_for<0, NumDTensor, 1>{}([&](auto i) {
-            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
-
-            // FIXME: layout
-            if constexpr(is_same_v<DLayout, ctc::G_NW_K> || is_same_v<DLayout, ctc::G_NHW_K> ||
-                         is_same_v<DLayout, ctc::G_NDHW_K> || is_same_v<DLayout, ctc::GNWK> ||
-                         is_same_v<DLayout, ctc::GNHWK> || is_same_v<DLayout, ctc::GNDHWK> ||
-                         is_same_v<DLayout, ctc::NWGK> || is_same_v<DLayout, ctc::NHWGK> ||
-                         is_same_v<DLayout, ctc::NDHWGK> || is_same_v<DLayout, ctc::GK> ||
-                         is_same_v<DLayout, ctc::G_K>)
-            {
-                const index_t K = arg.ds_g_n_k_wos_lengths_[i][2];
-
-                if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
-                {
-                    valid = false;
-                }
-            }
-            else
-            {
-                valid = false;
-            }
-        });
-
-        if(!valid)
-        {
-            return false;
-        }
-
-        // check vector access of E
-        if constexpr(is_same_v<ELayout, ctc::G_NW_K> || is_same_v<ELayout, ctc::G_NHW_K> ||
-                     is_same_v<ELayout, ctc::G_NDHW_K> || is_same_v<ELayout, ctc::GNWK> ||
-                     is_same_v<ELayout, ctc::GNHWK> || is_same_v<ELayout, ctc::GNDHWK> ||
-                     is_same_v<ELayout, ctc::NWGK> || is_same_v<ELayout, ctc::NHWGK> ||
-                     is_same_v<ELayout, ctc::NDHWGK>)
-        {
-            const index_t K = arg.e_g_n_k_wos_lengths_[2];
-
-            if(!(K % CDEBlockTransferScalarPerVector_NPerBlock == 0))
-            {
-                return false;
-            }
-        }
-        else
-        {
-            return false;
-        }
-
-        // check Gridwise GEMM
-        return GridwiseGemm::CheckValidity(arg.a_grid_desc_m_k_,
-                                           arg.b_grid_desc_n_k_,
-                                           arg.ds_grid_desc_m_n_,
-                                           arg.e_grid_desc_m_n_,
-                                           arg.block_2_etile_map_);
-    }
-
-    bool IsSupportedArgument(const BaseArgument* p_arg) override
-    {
-        return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
-    }
-
-    static auto MakeArgument(
-        const void* p_a,
-        const void* p_b,
-        const std::array<const void*, NumDTensor>& p_ds,
-        void* p_e,
-        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
-        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
-        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
-        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
-        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
-        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 AElementwiseOperation& a_element_op,
-        const BElementwiseOperation& b_element_op,
-        const CDEElementwiseOperation& cde_element_op)
-    {
-        return Argument{p_a,
-                        p_b,
-                        p_ds,
-                        p_e,
-                        a_g_n_c_wis_lengths,
-                        a_g_n_c_wis_strides,
-                        b_g_k_c_xs_lengths,
-                        b_g_k_c_xs_strides,
-                        ds_g_n_k_wos_lengths,
-                        ds_g_n_k_wos_strides,
-                        e_g_n_k_wos_lengths,
-                        e_g_n_k_wos_strides,
-                        conv_filter_strides,
-                        conv_filter_dilations,
-                        input_left_pads,
-                        input_right_pads,
-                        a_element_op,
-                        b_element_op,
-                        cde_element_op};
-    }
-
-    static auto MakeInvoker() { return Invoker{}; }
-
-    std::unique_ptr<BaseArgument> MakeArgumentPointer(
-        const void* p_a,
-        const void* p_b,
-        const std::array<const void*, NumDTensor>& p_ds,
-        void* p_e,
-        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_lengths,
-        const std::array<index_t, NDimSpatial + 3>& a_g_n_c_wis_strides,
-        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_lengths,
-        const std::array<index_t, NDimSpatial + 3>& b_g_k_c_xs_strides,
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_lengths,
-        const std::array<std::array<index_t, NDimSpatial + 3>, NumDTensor>& ds_g_n_k_wos_strides,
-        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_lengths,
-        const std::array<index_t, NDimSpatial + 3>& e_g_n_k_wos_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 AElementwiseOperation& a_element_op,
-        const BElementwiseOperation& b_element_op,
-        const CDEElementwiseOperation& cde_element_op) override
-    {
-        return std::make_unique<Argument>(p_a,
-                                          p_b,
-                                          p_ds,
-                                          p_e,
-                                          a_g_n_c_wis_lengths,
-                                          a_g_n_c_wis_strides,
-                                          b_g_k_c_xs_lengths,
-                                          b_g_k_c_xs_strides,
-                                          ds_g_n_k_wos_lengths,
-                                          ds_g_n_k_wos_strides,
-                                          e_g_n_k_wos_lengths,
-                                          e_g_n_k_wos_strides,
-                                          conv_filter_strides,
-                                          conv_filter_dilations,
-                                          input_left_pads,
-                                          input_right_pads,
-                                          a_element_op,
-                                          b_element_op,
-                                          cde_element_op);
-    }
-
-    std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
-    {
-        return std::make_unique<Invoker>(Invoker{});
-    }
-
-    std::string GetTypeString() const override
-    {
-        auto str = std::stringstream();
-
-        // clang-format off
-        str << "DeviceGroupedConvFwdMultipleD_Xdl_CShuffle"
-            << "<"
-            << BlockSize << ", "
-            << MPerBlock << ", "
-            << NPerBlock << ", "
-            << KPerBlock << ", "
-            << getConvForwardSpecializationString(ConvForwardSpecialization) << ", "
-            << MPerXDL << ", "
-            << NPerXDL << ", "
-            << MXdlPerWave << ", "
-            << NXdlPerWave << ", "
-            << ABlockTransferSrcScalarPerVector << ", "
-            << BBlockTransferSrcScalarPerVector << ", "
-            << CShuffleMXdlPerWavePerShuffle << ", "
-            << CShuffleNXdlPerWavePerShuffle
-            << ">";
-        // clang-format on
-
-        return str.str();
-    }
-};
+using DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle = DeviceGroupedConvFwdMultipleABD_Xdl_CShuffle<
+    NDimSpatial,
+    ALayout,
+    BLayout,
+    DsLayout,
+    ELayout,
+    ADataType,
+    BDataType,
+    AccDataType,
+    CShuffleDataType,
+    DsDataType,
+    EDataType,
+    AElementwiseOperation,
+    BElementwiseOperation,
+    CDEElementwiseOperation,
+    ConvForwardSpecialization,
+    GemmSpec,
+    NumGemmKPrefetchStage,
+    BlockSize,
+    MPerBlock,
+    NPerBlock,
+    KPerBlock,
+    AK1,
+    BK1,
+    MPerXDL,
+    NPerXDL,
+    MXdlPerWave,
+    NXdlPerWave,
+    ABlockTransferThreadClusterLengths_AK0_M_AK1,
+    ABlockTransferThreadClusterArrangeOrder,
+    ABlockTransferSrcAccessOrder,
+    ABlockTransferSrcVectorDim,
+    ABlockTransferSrcScalarPerVector,
+    ABlockTransferDstScalarPerVector_AK1,
+    ABlockLdsExtraM,
+    BBlockTransferThreadClusterLengths_BK0_N_BK1,
+    BBlockTransferThreadClusterArrangeOrder,
+    BBlockTransferSrcAccessOrder,
+    BBlockTransferSrcVectorDim,
+    BBlockTransferSrcScalarPerVector,
+    BBlockTransferDstScalarPerVector_BK1,
+    BBlockLdsExtraN,
+    CShuffleMXdlPerWavePerShuffle,
+    CShuffleNXdlPerWavePerShuffle,
+    CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
+    CDEBlockTransferScalarPerVector_NPerBlock,
+    ComputeDataType,
+    LoopSched>;
 
 } // namespace device
 } // namespace tensor_operation

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/utility/common_header.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <index_t NumATensor = 1, index_t NumBTensor = 1, index_t NumDTensor = 0, typename = void>
+struct ComputePtrOffsetOfStridedBatch
+{
+};
+
+template <index_t NumATensor, index_t NumBTensor, index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch<NumATensor,
+                                      NumBTensor,
+                                      NumDTensor,
+                                      ck::enable_if_t<(NumATensor > 1 || NumBTensor > 1)>>
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(Array<ck::index_t, NumATensor>& BatchStrideAs,
+                                   Array<ck::index_t, NumBTensor>& BatchStrideBs,
+                                   Array<ck::index_t, NumDTensor>& BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideAs),
+          BatchStrideB_(BatchStrideBs),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr auto GetAsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumATensor> as_offset;
+        static_for<0, NumATensor, 1>{}(
+            [&](auto i) { as_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideA_[i]); });
+        return as_offset;
+    }
+
+    __host__ __device__ constexpr auto GetBsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumBTensor> bs_offset;
+        static_for<0, NumBTensor, 1>{}(
+            [&](auto i) { bs_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideB_[i]); });
+        return bs_offset;
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    [[maybe_unused]] __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    // alias for kernels without multiple D
+    [[maybe_unused]] __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    Array<ck::index_t, NumATensor> BatchStrideA_;
+    Array<ck::index_t, NumBTensor> BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+    index_t& BatchStrideC_ = BatchStrideE_; // alias for kernels without multiple D
+};
+
+template <index_t NumATensor, index_t NumBTensor, index_t NumDTensor>
+struct ComputePtrOffsetOfStridedBatch<NumATensor,
+                                      NumBTensor,
+                                      NumDTensor,
+                                      ck::enable_if_t<(NumATensor == 1 && NumBTensor == 1)>>
+{
+    ComputePtrOffsetOfStridedBatch() = default;
+
+    ComputePtrOffsetOfStridedBatch(index_t BatchStrideA,
+                                   index_t BatchStrideB,
+                                   Array<ck::index_t, NumDTensor> BatchStrideDs,
+                                   index_t BatchStrideE)
+        : BatchStrideA_(BatchStrideA),
+          BatchStrideB_(BatchStrideB),
+          BatchStrideDs_(BatchStrideDs),
+          BatchStrideE_(BatchStrideE)
+    {
+    }
+
+    __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideA_);
+    }
+
+    __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideB_);
+    }
+
+    __host__ __device__ constexpr auto GetDsPtrOffset(index_t g_idx) const
+    {
+        Array<long_index_t, NumDTensor> ds_offset;
+        static_for<0, NumDTensor, 1>{}(
+            [&](auto i) { ds_offset(i) = g_idx * static_cast<long_index_t>(BatchStrideDs_[i]); });
+        return ds_offset;
+    }
+
+    [[maybe_unused]] __host__ __device__ constexpr long_index_t GetEPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    // alias for kernels without multiple D
+    [[maybe_unused]] __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const
+    {
+        return g_idx * static_cast<long_index_t>(BatchStrideE_);
+    }
+
+    ck::index_t BatchStrideA_;
+    ck::index_t BatchStrideB_;
+    Array<ck::index_t, NumDTensor> BatchStrideDs_;
+    index_t BatchStrideE_;
+    index_t& BatchStrideC_ = BatchStrideE_; // alias for kernels without multiple D
+};
+
+template <bool isTuple, typename Tensors>
+constexpr static auto GetNumABTensors()
+{
+    if constexpr(isTuple)
+    {
+        return Number<Tensors::Size()>{};
+    }
+    else
+    {
+        return Number<1>{};
+    }
+}
+
+template <bool isTuple, typename GridwiseGemm, typename DataType>
+constexpr static auto GetAGridPointer()
+{
+    if constexpr(isTuple)
+    {
+        return typename GridwiseGemm::AsGridPointer{};
+    }
+    else
+    {
+        return Tuple<const DataType*>{};
+    }
+}
+
+template <bool isTuple, typename GridwiseGemm, typename DataType>
+constexpr static auto GetBGridPointer()
+{
+    if constexpr(isTuple)
+    {
+        return typename GridwiseGemm::BsGridPointer{};
+    }
+    else
+    {
+        return Tuple<const DataType*>{};
+    }
+}
+
+template <bool isTuple, typename Id, typename Type>
+constexpr static auto UnpackDataType()
+{
+    if constexpr(isTuple)
+    {
+        // unpack if tuple
+        return tuple_element_t<Id{}, Type>{};
+    }
+    else
+    {
+        // if no, return Type
+        return Type{};
+    }
+}
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -193,6 +193,7 @@ template <typename ALayout,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
+          typename ComputeType    = ADataType,
           LoopScheduler LoopSched = make_default_loop_scheduler()>
 struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
                                                                         BLayout,
@@ -217,6 +218,8 @@ struct DeviceGroupedGemm_Xdl_Fixed_NK : public DeviceGroupedGemmFixedNK<ALayout,
     // GridwiseGemm
     using GridwiseGemm = GridwiseGemmMultipleD_xdl_splitk_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
+        BDataType,
+        ComputeType,
         AccDataType,
         CShuffleDataType,
         DsDataType,

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

@@ -265,10 +265,10 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                 const index_t stride_b = gemm_descs[i].stride_B_;
                 const index_t stride_c = gemm_descs[i].stride_C_;
 
-                const index_t m_padded = GridwiseGemm::CalculateMPadded(M);
-                const index_t n_padded = GridwiseGemm::CalculateNPadded(N);
-                const index_t k_padded = GridwiseGemm::CalculateKPadded(K, K_BATCH);
-                const index_t k0       = GridwiseGemm::CalculateK0(K, K_BATCH);
+                const index_t m_padded  = GridwiseGemm::CalculateMPadded(M);
+                const index_t n_padded  = GridwiseGemm::CalculateNPadded(N);
+                const index_t k_padded  = GridwiseGemm::CalculateKPadded(K, K_BATCH);
+                const index_t k0_padded = GridwiseGemm::CalculateK0Padded(K, K_BATCH);
 
                 const auto c_grid_desc_m_n = GridwiseGemm::MakeCGridDescriptor_M_N(M, N, stride_c);
 
@@ -297,7 +297,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                                            m_padded,
                                            n_padded,
                                            k_padded,
-                                           k0,
+                                           k0_padded,
                                            K_BATCH};
 
                 gemm_kernel_args_.emplace_back(
@@ -320,8 +320,8 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
 
                 auto& karg = gemm_kernel_args_[i].karg_;
 
-                const index_t k_padded = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
-                const index_t k0       = GridwiseGemm::CalculateK0(karg.K, K_BATCH);
+                const index_t k_padded  = GridwiseGemm::CalculateKPadded(karg.K, K_BATCH);
+                const index_t k0_padded = GridwiseGemm::CalculateK0Padded(karg.K, K_BATCH);
 
                 const auto c_grid_desc_m_n =
                     GridwiseGemm::MakeCGridDescriptor_M_N(karg.M, karg.N, karg.StrideC);
@@ -340,7 +340,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                     GroupedGemmBlock2ETileMap(local_b2c_tile_map, block_start);
 
                 karg.KPadded                            = k_padded;
-                karg.K0                                 = k0;
+                karg.K0Padded                           = k0_padded;
                 karg.k_batch                            = K_BATCH;
                 gemm_kernel_args_[i].block_2_ctile_map_ = grouped_block_2_ctile_map;
                 gemm_kernel_args_[i].block_start_       = block_start;
@@ -362,7 +362,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
     {
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
-            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0;
+            index_t K0                       = arg.gemm_kernel_args_[0].karg_.K0Padded;
             bool all_have_kbatch_gt_one      = arg.gemm_kernel_args_[0].karg_.k_batch > 1;
             bool all_have_main_k0_block_loop = GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
 
@@ -384,7 +384,7 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
                     throw std::runtime_error(err.str());
                 }
 
-                K0 = karg.K0;
+                K0 = karg.K0Padded;
                 bool not_all_have_main_k0_block_loop_same =
                     all_have_main_k0_block_loop xor GridwiseGemm::CalculateHasMainK0BlockLoop(K0);
                 bool not_all_have_kbatch_value_same = all_have_kbatch_gt_one xor (kbatch > 1);

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

@@ -5,65 +5,53 @@
 
 #include "ck/tensor_description/tensor_descriptor.hpp"
 #include "ck/tensor_description/tensor_descriptor_helper.hpp"
-#include "ck/tensor_operation/gpu/device/device_image_to_column.hpp"
-#include "ck/tensor_operation/gpu/grid/gridwise_image_to_column.hpp"
+#include "ck/tensor_operation/gpu/device/device_conv_tensor_rearrange.hpp"
+#include "ck/tensor_operation/gpu/grid/gridwise_tensor_rearrange.hpp"
 #include "ck/host_utility/kernel_launch.hpp"
 #include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
 #include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp"
 #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/conv_tensor_rearrange_op.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_utils.hpp"
 #include "ck/host_utility/io.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 
-template <typename InputGridDesc,
-          typename InputDataType,
-          typename OutputGridDesc,
-          typename OutputDataType,
-          typename Block2ETileMap,
-          typename GridwiseImageToColumnKernel>
-__global__ void
-#if CK_USE_LAUNCH_BOUNDS
-    __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
-#endif
-        kernel_image_to_column(const InputGridDesc in_grid_desc,
-                               const InputDataType* __restrict__ p_in_global,
-                               const OutputGridDesc out_grid_desc,
-                               OutputDataType* __restrict__ p_out_global,
-                               const Block2ETileMap block_2_tile_map)
-{
-#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx908__) ||             \
-    defined(__gfx90a__) || defined(__gfx940__) || defined(__gfx1030__) || defined(__gfx1100__) || \
-    defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx941__) || defined(__gfx942__))
-    GridwiseImageToColumnKernel::Run(
-        in_grid_desc, p_in_global, out_grid_desc, p_out_global, block_2_tile_map);
-#else
-    ignore = in_grid_desc;
-    ignore = p_in_global;
-    ignore = out_grid_desc;
-    ignore = p_out_global;
-    ignore = block_2_tile_map;
-#endif
-}
-
-// Image to column for input layout NDHWC:
-//   input : input image [N, Di, Hi, Wi, C],
-//   output : output image [N * Do * Ho * Wo, Z * Y * X * C]
+// Image to column:
+//   input : input image [G, N, Di, Hi, Wi, C]
+//   output : gemm form [G * N * Do * Ho * Wo, Z * Y * X * C]
+//   input : input image [N, Di, Hi, Wi, G, C]
+//   output : gemm form [N * Do * Ho * Wo * G, Z * Y * X * C]
 template <index_t NDimSpatial,
-          typename InputLayout,
+          typename ImageLayout,
           typename InputDataType,
           typename OutputDataType,
           index_t BlockSize,
           index_t MPerBlock,
           index_t KPerBlock,
           typename ThreadClusterLengths,
-          index_t ScalarPerVector>
+          index_t ScalarPerVector,
+          typename std::enable_if<NDimSpatial >= 1 && NDimSpatial <= 3, bool>::type = false>
 struct DeviceImageToColumnImpl
-    : public DeviceImageToColumn<NDimSpatial, InputLayout, InputDataType, OutputDataType>
+    : public DeviceConvTensorRearrange<NDimSpatial,
+                                       ImageLayout,
+                                       InputDataType,
+                                       OutputDataType,
+                                       conv_tensor_rearrange_op::ImageToColumn>
 {
+    static constexpr bool is_NSpatialGC =
+        std::is_same_v<ImageLayout, tensor_layout::convolution::NWGC> ||
+        std::is_same_v<ImageLayout, tensor_layout::convolution::NHWGC> ||
+        std::is_same_v<ImageLayout, tensor_layout::convolution::NDHWGC>;
+    static constexpr bool is_GNSpatialC =
+        std::is_same_v<ImageLayout, tensor_layout::convolution::GNWC> ||
+        std::is_same_v<ImageLayout, tensor_layout::convolution::GNHWC> ||
+        std::is_same_v<ImageLayout, tensor_layout::convolution::GNDHWC>;
 
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
@@ -83,7 +71,7 @@ struct DeviceImageToColumnImpl
                             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_g_n_c_wis_strides,
+                            const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_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,
@@ -110,9 +98,9 @@ struct DeviceImageToColumnImpl
         c_g_n_k_wos_lengths[I1] = N;
 
         const auto in_gemmmraw_gemmkraw_desc =
-            conv_to_gemm_transformer.template MakeADescriptor_M_K<InputLayout>(
+            conv_to_gemm_transformer.template MakeADescriptor_M_K<ImageLayout>(
                 a_g_n_c_wis_lengths,
-                input_g_n_c_wis_strides,
+                image_g_n_c_wis_strides,
                 b_g_k_c_xs_lengths,
                 {}, // not needed for A Descriptor
                 c_g_n_k_wos_lengths,
@@ -132,7 +120,7 @@ struct DeviceImageToColumnImpl
                           const ck::index_t C,
                           const std::array<index_t, NDimSpatial>& filter_spatial_lengths,
                           const std::array<index_t, NDimSpatial>& output_spatial_lengths,
-                          const std::array<index_t, 2>& output_m_k_strides)
+                          const std::array<index_t, 3>& gemm_g_m_k_strides)
     {
         const index_t NDoHoWo =
             N * ck::accumulate_n<index_t>(
@@ -140,51 +128,55 @@ struct DeviceImageToColumnImpl
         const index_t CZYX =
             C * ck::accumulate_n<index_t>(
                     filter_spatial_lengths.begin(), NDimSpatial, 1, std::multiplies<>());
-        const auto desc_mraw_kraw = make_naive_tensor_descriptor(
-            make_tuple(NDoHoWo, CZYX), make_tuple(output_m_k_strides[I0], output_m_k_strides[I1]));
 
-        const auto desc_m_k = matrix_padder.PadADescriptor_M_K(desc_mraw_kraw);
-        return desc_m_k;
+        const auto desc_mraw_kraw = make_naive_tensor_descriptor(
+            make_tuple(NDoHoWo, CZYX), make_tuple(gemm_g_m_k_strides[I1], gemm_g_m_k_strides[I2]));
+        return matrix_padder.PadADescriptor_M_K(desc_mraw_kraw);
     }
 
     using InputGridDesc =
         remove_cvref_t<decltype(MakeInputDescriptor_M_K(1, 1, {}, {}, {}, {}, {}, {}, {}, {}))>;
     using OutputGridDesc = remove_cvref_t<decltype(MakeOutDescriptor_M_K(1, 1, {}, {}, {}))>;
 
-    using Block2ETileMap = remove_cvref_t<decltype(
-        BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(OutputGridDesc{}))>;
-
-    using GridwiseImageToColumnKernel = GridwiseImageToColumn<InputGridDesc,
-                                                              InputDataType,
-                                                              OutputGridDesc,
-                                                              OutputDataType,
-                                                              BlockSize,
-                                                              MPerBlock,
-                                                              KPerBlock,
-                                                              ThreadClusterLengths,
-                                                              ScalarPerVector,
-                                                              Block2ETileMap>;
+    using Block2ETileMap = remove_cvref_t<
+        decltype(BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
+            OutputGridDesc{}))>;
+
+    using GridwiseTensorRearrangeKernel = GridwiseTensorRearrange<InputGridDesc,
+                                                                  InputDataType,
+                                                                  OutputGridDesc,
+                                                                  OutputDataType,
+                                                                  BlockSize,
+                                                                  MPerBlock,
+                                                                  KPerBlock,
+                                                                  ThreadClusterLengths,
+                                                                  ScalarPerVector,
+                                                                  InMemoryDataOperationEnum::Set,
+                                                                  Block2ETileMap,
+                                                                  ComputePtrOffsetOfStridedBatch<>>;
 
     struct Argument : public BaseArgument
     {
         Argument(const void* p_in, // input image
-                 void* p_out,      // output image
+                 void* p_out,      // gemm form
+                 const ck::index_t G,
                  const ck::index_t N,
                  const ck::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_g_n_c_wis_strides,
-                 const std::array<index_t, 2>& output_m_k_strides,
+                 const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
+                 const std::array<index_t, 3>& gemm_g_m_k_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)
-            : C_(C),
+            : G_(G),
+              C_(C),
               X_(filter_spatial_lengths[NDimSpatial - I1]),
               p_in_{static_cast<const InputDataType*>(p_in)},
               p_out_{static_cast<OutputDataType*>(p_out)},
-              input_g_n_c_wis_strides_{input_g_n_c_wis_strides},
+              image_g_n_c_wis_strides_{image_g_n_c_wis_strides},
               conv_filter_strides_{conv_filter_strides},
               conv_filter_dilations_{conv_filter_dilations},
               input_left_pads_{input_left_pads},
@@ -196,15 +188,17 @@ struct DeviceImageToColumnImpl
                                                         input_spatial_lengths,
                                                         filter_spatial_lengths,
                                                         output_spatial_lengths,
-                                                        input_g_n_c_wis_strides,
-
+                                                        image_g_n_c_wis_strides,
                                                         conv_filter_strides,
                                                         conv_filter_dilations,
                                                         input_left_pads,
                                                         input_right_pads);
 
             out_grid_desc_m_k_ = MakeOutDescriptor_M_K(
-                N, C, filter_spatial_lengths, output_spatial_lengths, output_m_k_strides);
+                N, C, filter_spatial_lengths, output_spatial_lengths, gemm_g_m_k_strides);
+
+            compute_ptr_offset_of_batch_.BatchStrideA_ = image_g_n_c_wis_strides[I0];
+            compute_ptr_offset_of_batch_.BatchStrideC_ = gemm_g_m_k_strides[I0];
         }
 
         void Print() const
@@ -213,13 +207,14 @@ struct DeviceImageToColumnImpl
             std::cout << out_grid_desc_m_k_ << std::endl;
         }
 
+        const ck::index_t G_;
         const ck::index_t C_;
         const ck::index_t X_;
 
         const InputDataType* p_in_;
         OutputDataType* p_out_;
 
-        const std::array<index_t, NDimSpatial + 3>& input_g_n_c_wis_strides_;
+        const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_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_;
@@ -227,6 +222,8 @@ struct DeviceImageToColumnImpl
 
         InputGridDesc in_grid_desc_m_k_;
         OutputGridDesc out_grid_desc_m_k_;
+
+        ComputePtrOffsetOfStridedBatch<> compute_ptr_offset_of_batch_;
     };
 
     struct Invoker : public BaseInvoker
@@ -241,13 +238,15 @@ struct DeviceImageToColumnImpl
             const auto block_2_tile_map =
                 BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, KPerBlock, OutputGridDesc>(
                     arg.out_grid_desc_m_k_);
-            const index_t grid_size = block_2_tile_map.CalculateGridSize(arg.out_grid_desc_m_k_);
-            const auto kernel       = kernel_image_to_column<InputGridDesc,
-                                                       InputDataType,
-                                                       OutputGridDesc,
-                                                       OutputDataType,
-                                                       Block2ETileMap,
-                                                       GridwiseImageToColumnKernel>;
+            const index_t grid_size =
+                block_2_tile_map.CalculateGridSize(arg.out_grid_desc_m_k_) * arg.G_;
+            const auto kernel = kernel_tensor_rearrange<InputGridDesc,
+                                                        InputDataType,
+                                                        OutputGridDesc,
+                                                        OutputDataType,
+                                                        Block2ETileMap,
+                                                        ComputePtrOffsetOfStridedBatch<>,
+                                                        GridwiseTensorRearrangeKernel>;
 
             float elapsed_time = launch_and_time_kernel(stream_config,
                                                         kernel,
@@ -258,7 +257,9 @@ struct DeviceImageToColumnImpl
                                                         arg.p_in_,
                                                         arg.out_grid_desc_m_k_,
                                                         arg.p_out_,
-                                                        block_2_tile_map);
+                                                        arg.G_,
+                                                        block_2_tile_map,
+                                                        arg.compute_ptr_offset_of_batch_);
             return elapsed_time;
         }
 
@@ -271,13 +272,7 @@ struct DeviceImageToColumnImpl
 
     bool IsSupportedArgument(const Argument& arg)
     {
-        using namespace tensor_layout::convolution;
-        if(!(std::is_same_v<InputLayout, GNWC> || std::is_same_v<InputLayout, GNHWC> ||
-             std::is_same_v<InputLayout, GNDHWC>))
-        {
-            return false;
-        }
-        if(!(NDimSpatial >= 1 && NDimSpatial <= 3))
+        if constexpr(!(is_NSpatialGC || is_GNSpatialC))
         {
             return false;
         }
@@ -286,8 +281,8 @@ struct DeviceImageToColumnImpl
         const auto w_pad_right = arg.input_right_pads_[NDimSpatial - I1];
         const auto dilation_x  = arg.conv_filter_dilations_[NDimSpatial - I1];
         const auto stride_x    = arg.conv_filter_strides_[NDimSpatial - I1];
-        bool is_w_packed       = arg.input_g_n_c_wis_strides_[NDimSpatial + I2] == arg.C_;
-        bool is_c_packed       = arg.input_g_n_c_wis_strides_[I2] == 1;
+        bool is_w_packed       = arg.image_g_n_c_wis_strides_[NDimSpatial + I2] == arg.C_;
+        bool is_c_packed       = arg.image_g_n_c_wis_strides_[I2] == 1;
 
         // check vector acces with c not packed
         if(!is_c_packed && ScalarPerVector != 1)
@@ -309,8 +304,8 @@ struct DeviceImageToColumnImpl
         if(dilation_x > 1 && arg.C_ % ScalarPerVector != 0)
             return false;
 
-        return GridwiseImageToColumnKernel::CheckValidity(arg.in_grid_desc_m_k_,
-                                                          arg.out_grid_desc_m_k_);
+        return GridwiseTensorRearrangeKernel::CheckValidity(arg.in_grid_desc_m_k_,
+                                                            arg.out_grid_desc_m_k_);
     }
 
     bool IsSupportedArgument(const BaseArgument* p_arg) override
@@ -319,14 +314,15 @@ struct DeviceImageToColumnImpl
     }
 
     static auto MakeArgument(const void* p_in, // input image
-                             void* p_out,      // output image
+                             void* p_out,      // gemm form
+                             const ck::index_t G,
                              const ck::index_t N,
                              const ck::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_g_n_c_wis_strides,
-                             const std::array<index_t, 2>& output_m_k_strides,
+                             const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
+                             const std::array<index_t, 3>& gemm_g_m_k_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,
@@ -334,13 +330,14 @@ struct DeviceImageToColumnImpl
     {
         return Argument{static_cast<const InputDataType*>(p_in),
                         static_cast<OutputDataType*>(p_out),
+                        G,
                         N,
                         C,
                         input_spatial_lengths,
                         filter_spatial_lengths,
                         output_spatial_lengths,
-                        input_g_n_c_wis_strides,
-                        output_m_k_strides,
+                        image_g_n_c_wis_strides,
+                        gemm_g_m_k_strides,
                         conv_filter_strides,
                         conv_filter_dilations,
                         input_left_pads,
@@ -351,14 +348,15 @@ struct DeviceImageToColumnImpl
 
     std::unique_ptr<BaseArgument>
     MakeArgumentPointer(const void* p_in, // input image
-                        void* p_out,      // output image
+                        void* p_out,      // gemm form
+                        const ck::index_t G,
                         const ck::index_t N,
                         const ck::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_g_n_c_wis_strides,
-                        const std::array<index_t, 2>& output_m_k_strides,
+                        const std::array<index_t, NDimSpatial + 3>& image_g_n_c_wis_strides,
+                        const std::array<index_t, 3>& gemm_g_m_k_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,
@@ -366,13 +364,14 @@ struct DeviceImageToColumnImpl
     {
         return std::make_unique<Argument>(static_cast<const InputDataType*>(p_in),
                                           static_cast<OutputDataType*>(p_out),
+                                          G,
                                           N,
                                           C,
                                           input_spatial_lengths,
                                           filter_spatial_lengths,
                                           output_spatial_lengths,
-                                          input_g_n_c_wis_strides,
-                                          output_m_k_strides,
+                                          image_g_n_c_wis_strides,
+                                          gemm_g_m_k_strides,
                                           conv_filter_strides,
                                           conv_filter_dilations,
                                           input_left_pads,

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <iostream>
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_normalization_bwd_gamma_beta.hpp"
+#include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_bwd_gamma_beta.hpp"
+#include "ck/tensor_description/tensor_descriptor.hpp"
+#include "ck/tensor_description/tensor_descriptor_helper.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
+
+// M is invarient dimension, K is reduced dimension
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+template <typename GridwiseReduction,
+          typename DYDataType,
+          typename XDataType,
+          typename MeanInvStdDataType,
+          typename DGammaDataType,
+          typename DBetaDataType,
+          typename GridDesc_M_K,
+          typename GridDesc_M>
+__global__ void
+kernel_normalization_bwd_gamma_beta(const GridDesc_M_K dy_grid_desc_m_k,
+                                    const GridDesc_M_K x_grid_desc_m_k,
+                                    const GridDesc_M_K mean_grid_desc_m_k,
+                                    const GridDesc_M_K inv_std_grid_desc_m_k,
+                                    const GridDesc_M dgamma_grid_desc_m,
+                                    const GridDesc_M dbeta_grid_desc_m,
+                                    index_t num_k_block_tile_iteration,
+                                    const DYDataType* const __restrict__ p_dy_global,
+                                    const XDataType* const __restrict__ p_x_global,
+                                    const MeanInvStdDataType* const __restrict__ p_mean_global,
+                                    const MeanInvStdDataType* const __restrict__ p_inv_std_global,
+                                    DGammaDataType* const __restrict__ p_dgamma_global,
+                                    DBetaDataType* const __restrict__ p_dbeta_global)
+{
+    GridwiseReduction::Run(dy_grid_desc_m_k,
+                           x_grid_desc_m_k,
+                           mean_grid_desc_m_k,
+                           inv_std_grid_desc_m_k,
+                           dgamma_grid_desc_m,
+                           dbeta_grid_desc_m,
+                           num_k_block_tile_iteration,
+                           p_dy_global,
+                           p_x_global,
+                           p_mean_global,
+                           p_inv_std_global,
+                           p_dgamma_global,
+                           p_dbeta_global);
+};
+
+template <typename DYDataType,
+          typename XDataType,
+          typename MeanInvStdDataType,
+          typename ComputeDataType,
+          typename DGammaDataType,
+          typename DBetaDataType,
+          index_t Rank,
+          index_t NumReduceDim,
+          index_t BlockSize,
+          index_t MThreadClusterSize,
+          index_t KThreadClusterSize,
+          index_t MThreadSliceSize,
+          index_t KThreadSliceSize,
+          bool IsDYFastestDimReduced,
+          index_t DYSrcVectorSize,
+          bool IsXFastestDimReduced,
+          index_t XSrcVectorSize,
+          bool IsMeanInvStdFastestDimReduced,
+          index_t MeanInvStdSrcVectorSize,
+          index_t DGammaDstVectorSize,
+          index_t DBetaDstVectorSize>
+struct DeviceNormalizationBwdGammaBetaImpl
+    : public DeviceNormalizationBwdGammaBeta<DYDataType,
+                                             XDataType,
+                                             MeanInvStdDataType,
+                                             DGammaDataType,
+                                             DBetaDataType,
+                                             Rank,
+                                             NumReduceDim>
+{
+
+    static constexpr index_t DYSrcVectorDim         = IsDYFastestDimReduced ? 1 : 0;
+    static constexpr index_t XSrcVectorDim          = IsXFastestDimReduced ? 1 : 0;
+    static constexpr index_t MeanInvStdSrcVectorDim = IsMeanInvStdFastestDimReduced ? 1 : 0;
+
+    static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize);
+
+    static_assert(((DYSrcVectorDim == 0 && MThreadSliceSize % DYSrcVectorSize == 0) ||
+                   (DYSrcVectorDim == 1 && KThreadSliceSize % DYSrcVectorSize == 0)),
+                  "Invalid thread slice sizes and/or dy vector sizes configuration, please check!");
+
+    static_assert(((XSrcVectorDim == 0 && MThreadSliceSize % XSrcVectorSize == 0) ||
+                   (XSrcVectorDim == 1 && KThreadSliceSize % XSrcVectorSize == 0)),
+                  "Invalid thread slice sizes and/or x vector sizes configuration, please check!");
+
+    static_assert(
+        ((MThreadSliceSize % DGammaDstVectorSize == 0) ||
+         (MThreadSliceSize % DBetaDstVectorSize == 0)),
+        "Invalid thread slice sizes and/or Gamma and beta vector sizes configuration, please "
+        "check!");
+
+    static_assert(
+        (MeanInvStdSrcVectorDim == 0 && MThreadSliceSize % MeanInvStdSrcVectorSize == 0) ||
+            (MeanInvStdSrcVectorDim == 1 && KThreadSliceSize % MeanInvStdSrcVectorSize == 0),
+        "Invalid thread slice sizes and/or mean and inverse std vector sizes configuration, please "
+        "check!");
+
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
+    static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
+    static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
+
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+    static_assert(!reduceAllDim);
+
+    static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
+                                    const std::vector<index_t>& inStrides,
+                                    int numBlockTileIteration)
+    {
+        const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<Rank>{});
+        const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<Rank>{});
+
+        const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto in_grid_desc_m_k = [&]() {
+            using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+            using ReduceDims    = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
+
+            const auto reduceDimLengths =
+                make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
+            const auto invariantDimLengths =
+                make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
+
+            return transform_tensor_descriptor(inDesc,
+                                               make_tuple(make_merge_transform(invariantDimLengths),
+                                                          make_merge_transform(reduceDimLengths)),
+                                               make_tuple(InvariantDims{}, ReduceDims{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }();
+
+        const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
+        const auto reduceLength    = in_grid_desc_m_k.GetLength(Number<1>{});
+
+        const auto inPad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+        const auto inPad_K = K_BlockTileSize * numBlockTileIteration - reduceLength;
+
+        auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
+            in_grid_desc_m_k,
+            make_tuple(make_right_pad_transform(invariantLength, inPad_M),
+                       make_right_pad_transform(reduceLength, inPad_K)),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        return in_grid_desc_m_k_padded;
+    }
+
+    static auto MakeDst1dDescriptor(const std::vector<index_t>& outLengths,
+                                    const std::vector<index_t>& outStrides)
+    {
+        const auto tupleDstLengths =
+            generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumInvariantDim>{});
+        const auto tupleDstStrides =
+            generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumInvariantDim>{});
+
+        auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
+
+        auto out_grid_desc_m = transform_tensor_descriptor(
+            outDesc,
+            make_tuple(make_merge_transform(tupleDstLengths)),
+            make_tuple(typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type{}),
+            make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
+
+        const auto outPad =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto out_grid_desc_m_padded = transform_tensor_descriptor(
+            out_grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, outPad)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+        return (out_grid_desc_m_padded);
+    };
+
+    using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1));
+    using GridDesc_M   = decltype(MakeDst1dDescriptor({1}, {1}));
+
+    using GridwiseNormalizationBwdGammaBeta =
+        GridwiseNormalizationBwdGammaBeta_mk_to_k<DYDataType,
+                                                  XDataType,
+                                                  MeanInvStdDataType,
+                                                  ComputeDataType,
+                                                  DGammaDataType,
+                                                  DBetaDataType,
+                                                  GridDesc_M_K,
+                                                  GridDesc_M,
+                                                  BlockSize,
+                                                  MThreadClusterSize,
+                                                  KThreadClusterSize,
+                                                  MThreadSliceSize,
+                                                  KThreadSliceSize,
+                                                  DYSrcVectorDim,
+                                                  DYSrcVectorSize,
+                                                  XSrcVectorDim,
+                                                  XSrcVectorSize,
+                                                  MeanInvStdSrcVectorDim,
+                                                  MeanInvStdSrcVectorSize,
+                                                  DGammaDstVectorSize,
+                                                  DBetaDstVectorSize>;
+
+    struct Argument : public BaseArgument
+    {
+        Argument(const std::vector<index_t> inLengths,
+                 const std::vector<index_t> dyStrides,
+                 const std::vector<index_t> xStrides,
+                 const std::vector<index_t> meanStrides,
+                 const std::vector<index_t> invStdStrides,
+                 const std::vector<index_t> outLengths,
+                 const std::vector<index_t> dgammaStrides,
+                 const std::vector<index_t> dbetaStrides,
+                 const std::vector<index_t> reduceDims,
+                 const DYDataType* p_dy,
+                 const XDataType* p_x,
+                 const MeanInvStdDataType* p_mean,
+                 const MeanInvStdDataType* p_invStd,
+                 DGammaDataType* p_dgamma,
+                 DBetaDataType* p_dbeta)
+            : p_dy_(p_dy),
+              p_x_(p_x),
+              p_mean_(p_mean),
+              p_invStd_(p_invStd),
+              p_dgamma_(p_dgamma),
+              p_dbeta_(p_dbeta),
+              outLengths_{outLengths},
+              dgammaStrides_{dgammaStrides},
+              dbetaStrides_{dbetaStrides}
+        {
+            inLengths_   = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
+            dyStrides_   = shuffle_tensor_dimensions<Rank, NumReduceDim>(dyStrides, reduceDims);
+            xStrides_    = shuffle_tensor_dimensions<Rank, NumReduceDim>(xStrides, reduceDims);
+            meanStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(meanStrides, reduceDims);
+            invStdStrides_ =
+                shuffle_tensor_dimensions<Rank, NumReduceDim>(invStdStrides, reduceDims);
+
+            std::tie(MRaw_, KRaw_) = get_2d_lengths<Rank, NumReduceDim>(inLengths_);
+
+            numBlockTileIteration_ = math::integer_divide_ceil(KRaw_, K_BlockTileSize);
+
+            gridSize_ = math::integer_divide_ceil(MRaw_, M_BlockTileSize);
+
+            dy_grid_desc_m_k_ = MakeSrc2dDescriptor(inLengths_, dyStrides_, numBlockTileIteration_);
+            x_grid_desc_m_k_  = MakeSrc2dDescriptor(inLengths_, xStrides_, numBlockTileIteration_);
+            mean_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(inLengths_, meanStrides_, numBlockTileIteration_);
+            inv_std_grid_desc_m_k_ =
+                MakeSrc2dDescriptor(inLengths_, invStdStrides_, numBlockTileIteration_);
+
+            dgamma_grid_desc_m_ = MakeDst1dDescriptor(outLengths_, dgammaStrides_);
+            dbeta_grid_desc_m_  = MakeDst1dDescriptor(outLengths_, dbetaStrides_);
+        }
+
+        const DYDataType* p_dy_;
+        const XDataType* p_x_;
+        const MeanInvStdDataType* p_mean_;
+        const MeanInvStdDataType* p_invStd_;
+        DGammaDataType* p_dgamma_;
+        DBetaDataType* p_dbeta_;
+
+        std::vector<index_t> inLengths_;
+        std::vector<index_t> dyStrides_;
+        std::vector<index_t> xStrides_;
+        std::vector<index_t> meanStrides_;
+        std::vector<index_t> invStdStrides_;
+        std::vector<index_t> outLengths_;
+        std::vector<index_t> dgammaStrides_;
+        std::vector<index_t> dbetaStrides_;
+
+        int numBlockTileIteration_;
+        size_t gridSize_;
+
+        // Source descriptor
+        GridDesc_M_K dy_grid_desc_m_k_;
+        GridDesc_M_K x_grid_desc_m_k_;
+        GridDesc_M_K mean_grid_desc_m_k_;
+        GridDesc_M_K inv_std_grid_desc_m_k_;
+
+        // Destination descriptor
+        GridDesc_M dgamma_grid_desc_m_;
+        GridDesc_M dbeta_grid_desc_m_;
+
+        index_t MRaw_; // invarient length
+        index_t KRaw_; // reduce length
+    };
+
+    struct Invoker : public BaseInvoker
+    {
+        float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
+        {
+            const auto kernel_main =
+                kernel_normalization_bwd_gamma_beta<GridwiseNormalizationBwdGammaBeta,
+                                                    DYDataType,
+                                                    XDataType,
+                                                    MeanInvStdDataType,
+                                                    DGammaDataType,
+                                                    DBetaDataType,
+                                                    GridDesc_M_K,
+                                                    GridDesc_M>;
+
+            return launch_and_time_kernel(stream_config,
+                                          kernel_main,
+                                          dim3(arg.gridSize_),
+                                          dim3(BlockSize),
+                                          0,
+                                          arg.dy_grid_desc_m_k_,
+                                          arg.x_grid_desc_m_k_,
+                                          arg.mean_grid_desc_m_k_,
+                                          arg.inv_std_grid_desc_m_k_,
+                                          arg.dgamma_grid_desc_m_,
+                                          arg.dbeta_grid_desc_m_,
+                                          arg.numBlockTileIteration_,
+                                          arg.p_dy_,
+                                          arg.p_x_,
+                                          arg.p_mean_,
+                                          arg.p_invStd_,
+                                          arg.p_dgamma_,
+                                          arg.p_dbeta_);
+        }
+
+        float Run(const BaseArgument* p_arg,
+                  const StreamConfig& stream_config = StreamConfig{}) override
+        {
+            return Run(*dynamic_cast<const Argument*>(p_arg), stream_config);
+        }
+    };
+
+    template <index_t SrcVectorDim, index_t SrcVectorSize>
+    bool IsSrcVectorDimSizeValid(const std::vector<index_t>& lengths,
+                                 const std::vector<index_t>& strides)
+    {
+        if constexpr(SrcVectorSize == 1)
+            return true;
+
+        // Fastest dimension is not reduced
+        if constexpr(SrcVectorDim == 0)
+        {
+            if constexpr(NumInvariantDim == 0)
+                return false;
+
+            if(strides[NumInvariantDim - 1] != 1)
+                return false;
+
+            if(lengths[NumInvariantDim - 1] % SrcVectorSize != 0)
+                return false;
+        }
+        else // Fastest dimension is reduced
+        {
+            if(strides[Rank - 1] != 1)
+                return false;
+
+            if(lengths[Rank - 1] % SrcVectorSize != 0)
+                return false;
+        };
+
+        return true;
+    }
+
+    template <index_t DstVectorSize>
+    bool IsDstVectorSizeValid(const std::vector<index_t>& lengths,
+                              const std::vector<index_t>& strides)
+    {
+        if constexpr(DstVectorSize == 1)
+            return true;
+
+        if(strides[NumInvariantDim - 1] != 1)
+            return false;
+
+        if(lengths[NumInvariantDim - 1] % DstVectorSize != 0)
+            return false;
+
+        return true;
+    }
+
+    bool IsSupportedArgument(const BaseArgument* p_arg) override
+    {
+        const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
+
+        bool pass = true;
+        pass &= IsSrcVectorDimSizeValid<DYSrcVectorDim, DYSrcVectorSize>(p_arg_->inLengths_,
+                                                                         p_arg_->dyStrides_);
+        pass &= IsSrcVectorDimSizeValid<XSrcVectorDim, XSrcVectorSize>(p_arg_->inLengths_,
+                                                                       p_arg_->xStrides_);
+        pass &= IsSrcVectorDimSizeValid<MeanInvStdSrcVectorDim, MeanInvStdSrcVectorSize>(
+            p_arg_->inLengths_, p_arg_->meanStrides_);
+        pass &= IsSrcVectorDimSizeValid<MeanInvStdSrcVectorDim, MeanInvStdSrcVectorSize>(
+            p_arg_->inLengths_, p_arg_->invStdStrides_);
+
+        pass &=
+            IsDstVectorSizeValid<DGammaDstVectorSize>(p_arg_->outLengths_, p_arg_->dgammaStrides_);
+        pass &=
+            IsDstVectorSizeValid<DBetaDstVectorSize>(p_arg_->outLengths_, p_arg_->dbetaStrides_);
+
+        return pass;
+    }
+
+    std::unique_ptr<BaseArgument> MakeArgumentPointer(const std::vector<index_t> inLengths,
+                                                      const std::vector<index_t> dyStrides,
+                                                      const std::vector<index_t> xStrides,
+                                                      const std::vector<index_t> meanStrides,
+                                                      const std::vector<index_t> invStdStrides,
+                                                      const std::vector<index_t> outLengths,
+                                                      const std::vector<index_t> dgammaStrides,
+                                                      const std::vector<index_t> dbetaStrides,
+                                                      const std::vector<index_t> reduceDims,
+                                                      const void* p_dy,
+                                                      const void* p_x,
+                                                      const void* p_mean,
+                                                      const void* p_invStd,
+                                                      void* p_dgamma,
+                                                      void* p_dbeta) override
+    {
+        if(inLengths.size() != Rank || dyStrides.size() != Rank || xStrides.size() != Rank ||
+           meanStrides.size() != Rank || invStdStrides.size() != Rank)
+            throw std::runtime_error("dimension is incorrect");
+
+        if(outLengths.size() != NumInvariantDim || dgammaStrides.size() != NumInvariantDim ||
+           dbetaStrides.size() != NumInvariantDim)
+            throw std::runtime_error("dimension is incorrect");
+
+        return std::make_unique<Argument>(inLengths,
+                                          dyStrides,
+                                          xStrides,
+                                          meanStrides,
+                                          invStdStrides,
+                                          outLengths,
+                                          dgammaStrides,
+                                          dbetaStrides,
+                                          reduceDims,
+                                          static_cast<const DYDataType*>(p_dy),
+                                          static_cast<const XDataType*>(p_x),
+                                          static_cast<const MeanInvStdDataType*>(p_mean),
+                                          static_cast<const MeanInvStdDataType*>(p_invStd),
+                                          static_cast<DGammaDataType*>(p_dgamma),
+                                          static_cast<DBetaDataType*>(p_dbeta));
+    }
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
+    {
+        return std::make_unique<Invoker>();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp → include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_impl.hpp

@@ -7,7 +7,7 @@
 #include <sstream>
 
 #include "ck/utility/reduction_operator.hpp"
-#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_selector.hpp"
@@ -28,6 +28,7 @@ template <typename XDataType,
           typename BetaDataType,
           typename ComputeDataType,
           typename YDataType,
+          typename SaveMeanInvStdDataType,
           typename YElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim,
@@ -43,15 +44,16 @@ template <typename XDataType,
           index_t BetaSrcVectorDim,
           index_t BetaSrcVectorSize,
           index_t YDstVectorSize,
+          index_t SaveMeanInvStdDstVectorSize,
           bool UseWelford = true>
-struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
-                                                            GammaDataType,
-                                                            BetaDataType,
-                                                            ComputeDataType,
-                                                            YDataType,
-                                                            YElementwiseOperation,
-                                                            Rank,
-                                                            NumReduceDim>
+struct DeviceNormalizationFwdImpl : public DeviceNormalizationFwd<XDataType,
+                                                                  GammaDataType,
+                                                                  BetaDataType,
+                                                                  YDataType,
+                                                                  SaveMeanInvStdDataType,
+                                                                  YElementwiseOperation,
+                                                                  Rank,
+                                                                  NumReduceDim>
 {
     static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize);
     static_assert(
@@ -64,18 +66,24 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
          (BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
         "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
 
+    static_assert(MThreadSliceSize % SaveMeanInvStdDstVectorSize == 0,
+                  "Invalid thread slice sizes and/or save mean and inverse std vector sizes "
+                  "configuration, please check!");
+
     using PassThrough = tensor_operation::element_wise::PassThrough;
 
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
     static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
     static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
 
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+    static_assert(!reduceAllDim); // TODO
+
     static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
                                     const std::vector<index_t>& inStrides,
                                     int numBlockTileIteration)
     {
-        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
         static constexpr index_t numSrcDim = Rank;
-        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
 
         const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
         const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
@@ -133,7 +141,37 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         return (in_grid_desc_m_k_padded);
     };
 
+    static auto MakeSaveMeanInvStdDescriptor_M(const std::vector<index_t>& lengths,
+                                               const std::vector<index_t>& strides)
+    {
+        using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+
+        const auto tupleSrcLengths = make_tuple_from_array_and_index_seq(lengths, InvariantDims{});
+        const auto tupleSrcStrides = make_tuple_from_array_and_index_seq(strides, InvariantDims{});
+
+        const auto desc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto grid_desc_m =
+            transform_tensor_descriptor(desc,
+                                        make_tuple(make_merge_transform(tupleSrcLengths)),
+                                        make_tuple(InvariantDims{}),
+                                        make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+        const auto pad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded = transform_tensor_descriptor(
+            grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, pad_M)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+
+        return grid_desc_m_padded;
+    }
+
     using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1));
+    using GridDesc_M   = decltype(MakeSaveMeanInvStdDescriptor_M({1}, {1}));
 
     struct Argument : public BaseArgument
     {
@@ -142,17 +180,23 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                  const std::vector<index_t> gammaStrides,
                  const std::vector<index_t> betaStrides,
                  const std::vector<index_t> yStrides,
+                 const std::vector<index_t> saveMeanStrides,
+                 const std::vector<index_t> saveInvStdStrides,
                  const std::vector<index_t> reduceDims,
                  YElementwiseOperation y_elementwise_op,
                  double epsilon,
                  const XDataType* p_x,
                  const GammaDataType* p_gamma,
                  const BetaDataType* p_beta,
-                 YDataType* p_y)
+                 YDataType* p_y,
+                 SaveMeanInvStdDataType* p_saveMean,
+                 SaveMeanInvStdDataType* p_saveInvStd)
             : p_x_(p_x),
               p_gamma_(p_gamma),
               p_beta_(p_beta),
               p_y_(p_y),
+              p_saveMean_(p_saveMean),
+              p_saveInvStd_(p_saveInvStd),
               y_elementwise_op_(y_elementwise_op)
         {
             epsilon_ = static_cast<ComputeDataType>(epsilon);
@@ -162,16 +206,14 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
             yStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
             gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
             betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+            saveMeanStrides_   = saveMeanStrides;
+            saveInvStdStrides_ = saveInvStdStrides;
 
-            long_index_t invariant_length;
-            long_index_t reduce_length;
-
-            std::tie(invariant_length, reduce_length) =
-                get_2d_lengths<Rank, NumReduceDim>(Lengths_);
+            std::tie(MRaw_, KRaw_) = get_2d_lengths<Rank, NumReduceDim>(Lengths_);
 
-            numBlockTileIteration_ = math::integer_divide_ceil(reduce_length, K_BlockTileSize);
+            numBlockTileIteration_ = math::integer_divide_ceil(KRaw_, K_BlockTileSize);
 
-            gridSize_ = math::integer_divide_ceil(invariant_length, M_BlockTileSize);
+            gridSize_ = math::integer_divide_ceil(MRaw_, M_BlockTileSize);
 
             x_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, xStrides_, numBlockTileIteration_);
             gamma_grid_desc_m_k_ =
@@ -179,9 +221,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
             beta_grid_desc_m_k_ =
                 MakeSrc2dDescriptor(Lengths_, betaStrides_, numBlockTileIteration_);
             y_grid_desc_m_k_ = MakeSrc2dDescriptor(Lengths_, yStrides_, numBlockTileIteration_);
+            save_mean_grid_desc_m_    = MakeSaveMeanInvStdDescriptor_M(Lengths_, saveMeanStrides);
+            save_inv_std_grid_desc_m_ = MakeSaveMeanInvStdDescriptor_M(Lengths_, saveInvStdStrides);
 
             isSweeponce_ =
                 x_grid_desc_m_k_.GetLength(Number<1>{}) <= KThreadClusterSize * KThreadSliceSize;
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length_ = 1;
+            else
+                invariant_lowest_length_ = Lengths_[NumInvariantDim - 1];
         }
 
         ComputeDataType epsilon_;
@@ -190,12 +239,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         const GammaDataType* p_gamma_;
         const BetaDataType* p_beta_;
         YDataType* p_y_;
+        SaveMeanInvStdDataType* p_saveMean_;
+        SaveMeanInvStdDataType* p_saveInvStd_;
 
         std::vector<index_t> Lengths_;
         std::vector<index_t> xStrides_;
         std::vector<index_t> gammaStrides_;
         std::vector<index_t> betaStrides_;
         std::vector<index_t> yStrides_;
+        std::vector<index_t> saveMeanStrides_;
+        std::vector<index_t> saveInvStdStrides_;
 
         YElementwiseOperation y_elementwise_op_;
 
@@ -206,7 +259,14 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         GridDesc_M_K gamma_grid_desc_m_k_;
         GridDesc_M_K beta_grid_desc_m_k_;
         GridDesc_M_K y_grid_desc_m_k_;
+        GridDesc_M save_mean_grid_desc_m_;
+        GridDesc_M save_inv_std_grid_desc_m_;
         bool isSweeponce_;
+
+        index_t MRaw_; // invarient length
+        index_t KRaw_; // reduce length
+
+        index_t invariant_lowest_length_;
     };
 
     struct Invoker : public BaseInvoker
@@ -217,9 +277,11 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                                                            GammaDataType,
                                                            BetaDataType,
                                                            YDataType,
+                                                           SaveMeanInvStdDataType,
                                                            ComputeDataType,
                                                            YElementwiseOperation,
                                                            GridDesc_M_K,
+                                                           GridDesc_M,
                                                            BlockSize,
                                                            MThreadClusterSize,
                                                            KThreadClusterSize,
@@ -233,6 +295,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                                                            BetaSrcVectorSize,
                                                            XYSrcVectorDim,
                                                            YDstVectorSize,
+                                                           SaveMeanInvStdDstVectorSize,
                                                            UseWelford>(arg.isSweeponce_);
 
             float avg_time = 0;
@@ -245,12 +308,16 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                                                arg.gamma_grid_desc_m_k_,
                                                arg.beta_grid_desc_m_k_,
                                                arg.y_grid_desc_m_k_,
+                                               arg.save_mean_grid_desc_m_,
+                                               arg.save_inv_std_grid_desc_m_,
                                                arg.numBlockTileIteration_,
                                                arg.epsilon_,
                                                arg.p_x_,
                                                arg.p_gamma_,
                                                arg.p_beta_,
                                                arg.p_y_,
+                                               arg.p_saveMean_,
+                                               arg.p_saveInvStd_,
                                                arg.y_elementwise_op_);
 
             return (avg_time);
@@ -267,8 +334,6 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
     {
         const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
 
-        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
-
         if constexpr(XYSrcVectorDim == 0)
         {
             if constexpr(NumInvariantDim == 0)
@@ -277,13 +342,15 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
             }
             else
             {
+                printf("!!!! %d\n", p_arg_->invariant_lowest_length_);
+
                 if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
                     return false;
 
-                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                if(p_arg_->invariant_lowest_length_ % XSrcVectorSize != 0)
                     return false;
 
-                if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
+                if(p_arg_->invariant_lowest_length_ % YDstVectorSize != 0)
                     return false;
             };
         }
@@ -325,7 +392,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
             if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
                 return (false);
 
-            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+            if(p_arg_->invariant_lowest_length_ % BetaSrcVectorSize != 0)
                 return (false);
         }
         else // if fastest dim is reduced
@@ -337,6 +404,9 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                 return (false);
         }
 
+        if(p_arg_->invariant_lowest_length_ % SaveMeanInvStdDstVectorSize != 0)
+            return false;
+
         return true;
     };
 
@@ -346,6 +416,8 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                         const std::vector<index_t> gammaStrides,
                         const std::vector<index_t> betaStrides,
                         const std::vector<index_t> yStrides,
+                        const std::vector<index_t> saveMeanStrides,
+                        const std::vector<index_t> saveInvStdStrides,
                         const std::vector<index_t> reduceDims,
                         double epsilon,
                         const void* p_x,
@@ -353,27 +425,30 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
                         const void* p_beta,
                         void* p_y,
                         void* p_saveMean,
-                        void* p_saveInvVar,
+                        void* p_saveInvStd,
                         YElementwiseOperation y_elementwise_op) override
     {
-        // TODO
-        // Optional cache of the intermediate results (mean and InvVariance) during the
-        // forward pass could speedup in the backward
-        ignore = p_saveMean;
-        ignore = p_saveInvVar;
+        if(lengths.size() != Rank || xStrides.size() != Rank || gammaStrides.size() != Rank ||
+           betaStrides.size() != Rank || yStrides.size() != Rank ||
+           saveMeanStrides.size() != NumInvariantDim || saveInvStdStrides.size() != NumInvariantDim)
+            throw std::runtime_error("dimension is incorrect");
 
         return std::make_unique<Argument>(lengths,
                                           xStrides,
                                           gammaStrides,
                                           betaStrides,
                                           yStrides,
+                                          saveMeanStrides,
+                                          saveInvStdStrides,
                                           reduceDims,
                                           y_elementwise_op,
                                           epsilon,
                                           static_cast<const XDataType*>(p_x),
                                           static_cast<const GammaDataType*>(p_gamma),
                                           static_cast<const BetaDataType*>(p_beta),
-                                          static_cast<YDataType*>(p_y));
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<SaveMeanInvStdDataType*>(p_saveMean),
+                                          static_cast<SaveMeanInvStdDataType*>(p_saveInvStd));
     };
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
@@ -386,7 +461,7 @@ struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceNormalizationImpl<" << BlockSize << ",";
+        str << "DeviceNormalizationFwdImpl<" << BlockSize << ",";
         str << "Cluster_MK_" << MThreadClusterSize << "_" << KThreadClusterSize << ",";
         str << "Slice_MK_" << MThreadSliceSize << "_" << KThreadSliceSize << ",";
         str << "XYSrcVectorDim_" << XYSrcVectorDim  << ",";

include/ck/tensor_operation/gpu/device/impl/device_normalization_splitk_impl.hpp → include/ck/tensor_operation/gpu/device/impl/device_normalization_fwd_splitk_impl.hpp

@@ -8,7 +8,7 @@
 
 #include "ck/utility/reduction_operator.hpp"
 #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
-#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
+#include "ck/tensor_operation/gpu/device/device_normalization_fwd.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
 #include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp"
@@ -19,7 +19,7 @@
 namespace ck {
 template <typename GridwiseWelford,
           typename XDataType,
-          typename MeanVarDataType,
+          typename WorkspaceMeanVarDataType,
           typename ComputeDataType,
           typename XGridDesc_M_K,
           typename MeanVarGridDesc_M_KBlock>
@@ -28,8 +28,8 @@ kernel_normalizationSplitK1st(const XGridDesc_M_K x_grid_desc_m_k,
                               const MeanVarGridDesc_M_KBlock mean_var_grid_desc_m_kblock,
                               index_t num_k_block_tile_iteration,
                               const XDataType* const __restrict__ p_x_global,
-                              MeanVarDataType* const __restrict__ p_welford_mean,
-                              MeanVarDataType* const __restrict__ p_welford_variance,
+                              WorkspaceMeanVarDataType* const __restrict__ p_welford_mean,
+                              WorkspaceMeanVarDataType* const __restrict__ p_welford_variance,
                               int32_t* const __restrict__ p_welford_count)
 {
     GridwiseWelford::Run(x_grid_desc_m_k,
@@ -42,16 +42,18 @@ kernel_normalizationSplitK1st(const XGridDesc_M_K x_grid_desc_m_k,
 };
 
 template <typename GridwiseWelfordNormalization,
-          typename MeanVarDataType,
+          typename WorkspaceMeanVarDataType,
           typename XDataType,
           typename GammaDataType,
           typename BetaDataType,
           typename YDataType,
+          typename SaveMeanInvStdDataType,
           typename ComputeDataType,
           typename YElementwiseOperation,
           typename MeanVarGridDesc_M_KBlock,
           typename CountGridDesc_M_KBlock,
-          typename XYGammaBetaGridDesc_M_K>
+          typename XYGammaBetaGridDesc_M_K,
+          typename SaveMeanInvStdGridDesc_M>
 __global__ void
 kernel_normalizationSplitK2nd(const MeanVarGridDesc_M_KBlock mean_var_grid_desc_m_kblock,
                               const CountGridDesc_M_KBlock count_grid_desc_m_kblock,
@@ -59,17 +61,21 @@ kernel_normalizationSplitK2nd(const MeanVarGridDesc_M_KBlock mean_var_grid_desc_
                               const XYGammaBetaGridDesc_M_K gamma_grid_desc_m_k,
                               const XYGammaBetaGridDesc_M_K beta_grid_desc_m_k,
                               const XYGammaBetaGridDesc_M_K y_grid_desc_m_k,
+                              const SaveMeanInvStdGridDesc_M save_mean_grid_desc_m,
+                              const SaveMeanInvStdGridDesc_M save_inv_std_grid_desc_m,
                               index_t num_k_mean_var_count_iteration,
                               index_t num_k_block_tile_iteration,
                               index_t k_grid_size,
                               ComputeDataType epsilon,
-                              const MeanVarDataType* const p_mean_global,
-                              const MeanVarDataType* const p_variance_global,
+                              const WorkspaceMeanVarDataType* const p_mean_global,
+                              const WorkspaceMeanVarDataType* const p_variance_global,
                               const int32_t* const p_welford_count_global,
                               const XDataType* const __restrict__ p_x_global,
                               const GammaDataType* const __restrict__ p_gamma_global,
                               const BetaDataType* const __restrict__ p_beta_global,
                               YDataType* const __restrict__ p_y_global,
+                              SaveMeanInvStdDataType* const __restrict__ p_save_mean_global,
+                              SaveMeanInvStdDataType* const __restrict__ p_save_inv_std_global,
                               const YElementwiseOperation y_elementwise_op)
 {
     GridwiseWelfordNormalization::Run(mean_var_grid_desc_m_kblock,
@@ -78,6 +84,8 @@ kernel_normalizationSplitK2nd(const MeanVarGridDesc_M_KBlock mean_var_grid_desc_
                                       gamma_grid_desc_m_k,
                                       beta_grid_desc_m_k,
                                       y_grid_desc_m_k,
+                                      save_mean_grid_desc_m,
+                                      save_inv_std_grid_desc_m,
                                       num_k_mean_var_count_iteration,
                                       num_k_block_tile_iteration,
                                       k_grid_size,
@@ -89,6 +97,8 @@ kernel_normalizationSplitK2nd(const MeanVarGridDesc_M_KBlock mean_var_grid_desc_
                                       p_gamma_global,
                                       p_beta_global,
                                       p_y_global,
+                                      p_save_mean_global,
+                                      p_save_inv_std_global,
                                       y_elementwise_op);
 };
 } // namespace ck
@@ -107,6 +117,7 @@ template <typename XDataType,
           typename BetaDataType,
           typename ComputeDataType,
           typename YDataType,
+          typename SaveMeanInvStdDataType,
           typename YElementwiseOperation,
           index_t Rank,
           index_t NumReduceDim,
@@ -121,17 +132,18 @@ template <typename XDataType,
           index_t GammaSrcVectorSize,
           index_t BetaSrcVectorDim,
           index_t BetaSrcVectorSize,
-          index_t YDstVectorSize>
-struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
-                                                                  GammaDataType,
-                                                                  BetaDataType,
-                                                                  ComputeDataType,
-                                                                  YDataType,
-                                                                  YElementwiseOperation,
-                                                                  Rank,
-                                                                  NumReduceDim>
+          index_t YDstVectorSize,
+          index_t SaveMeanInvStdDstVectorSize>
+struct DeviceNormalizationFwdSplitKImpl : public DeviceNormalizationFwd<XDataType,
+                                                                        GammaDataType,
+                                                                        BetaDataType,
+                                                                        YDataType,
+                                                                        SaveMeanInvStdDataType,
+                                                                        YElementwiseOperation,
+                                                                        Rank,
+                                                                        NumReduceDim>
 {
-    using MeanVarDataType = ComputeDataType;
+    using WorkspaceMeanVarDataType = SaveMeanInvStdDataType;
 
     static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize);
     static_assert(
@@ -144,22 +156,28 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
          (BetaSrcVectorDim == 1 && KThreadSliceSize % BetaSrcVectorSize == 0)),
         "Invalid thread slice sizes and/or beta vector sizes configuration, please check!");
 
+    static_assert(MThreadSliceSize % SaveMeanInvStdDstVectorSize == 0,
+                  "Invalid thread slice sizes and/or save mean and inverse std vector sizes "
+                  "configuration, please check!");
+
     using PassThrough = tensor_operation::element_wise::PassThrough;
 
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
 
+    static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
     static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
     static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
 
+    static constexpr bool reduceAllDim = (NumInvariantDim == 0);
+    static_assert(!reduceAllDim); // TODO
+
     static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
                                     const std::vector<index_t>& inStrides,
                                     int kBlockSize,
                                     int numBlockTileIteration)
     {
-        constexpr index_t NumInvariantDim  = Rank - NumReduceDim;
         static constexpr index_t numSrcDim = Rank;
-        static constexpr bool reduceAllDim = (NumInvariantDim == 0);
 
         const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
         const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
@@ -219,7 +237,7 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
     };
 
     template <typename DoPads, index_t MPerTile, index_t KPerTile>
-    static auto MakeMeanVarDescriptor_M_K(index_t M, index_t K)
+    static auto MakeWorkspaceMeanVarDescriptor_M_K(index_t M, index_t K)
     {
         const auto grid_desc_m_k =
             make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(K, I1));
@@ -227,26 +245,57 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
     }
 
     template <typename DoPads, index_t MPerTile, index_t KPerTile>
-    static auto MakeCountDescriptor_M_K(index_t M, index_t K)
+    static auto MakeWorkspaceCountDescriptor_M_K(index_t M, index_t K)
     {
         const auto grid_desc_m_k =
             make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I0, I1));
         return PadTensorDescriptor(grid_desc_m_k, make_tuple(MPerTile, KPerTile), DoPads{});
     }
 
+    static auto MakeSaveMeanInvStdDescriptor_M(const std::vector<index_t>& lengths,
+                                               const std::vector<index_t>& strides)
+    {
+        using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
+
+        const auto tupleSrcLengths = make_tuple_from_array_and_index_seq(lengths, InvariantDims{});
+        const auto tupleSrcStrides = make_tuple_from_array_and_index_seq(strides, InvariantDims{});
+
+        const auto desc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
+
+        const auto grid_desc_m =
+            transform_tensor_descriptor(desc,
+                                        make_tuple(make_merge_transform(tupleSrcLengths)),
+                                        make_tuple(InvariantDims{}),
+                                        make_tuple(Sequence<0>{}));
+
+        const auto invariantLength = grid_desc_m.GetLength(Number<0>{});
+        const auto pad_M =
+            math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
+
+        auto grid_desc_m_padded = transform_tensor_descriptor(
+            grid_desc_m,
+            make_tuple(make_right_pad_transform(invariantLength, pad_M)),
+            make_tuple(Sequence<0>{}),
+            make_tuple(Sequence<0>{}));
+
+        return grid_desc_m_padded;
+    }
+
     using SrcGridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
     using Kernel1MeanVarGridDesc_M_KBlock =
-        decltype(MakeMeanVarDescriptor_M_K<Sequence<true, false>, 1, 1>(1, 1));
+        decltype(MakeWorkspaceMeanVarDescriptor_M_K<Sequence<true, false>, 1, 1>(1, 1));
 
     using Kernel2MeanVarGridDesc_M_KBlock =
-        decltype(MakeMeanVarDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1));
+        decltype(MakeWorkspaceMeanVarDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1));
 
     using Kernel2CountGridDesc_M_KBlock =
-        decltype(MakeCountDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1));
+        decltype(MakeWorkspaceCountDescriptor_M_K<Sequence<true, true>, 1, 1>(1, 1));
+
+    using SaveMeanInvStdGridDesc_M = decltype(MakeSaveMeanInvStdDescriptor_M({1}, {1}));
 
     using GridwiseWelford = GridwiseNormalizationSplitK1st<XDataType,
                                                            ComputeDataType,
-                                                           MeanVarDataType,
+                                                           WorkspaceMeanVarDataType,
                                                            SrcGridDesc_M_K,
                                                            Kernel1MeanVarGridDesc_M_KBlock,
                                                            BlockSize,
@@ -258,16 +307,18 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
                                                            XSrcVectorSize>;
 
     using GridwiseWelfordNormalization =
-        GridwiseNormalizationSplitK2nd<MeanVarDataType,
+        GridwiseNormalizationSplitK2nd<WorkspaceMeanVarDataType,
                                        XDataType,
                                        GammaDataType,
                                        BetaDataType,
                                        YDataType,
+                                       SaveMeanInvStdDataType,
                                        ComputeDataType,
                                        YElementwiseOperation,
                                        Kernel2MeanVarGridDesc_M_KBlock,
                                        Kernel2CountGridDesc_M_KBlock,
                                        SrcGridDesc_M_K,
+                                       SaveMeanInvStdGridDesc_M,
                                        BlockSize,
                                        MThreadClusterSize,
                                        KThreadClusterSize,
@@ -280,7 +331,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
                                        BetaSrcVectorDim,
                                        BetaSrcVectorSize,
                                        XYVectorDim,
-                                       YDstVectorSize>;
+                                       YDstVectorSize,
+                                       SaveMeanInvStdDstVectorSize>;
 
     struct Argument : public BaseArgument
     {
@@ -289,17 +341,23 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
                  const std::vector<index_t> gammaStrides,
                  const std::vector<index_t> betaStrides,
                  const std::vector<index_t> yStrides,
+                 const std::vector<index_t> saveMeanStrides,
+                 const std::vector<index_t> saveInvStdStrides,
                  const std::vector<index_t> reduceDims,
                  YElementwiseOperation y_elementwise_op,
                  double epsilon,
                  const XDataType* p_x,
                  const GammaDataType* p_gamma,
                  const BetaDataType* p_beta,
-                 YDataType* p_y)
+                 YDataType* p_y,
+                 SaveMeanInvStdDataType* p_saveMean,
+                 SaveMeanInvStdDataType* p_saveInvStd)
             : p_x_(p_x),
               p_gamma_(p_gamma),
               p_beta_(p_beta),
               p_y_(p_y),
+              p_saveMean_(p_saveMean),
+              p_saveInvStd_(p_saveInvStd),
               p_workspace_mean_{nullptr},
               p_workspace_var_{nullptr},
               p_workspace_count_{nullptr},
@@ -312,6 +370,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
             yStrides_     = shuffle_tensor_dimensions<Rank, NumReduceDim>(yStrides, reduceDims);
             gammaStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(gammaStrides, reduceDims);
             betaStrides_  = shuffle_tensor_dimensions<Rank, NumReduceDim>(betaStrides, reduceDims);
+            saveMeanStrides_   = saveMeanStrides;
+            saveInvStdStrides_ = saveInvStdStrides;
 
             std::tie(MRaw_, KRaw_) = get_2d_lengths<Rank, NumReduceDim>(Lengths_);
 
@@ -346,20 +406,28 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
             y_grid_desc_m_k_ =
                 MakeSrc2dDescriptor(Lengths_, yStrides_, kGridSize_, numBlockTileIteration_);
 
+            save_mean_grid_desc_m_    = MakeSaveMeanInvStdDescriptor_M(Lengths_, saveMeanStrides);
+            save_inv_std_grid_desc_m_ = MakeSaveMeanInvStdDescriptor_M(Lengths_, saveInvStdStrides);
+
             // We don't need to pad in K dimension for Welford1. Set KPerTile 1.
             kernel1_mean_var_grid_desc_m_kblock_ =
-                MakeMeanVarDescriptor_M_K<Sequence<true, false>, M_BlockTileSize, 1>(MRaw_,
-                                                                                     kGridSize_);
+                MakeWorkspaceMeanVarDescriptor_M_K<Sequence<true, false>, M_BlockTileSize, 1>(
+                    MRaw_, kGridSize_);
 
             kernel2_mean_var_grid_desc_m_kblock_ =
-                MakeMeanVarDescriptor_M_K<Sequence<true, true>,
-                                          M_BlockTileSize,
-                                          K_MeanVarCountBlockTileSize>(MRaw_, kGridSize_);
+                MakeWorkspaceMeanVarDescriptor_M_K<Sequence<true, true>,
+                                                   M_BlockTileSize,
+                                                   K_MeanVarCountBlockTileSize>(MRaw_, kGridSize_);
 
             kernel2_count_grid_desc_m_kblock_ =
-                MakeCountDescriptor_M_K<Sequence<true, true>,
-                                        M_BlockTileSize,
-                                        K_MeanVarCountBlockTileSize>(MRaw_, kGridSize_);
+                MakeWorkspaceCountDescriptor_M_K<Sequence<true, true>,
+                                                 M_BlockTileSize,
+                                                 K_MeanVarCountBlockTileSize>(MRaw_, kGridSize_);
+
+            if constexpr(NumInvariantDim == 0)
+                invariant_lowest_length_ = 1;
+            else
+                invariant_lowest_length_ = Lengths_[NumInvariantDim - 1];
         }
 
         ComputeDataType epsilon_;
@@ -368,6 +436,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         const GammaDataType* p_gamma_;
         const BetaDataType* p_beta_;
         YDataType* p_y_;
+        SaveMeanInvStdDataType* p_saveMean_;
+        SaveMeanInvStdDataType* p_saveInvStd_;
         void* p_workspace_mean_;
         void* p_workspace_var_;
         void* p_workspace_count_;
@@ -377,6 +447,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         std::vector<index_t> gammaStrides_;
         std::vector<index_t> betaStrides_;
         std::vector<index_t> yStrides_;
+        std::vector<index_t> saveMeanStrides_;
+        std::vector<index_t> saveInvStdStrides_;
 
         YElementwiseOperation y_elementwise_op_;
 
@@ -389,6 +461,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         SrcGridDesc_M_K gamma_grid_desc_m_k_;
         SrcGridDesc_M_K beta_grid_desc_m_k_;
         SrcGridDesc_M_K y_grid_desc_m_k_;
+        SaveMeanInvStdGridDesc_M save_mean_grid_desc_m_;
+        SaveMeanInvStdGridDesc_M save_inv_std_grid_desc_m_;
 
         Kernel1MeanVarGridDesc_M_KBlock kernel1_mean_var_grid_desc_m_kblock_;
         Kernel2MeanVarGridDesc_M_KBlock kernel2_mean_var_grid_desc_m_kblock_;
@@ -396,6 +470,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
 
         index_t MRaw_; // invarient length
         index_t KRaw_; // reduce length
+
+        index_t invariant_lowest_length_;
     };
 
     struct Invoker : public BaseInvoker
@@ -408,60 +484,68 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
 
             auto kernel1 = kernel_normalizationSplitK1st<GridwiseWelford,
                                                          XDataType,
-                                                         MeanVarDataType,
+                                                         WorkspaceMeanVarDataType,
                                                          ComputeDataType,
                                                          SrcGridDesc_M_K,
                                                          Kernel1MeanVarGridDesc_M_KBlock>;
 
             auto kernel2 = kernel_normalizationSplitK2nd<GridwiseWelfordNormalization,
-                                                         MeanVarDataType,
+                                                         WorkspaceMeanVarDataType,
                                                          XDataType,
                                                          GammaDataType,
                                                          BetaDataType,
                                                          YDataType,
+                                                         SaveMeanInvStdDataType,
                                                          ComputeDataType,
                                                          YElementwiseOperation,
                                                          Kernel2MeanVarGridDesc_M_KBlock,
                                                          Kernel2CountGridDesc_M_KBlock,
-                                                         SrcGridDesc_M_K>;
+                                                         SrcGridDesc_M_K,
+                                                         SaveMeanInvStdGridDesc_M>;
 
             float avg_time = 0;
-            avg_time += launch_and_time_kernel(stream_config,
-                                               kernel1,
-                                               dim3(arg.gridSize_),
-                                               dim3(BlockSize),
-                                               0,
-                                               arg.x_grid_desc_m_k_,
-                                               arg.kernel1_mean_var_grid_desc_m_kblock_,
-                                               arg.numBlockTileIteration_,
-                                               arg.p_x_,
-                                               static_cast<MeanVarDataType*>(arg.p_workspace_mean_),
-                                               static_cast<MeanVarDataType*>(arg.p_workspace_var_),
-                                               static_cast<int32_t*>(arg.p_workspace_count_));
-
-            avg_time += launch_and_time_kernel(stream_config,
-                                               kernel2,
-                                               dim3(arg.gridSize_),
-                                               dim3(BlockSize),
-                                               0,
-                                               arg.kernel2_mean_var_grid_desc_m_kblock_,
-                                               arg.kernel2_count_grid_desc_m_kblock_,
-                                               arg.x_grid_desc_m_k_,
-                                               arg.gamma_grid_desc_m_k_,
-                                               arg.beta_grid_desc_m_k_,
-                                               arg.y_grid_desc_m_k_,
-                                               arg.numMeanVarCountIteration_,
-                                               arg.numBlockTileIteration_,
-                                               arg.kGridSize_,
-                                               arg.epsilon_,
-                                               static_cast<MeanVarDataType*>(arg.p_workspace_mean_),
-                                               static_cast<MeanVarDataType*>(arg.p_workspace_var_),
-                                               static_cast<int32_t*>(arg.p_workspace_count_),
-                                               arg.p_x_,
-                                               arg.p_gamma_,
-                                               arg.p_beta_,
-                                               arg.p_y_,
-                                               arg.y_elementwise_op_);
+            avg_time += launch_and_time_kernel(
+                stream_config,
+                kernel1,
+                dim3(arg.gridSize_),
+                dim3(BlockSize),
+                0,
+                arg.x_grid_desc_m_k_,
+                arg.kernel1_mean_var_grid_desc_m_kblock_,
+                arg.numBlockTileIteration_,
+                arg.p_x_,
+                static_cast<WorkspaceMeanVarDataType*>(arg.p_workspace_mean_),
+                static_cast<WorkspaceMeanVarDataType*>(arg.p_workspace_var_),
+                static_cast<int32_t*>(arg.p_workspace_count_));
+
+            avg_time += launch_and_time_kernel(
+                stream_config,
+                kernel2,
+                dim3(arg.gridSize_),
+                dim3(BlockSize),
+                0,
+                arg.kernel2_mean_var_grid_desc_m_kblock_,
+                arg.kernel2_count_grid_desc_m_kblock_,
+                arg.x_grid_desc_m_k_,
+                arg.gamma_grid_desc_m_k_,
+                arg.beta_grid_desc_m_k_,
+                arg.y_grid_desc_m_k_,
+                arg.save_mean_grid_desc_m_,
+                arg.save_inv_std_grid_desc_m_,
+                arg.numMeanVarCountIteration_,
+                arg.numBlockTileIteration_,
+                arg.kGridSize_,
+                arg.epsilon_,
+                static_cast<const WorkspaceMeanVarDataType*>(arg.p_workspace_mean_),
+                static_cast<const WorkspaceMeanVarDataType*>(arg.p_workspace_var_),
+                static_cast<const int32_t*>(arg.p_workspace_count_),
+                arg.p_x_,
+                arg.p_gamma_,
+                arg.p_beta_,
+                arg.p_y_,
+                arg.p_saveMean_,
+                arg.p_saveInvStd_,
+                arg.y_elementwise_op_);
 
             return avg_time;
         };
@@ -482,10 +566,10 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         int welford_size = pArg_->MRaw_ * pArg_->kGridSize_;
 
         // workspace for welford intermediate mean
-        workspace_size += welford_size * sizeof(MeanVarDataType) + 64;
+        workspace_size += welford_size * sizeof(WorkspaceMeanVarDataType) + 64;
 
         // workspace for welford intermediate variance
-        workspace_size += welford_size * sizeof(MeanVarDataType) + 64;
+        workspace_size += welford_size * sizeof(WorkspaceMeanVarDataType) + 64;
 
         // workspace for welford intermediate count
         workspace_size += pArg_->kGridSize_ * sizeof(int32_t) + 64;
@@ -504,13 +588,13 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         // setup buffer used for intermediate welford mean
         pArg_->p_workspace_mean_ = static_cast<char*>(pArg_->p_workspace_);
 
-        index_t mean_space_sz = welford_size * sizeof(MeanVarDataType);
+        index_t mean_space_sz = welford_size * sizeof(WorkspaceMeanVarDataType);
         mean_space_sz         = math::integer_least_multiple(mean_space_sz, 64);
 
         // setup buffer used for intermediate welford varirance
         pArg_->p_workspace_var_ = reinterpret_cast<char*>(pArg_->p_workspace_mean_) + mean_space_sz;
 
-        index_t variance_space_sz = welford_size * sizeof(MeanVarDataType);
+        index_t variance_space_sz = welford_size * sizeof(WorkspaceMeanVarDataType);
         variance_space_sz         = math::integer_least_multiple(variance_space_sz, 64);
 
         // setup buffer used for intermediate welford count
@@ -522,8 +606,6 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
     {
         const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
 
-        constexpr index_t NumInvariantDim = Rank - NumReduceDim;
-
         if constexpr(XYVectorDim == 0)
         {
             if constexpr(NumInvariantDim == 0)
@@ -535,10 +617,10 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
                 if(p_arg_->xStrides_[NumInvariantDim - 1] != 1)
                     return false;
 
-                if(p_arg_->invariant_lowest_length % XSrcVectorSize != 0)
+                if(p_arg_->invariant_lowest_length_ % XSrcVectorSize != 0)
                     return false;
 
-                if(p_arg_->invariant_lowest_length % YDstVectorSize != 0)
+                if(p_arg_->invariant_lowest_length_ % YDstVectorSize != 0)
                     return false;
             };
         }
@@ -578,7 +660,7 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
             if(p_arg_->betaStrides_[NumInvariantDim - 1] != 1)
                 return false;
 
-            if(p_arg_->invariant_lowest_length % BetaSrcVectorSize != 0)
+            if(p_arg_->invariant_lowest_length_ % BetaSrcVectorSize != 0)
                 return false;
         }
         else // if fastest dim is reduced
@@ -593,6 +675,9 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         if(p_arg_->kGridSize_ <= 1)
             return false;
 
+        if(p_arg_->invariant_lowest_length_ % SaveMeanInvStdDstVectorSize != 0)
+            return false;
+
         return true;
     };
 
@@ -602,6 +687,8 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
                         const std::vector<index_t> gammaStrides,
                         const std::vector<index_t> betaStrides,
                         const std::vector<index_t> yStrides,
+                        const std::vector<index_t> saveMeanStrides,
+                        const std::vector<index_t> saveInvStdStrides,
                         const std::vector<index_t> reduceDims,
                         double epsilon,
                         const void* p_x,
@@ -609,27 +696,30 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
                         const void* p_beta,
                         void* p_y,
                         void* p_saveMean,
-                        void* p_saveInvVar,
+                        void* p_saveInvStd,
                         YElementwiseOperation y_elementwise_op) override
     {
-        // TODO
-        // Optional cache of the intermediate results (mean and InvVariance) during the
-        // forward pass could speedup in the backward
-        ignore = p_saveMean;
-        ignore = p_saveInvVar;
+        if(lengths.size() != Rank || xStrides.size() != Rank || gammaStrides.size() != Rank ||
+           betaStrides.size() != Rank || yStrides.size() != Rank ||
+           saveMeanStrides.size() != NumInvariantDim || saveInvStdStrides.size() != NumInvariantDim)
+            throw std::runtime_error("dimension is incorrect");
 
         return std::make_unique<Argument>(lengths,
                                           xStrides,
                                           gammaStrides,
                                           betaStrides,
                                           yStrides,
+                                          saveMeanStrides,
+                                          saveInvStdStrides,
                                           reduceDims,
                                           y_elementwise_op,
                                           epsilon,
                                           static_cast<const XDataType*>(p_x),
                                           static_cast<const GammaDataType*>(p_gamma),
                                           static_cast<const BetaDataType*>(p_beta),
-                                          static_cast<YDataType*>(p_y));
+                                          static_cast<YDataType*>(p_y),
+                                          static_cast<SaveMeanInvStdDataType*>(p_saveMean),
+                                          static_cast<SaveMeanInvStdDataType*>(p_saveInvStd));
     };
 
     std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
@@ -642,7 +732,7 @@ struct DeviceNormalizationSplitKImpl : public DeviceNormalization<XDataType,
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceNormalizationSplitKImpl<" << BlockSize << ",";
+        str << "DeviceNormalizationFwdSplitKImpl<" << BlockSize << ",";
         str << "Cluster_MK_" << MThreadClusterSize << "_" << KThreadClusterSize << ",";
         str << "Slice_MK_" << MThreadSliceSize << "_" << KThreadSliceSize << ",";
         str << "XYSrcVectorDim_" << XYVectorDim  << ",";

include/ck/tensor_operation/gpu/element/binary_element_wise_operation.hpp

@@ -85,10 +85,13 @@ struct Add
 
 struct ScaleAdd
 {
-    __host__ __device__ ScaleAdd(float scale) : scale_(scale) {}
+    __host__ __device__ ScaleAdd(float scale = 1.f) : scale_(scale) {}
 
     template <typename Y, typename X0, typename X1>
-    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const
+    {
+        y = ck::type_convert<Y>(scale_ * ck::type_convert<float>(x0) + ck::type_convert<float>(x1));
+    }
 
     template <>
     __host__ __device__ void
@@ -186,6 +189,32 @@ struct Bilinear
         y = type_convert<half_t>(alpha_ * x0 + beta_ * ck::type_convert<float>(x1));
     };
 
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t, bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x0, const bhalf_t& x1) const
+    {
+        const float x0_tmp = type_convert<float>(x0);
+        const float x1_tmp = type_convert<float>(x1);
+        const float y_tmp  = alpha_ * x0_tmp + beta_ * x1_tmp;
+        y                  = type_convert<bhalf_t>(y_tmp);
+    };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<bhalf_t, float, bhalf_t>(bhalf_t& y, const float& x0, const bhalf_t& x1) const
+    {
+        const float x1_tmp = ck::type_convert<float>(x1);
+        const float y_tmp  = alpha_ * x0 + beta_ * x1_tmp;
+        y                  = y_tmp;
+    };
+
+    template <>
+    __host__ __device__ constexpr void operator()<std::int8_t, std::int32_t, std::int8_t>(
+        std::int8_t& y, const std::int32_t& x0, const std::int8_t& x1) const
+    {
+        y = type_convert<std::int8_t>(x0 + ck::type_convert<std::int32_t>(x1));
+    };
+
     float alpha_;
     float beta_;
 };

include/ck/tensor_operation/gpu/element/element_wise_operation.hpp

@@ -311,6 +311,71 @@ struct AddAddFastGelu
     }
 };
 
+// E = Relu(alpha1 * C + alpha2 * D0 + D1)
+struct ScaleAddScaleAddRelu
+{
+
+    ScaleAddScaleAddRelu(const float alpha1 = 1.f, const float alpha2 = 1.f)
+        : alpha1_(alpha1), alpha2_(alpha2)
+    {
+    }
+
+    template <typename E, typename C, typename D0, typename D1>
+    __host__ __device__ constexpr void
+    operator()(E& e, const C& c, const D0& d0, const D1& d1) const;
+
+    template <>
+    __host__ __device__ constexpr void operator()<float, float, float, float>(float& e,
+                                                                              const float& c,
+                                                                              const float& d0,
+                                                                              const float& d1) const
+    {
+        const float x = c * alpha1_ + alpha2_ * d0 + d1;
+        Relu{}.template operator()<float>(e, x);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<half_t, half_t, half_t, half_t>(
+        half_t& e, const half_t& c, const half_t& d0, const half_t& d1) const
+    {
+        const float x = type_convert<float>(c) * alpha1_ + alpha2_ * type_convert<float>(d0) +
+                        type_convert<float>(d1);
+
+        float result = 0;
+        Relu{}.template operator()<float>(result, x);
+
+        e = type_convert<half_t>(result);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<bhalf_t, bhalf_t, bhalf_t, bhalf_t>(
+        bhalf_t& e, const bhalf_t& c, const bhalf_t& d0, const bhalf_t& d1) const
+    {
+        const float x = type_convert<float>(c) * alpha1_ + alpha2_ * type_convert<float>(d0) +
+                        type_convert<float>(d1);
+
+        float result = 0;
+        Relu{}.template operator()<float>(result, x);
+
+        e = type_convert<bhalf_t>(result);
+    }
+
+    template <>
+    __host__ __device__ constexpr void operator()<int8_t, int8_t, float, float>(
+        int8_t& e, const int8_t& c, const float& d0, const float& d1) const
+    {
+        const float x = type_convert<float>(c) * alpha1_ + alpha2_ * d0 + d1;
+
+        float result = 0;
+        Relu{}.template operator()<float>(result, x);
+
+        e = type_convert<int8_t>(result);
+    }
+
+    const float alpha1_;
+    const float alpha2_;
+};
+
 struct Normalize
 {
     // FIXME: is double absolutely necessary?

include/ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp

@@ -16,6 +16,57 @@ namespace element_wise {
 extern "C" __device__ float __ocml_native_recip_f32(float);
 #endif
 
+struct PassThroughPack2
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    __host__ __device__ constexpr void operator()(ck::f8x2_t& y, const ck::half2_t& x) const
+    {
+        // fake conversion
+        uint16_t t = ck::bit_cast<uint32_t>(x);
+        y          = ck::bit_cast<ck::f8x2_t>(t);
+    }
+
+    __host__ __device__ constexpr void operator()(ck::half2_t& y, const ck::f8x2_t& x) const
+    {
+        auto t = type_convert<float2_t>(x);
+        y      = type_convert<half2_t>(t);
+    }
+
+    __host__ __device__ constexpr void operator()(ck::half2_t& y, const ck::half2_t& x) const
+    {
+        y = x;
+    }
+
+    __host__ __device__ constexpr void operator()(ck::f8x2_t& y, const ck::f8x2_t& x) const
+    {
+        y = x;
+    }
+
+    __host__ __device__ constexpr void operator()(ck::float2_t& y, const ck::float2_t& x) const
+    {
+        y = x;
+    }
+
+    __host__ __device__ constexpr void operator()(ck::int8x2_t& y, const ck::int8x2_t& x) const
+    {
+        y = x;
+    }
+
+    __host__ __device__ constexpr void operator()(ck::bhalf2_t& y, const ck::bhalf2_t& x) const
+    {
+        y = x;
+    }
+
+    __host__ __device__ constexpr void operator()(ck::double2_t& y, const ck::double2_t& x) const
+    {
+        y = x;
+    }
+
+    constexpr const static bool is_pack2_invocable = true;
+};
+
 struct PassThrough
 {
     template <typename Y, typename X>
@@ -27,6 +78,18 @@ struct PassThrough
         y = x;
     }
 
+    template <>
+    __host__ __device__ void operator()<float, double>(float& y, const double& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<double, float>(double& y, const float& x) const
+    {
+        y = type_convert<double>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<float, float>(float& y, const float& x) const
     {
@@ -63,24 +126,48 @@ struct PassThrough
         y = type_convert<bhalf_t>(x);
     }
 
+    template <>
+    __host__ __device__ void operator()<float, bhalf_t>(float& y, const bhalf_t& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<bhalf_t, half_t>(bhalf_t& y, const half_t& x) const
     {
         y = type_convert<bhalf_t>(x);
     }
 
+    template <>
+    __host__ __device__ void operator()<float, half_t>(float& y, const half_t& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<int8_t, int8_t>(int8_t& y, const int8_t& x) const
     {
         y = x;
     }
 
+    template <>
+    __host__ __device__ void operator()<half_t, int8_t>(half_t& y, const int8_t& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
     template <>
     __host__ __device__ void operator()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
     {
         y = type_convert<int8_t>(x);
     }
 
+    template <>
+    __host__ __device__ void operator()<int8_t, float>(int8_t& y, const float& x) const
+    {
+        y = type_convert<int8_t>(x);
+    }
+
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
     template <>
     __host__ __device__ void operator()<int4_t, int4_t>(int4_t& y, const int4_t& x) const
@@ -118,6 +205,36 @@ struct PassThrough
     {
         y = type_convert<f8_t>(x);
     }
+
+    template <>
+    __host__ __device__ void operator()<bf8_t, bf8_t>(bf8_t& y, const bf8_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<float, bf8_t>(float& y, const bf8_t& x) const
+    {
+        y = type_convert<float>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<bf8_t, float>(bf8_t& y, const float& x) const
+    {
+        y = type_convert<bf8_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<half_t, bf8_t>(half_t& y, const bf8_t& x) const
+    {
+        y = type_convert<half_t>(x);
+    }
+
+    template <>
+    __host__ __device__ void operator()<bf8_t, half_t>(bf8_t& y, const half_t& x) const
+    {
+        y = ck::type_convert<bf8_t>(x);
+    }
 };
 
 struct UnaryConvert
@@ -153,7 +270,8 @@ struct ConvertF8SR
     __host__ __device__ void operator()(Y& y, const X& x) const
     {
         // check Y datatype
-        static_assert(is_same<Y, f8_t>::value, "Data type is not supported by this operation!");
+        static_assert(is_same<Y, f8_t>::value || is_same<Y, bf8_t>::value,
+                      "Data type is not supported by this operation!");
 
         // check X datatype
         static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
@@ -163,6 +281,24 @@ struct ConvertF8SR
     }
 };
 
+struct ConvertF8RNE
+{
+    // convert to fp8 using rounding to nearest even
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        // check Y datatype
+        static_assert(is_same<Y, f8_t>::value || is_same<Y, bf8_t>::value,
+                      "Data type is not supported by this operation!");
+
+        // check X datatype
+        static_assert(is_same<X, float>::value || is_same<X, half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        y = f8_convert_rne<Y>(x);
+    }
+};
+
 struct Scale
 {
     __host__ __device__ Scale(float scale) : scale_(scale) {}
@@ -170,6 +306,20 @@ struct Scale
     template <typename Y, typename X>
     __host__ __device__ void operator()(Y& y, const X& x) const;
 
+    template <>
+    __host__ __device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        y = ck::type_convert<half_t>(scale_) * x;
+    };
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
+    {
+        const float x_tmp = ck::type_convert<float>(x);
+        const float y_tmp = scale_ * x_tmp;
+        y                 = ck::type_convert<bhalf_t>(y_tmp);
+    };
+
     template <>
     __host__ __device__ void operator()<float, float>(float& y, const float& x) const
     {
@@ -223,8 +373,8 @@ struct UnarySquare
     template <typename T>
     __host__ __device__ void operator()(T& y, const T& x) const
     {
-        static_assert(is_same_v<T, float> || is_same_v<T, double> || is_same_v<T, int32_t> ||
-                          is_same_v<T, int8_t>
+        static_assert(is_same_v<T, float> || is_same_v<T, half_t> || is_same_v<T, double> ||
+                          is_same_v<T, int32_t> || is_same_v<T, int8_t>
 #ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
                           || is_same_v<T, int4_t>
 #endif
@@ -388,10 +538,11 @@ struct Sigmoid
     __host__ __device__ void operator()(T& y, const T& x) const
     {
         static_assert(is_same<T, float>::value || is_same<T, double>::value ||
-                          is_same<T, ck::half_t>::value,
+                          is_same<T, ck::half_t>::value || is_same<T, int8_t>::value ||
+                          is_same<T, int32_t>::value,
                       "Data type is not supported by this operation!");
-
-        y = 1 / (ck::type_convert<T>(1) + exp(-x));
+        constexpr T one = type_convert<T>(1);
+        y               = one / (one + ck::math::exp(-x));
     };
 };
 
@@ -401,7 +552,8 @@ struct TanH
     __host__ __device__ void operator()(T& y, const T& x) const
     {
         static_assert(is_same<T, float>::value || is_same<T, double>::value ||
-                          is_same<T, ck::half_t>::value,
+                          is_same<T, ck::half_t>::value || is_same<T, int8_t>::value ||
+                          is_same<T, int32_t>::value,
                       "Data type is not supported by this operation!");
 
         y = ck::math::tanh(x);
@@ -412,17 +564,116 @@ struct Swish
 {
     Swish(float beta = 1.0f) : beta_(beta) {}
 
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        static_assert(is_same<X, float>::value || is_same<X, double>::value ||
+                          is_same<X, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        static_assert(is_same<Y, float>::value || is_same<Y, double>::value ||
+                          is_same<Y, ck::half_t>::value,
+                      "Data type is not supported by this operation!");
+
+        float bx = -beta_ * type_convert<float>(x);
+        y        = type_convert<Y>(x / (1.f + ck::math::exp(bx)));
+    };
+
+    const float beta_;
+};
+
+struct SoftRelu
+{
+    SoftRelu(float alpha = 1.f) : alpha_(alpha){};
+
     template <typename T>
     __host__ __device__ void operator()(T& y, const T& x) const
     {
         static_assert(is_same<T, float>::value || is_same<T, double>::value ||
-                          is_same<T, ck::half_t>::value,
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
                       "Data type is not supported by this operation!");
+        T casted_alpha  = type_convert<T>(alpha_);
+        constexpr T one = type_convert<T>(1);
+        y               = ck::math::log(one + ck::math::exp(x * casted_alpha)) / casted_alpha;
+    }
+    const float alpha_;
+};
 
-        y = x / (ck::type_convert<T>(1) + ck::math::exp(-beta_ * x));
-    };
+struct Power
+{
+    Power(float alpha = 0.f, float beta = 1.f, float gamma = 2.f)
+        : alpha_(alpha), beta_(beta), gamma_(gamma){};
 
-    float beta_ = 1.0f;
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        T casted_alpha     = type_convert<T>(alpha_);
+        T casted_beta      = type_convert<T>(beta_);
+        T casted_gamma     = type_convert<T>(gamma_);
+        T shifted_scaled_x = casted_alpha + casted_beta * x;
+        y                  = ck::math::pow(shifted_scaled_x, casted_gamma);
+    }
+    const float alpha_;
+    const float beta_;
+    const float gamma_;
+};
+
+struct ClippedRelu
+{
+    ClippedRelu(float alpha = 0.f, float beta = 1.f) : alpha_(alpha), beta_(beta){};
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        T casted_alpha = type_convert<T>(alpha_);
+        T casted_beta  = type_convert<T>(beta_);
+        y              = ck::math::min(casted_beta, ck::math::max(casted_alpha, x));
+    }
+    const float alpha_;
+    const float beta_;
+};
+
+struct LeakyRelu
+{
+    LeakyRelu(float alpha = 0.01f) : alpha_(alpha){};
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        T casted_alpha = type_convert<T>(alpha_);
+        y              = x >= 0 ? x : x * casted_alpha;
+    }
+    const float alpha_;
+};
+
+struct Elu
+{
+    Elu(float alpha = 1.f) : alpha_(alpha){};
+
+    template <typename T>
+    __host__ __device__ void operator()(T& y, const T& x) const
+    {
+        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
+                          is_same<T, half_t>::value || is_same<T, int32_t>::value ||
+                          is_same<T, int8_t>::value,
+                      "Data type is not supported by this operation!");
+        T casted_alpha = type_convert<T>(alpha_);
+        y              = x > 0 ? x : casted_alpha * ck::math::expm1(x);
+    }
+    const float alpha_;
 };
 
 } // namespace element_wise

include/ck/tensor_operation/gpu/grid/gemm_layernorm/gridwise_gemm_multiple_d_welford_first_half_xdl_cshuffle.hpp

@@ -522,6 +522,7 @@ struct GridwiseGemmMultipleDWelfordFirstHalf_xdl_cshuffle
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
             ABDataType,
+            ABDataType,
             AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_gemm_xdl_cshuffle_v1.hpp

@@ -628,7 +628,8 @@ struct GridwiseBatchedGemmGemm_Xdl_CShuffle
             Gemm1KPack,
             false,      // TransposeC
             Gemm1KPack, // AMmaKStride
-            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            Gemm1KPack *
+                XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, FloatAB, false>{}.K0PerXdlops>{
             // BMmaKStride
             make_multi_index(0, 0, 0, 0)}; // A_origin
 

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_gemm_multiple_d_xdl_cshuffle_v1.hpp

@@ -880,7 +880,12 @@ struct GridwiseBatchedGemmMultipleDGemmMultipleD_Xdl_CShuffle
             Gemm1KPack,
             false,      // TransposeC
             Gemm1KPack, // AMmaKStride
-            Gemm1KPack * XdlopsGemm<A0B0B1DataType, Gemm0MPerXdl, Gemm0NPerXdl, Gemm1KPack, false>{}
+            Gemm1KPack * XdlopsGemm<A0B0B1DataType,
+                                    Gemm0MPerXdl,
+                                    Gemm0NPerXdl,
+                                    Gemm1KPack,
+                                    A0B0B1DataType,
+                                    false>{}
                              .K0PerXdlops>{                               // BMmaKStride
                                            make_multi_index(0, 0, 0, 0)}; // A_origin
 

include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_multiple_d_softmax_gemm_xdl_cshuffle_v1.hpp

@@ -794,7 +794,8 @@ struct GridwiseBatchedGemmMultipleDSoftmaxGemm_Xdl_CShuffle
             Gemm1KPack,
             true,       // TransposeC
             Gemm1KPack, // AMmaKStride
-            Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
+            Gemm1KPack *
+                XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, FloatAB, false>{}.K0PerXdlops>{
             // BMmaKStride
             make_multi_index(0, 0, 0, 0)}; // A_origin