Clean up conv example, Instances, profiler and test (#324)

* convnd_fwd fp16 example

* update example

* update example

* update instance

* updating refernce conv

* update reference conv

* update conv fwd profiler

* update conv 1d and 3d instance

* update include path

* clean

* update profiler for conv bwd data and weight

* update conv bwd weight

* clean

* update conv example

* update profiler for conv bwd weight

* update ckprofiler for conv bwd data

* fix reference conv bwd data bug; update conv bwd data test

* update examples

* fix initialization issue

* update test for conv fwd

* clean

* clean

* remove test case too sensitive to error threshhold

* fix test

* clean

* fix build

* adding conv multiple d

* adding conv multiple D

* add matrix padder

* add gemm padding to convnd

* adding group conv

* update gemm multi-d

* refactor

* refactor

* refactor

* clean

* clean

* refactor

* refactor

* reorg

* add ds

* add bias

* clean

* add G

* adding group

* adding group

* adding group

* update Tensor

* clean

* update example

* update DeviceGemmMultipleD_Xdl_CShuffle

* update conv bwd-data and bwd-weight

* upate contraction example

* update gemm and batch gemm with e permute

* fix example build

* instance for grouped conv1d

* update example

* adding group conv instance

* update gemm bilinear instance

* update gemm+add+add+fastgelu instance

* update profiler

* update profiler

* update test

* update test and client example

* clean

* add grouped conv into profiler

* update profiler

* clean

* add test grouped conv, update all conv test to gtest

* update test

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

@@ -13,8 +13,8 @@
 #include "ck/tensor_operation/gpu/device/device_gemm.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

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

@@ -13,8 +13,8 @@
 #include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

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

@@ -13,8 +13,8 @@
 #include "ck/tensor_operation/gpu/device/device_gemm_splitk.hpp"
 #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include <vector>
+
+#include "ck/tensor_operation/gpu/device/device_base.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// Convolution Forward:
+//   input : input image A[G, N, C, Hi, Wi],
+//   input : weight B[G, K, C, Y, X],
+//   input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ...
+//   output : output image E[G, N, K, Ho, Wo]
+//   C = a_op(A) * b_op(B)
+//   E = cde_op(C, D0, D1, ...)
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          typename ADataType,
+          typename BDataType,
+          typename DsDataType,
+          typename EDataType,
+          typename AElementwiseOperation,
+          typename BElementwiseOperation,
+          typename CDEElementwiseOperation>
+struct DeviceGroupedConvFwdMultipleD : public BaseOperator
+{
+    static constexpr index_t NumDTensor = DsDataType::Size();
+
+    virtual 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) = 0;
+
+    virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, 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/device_grouped_conv_fwd_multiple_d.hpp"
+#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.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"
+
+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
+ * 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_batch_gemm_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__))
+
+#if 1
+    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
+    __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
+
+    GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
+                                                  p_b_grid,
+                                                  p_ds_grid,
+                                                  p_e_grid,
+                                                  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);
+#endif
+
+#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.
+//
+// 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,
+          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 matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    template <typename ALay,
+              typename std::enable_if<NDimSpatial == 1 &&
+                                          is_same_v<ALay, tensor_layout::convolution::GNWC>,
+                                      bool>::type = false>
+    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 index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Wi = a_g_n_c_wis_lengths[3];
+
+        const index_t Wo = e_g_n_k_wos_lengths[3];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NWo = N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                                    e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                                    index_t{1},
+                                                    std::multiplies<index_t>());
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NWo, C));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmmraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+                in_n_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else
+        {
+            const index_t X             = b_g_k_c_xs_lengths[3];
+            const index_t ConvDilationW = conv_filter_dilations[0];
+            const index_t InLeftPadW    = input_left_pads[0];
+            const index_t InRightPadW   = input_right_pads[0];
+
+            const auto in_n_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Wi, C));
+
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Wo)),
+                                                       make_merge_transform(make_tuple(X, C))),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+    }
+
+    template <typename ALay,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          is_same_v<ALay, tensor_layout::convolution::GNHWC>,
+                                      bool>::type = false>
+    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 index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Hi = a_g_n_c_wis_lengths[3];
+        const index_t Wi = a_g_n_c_wis_lengths[4];
+
+        const index_t Ho = e_g_n_k_wos_lengths[3];
+        const index_t Wo = e_g_n_k_wos_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo = N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                                      e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                                      index_t{1},
+                                                      std::multiplies<index_t>());
+
+            const auto in_gemmmraw_gemmkraw_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NHoWo, C));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                transform_tensor_descriptor(in_n_ho_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else
+        {
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const index_t InLeftPadH = input_left_pads[0];
+            const index_t InLeftPadW = input_left_pads[1];
+
+            const index_t InRightPadH = input_right_pads[0];
+            const index_t InRightPadW = input_right_pads[1];
+
+            const auto in_n_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_merge_transform(make_tuple(Y, X, C))),
+                                            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+    }
+
+    template <typename ALay,
+              typename std::enable_if<NDimSpatial == 3 &&
+                                          is_same_v<ALay, tensor_layout::convolution::GNDHWC>,
+                                      bool>::type = false>
+    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 index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Di = a_g_n_c_wis_lengths[3];
+        const index_t Hi = a_g_n_c_wis_lengths[4];
+        const index_t Wi = a_g_n_c_wis_lengths[5];
+
+        const index_t Do = e_g_n_k_wos_lengths[3];
+        const index_t Ho = e_g_n_k_wos_lengths[4];
+        const index_t Wo = e_g_n_k_wos_lengths[5];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NDoHoWo =
+                N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                    e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                    index_t{1},
+                                    std::multiplies<index_t>());
+
+            const auto in_gemmmraw_gemmkraw_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(NDoHoWo, C));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            const auto in_n_di_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            const auto in_n_do_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmmraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+                in_n_do_ho_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else
+        {
+            const index_t Z = b_g_k_c_xs_lengths[3];
+            const index_t Y = b_g_k_c_xs_lengths[4];
+            const index_t X = b_g_k_c_xs_lengths[5];
+
+            const index_t ConvDilationD = conv_filter_dilations[0];
+            const index_t ConvDilationH = conv_filter_dilations[1];
+            const index_t ConvDilationW = conv_filter_dilations[2];
+
+            const index_t InLeftPadD = input_left_pads[0];
+            const index_t InLeftPadH = input_left_pads[1];
+            const index_t InLeftPadW = input_left_pads[2];
+
+            const index_t InRightPadD = input_right_pads[0];
+            const index_t InRightPadH = input_right_pads[1];
+            const index_t InRightPadW = input_right_pads[2];
+
+            const auto in_n_di_hi_wi_c_grid_desc =
+                make_naive_tensor_descriptor_packed(make_tuple(N, Di, Hi, Wi, C));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmmraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_merge_transform(make_tuple(Z, Y, X, C))),
+                make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+    }
+
+    // TODO: implement ck::tensor_layout::convolution that describe packed/strided dimemsion as
+    // properties
+    template <typename ALay,
+              typename std::enable_if<NDimSpatial == 1 &&
+                                          (is_same_v<ALay, tensor_layout::convolution::G_NW_C> ||
+                                           is_same_v<ALay, tensor_layout::convolution::NWGC>),
+                                      bool>::type = false>
+    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 index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Wi = a_g_n_c_wis_lengths[3];
+
+        const index_t Wo = e_g_n_k_wos_lengths[3];
+
+        const index_t ConvStrideW = conv_filter_strides[0];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo = N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                                      e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                                      index_t{1},
+                                                      std::multiplies<index_t>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                make_naive_tensor_descriptor(make_tuple(NHoWo, C), make_tuple(WiStride, CStride));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t WiStride = a_g_n_c_wis_strides[3];
+            const auto CStride     = I1;
+
+            const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(NStride, WiStride, CStride));
+
+            const auto in_n_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_gemmmraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+                in_n_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Wo)), make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else
+        {
+            const index_t X             = b_g_k_c_xs_lengths[3];
+            const index_t ConvDilationW = conv_filter_dilations[0];
+            const index_t InLeftPadW    = input_left_pads[0];
+            const index_t InRightPadW   = input_right_pads[0];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t WiStride = a_g_n_c_wis_strides[3];
+            const auto CStride     = I1;
+
+            const auto in_n_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Wi, C), make_tuple(NStride, WiStride, CStride));
+
+            const auto in_n_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+            const auto in_n_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                transform_tensor_descriptor(in_n_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Wo)),
+                                                       make_merge_transform(make_tuple(X, C))),
+                                            make_tuple(Sequence<0, 2>{}, Sequence<1, 3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+    }
+
+    template <typename ALay,
+              typename std::enable_if<NDimSpatial == 2 &&
+                                          (is_same_v<ALay, tensor_layout::convolution::G_NHW_C> ||
+                                           is_same_v<ALay, tensor_layout::convolution::NHWGC>),
+                                      bool>::type = false>
+    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 index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Hi = a_g_n_c_wis_lengths[3];
+        const index_t Wi = a_g_n_c_wis_lengths[4];
+
+        const index_t Ho = e_g_n_k_wos_lengths[3];
+        const index_t Wo = e_g_n_k_wos_lengths[4];
+
+        const index_t ConvStrideH = conv_filter_strides[0];
+        const index_t ConvStrideW = conv_filter_strides[1];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NHoWo = N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                                      e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                                      index_t{1},
+                                                      std::multiplies<index_t>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmmraw_gemmkraw_grid_desc =
+                make_naive_tensor_descriptor(make_tuple(NHoWo, C), make_tuple(WiStride, CStride));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t HiStride = a_g_n_c_wis_strides[3];
+            const index_t WiStride = a_g_n_c_wis_strides[4];
+            const auto CStride     = I1;
+
+            const auto in_n_hi_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
+
+            const auto in_n_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                transform_tensor_descriptor(in_n_ho_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_pass_through_transform(C)),
+                                            make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else
+        {
+            const index_t Y = b_g_k_c_xs_lengths[3];
+            const index_t X = b_g_k_c_xs_lengths[4];
+
+            const index_t ConvDilationH = conv_filter_dilations[0];
+            const index_t ConvDilationW = conv_filter_dilations[1];
+
+            const index_t InLeftPadH = input_left_pads[0];
+            const index_t InLeftPadW = input_left_pads[1];
+
+            const index_t InRightPadH = input_right_pads[0];
+            const index_t InRightPadW = input_right_pads[1];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t HiStride = a_g_n_c_wis_strides[3];
+            const index_t WiStride = a_g_n_c_wis_strides[4];
+            const auto CStride     = I1;
+
+            const auto in_n_hi_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Hi, Wi, C), make_tuple(NStride, HiStride, WiStride, CStride));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+            const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+            const auto in_gemmmraw_gemmk_grid_desc =
+                transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                            make_tuple(make_merge_transform(make_tuple(N, Ho, Wo)),
+                                                       make_merge_transform(make_tuple(Y, X, C))),
+                                            make_tuple(Sequence<0, 2, 4>{}, Sequence<1, 3, 5>{}),
+                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmk_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+    }
+
+    template <typename ALay,
+              typename std::enable_if<NDimSpatial == 3 &&
+                                          (is_same_v<ALay, tensor_layout::convolution::G_NDHW_C> ||
+                                           is_same_v<ALay, tensor_layout::convolution::NDHWGC>),
+                                      bool>::type = false>
+    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 index_t N = a_g_n_c_wis_lengths[1];
+        const index_t C = a_g_n_c_wis_lengths[2];
+
+        const index_t Di = a_g_n_c_wis_lengths[3];
+        const index_t Hi = a_g_n_c_wis_lengths[4];
+        const index_t Wi = a_g_n_c_wis_lengths[5];
+
+        const index_t Do = e_g_n_k_wos_lengths[3];
+        const index_t Ho = e_g_n_k_wos_lengths[4];
+        const index_t Wo = e_g_n_k_wos_lengths[5];
+
+        const index_t ConvStrideD = conv_filter_strides[0];
+        const index_t ConvStrideH = conv_filter_strides[1];
+        const index_t ConvStrideW = conv_filter_strides[2];
+
+        if constexpr(ConvForwardSpecialization ==
+                     ConvolutionForwardSpecialization::Filter1x1Stride1Pad0)
+        {
+            const index_t NDoHoWo =
+                N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                    e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                    index_t{1},
+                                    std::multiplies<index_t>());
+
+            // This is different
+            const index_t WiStride = a_g_n_c_wis_strides[2 + NDimSpatial];
+            const auto CStride     = I1;
+
+            const auto in_gemmmraw_gemmkraw_grid_desc =
+                make_naive_tensor_descriptor(make_tuple(NDoHoWo, C), make_tuple(WiStride, CStride));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else if constexpr(ConvForwardSpecialization ==
+                          ConvolutionForwardSpecialization::Filter1x1Pad0)
+        {
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t DiStride = a_g_n_c_wis_strides[3];
+            const index_t HiStride = a_g_n_c_wis_strides[4];
+            const index_t WiStride = a_g_n_c_wis_strides[5];
+            const auto CStride     = I1;
+
+            const auto in_n_di_hi_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+
+            const auto in_n_do_ho_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_embed_transform(make_tuple(Do), make_tuple(ConvStrideD)),
+                           make_embed_transform(make_tuple(Ho), make_tuple(ConvStrideH)),
+                           make_embed_transform(make_tuple(Wo), make_tuple(ConvStrideW)),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_gemmmraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+                in_n_do_ho_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_pass_through_transform(C)),
+                make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+        else
+        {
+            const index_t Z = b_g_k_c_xs_lengths[3];
+            const index_t Y = b_g_k_c_xs_lengths[4];
+            const index_t X = b_g_k_c_xs_lengths[5];
+
+            const index_t ConvDilationD = conv_filter_dilations[0];
+            const index_t ConvDilationH = conv_filter_dilations[1];
+            const index_t ConvDilationW = conv_filter_dilations[2];
+
+            const index_t InLeftPadD = input_left_pads[0];
+            const index_t InLeftPadH = input_left_pads[1];
+            const index_t InLeftPadW = input_left_pads[2];
+
+            const index_t InRightPadD = input_right_pads[0];
+            const index_t InRightPadH = input_right_pads[1];
+            const index_t InRightPadW = input_right_pads[2];
+
+            // This is different
+            const index_t NStride  = a_g_n_c_wis_strides[1];
+            const index_t DiStride = a_g_n_c_wis_strides[3];
+            const index_t HiStride = a_g_n_c_wis_strides[4];
+            const index_t WiStride = a_g_n_c_wis_strides[5];
+            const auto CStride     = I1;
+
+            const auto in_n_di_hi_wi_c_grid_desc = make_naive_tensor_descriptor(
+                make_tuple(N, Di, Hi, Wi, C),
+                make_tuple(NStride, DiStride, HiStride, WiStride, CStride));
+
+            const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                in_n_di_hi_wi_c_grid_desc,
+                make_tuple(make_pass_through_transform(N),
+                           make_pad_transform(Di, InLeftPadD, InRightPadD),
+                           make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                           make_pad_transform(Wi, InLeftPadW, InRightPadW),
+                           make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}));
+
+            const auto in_n_z_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                in_n_hip_wip_c_grid_desc,
+                make_tuple(
+                    make_pass_through_transform(N),
+                    make_embed_transform(make_tuple(Z, Do), make_tuple(ConvDilationD, ConvStrideD)),
+                    make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                    make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                    make_pass_through_transform(C)),
+                make_tuple(
+                    Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}, Sequence<4>{}),
+                make_tuple(Sequence<0>{},
+                           Sequence<1, 2>{},
+                           Sequence<3, 4>{},
+                           Sequence<5, 6>{},
+                           Sequence<7>{}));
+
+            const auto in_gemmmraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+                in_n_z_do_y_ho_x_wo_c_grid_desc,
+                make_tuple(make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                           make_merge_transform(make_tuple(Z, Y, X, C))),
+                make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+            const auto in_gemmm_gemmk_grid_desc =
+                matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_grid_desc);
+
+            return in_gemmm_gemmk_grid_desc;
+        }
+    }
+
+    template <typename BLay,
+              typename std::enable_if<is_same_v<BLay, tensor_layout::convolution::GKXC> ||
+                                          is_same_v<BLay, tensor_layout::convolution::GKYXC> ||
+                                          is_same_v<BLay, tensor_layout::convolution::GKZYXC>,
+                                      bool>::type = false>
+    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 index_t K = b_g_k_c_xs_lengths[1];
+        const index_t C = b_g_k_c_xs_lengths[2];
+
+        const index_t YX = std::accumulate(b_g_k_c_xs_lengths.begin() + 3,
+                                           b_g_k_c_xs_lengths.begin() + 3 + NDimSpatial,
+                                           index_t{1},
+                                           std::multiplies<index_t>());
+
+        const auto wei_k_yxc_grid_desc = make_naive_tensor_descriptor_packed(make_tuple(K, YX * C));
+
+        const auto wei_gemmn_gemmk_grid_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_k_yxc_grid_desc);
+
+        return wei_gemmn_gemmk_grid_desc;
+    }
+
+    template <typename BLay,
+              typename std::enable_if<is_same_v<BLay, tensor_layout::convolution::G_K_X_C> ||
+                                          is_same_v<BLay, tensor_layout::convolution::G_K_YX_C> ||
+                                          is_same_v<BLay, tensor_layout::convolution::G_K_ZYX_C> ||
+                                          is_same_v<BLay, tensor_layout::convolution::KXGC> ||
+                                          is_same_v<BLay, tensor_layout::convolution::KYXGC> ||
+                                          is_same_v<BLay, tensor_layout::convolution::KZYXGC>,
+                                      bool>::type = false>
+    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 index_t K = b_g_k_c_xs_lengths[1];
+        const index_t C = b_g_k_c_xs_lengths[2];
+
+        const index_t YX = std::accumulate(b_g_k_c_xs_lengths.begin() + 3,
+                                           b_g_k_c_xs_lengths.begin() + 3 + NDimSpatial,
+                                           index_t{1},
+                                           std::multiplies<index_t>());
+
+        const index_t KStride = b_g_k_c_xs_strides[1];
+        const index_t XStride = b_g_k_c_xs_strides[2 + NDimSpatial];
+        const auto CStride    = I1;
+
+        const auto wei_k_yx_c_grid_desc = make_naive_tensor_descriptor(
+            make_tuple(K, YX, C), make_tuple(KStride, XStride, CStride));
+
+        const auto wei_gemmnraw_gemmkraw_grid_desc = transform_tensor_descriptor(
+            wei_k_yx_c_grid_desc,
+            make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(YX, C))),
+            make_tuple(Sequence<0>{}, Sequence<1, 2>{}),
+            make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+        const auto wei_gemmn_gemmk_grid_desc =
+            matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_grid_desc);
+
+        return wei_gemmn_gemmk_grid_desc;
+    }
+
+    template <typename ELay,
+              typename std::enable_if<is_same_v<ELay, tensor_layout::convolution::GNWK> ||
+                                          is_same_v<ELay, tensor_layout::convolution::GNHWK> ||
+                                          is_same_v<ELay, tensor_layout::convolution::GNDHWK>,
+                                      bool>::type = false>
+    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 index_t N = e_g_n_k_wos_lengths[1];
+        const index_t K = e_g_n_k_wos_lengths[2];
+
+        const index_t NHoWo = N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                                  e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                                  index_t{1},
+                                                  std::multiplies<index_t>());
+
+        const auto out_gemmmraw_gemmnraw_grid_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(NHoWo, K));
+
+        const auto out_gemmm_gemmn_grid_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_grid_desc);
+
+        return out_gemmm_gemmn_grid_desc;
+    }
+
+    template <typename ELay,
+              typename std::enable_if<is_same_v<ELay, tensor_layout::convolution::G_NW_K> ||
+                                          is_same_v<ELay, tensor_layout::convolution::G_NHW_K> ||
+                                          is_same_v<ELay, tensor_layout::convolution::G_NDHW_K> ||
+                                          is_same_v<ELay, tensor_layout::convolution::NWGK> ||
+                                          is_same_v<ELay, tensor_layout::convolution::NHWGK> ||
+                                          is_same_v<ELay, tensor_layout::convolution::NDHWGK>,
+                                      bool>::type = false>
+    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 index_t N = e_g_n_k_wos_lengths[1];
+        const index_t K = e_g_n_k_wos_lengths[2];
+
+        const auto KStride     = I1;
+        const index_t WoStride = e_g_n_k_wos_strides[NDimSpatial + 2];
+
+        const index_t NHoWo = N * std::accumulate(e_g_n_k_wos_lengths.begin() + 3,
+                                                  e_g_n_k_wos_lengths.begin() + 3 + NDimSpatial,
+                                                  index_t{1},
+                                                  std::multiplies<index_t>());
+
+        const auto out_gemmmraw_gemmnraw_grid_desc =
+            make_naive_tensor_descriptor(make_tuple(NHoWo, K), make_tuple(WoStride, KStride));
+
+        const auto out_gemmm_gemmn_grid_desc =
+            matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_grid_desc);
+
+        return out_gemmm_gemmn_grid_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>{});
+    }
+
+    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>({}, {}))>;
+
+    // GridwiseGemm
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
+        ADataType, // TODO: distinguish A/B datatype
+        AccDataType,
+        CShuffleDataType,
+        DsDataType,
+        EDataType,
+        AElementwiseOperation,
+        BElementwiseOperation,
+        CDEElementwiseOperation,
+        InMemoryDataOperationEnum::Set,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
+        EGridDesc_M_N,
+        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>;
+
+    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 Block2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
+    // 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)},
+              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
+        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_;
+        typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        Block2ETileMap block_2_etile_map_;
+
+        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 1
+            arg.Print();
+#endif
+            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.a_g_n_c_wis_lengths_[0]; // Group count
+
+            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_batch_gemm_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,
+                    typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
+                    typename GridwiseGemm::EGridDescriptor_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")
+        {
+            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>)
+            {
+                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)
+            << ">";
+        // clang-format on
+
+        return str.str();
+    }
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -18,7 +18,8 @@ struct GemmDesc
 
 template <typename ALayout,
           typename BLayout,
-          typename DELayout,
+          typename DsLayout,
+          typename ELayout,
           typename ADataType,
           typename BDataType,
           typename DsDataType,
@@ -30,6 +31,8 @@ struct DeviceGroupedGemm : public BaseOperator
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
 
+    static_assert(DsLayout::Size() == DsDataType::Size(), "wrong! inconsisiten NumDTensor");
+
     virtual std::unique_ptr<BaseArgument>
     MakeArgumentPointer(std::vector<const void*>& p_a,
                         std::vector<const void*>& p_b,
@@ -43,27 +46,6 @@ struct DeviceGroupedGemm : public BaseOperator
     virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
 };
 
-template <typename ALayout,
-          typename BLayout,
-          typename DELayout,
-          typename ADataType,
-          typename BDataType,
-          typename DsDataType,
-          typename EDataType,
-          typename AElementwiseOperation,
-          typename BElementwiseOperation,
-          typename CElementwiseOperation>
-using DeviceGroupedGemmPtr = std::unique_ptr<DeviceGroupedGemm<ALayout,
-                                                               BLayout,
-                                                               DELayout,
-                                                               ADataType,
-                                                               BDataType,
-                                                               DsDataType,
-                                                               EDataType,
-                                                               AElementwiseOperation,
-                                                               BElementwiseOperation,
-                                                               CElementwiseOperation>>;
-
 } // namespace device
 } // namespace tensor_operation
 } // namespace ck

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

@@ -13,9 +13,10 @@
 #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
 #include "ck/tensor_operation/gpu/device/device_grouped_gemm.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/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {
@@ -72,11 +73,11 @@ __global__ void
         a_element_op,
         b_element_op,
         c_element_op,
-        gemm_desc_ptr[group_id].a_grid_desc_k0_m_k1_,
-        gemm_desc_ptr[group_id].b_grid_desc_k0_n_k1_,
+        gemm_desc_ptr[group_id].a_grid_desc_ak0_m_ak1_,
+        gemm_desc_ptr[group_id].b_grid_desc_bk0_n_bk1_,
         gemm_desc_ptr[group_id].ds_grid_desc_mblock_mperblock_nblock_nperblock_,
         gemm_desc_ptr[group_id].e_grid_desc_mblock_mperblock_nblock_nperblock_,
-        gemm_desc_ptr[group_id].block_2_ctile_map_);
+        gemm_desc_ptr[group_id].block_2_etile_map_);
 #else
     ignore = gemm_descs_const;
     ignore = group_count;
@@ -88,10 +89,11 @@ __global__ void
 
 template <typename ALayout,
           typename BLayout,
-          typename DELayout,
+          typename DsLayout,
+          typename ELayout,
           typename ADataType,
           typename BDataType,
-          typename GemmAccDataType,
+          typename AccDataType,
           typename CShuffleDataType,
           typename DsDataType,
           typename EDataType,
@@ -116,37 +118,43 @@ template <typename ALayout,
           ck::index_t ABlockTransferSrcVectorDim,
           ck::index_t ABlockTransferSrcScalarPerVector,
           ck::index_t ABlockTransferDstScalarPerVector_K1,
-          bool ABlockLdsAddExtraM,
+          bool ABlockLdsExtraM,
           typename BBlockTransferThreadClusterLengths_K0_N_K1,
           typename BBlockTransferThreadClusterArrangeOrder,
           typename BBlockTransferSrcAccessOrder,
           ck::index_t BBlockTransferSrcVectorDim,
           ck::index_t BBlockTransferSrcScalarPerVector,
           ck::index_t BBlockTransferDstScalarPerVector_K1,
-          bool BBlockLdsAddExtraN,
+          bool BBlockLdsExtraN,
           index_t CShuffleMXdlPerWavePerShuffle,
           index_t CShuffleNXdlPerWavePerShuffle,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched = make_default_loop_scheduler()>
-struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
-                                                       BLayout,
-                                                       DELayout,
-                                                       ADataType,
-                                                       BDataType,
-                                                       DsDataType,
-                                                       EDataType,
-                                                       AElementwiseOperation,
-                                                       BElementwiseOperation,
-                                                       CDEElementwiseOperation>
+struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
+                                                        BLayout,
+                                                        DsLayout,
+                                                        ELayout,
+                                                        ADataType,
+                                                        BDataType,
+                                                        DsDataType,
+                                                        EDataType,
+                                                        AElementwiseOperation,
+                                                        BElementwiseOperation,
+                                                        CDEElementwiseOperation>
 {
+    using DeviceOp = DeviceGroupedGemm_Xdl;
+
     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 auto MakeAGridDescriptor_AK0_M_AK1(index_t MRaw, index_t KRaw, index_t StrideA)
+    static constexpr auto matrix_padder =
+        MatrixPadder<GemmSpec, index_t, index_t, index_t>{MPerBlock, NPerBlock, KPerBlock};
+
+    static auto MakeAGridDescriptor_M_K(index_t MRaw, index_t KRaw, index_t StrideA)
     {
         const auto a_grid_desc_mraw_kraw = [&]() {
             if constexpr(is_same_v<tensor_layout::gemm::RowMajor, ALayout>)
@@ -161,95 +169,10 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
             }
         }();
 
-        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
-        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
-
-        const auto MPad = M - MRaw;
-        const auto KPad = K - KRaw;
-
-        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
-                     GemmSpec == GemmSpecialization::MNKPadding)
-        {
-            // pad both M and K
-            assert(K % AK1 == 0);
-
-            const auto AK0 = K / AK1;
-
-            const auto a_grid_desc_m_k =
-                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
-                                            make_tuple(make_right_pad_transform(MRaw, MPad),
-                                                       make_right_pad_transform(KRaw, KPad)),
-                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-            const auto a_grid_desc_ak0_m_ak1 =
-                transform_tensor_descriptor(a_grid_desc_m_k,
-                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
-                                                       make_pass_through_transform(M)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return a_grid_desc_ak0_m_ak1;
-        }
-        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
-                          GemmSpec == GemmSpecialization::MNPadding)
-        {
-            // pad M, but not K
-            assert(KRaw % AK1 == 0);
-
-            const auto AK0 = KRaw / AK1;
-
-            const auto a_grid_desc_ak0_m_ak1 =
-                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
-                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
-                                                       make_right_pad_transform(MRaw, MPad)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return a_grid_desc_ak0_m_ak1;
-        }
-        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
-                          GemmSpec == GemmSpecialization::NKPadding)
-        {
-            // pad K, but not M
-            assert(K % AK1 == 0);
-
-            const auto AK0 = K / AK1;
-
-            const auto a_grid_desc_m_k = transform_tensor_descriptor(
-                a_grid_desc_mraw_kraw,
-                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-            const auto a_grid_desc_ak0_m_ak1 =
-                transform_tensor_descriptor(a_grid_desc_m_k,
-                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
-                                                       make_pass_through_transform(MRaw)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return a_grid_desc_ak0_m_ak1;
-        }
-        else
-        {
-            // not pad M or K
-            assert(KRaw % AK1 == 0);
-
-            const auto AK0 = KRaw / AK1;
-
-            const auto a_grid_desc_ak0_m_ak1 =
-                transform_tensor_descriptor(a_grid_desc_mraw_kraw,
-                                            make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
-                                                       make_pass_through_transform(MRaw)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return a_grid_desc_ak0_m_ak1;
-        }
+        return matrix_padder.PadADescriptor_M_K(a_grid_desc_mraw_kraw);
     }
 
-    static auto MakeBGridDescriptor_BK0_N_BK1(index_t KRaw, index_t NRaw, index_t StrideB)
+    static auto MakeBGridDescriptor_N_K(index_t KRaw, index_t NRaw, index_t StrideB)
     {
         const auto b_grid_desc_nraw_kraw = [&]() {
             if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
@@ -264,160 +187,50 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
             }
         }();
 
-        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
-        const auto K = math::integer_divide_ceil(KRaw, KPerBlock) * KPerBlock;
-
-        const auto NPad = N - NRaw;
-        const auto KPad = K - KRaw;
-
-        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
-                     GemmSpec == GemmSpecialization::MNKPadding)
-        {
-            // pad both N and K
-            assert(K % BK1 == 0);
-
-            const auto BK0 = K / BK1;
-
-            const auto b_grid_desc_n_k =
-                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
-                                            make_tuple(make_right_pad_transform(NRaw, NPad),
-                                                       make_right_pad_transform(KRaw, KPad)),
-                                            make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                            make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-            const auto b_grid_desc_bk0_n_bk1 =
-                transform_tensor_descriptor(b_grid_desc_n_k,
-                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                       make_pass_through_transform(N)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return b_grid_desc_bk0_n_bk1;
-        }
-        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
-                          GemmSpec == GemmSpecialization::MNPadding)
-        {
-            // pad N, but not K
-            assert(KRaw % BK1 == 0);
-
-            const auto BK0 = KRaw / BK1;
-
-            const auto b_grid_desc_bk0_n_bk1 =
-                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
-                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                       make_right_pad_transform(NRaw, NPad)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return b_grid_desc_bk0_n_bk1;
-        }
-        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
-                          GemmSpec == GemmSpecialization::MKPadding)
-        {
-            // pad K, but not N
-            assert(K % BK1 == 0);
-
-            const auto BK0 = K / BK1;
-
-            const auto b_grid_desc_n_k = transform_tensor_descriptor(
-                b_grid_desc_nraw_kraw,
-                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-            const auto b_grid_desc_bk0_n_bk1 =
-                transform_tensor_descriptor(b_grid_desc_n_k,
-                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                       make_pass_through_transform(NRaw)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return b_grid_desc_bk0_n_bk1;
-        }
-        else
-        {
-            // not pad N or K
-            assert(KRaw % BK1 == 0);
-
-            const auto BK0 = KRaw / BK1;
-
-            const auto b_grid_desc_bk0_n_bk1 =
-                transform_tensor_descriptor(b_grid_desc_nraw_kraw,
-                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                       make_pass_through_transform(NRaw)),
-                                            make_tuple(Sequence<1>{}, Sequence<0>{}),
-                                            make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-
-            return b_grid_desc_bk0_n_bk1;
-        }
+        return matrix_padder.PadBDescriptor_N_K(b_grid_desc_nraw_kraw);
     }
 
+    template <typename ELay>
     static auto MakeEGridDescriptor_M_N(index_t MRaw, index_t NRaw, index_t StrideE)
     {
-        const auto c_grid_desc_mraw_nraw = [&]() {
-            if constexpr(is_same<tensor_layout::gemm::RowMajor, DELayout>::value)
+        const auto e_grid_desc_mraw_nraw = [&]() {
+            if constexpr(is_same<tensor_layout::gemm::RowMajor, ELay>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
                                                     make_tuple(StrideE, I1));
             }
-            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, DELayout>::value)
+            else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ELay>::value)
             {
                 return make_naive_tensor_descriptor(make_tuple(MRaw, NRaw),
                                                     make_tuple(I1, StrideE));
             }
         }();
 
-        const auto M = math::integer_divide_ceil(MRaw, MPerBlock) * MPerBlock;
-        const auto N = math::integer_divide_ceil(NRaw, NPerBlock) * NPerBlock;
+        return matrix_padder.PadCDescriptor_M_N(e_grid_desc_mraw_nraw);
+    }
 
-        const auto MPad = M - MRaw;
-        const auto NPad = N - NRaw;
+    static auto MakeDsGridDescriptor_M_N(const std::array<index_t, NumDTensor>& MRaws,
+                                         const std::array<index_t, NumDTensor>& NRaws,
+                                         const std::array<index_t, NumDTensor>& DsStride)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
 
-        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
-                     GemmSpec == GemmSpecialization::MNKPadding)
-        {
-            // pad M and N
-            return transform_tensor_descriptor(c_grid_desc_mraw_nraw,
-                                               make_tuple(make_right_pad_transform(MRaw, MPad),
-                                                          make_right_pad_transform(NRaw, NPad)),
-                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
-                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
-        }
-        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
-                          GemmSpec == GemmSpecialization::MKPadding)
-        {
-            // pad M, but not N
-            return transform_tensor_descriptor(
-                c_grid_desc_mraw_nraw,
-                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
-        }
-        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
-                          GemmSpec == GemmSpecialization::NKPadding)
-        {
-            // pad N, but not M
-            return transform_tensor_descriptor(
-                c_grid_desc_mraw_nraw,
-                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
-        }
-        else
-        {
-            // not pad M or N
-            return c_grid_desc_mraw_nraw;
-        }
+                return DeviceOp::MakeEGridDescriptor_M_N<DLayout>(MRaws[i], NRaws[i], DsStride[i]);
+            },
+            Number<NumDTensor>{});
     }
 
-    using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1(1, 1, 1));
-    using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1(1, 1, 1));
-    using EGridDesc_M_N       = decltype(MakeEGridDescriptor_M_N(1, 1, 1));
+    using AGridDesc_M_K  = decltype(MakeAGridDescriptor_M_K(1, 1, 1));
+    using BGridDesc_N_K  = decltype(MakeBGridDescriptor_N_K(1, 1, 1));
+    using DsGridDesc_M_N = remove_cvref_t<decltype(MakeDsGridDescriptor_M_N({}, {}, {}))>;
+    using EGridDesc_M_N  = decltype(MakeEGridDescriptor_M_N<ELayout>(1, 1, 1));
 
     // GridwiseGemm
-    using GridwiseGemm = GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle<
+    using GridwiseGemm = GridwiseGemmMultipleD_xdl_cshuffle<
         ADataType, // TODO: distinguish A/B datatype
-        GemmAccDataType,
+        AccDataType,
         CShuffleDataType,
         DsDataType,
         EDataType,
@@ -425,8 +238,9 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
         BElementwiseOperation,
         CDEElementwiseOperation,
         InMemoryDataOperationEnum::Set,
-        AGridDesc_AK0_M_AK1,
-        BGridDesc_BK0_N_BK1,
+        AGridDesc_M_K,
+        BGridDesc_N_K,
+        DsGridDesc_M_N,
         EGridDesc_M_N,
         NumPrefetch, // NumGemmKPrefetchStage
         BlockSize,
@@ -446,7 +260,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
         ABlockTransferSrcScalarPerVector,
         ABlockTransferDstScalarPerVector_K1,
         false, // AThreadTransferSrcResetCoordinateAfterRun,
-        ABlockLdsAddExtraM,
+        ABlockLdsExtraM,
         BBlockTransferThreadClusterLengths_K0_N_K1,
         BBlockTransferThreadClusterArrangeOrder,
         BBlockTransferSrcAccessOrder,
@@ -454,76 +268,87 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
         BBlockTransferSrcScalarPerVector,
         BBlockTransferDstScalarPerVector_K1,
         false, // BThreadTransferSrcResetCoordinateAfterRun,
-        BBlockLdsAddExtraN,
+        BBlockLdsExtraN,
         CShuffleMXdlPerWavePerShuffle,
         CShuffleNXdlPerWavePerShuffle,
         CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
         CDEBlockTransferScalarPerVector_NPerBlock,
         LoopSched>;
 
+    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{}))>;
+
     struct GroupedGemmBlock2ETileMap
     {
-        using UnderlyingBlock2CTileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+        using UnderlyingBlock2ETileMap = typename GridwiseGemm::DefaultBlock2ETileMap;
+
         static_assert(
             std::is_same<decltype(GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{})),
                          typename GridwiseGemm::DefaultBlock2ETileMap>::value,
             "Wrong! Should be the same type name");
+
         GroupedGemmBlock2ETileMap()
         {
-            block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{});
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(EGridDesc_M_N{});
             BlockStart_        = -1;
         }
 
-        GroupedGemmBlock2ETileMap(const EGridDesc_M_N& c_grid_desc_m_n, ck::index_t BlockStart)
+        GroupedGemmBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, ck::index_t BlockStart)
         {
-            block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(c_grid_desc_m_n);
+            block_2_etile_map_ = GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
             BlockStart_        = BlockStart;
         }
 
         template <typename TopIdx>
         __host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
         {
-            return block_2_ctile_map_.CalculateBottomIndex(
+            return block_2_etile_map_.CalculateBottomIndex(
                 make_multi_index(idx_top[I0] - BlockStart_));
         }
 
+        // it's actually E-Tile
         template <typename CTileIdx, typename CTileDim>
         __host__ __device__ bool ValidCTileIndex(const CTileIdx& c_tile_idx,
                                                  const CTileDim& c_tile_dim) const
         {
-            return block_2_ctile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
+            return block_2_etile_map_.ValidCTileIndex(c_tile_idx, c_tile_dim);
         }
 
-        __host__ bool CheckValidity(const EGridDesc_M_N& c_grid_desc_m_n) const
+        __host__ bool CheckValidity(const EGridDesc_M_N& e_grid_desc_m_n) const
         {
-            return block_2_ctile_map_.CheckValidity(c_grid_desc_m_n);
+            return block_2_etile_map_.CheckValidity(e_grid_desc_m_n);
         }
 
-        typename GridwiseGemm::DefaultBlock2ETileMap block_2_ctile_map_;
+        typename GridwiseGemm::DefaultBlock2ETileMap block_2_etile_map_;
         ck::index_t BlockStart_;
     };
 
     struct GemmBiasTransKernelArg
     {
-        AGridDesc_AK0_M_AK1 a_grid_desc_k0_m_k1_;
-        BGridDesc_BK0_N_BK1 b_grid_desc_k0_n_k1_;
-        EGridDesc_M_N e_grid_desc_m_n_;
-
-        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
-            e_grid_desc_mblock_mperblock_nblock_nperblock_;
-
-        StaticallyIndexedArray<
-            typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
-            NumDTensor>
-            ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of different
-
-        GroupedGemmBlock2ETileMap block_2_ctile_map_;
-
+        // pointers
         const ADataType* a_ptr_;
         const BDataType* b_ptr_;
         typename GridwiseGemm::DsGridPointer ds_ptr_;
         EDataType* e_ptr_;
 
+        // tensor descriptors for problem definiton
+        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_;
+        typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            ds_grid_desc_mblock_mperblock_nblock_nperblock_;
+        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+            e_grid_desc_mblock_mperblock_nblock_nperblock_;
+
+        // block-to-e-tile map
+        GroupedGemmBlock2ETileMap block_2_etile_map_;
         ck::index_t BlockStart_, BlockEnd_;
     };
 
@@ -563,66 +388,85 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
                 const index_t StrideB = gemm_descs[i].stride_B_;
                 const index_t StrideC = gemm_descs[i].stride_C_;
 
-                const auto a_grid_desc_k0_m_k1_ =
-                    DeviceGroupedGemmXdl::MakeAGridDescriptor_AK0_M_AK1(M, K, StrideA);
-                const auto b_grid_desc_k0_n_k1_ =
-                    DeviceGroupedGemmXdl::MakeBGridDescriptor_BK0_N_BK1(K, N, StrideB);
+                // pointer
+                typename GridwiseGemm::DsGridPointer p_ds_grid{};
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
+
+                    p_ds_grid(j) = static_cast<const DDataType*>(p_Ds[i][j]);
+                });
 
-                const auto e_grid_desc_m_n_ =
-                    DeviceGroupedGemmXdl::MakeEGridDescriptor_M_N(M, N, StrideC);
+                // tensor descriptors for problem definiton
+                const auto a_grid_desc_m_k = DeviceOp::MakeAGridDescriptor_M_K(M, K, StrideA);
+                const auto b_grid_desc_n_k = DeviceOp::MakeBGridDescriptor_N_K(K, N, StrideB);
+
+                DsGridDesc_M_N ds_grid_desc_m_n;
+
+                static_for<0, NumDTensor, 1>{}([&](auto j) {
+                    using DLayout = remove_cvref_t<tuple_element_t<j.value, DsLayout>>;
+
+                    ds_grid_desc_m_n(j) = DeviceOp::MakeEGridDescriptor_M_N<DLayout>(
+                        M, N, gemm_descs[i].stride_Ds_[j]);
+                });
+
+                const auto e_grid_desc_m_n =
+                    DeviceOp::MakeEGridDescriptor_M_N<ELayout>(M, N, StrideC);
+
+                // tensor descriptors for block/thread-wise copy
+                const auto a_grid_desc_ak0_m_ak1 =
+                    GridwiseGemm::MakeDefaultAGridDescriptor_AK0_M_AK1(a_grid_desc_m_k);
+
+                const auto b_grid_desc_bk0_n_bk1 =
+                    GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k);
 
                 const index_t grid_size_grp =
-                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n_, 0)
-                        .block_2_ctile_map_.CalculateGridSize(e_grid_desc_m_n_);
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, 0)
+                        .block_2_etile_map_.CalculateGridSize(e_grid_desc_m_n);
 
                 const index_t BlockStart = grid_size_;
                 const index_t BlockEnd   = grid_size_ + grid_size_grp;
 
                 grid_size_ += grid_size_grp;
 
-                const auto block_2_ctile_map_ =
-                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n_, BlockStart);
+                // block-to-e-tile map
+                const auto block_2_etile_map =
+                    GroupedGemmBlock2ETileMap(e_grid_desc_m_n, BlockStart);
 
-                if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
-                                               b_grid_desc_k0_n_k1_,
-                                               e_grid_desc_m_n_,
-                                               block_2_ctile_map_))
+                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))
                 {
-                    auto e_grid_desc_mblock_mperblock_nblock_nperblock_ =
-                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                            e_grid_desc_m_n_);
-                    StaticallyIndexedArray<
-                        typename GridwiseGemm::EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
-                        NumDTensor>
-                        ds_grid_desc_mblock_mperblock_nblock_nperblock_; // FIXME: Ds desc may be of
-                                                                         // different
-
-                    typename GridwiseGemm::DsGridPointer p_ds_grid_{};
+                    // tensor descriptors for block/thread-wise copy
+                    typename GridwiseGemm::DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock;
 
                     static_for<0, NumDTensor, 1>{}([&](auto j) {
-                        using DDataType = remove_cvref_t<tuple_element_t<j.value, DsDataType>>;
-
-                        p_ds_grid_(j) = static_cast<const DDataType*>(p_Ds[i][j]);
-
-                        const auto d_grid_desc_m_n = DeviceGroupedGemmXdl::MakeEGridDescriptor_M_N(
-                            M, N, gemm_descs[i].stride_Ds_[j]);
-
-                        ds_grid_desc_mblock_mperblock_nblock_nperblock_(j) =
+                        ds_grid_desc_mblock_mperblock_nblock_nperblock(j) =
                             GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                                d_grid_desc_m_n);
+                                ds_grid_desc_m_n[j]);
                     });
 
+                    const auto e_grid_desc_mblock_mperblock_nblock_nperblock =
+                        GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                            e_grid_desc_m_n);
+
                     gemm_desc_kernel_arg_.push_back(
-                        GemmBiasTransKernelArg{a_grid_desc_k0_m_k1_,
-                                               b_grid_desc_k0_n_k1_,
-                                               e_grid_desc_m_n_,
-                                               e_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                               ds_grid_desc_mblock_mperblock_nblock_nperblock_,
-                                               block_2_ctile_map_,
-                                               static_cast<const ADataType*>(p_As[i]),
+                        GemmBiasTransKernelArg{static_cast<const ADataType*>(p_As[i]),
                                                static_cast<const BDataType*>(p_Bs[i]),
-                                               p_ds_grid_,
+                                               p_ds_grid,
                                                static_cast<EDataType*>(p_Es[i]),
+                                               a_grid_desc_m_k,
+                                               b_grid_desc_n_k,
+                                               ds_grid_desc_m_n,
+                                               e_grid_desc_m_n,
+                                               a_grid_desc_ak0_m_ak1,
+                                               b_grid_desc_bk0_n_bk1,
+                                               ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               e_grid_desc_mblock_mperblock_nblock_nperblock,
+                                               block_2_etile_map,
                                                BlockStart,
                                                BlockEnd});
                 }
@@ -643,7 +487,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
     // Invoker
     struct Invoker : public BaseInvoker
     {
-        using Argument = DeviceGroupedGemmXdl::Argument;
+        using Argument = DeviceOp::Argument;
 
         float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
         {
@@ -651,32 +495,39 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
 
             for(std::size_t i = 0; i < arg.gemm_desc_kernel_arg_.size(); i++)
             {
-                std::cout << "group: " << i << " arg.a_grid_desc_k0_m_k1_{"
-                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I0) << ", "
-                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
-                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I2) << "}";
-
-                std::cout << ", arg.b_grid_desc_k0_n_k1_{"
-                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_.GetLength(I0) << ", "
-                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
-                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_.GetLength(I2) << "}";
+                std::cout << "group: " << i << " arg.a_grid_desc_ak0_m_ak1_{"
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2)
+                          << "}";
+
+                std::cout << ", arg.b_grid_desc_bk0_n_bk1_{"
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I0)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I1)
+                          << ", "
+                          << arg.gemm_desc_kernel_arg_[i].b_grid_desc_bk0_n_bk1_.GetLength(I2)
+                          << "}";
 
                 std::cout << ", arg.e_grid_desc_m_n_{ "
                           << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I0) << ", "
                           << arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_.GetLength(I1) << "}"
                           << std::endl;
 
-                if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_,
-                                                arg.gemm_desc_kernel_arg_[i].b_grid_desc_k0_n_k1_,
+                if(!GridwiseGemm::CheckValidity(arg.gemm_desc_kernel_arg_[i].a_grid_desc_m_k_,
+                                                arg.gemm_desc_kernel_arg_[i].b_grid_desc_n_k_,
+                                                arg.gemm_desc_kernel_arg_[i].ds_grid_desc_m_n_,
                                                 arg.gemm_desc_kernel_arg_[i].e_grid_desc_m_n_,
-                                                arg.gemm_desc_kernel_arg_[i].block_2_ctile_map_))
+                                                arg.gemm_desc_kernel_arg_[i].block_2_etile_map_))
                 {
                     throw std::runtime_error(
                         "wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
                 }
 
-                const auto K = arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I0) *
-                               arg.gemm_desc_kernel_arg_[i].a_grid_desc_k0_m_k1_.GetLength(I2);
+                const auto K = arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I0) *
+                               arg.gemm_desc_kernel_arg_[i].a_grid_desc_ak0_m_ak1_.GetLength(I2);
 
                 if(GridwiseGemm::CalculateHasMainKBlockLoop(K) != has_main_k_block_loop)
                 {
@@ -733,18 +584,14 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
         }
     };
 
-    static constexpr bool IsValidCompilationParameter()
-    {
-        // TODO: properly implement this check
-        return true;
-    }
-
     static bool IsSupportedArgument(const Argument& arg)
     {
         if(ck::type_convert<ck::index_t>(arg.gemm_desc_kernel_arg_.size()) != arg.group_count_)
+        {
             return false;
-        else
-            return true;
+        }
+
+        return true;
     }
 
     // polymorphic
@@ -795,7 +642,7 @@ struct DeviceGroupedGemmXdl : public DeviceGroupedGemm<ALayout,
         auto str = std::stringstream();
 
         // clang-format off
-        str << "DeviceGroupedGemmXdl"
+        str << "DeviceGroupedGemm_Xdl"
             << "<"
             << BlockSize << ", "
             << MPerBlock << ", "

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

@@ -13,8 +13,8 @@
 #include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_layernorm.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

include/ck/tensor_operation/gpu/device/device_pool2d_fwd_nhwc_nhwc.hpp

@@ -11,8 +11,8 @@
 #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
 #include "ck/tensor_operation/gpu/device/device_pool2d_fwd.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

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

@@ -14,8 +14,8 @@
 #include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

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

@@ -6,8 +6,8 @@
 #include <iostream>
 #include <sstream>
 
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
 #include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"

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

@@ -14,8 +14,8 @@
 #include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_softmax.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 
 namespace ck {
 namespace tensor_operation {

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

@@ -6,8 +6,8 @@
 #include <iostream>
 #include <vector>
 
-#include "ck/device_utility/device_prop.hpp"
-#include "ck/device_utility/kernel_launch.hpp"
+#include "ck/host_utility/device_prop.hpp"
+#include "ck/host_utility/kernel_launch.hpp"
 #include "ck/tensor_operation/gpu/device/device_base.hpp"
 #include "ck/tensor_operation/gpu/grid/gridwise_unary_elementwise_1d.hpp"
 

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

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#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/gemm_specialization.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+
+// M/N/KPerTileType could be index_t or Number<>
+template <GemmSpecialization GemmSpec,
+          typename MPerTileType,
+          typename NPerTileType,
+          typename KPerTileType>
+struct MatrixPadder
+{
+    static constexpr auto I0 = Number<0>{};
+    static constexpr auto I1 = Number<1>{};
+    static constexpr auto I2 = Number<2>{};
+    static constexpr auto I3 = Number<3>{};
+
+    template <typename ADesc_MRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadADescriptor_M_K(const ADesc_MRaw_KRaw& a_desc_mraw_kraw) const
+    {
+        const auto MRaw = a_desc_mraw_kraw.GetLength(I0);
+        const auto KRaw = a_desc_mraw_kraw.GetLength(I1);
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerTile_) * MPerTile_;
+        const auto K = math::integer_divide_ceil(KRaw, KPerTile_) * KPerTile_;
+
+        const auto MPad = M - MRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both M and K
+            return transform_tensor_descriptor(a_desc_mraw_kraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(KRaw, KPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad M, but not K
+            return transform_tensor_descriptor(
+                a_desc_mraw_kraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(KRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad K, but not M
+            return transform_tensor_descriptor(
+                a_desc_mraw_kraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or K
+            return a_desc_mraw_kraw;
+        }
+    }
+
+    template <typename BDesc_NRaw_KRaw>
+    __host__ __device__ constexpr auto
+    PadBDescriptor_N_K(const BDesc_NRaw_KRaw& b_desc_nraw_kraw) const
+    {
+        const auto NRaw = b_desc_nraw_kraw.GetLength(I0);
+        const auto KRaw = b_desc_nraw_kraw.GetLength(I1);
+
+        const auto N = math::integer_divide_ceil(NRaw, NPerTile_) * NPerTile_;
+        const auto K = math::integer_divide_ceil(KRaw, KPerTile_) * KPerTile_;
+
+        const auto NPad = N - NRaw;
+        const auto KPad = K - KRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::NKPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad both N and K
+            return transform_tensor_descriptor(b_desc_nraw_kraw,
+                                               make_tuple(make_right_pad_transform(NRaw, NPad),
+                                                          make_right_pad_transform(KRaw, KPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::MNPadding)
+        {
+            // pad N, but not K
+            return transform_tensor_descriptor(
+                b_desc_nraw_kraw,
+                make_tuple(make_right_pad_transform(NRaw, NPad), make_pass_through_transform(KRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::KPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad K, but not N
+            return transform_tensor_descriptor(
+                b_desc_nraw_kraw,
+                make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad N or K
+            return b_desc_nraw_kraw;
+        }
+    }
+
+    template <typename CDesc_MRaw_NRaw>
+    __host__ __device__ constexpr auto
+    PadCDescriptor_M_N(const CDesc_MRaw_NRaw& c_desc_mraw_nraw) const
+    {
+        const auto MRaw = c_desc_mraw_nraw.GetLength(I0);
+        const auto NRaw = c_desc_mraw_nraw.GetLength(I1);
+
+        const auto M = math::integer_divide_ceil(MRaw, MPerTile_) * MPerTile_;
+        const auto N = math::integer_divide_ceil(NRaw, NPerTile_) * NPerTile_;
+
+        const auto MPad = M - MRaw;
+        const auto NPad = N - NRaw;
+
+        if constexpr(GemmSpec == GemmSpecialization::MNPadding ||
+                     GemmSpec == GemmSpecialization::MNKPadding)
+        {
+            // pad M and N
+            return transform_tensor_descriptor(c_desc_mraw_nraw,
+                                               make_tuple(make_right_pad_transform(MRaw, MPad),
+                                                          make_right_pad_transform(NRaw, NPad)),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                               make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::MPadding ||
+                          GemmSpec == GemmSpecialization::MKPadding)
+        {
+            // pad M, but not N
+            return transform_tensor_descriptor(
+                c_desc_mraw_nraw,
+                make_tuple(make_right_pad_transform(MRaw, MPad), make_pass_through_transform(NRaw)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else if constexpr(GemmSpec == GemmSpecialization::NPadding ||
+                          GemmSpec == GemmSpecialization::NKPadding)
+        {
+            // pad N, but not M
+            return transform_tensor_descriptor(
+                c_desc_mraw_nraw,
+                make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(NRaw, NPad)),
+                make_tuple(Sequence<0>{}, Sequence<1>{}),
+                make_tuple(Sequence<0>{}, Sequence<1>{}));
+        }
+        else
+        {
+            // not pad M or N
+            return c_desc_mraw_nraw;
+        }
+    }
+
+    MPerTileType MPerTile_;
+    NPerTileType NPerTile_;
+    KPerTileType KPerTile_;
+};
+
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -25,41 +25,146 @@ struct ColumnMajor : public BaseTensorLayout
 
 namespace convolution {
 
-// 1D Conv
+// input tensor
+// packed NCW/NCHW/NCDHW
+struct NCW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCW";
+};
+
+struct NCHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCHW";
+};
+
+struct NCDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NCDHW";
+};
+
+// packed GNCW/GNCHW/GNCDHW
+struct GNCW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCW";
+};
+
+struct GNCHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCHW";
+};
+
+struct GNCDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNCDHW";
+};
+
+// input tensor
+// packed NWC/NHWC/NDHWC
 struct NWC : public BaseTensorLayout
 {
     static constexpr const char* name = "NWC";
 };
 
-struct KXC : public BaseTensorLayout
+struct NHWC : public BaseTensorLayout
 {
-    static constexpr const char* name = "KXC";
+    static constexpr const char* name = "NHWC";
 };
 
-struct NWK : public BaseTensorLayout
+struct NDHWC : public BaseTensorLayout
 {
-    static constexpr const char* name = "NWK";
+    static constexpr const char* name = "NDHWC";
 };
 
-struct NCW : public BaseTensorLayout
+// input tensor
+// packed GNWC/GNHWC/GNDHWC
+struct GNWC : public BaseTensorLayout
 {
-    static constexpr const char* name = "NCW";
+    static constexpr const char* name = "GNWC";
+};
+
+struct GNHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNHWC";
 };
 
+struct GNDHWC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNDHWC";
+};
+
+// input tensor
+// packed GNWC/GNHWC/GNDHWC
+struct NWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWGC";
+};
+
+struct NHWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWGC";
+};
+
+struct NDHWGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWGC";
+};
+
+// input tensor
+// strided layout
+struct G_NW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW_C";
+};
+
+struct G_NHW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW_C";
+};
+
+struct G_NDHW_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW_C";
+};
+
+// weight tensor
+// packed KCX/KCYX/KCZYX
 struct KCX : public BaseTensorLayout
 {
     static constexpr const char* name = "KCX";
 };
 
-struct NKW : public BaseTensorLayout
+struct KCYX : public BaseTensorLayout
 {
-    static constexpr const char* name = "NKW";
+    static constexpr const char* name = "KCYX";
 };
 
-// 2D Conv
-struct NHWC : public BaseTensorLayout
+struct KCZYX : public BaseTensorLayout
 {
-    static constexpr const char* name = "NHWC";
+    static constexpr const char* name = "KCZYX";
+};
+
+// weight tensor
+// packed KCX/KCYX/KCZYX
+struct GKCX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCX";
+};
+
+struct GKCYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCYX";
+};
+
+struct GKCZYX : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKCZYX";
+};
+
+// weight tensor
+// packed KXC/KYXC/KZYXC
+struct KXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KXC";
 };
 
 struct KYXC : public BaseTensorLayout
@@ -67,19 +172,67 @@ struct KYXC : public BaseTensorLayout
     static constexpr const char* name = "KYXC";
 };
 
-struct NHWK : public BaseTensorLayout
+struct KZYXC : public BaseTensorLayout
 {
-    static constexpr const char* name = "NHWK";
+    static constexpr const char* name = "KZYXC";
 };
 
-struct NCHW : public BaseTensorLayout
+// weight tensor
+// packed GKXC/GKYXC/GKZYXC
+struct GKXC : public BaseTensorLayout
 {
-    static constexpr const char* name = "NCHW";
+    static constexpr const char* name = "GKXC";
 };
 
-struct KCYX : public BaseTensorLayout
+struct GKYXC : public BaseTensorLayout
 {
-    static constexpr const char* name = "KCYX";
+    static constexpr const char* name = "GKYXC";
+};
+
+struct GKZYXC : public BaseTensorLayout
+{
+    static constexpr const char* name = "GKZYXC";
+};
+
+// weight tensor
+// packed KXGC/KYXGC/KZYXGC
+struct KXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KXGC";
+};
+
+struct KYXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KYXGC";
+};
+
+struct KZYXGC : public BaseTensorLayout
+{
+    static constexpr const char* name = "KZYXGC";
+};
+
+// weight tensor
+// strided
+struct G_K_X_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_X_C";
+};
+
+struct G_K_YX_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_YX_C";
+};
+
+struct G_K_ZYX_C : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_K_ZYX_C";
+};
+
+// output tensor
+// packed NKW/NKHW/NKDHW
+struct NKW : public BaseTensorLayout
+{
+    static constexpr const char* name = "NKW";
 };
 
 struct NKHW : public BaseTensorLayout
@@ -87,34 +240,94 @@ struct NKHW : public BaseTensorLayout
     static constexpr const char* name = "NKHW";
 };
 
-// 3D Conv
-struct NDHWC : public BaseTensorLayout
+struct NKDHW : public BaseTensorLayout
 {
-    static constexpr const char* name = "NDHWC";
+    static constexpr const char* name = "NKDHW";
 };
 
-struct KZYXC : public BaseTensorLayout
+// output tensor
+// packed GNKW/GNKHW/GNKDHW
+struct GNKW : public BaseTensorLayout
 {
-    static constexpr const char* name = "KZYXC";
+    static constexpr const char* name = "GNKW";
+};
+
+struct GNKHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKHW";
+};
+
+struct GNKDHW : public BaseTensorLayout
+{
+    static constexpr const char* name = "GNKDHW";
+};
+
+// output tensor
+// packed NWK/NHWK/NDHWK
+struct NWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWK";
+};
+
+struct NHWK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWK";
 };
 
 struct NDHWK : public BaseTensorLayout
 {
     static constexpr const char* name = "NDHWK";
 };
-struct NCDHW : public BaseTensorLayout
+
+// output tensor
+// packed GNWK/GNHWK/GNDHWK
+struct GNWK : public BaseTensorLayout
 {
-    static constexpr const char* name = "NCDHW";
+    static constexpr const char* name = "GNWK";
 };
 
-struct KCZYX : public BaseTensorLayout
+struct GNHWK : public BaseTensorLayout
 {
-    static constexpr const char* name = "KCZYX";
+    static constexpr const char* name = "GNHWK";
 };
 
-struct NKDHW : public BaseTensorLayout
+struct GNDHWK : public BaseTensorLayout
 {
-    static constexpr const char* name = "NKDHW";
+    static constexpr const char* name = "GNDHWK";
+};
+
+// output tensor
+// packed NWGK/NHWGK/NDHWGK
+struct NWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NWGK";
+};
+
+struct NHWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NHWGK";
+};
+
+struct NDHWGK : public BaseTensorLayout
+{
+    static constexpr const char* name = "NDHWGK";
+};
+
+// output tensor
+// strided layout
+struct G_NW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NW_K";
+};
+
+struct G_NHW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NHW_K";
+};
+
+struct G_NDHW_K : public BaseTensorLayout
+{
+    static constexpr const char* name = "G_NDHW_K";
 };
 
 } // namespace convolution

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

@@ -104,6 +104,13 @@ struct Bilinear
         y = alpha_ * x0 + beta_ * x1;
     };
 
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    {
+        y = type_convert<half_t>(alpha_) * x0 + type_convert<half_t>(beta_) * x1;
+    };
+
     template <>
     __host__ __device__ constexpr void
     operator()<half_t, float, half_t>(half_t& y, const float& x0, const half_t& x1) const
@@ -117,12 +124,12 @@ struct Bilinear
 
 struct AddRelu
 {
-    template <typename T>
-    __host__ __device__ constexpr void operator()(T& y, const T& x0, const T& x1) const;
+    template <typename Y, typename X0, typename X1>
+    __host__ __device__ constexpr void operator()(Y& y, const X0& x0, const X1& x1) const;
 
     template <>
     __host__ __device__ constexpr void
-    operator()<float>(float& y, const float& x0, const float& x1) const
+    operator()<float, float, float>(float& y, const float& x0, const float& x1) const
     {
         const float a = x0 + x1;
         y             = a > 0.0f ? a : 0.0f;
@@ -130,7 +137,7 @@ struct AddRelu
 
     template <>
     __host__ __device__ constexpr void
-    operator()<double>(double& y, const double& x0, const double& x1) const
+    operator()<double, double, double>(double& y, const double& x0, const double& x1) const
     {
         const double a = x0 + x1;
         y              = a > 0.0 ? a : 0.0;
@@ -138,11 +145,19 @@ struct AddRelu
 
     template <>
     __host__ __device__ constexpr void
-    operator()<half_t>(half_t& y, const half_t& x0, const half_t& x1) const
+    operator()<half_t, half_t, half_t>(half_t& y, const half_t& x0, const half_t& x1) const
     {
         const half_t a = x0 + x1;
         y              = a > type_convert<half_t>(0.0f) ? a : type_convert<half_t>(0.0f);
     };
+
+    template <>
+    __host__ __device__ constexpr void
+    operator()<half_t, float, half_t>(half_t& y, const float& x0, const half_t& x1) const
+    {
+        const float a = x0 + x1;
+        y             = a > type_convert<half_t>(0.0f) ? a : type_convert<half_t>(0.0f);
+    };
 };
 
 struct AddHardswish

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

@@ -12,16 +12,65 @@ namespace element_wise {
 
 struct PassThrough
 {
-    template <typename T>
-    __host__ __device__ void operator()(T& y, const T& x) const
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const;
+
+    template <>
+    __host__ __device__ void operator()<double, double>(double& y, const double& x) const
     {
-        static_assert(is_same<T, float>::value || is_same<T, double>::value ||
-                          is_same<T, half_t>::value || is_same<T, bhalf_t>::value ||
-                          is_same<T, int32_t>::value || is_same<T, int8_t>::value,
-                      "Data type is not supported by this operation!");
+        y = x;
+    }
 
+    template <>
+    __host__ __device__ void operator()<float, float>(float& y, const float& x) const
+    {
         y = x;
-    };
+    }
+
+    template <>
+    __host__ __device__ void operator()<half_t, half_t>(half_t& y, const half_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, bhalf_t>(bhalf_t& y, const bhalf_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<int32_t, int32_t>(int32_t& y, const int32_t& x) const
+    {
+        y = x;
+    }
+
+    template <>
+    __host__ __device__ void operator()<bhalf_t, float>(bhalf_t& y, const float& x) const
+    {
+        y = type_convert<bhalf_t>(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()<int8_t, int32_t>(int8_t& y, const int32_t& x) const
+    {
+        y = type_convert<int8_t>(x);
+    }
+};
+
+struct UnaryConvert
+{
+    template <typename Y, typename X>
+    __host__ __device__ void operator()(Y& y, const X& x) const
+    {
+        y = type_convert<Y>(x);
+    }
 };
 
 struct Scale

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

@@ -18,25 +18,26 @@
 namespace ck {
 
 // GEMM:
-//   input : A[AK0, M, AK1]
-//   input : B[AK0, N, AK1]
+//   input : A[M, K]
+//   input : B[N, K]
 //   input : D0[M, N], D1[M, N], ...
 //   output : E[M, N]
 //   C = a_op(A) * b_op(B)
 //   E = cde_op(C, D0, D1, ...)
 // Assume:
 //   D0, D1, ... and E have the same layout
-template <typename FloatAB,
-          typename FloatGemmAcc,
-          typename FloatCShuffle,
+template <typename ABDataType, // FIXME: don't assume A/B have same datatype
+          typename AccDataType,
+          typename CShuffleDataType,
           typename DsDataType,
-          typename FloatE,
+          typename EDataType,
           typename AElementwiseOperation,
           typename BElementwiseOperation,
           typename CDEElementwiseOperation,
           InMemoryDataOperationEnum EGlobalMemoryDataOperation,
-          typename AGridDesc_AK0_M_AK1,
-          typename BGridDesc_BK0_N_BK1,
+          typename AGridDesc_M_K,
+          typename BGridDesc_N_K,
+          typename DsGridDesc_M_N,
           typename EGridDesc_M_N,
           index_t NumGemmKPrefetchStage,
           index_t BlockSize,
@@ -70,7 +71,7 @@ template <typename FloatAB,
           typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
           index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
           LoopScheduler LoopSched>
-struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
+struct GridwiseGemmMultipleD_xdl_cshuffle
 {
     static constexpr index_t NumDTensor = DsDataType::Size();
 
@@ -84,10 +85,10 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
     static constexpr auto I7 = Number<7>{};
 
     // K1 should be Number<...>
-    static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
-    static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
-    static constexpr auto AK1 = Number<AK1Value>{};
-    static constexpr auto BK1 = Number<BK1Value>{};
+    static constexpr auto AK1         = Number<AK1Value>{};
+    static constexpr auto BK1         = Number<BK1Value>{};
+    static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
+    static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
@@ -97,7 +98,7 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
     {
         // A matrix in LDS memory, dst of blockwise copy
         return make_naive_tensor_descriptor(
-            make_tuple(AK0, Number<MPerBlock>{}, AK1),
+            make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
             make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
     }
 
@@ -105,7 +106,7 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
     {
         // B matrix in LDS memory, dst of blockwise copy
         return make_naive_tensor_descriptor(
-            make_tuple(BK0, Number<NPerBlock>{}, BK1),
+            make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
             make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
     }
 
@@ -160,31 +161,123 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
             c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
 
         return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
-                             sizeof(FloatAB),
-                         c_block_size * sizeof(FloatCShuffle));
+                             sizeof(ABDataType),
+                         c_block_size * sizeof(CShuffleDataType));
+    }
+
+    // A desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
+    {
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
+
+        const auto AK0 = K / AK1;
+
+        return transform_tensor_descriptor(a_grid_desc_m_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                                                      make_pass_through_transform(M)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // B desc for source in blockwise copy
+    __host__ __device__ static constexpr auto
+    MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
+    {
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = b_grid_desc_n_k.GetLength(I1);
+
+        const auto BK0 = K / BK1;
+
+        return transform_tensor_descriptor(b_grid_desc_n_k,
+                                           make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                                      make_pass_through_transform(N)),
+                                           make_tuple(Sequence<1>{}, Sequence<0>{}),
+                                           make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+    }
+
+    // E desc for destination in blockwise copy
+    template <typename EGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const EGridDescriptor_M_N& e_grid_desc_m_n)
+    {
+        const auto M = e_grid_desc_m_n.GetLength(I0);
+        const auto N = e_grid_desc_m_n.GetLength(I1);
+
+        const auto MBlock = M / MPerBlock;
+        const auto NBlock = N / NPerBlock;
+
+        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
+            e_grid_desc_m_n,
+            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
+                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
+            make_tuple(Sequence<0>{}, Sequence<1>{}),
+            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
+
+        return e_grid_desc_mblock_mperblock_nblock_nperblock;
+    }
+
+    // Ds desc for source in blockwise copy
+    template <typename DsGridDescriptor_M_N>
+    __host__ __device__ static constexpr auto
+    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+        const DsGridDescriptor_M_N& ds_grid_desc_m_n)
+    {
+        return generate_tuple(
+            [&](auto i) {
+                return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
+            },
+            Number<NumDTensor>{});
+    }
+
+    // return block_id to E matrix tile idx (m0, n0) mapping
+    __host__ __device__ static constexpr auto
+    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
+    {
+        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
+            e_grid_desc_m_n);
     }
 
     // block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
     template <typename Block2ETileMap>
-    __host__ __device__ static constexpr bool
-    CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
-                  const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
-                  const EGridDesc_M_N& e_grid_desc_m_n,
-                  const Block2ETileMap& block_2_etile_map)
+    __host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
+                                                            const BGridDesc_N_K& b_grid_desc_n_k,
+                                                            const DsGridDesc_M_N& ds_grid_desc_m_n,
+                                                            const EGridDesc_M_N& e_grid_desc_m_n,
+                                                            const Block2ETileMap& block_2_etile_map)
     {
         static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
                           (NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
                       "Invalid tuning param!");
 
-        const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
-        const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
-        const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
+        const auto M = a_grid_desc_m_k.GetLength(I0);
+        const auto N = b_grid_desc_n_k.GetLength(I0);
+        const auto K = a_grid_desc_m_k.GetLength(I1);
 
+        // check consistency of desc
         if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
+        {
             return false;
+        }
+
+        bool valid = true;
+
+        static_for<0, NumDTensor, 1>{}([&](auto i) {
+            valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
+                              N == ds_grid_desc_m_n[i].GetLength(I1));
+        });
 
+        if(!valid)
+        {
+            return false;
+        }
+
+        // check tile size
         if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
+        {
             return false;
+        }
 
         // check gridwise gemm pipeline
         const auto num_k_loop = K / KPerBlock;
@@ -194,12 +287,23 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
             return false;
         }
 
+        // check block-to-E-tile
         if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
         {
             return false;
         }
 
         // TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
+        // check tensor size: cannot be larger than 2GB each
+        constexpr long_index_t TwoGB = (long_index_t{1} << 31);
+
+        if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
+             e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
+        {
+            return false;
+        }
+
         return true;
     }
 
@@ -210,60 +314,39 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
         return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
     }
 
-    __host__ __device__ static constexpr auto
-    MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(const EGridDesc_M_N& e_grid_desc_m_n)
-    {
-        const auto M = e_grid_desc_m_n.GetLength(I0);
-        const auto N = e_grid_desc_m_n.GetLength(I1);
-
-        const auto MBlock = M / MPerBlock;
-        const auto NBlock = N / NPerBlock;
-
-        const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
-            e_grid_desc_m_n,
-            make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
-                       make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
-            make_tuple(Sequence<0>{}, Sequence<1>{}),
-            make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
-
-        return e_grid_desc_mblock_mperblock_nblock_nperblock;
-    }
-
-    // return block_id to E matrix tile idx (m0, n0) mapping
-    __host__ __device__ static constexpr auto
-    MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
-    {
-        return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
-            e_grid_desc_m_n);
-    }
-
-    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+    using DefaultAGridDesc_AK0_M_AK1 =
+        remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
+    using DefaultBGridDesc_BK0_N_BK1 =
+        remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
+    using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock  = remove_cvref_t<decltype(
         MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
+    using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
+        MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
 
     using DefaultBlock2ETileMap =
         remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}))>;
 
     using DsGridPointer = decltype(MakeDsGridPointer());
 
-    template <bool HasMainKBlockLoop, typename Block2ETileMap>
-    __device__ static void
-    Run(const FloatAB* __restrict__ p_a_grid,
-        const FloatAB* __restrict__ p_b_grid,
-        DsGridPointer p_ds_grid,
-        FloatE* __restrict__ p_e_grid,
-        void* __restrict__ p_shared,
-        const AElementwiseOperation& a_element_op,
-        const BElementwiseOperation& b_element_op,
-        const CDEElementwiseOperation& cde_element_op,
-        const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
-        const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
-        const StaticallyIndexedArray<EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
-                                     NumDTensor>&
-            ds_grid_desc_mblock_mperblock_nblock_nperblock, // FIXME: Ds desc may be of different
-                                                            // type from E
-        const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
-            e_grid_desc_mblock_mperblock_nblock_nperblock,
-        const Block2ETileMap& block_2_etile_map)
+    template <bool HasMainKBlockLoop,
+              typename AGridDesc_AK0_M_AK1,
+              typename BGridDesc_BK0_N_BK1,
+              typename Block2ETileMap>
+    __device__ static void Run(const ABDataType* __restrict__ p_a_grid,
+                               const ABDataType* __restrict__ p_b_grid,
+                               DsGridPointer p_ds_grid,
+                               EDataType* __restrict__ p_e_grid,
+                               void* __restrict__ p_shared,
+                               const AElementwiseOperation& a_element_op,
+                               const BElementwiseOperation& b_element_op,
+                               const CDEElementwiseOperation& cde_element_op,
+                               const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
+                               const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
+                               const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   ds_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
+                                   e_grid_desc_mblock_mperblock_nblock_nperblock,
+                               const Block2ETileMap& block_2_etile_map)
     {
         const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
             p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
@@ -316,11 +399,11 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
                                                 AElementwiseOperation,
                                                 ck::tensor_operation::element_wise::PassThrough,
                                                 InMemoryDataOperationEnum::Set,
-                                                Sequence<AK0, MPerBlock, AK1>,
+                                                Sequence<AK0PerBlock, MPerBlock, AK1>,
                                                 ABlockTransferThreadClusterLengths_AK0_M_AK1,
                                                 ABlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatAB,
+                                                ABDataType,
+                                                ABDataType,
                                                 decltype(a_grid_desc_ak0_m_ak1),
                                                 decltype(a_block_desc_ak0_m_ak1),
                                                 ABlockTransferSrcAccessOrder,
@@ -347,11 +430,11 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
                                                 BElementwiseOperation,
                                                 ck::tensor_operation::element_wise::PassThrough,
                                                 InMemoryDataOperationEnum::Set,
-                                                Sequence<BK0, NPerBlock, BK1>,
+                                                Sequence<BK0PerBlock, NPerBlock, BK1>,
                                                 BBlockTransferThreadClusterLengths_BK0_N_BK1,
                                                 BBlockTransferThreadClusterArrangeOrder,
-                                                FloatAB,
-                                                FloatAB,
+                                                ABDataType,
+                                                ABDataType,
                                                 decltype(b_grid_desc_bk0_n_bk1),
                                                 decltype(b_block_desc_bk0_n_bk1),
                                                 BBlockTransferSrcAccessOrder,
@@ -379,13 +462,14 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
         //     c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
         //       register
         // sanity check
-        constexpr index_t KPack = math::max(
-            math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
+        constexpr index_t KPack =
+            math::max(math::lcm(AK1, BK1),
+                      MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
 
         auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
             BlockSize,
-            FloatAB,
-            FloatGemmAcc,
+            ABDataType,
+            AccDataType,
             decltype(a_block_desc_ak0_m_ak1),
             decltype(b_block_desc_bk0_n_bk1),
             MPerXdl,
@@ -402,10 +486,10 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
             a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
 
         auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
+            static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
 
         auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-            static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
+            static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
             b_block_desc_bk0_n_bk1.GetElementSpaceSize());
 
         constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
@@ -466,7 +550,7 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
                 GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
 
             auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
-                static_cast<FloatCShuffle*>(p_shared),
+                static_cast<CShuffleDataType*>(p_shared),
                 c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
 
             constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
@@ -518,8 +602,8 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
 
             // shuffle: threadwise copy C from VGPR to LDS
             auto c_thread_copy_vgpr_to_lds =
-                ThreadwiseTensorSliceTransfer_v1r3<FloatGemmAcc,
-                                                   FloatCShuffle,
+                ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
+                                                   CShuffleDataType,
                                                    decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
                                                    decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
                                                    ck::tensor_operation::element_wise::PassThrough,
@@ -576,8 +660,8 @@ struct GridwiseGemmMultipleD_k0mk1_k0nk1_mn_xdl_cshuffle
             // blockwise copy C/D/E between LDS and global
             auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
                 ThisThreadBlock,
-                decltype(container_concat(make_tuple(FloatCShuffle{}), DsDataType{})),
-                Tuple<FloatE>,
+                decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
+                Tuple<EDataType>,
                 decltype(c_ds_desc_refs),
                 decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
                 CDEElementwiseOperation,

include/ck/utility/tuple.hpp

@@ -21,6 +21,8 @@ struct TupleElementKey
 template <typename Key, typename Data>
 struct TupleElementKeyData
 {
+    using DataType = Data;
+
 #if 0 // workaround compiler complaint about implicitly-deleted default constructor
     __host__ __device__ constexpr TupleElementKeyData() = default;
 #else
@@ -34,29 +36,40 @@ struct TupleElementKeyData
     {
     }
 
-    Data mData;
+    DataType mData;
 };
 
+// for read access of tuple element
 template <typename Key, typename Data>
 __host__ __device__ constexpr const Data&
-get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
+get_tuple_element_data_reference(const TupleElementKeyData<Key, Data>& x)
 {
     return static_cast<const Data&>(x.mData);
 }
 
+// for write access of tuple element
 template <typename Key, typename Data>
-__host__ __device__ constexpr Data& get_tuple_element_data(TupleElementKeyData<Key, Data>& x)
+__host__ __device__ constexpr Data&
+get_tuple_element_data_reference(TupleElementKeyData<Key, Data>& x)
 {
     return x.mData;
 }
 
 // TODO: not sure the use of reference is correct
 template <typename Key, typename Data>
-__host__ __device__ constexpr Data&& get_tuple_element_data(TupleElementKeyData<Key, Data>&& x)
+__host__ __device__ constexpr Data&&
+get_tuple_element_data_reference(TupleElementKeyData<Key, Data>&& x)
 {
     return static_cast<Data&&>(x.mData);
 }
 
+// for infering type of tuple element
+template <typename Key, typename Data>
+__host__ __device__ constexpr Data get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
+{
+    return std::forward(x.mData);
+}
+
 template <typename Indices, typename... Xs>
 struct TupleImpl;
 
@@ -87,13 +100,13 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<I
     template <index_t I>
     __host__ __device__ constexpr const auto& GetElementDataByKey(TupleElementKey<I>) const
     {
-        return get_tuple_element_data<TupleElementKey<I>>(*this);
+        return get_tuple_element_data_reference<TupleElementKey<I>>(*this);
     }
 
     template <index_t I>
     __host__ __device__ constexpr auto& GetElementDataByKey(TupleElementKey<I>)
     {
-        return get_tuple_element_data<TupleElementKey<I>>(*this);
+        return get_tuple_element_data_reference<TupleElementKey<I>>(*this);
     }
 };
 
@@ -185,7 +198,8 @@ struct Tuple<>
 template <index_t I, typename TTuple>
 struct tuple_element
 {
-    using type = decltype(TTuple{}.At(Number<I>{}));
+    // type should keep the cv/ref qualifier of original tuple element
+    using type = decltype(detail::get_tuple_element_data<detail::TupleElementKey<I>>(TTuple{}));
 };
 
 template <index_t I, typename TTuple>

library/CMakeLists.txt

 add_subdirectory(src/tensor_operation_instance/gpu)
-add_subdirectory(src/host_tensor)
 add_subdirectory(src/utility)