Grouped 3d conv backward data support (#799)

* Grouped 3d conv backward data support

* Fix comments

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

@@ -19,8 +19,7 @@ getConvBackwardDataSpecializationString(const ConvolutionBackwardDataSpecializat
     switch(s)
     {
     case ConvolutionBackwardDataSpecialization::Default: return "Default";
-    case ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0:
-        return "FFilter1x1Stride1Pad0";
+    case ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0: return "Filter1x1Stride1Pad0";
     default: return "Unrecognized specialization!";
     }
 }

include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp

@@ -258,7 +258,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                                                CDEElementwiseOp>
 {
     // FIXME
-    static_assert(NDimSpatial == 2, "wrong! only implemented for 2D now");
+    static_assert(NDimSpatial == 2 || NDimSpatial == 3,
+                  "wrong! only implemented for 2D and 3D now");
 
     using DeviceOp = DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1;
 
@@ -491,130 +492,172 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
                 compute_ptr_offset_of_batch_.BatchStrideDs_(i) = ds_g_n_c_wis_strides[i][0];
             });
 
+            static constexpr auto NonSpatialDimsNum = Number<3>{};
+
+            static constexpr auto DIdx = Number<NonSpatialDimsNum>{};
+            static constexpr auto HIdx =
+                NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+            static constexpr auto WIdx = NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{}
+                                                          : Number<NonSpatialDimsNum + 2>{};
+
+            static constexpr auto ZIdx = Number<NonSpatialDimsNum>{};
+            static constexpr auto YIdx =
+                NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+            static constexpr auto XIdx = NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{}
+                                                          : Number<NonSpatialDimsNum + 2>{};
+
             // problem definition
-            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 Z = b_g_k_c_xs_lengths[ZIdx];
+            const index_t Y = b_g_k_c_xs_lengths[YIdx];
+            const index_t X = b_g_k_c_xs_lengths[XIdx];
 
-            const index_t ConvStrideH = conv_filter_strides_[0];
-            const index_t ConvStrideW = conv_filter_strides_[1];
+            const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+            const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+            const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-            const index_t ConvDilationH = conv_filter_dilations_[0];
-            const index_t ConvDilationW = conv_filter_dilations_[1];
+            const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+            const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+            const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = NDimSpatial == 3 ? ConvStrideD / GcdStrideDilationD : 1;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
-            for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+            for(index_t i_ztilde = 0; i_ztilde < ZTilde; ++i_ztilde)
             {
-                for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
-                {
-                    // check slice is valid
-                    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
-                    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
-                    if(YDotSlice * XDotSlice <= 0)
+                for(index_t i_ytilde = 0; i_ytilde < YTilde; ++i_ytilde)
+                {
+                    for(index_t i_xtilde = 0; i_xtilde < XTilde; ++i_xtilde)
                     {
-                        continue;
-                    }
-
-                    const auto a_grid_desc_ak0_m_ak1 =
-                        transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
-                            a_g_n_k_wos_lengths,
-                            a_g_n_k_wos_strides,
-                            b_g_k_c_xs_lengths,
-                            b_g_k_c_xs_strides,
-                            e_g_n_c_wis_lengths,
-                            e_g_n_c_wis_strides,
-                            conv_filter_strides,
-                            conv_filter_dilations,
-                            input_left_pads,
-                            input_right_pads,
-                            {i_ytilde, i_xtilde});
-
-                    const auto b_grid_desc_bk0_n_bk1 =
-                        transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
-                            a_g_n_k_wos_lengths,
-                            a_g_n_k_wos_strides,
-                            b_g_k_c_xs_lengths,
-                            b_g_k_c_xs_strides,
-                            e_g_n_c_wis_lengths,
-                            e_g_n_c_wis_strides,
-                            conv_filter_strides,
-                            conv_filter_dilations,
-                            input_left_pads,
-                            input_right_pads,
-                            {i_ytilde, i_xtilde});
-
-                    DsGridDesc_M_N ds_grid_desc_m_n;
-
-                    // populate Ds desc
-                    static_for<0, NumDTensor, 1>{}([&](auto i) {
-                        using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
-
-                        ds_grid_desc_m_n(i) =
-                            transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
+                        // check slice is valid
+                        const auto ZDotSlice =
+                            NDimSpatial == 3 ? math::integer_divide_ceil(Z - i_ztilde, ZTilde) : 1;
+                        const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
+                        const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
+
+                        if(YDotSlice * XDotSlice * ZDotSlice <= 0)
+                        {
+                            continue;
+                        }
+
+                        std::array<index_t, NDimSpatial> tildes;
+                        if constexpr(NDimSpatial == 2)
+                        {
+                            tildes = {i_ytilde, i_xtilde};
+                        }
+                        else if constexpr(NDimSpatial == 3)
+                        {
+                            tildes = {i_ztilde, i_ytilde, i_xtilde};
+                        }
+                        else
+                        {
+                            throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+                        }
+
+                        const auto a_grid_desc_ak0_m_ak1 =
+                            transform_conv_to_gemm.template MakeADescriptor_AK0_M_AK1<ALayout>(
                                 a_g_n_k_wos_lengths,
                                 a_g_n_k_wos_strides,
                                 b_g_k_c_xs_lengths,
                                 b_g_k_c_xs_strides,
-                                ds_g_n_c_wis_lengths[i],
-                                ds_g_n_c_wis_strides[i],
+                                e_g_n_c_wis_lengths,
+                                e_g_n_c_wis_strides,
                                 conv_filter_strides,
                                 conv_filter_dilations,
                                 input_left_pads,
                                 input_right_pads,
-                                {i_ytilde, i_xtilde});
-                    });
-
-                    const auto e_grid_desc_m_n =
-                        transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(
-                            a_g_n_k_wos_lengths,
-                            a_g_n_k_wos_strides,
-                            b_g_k_c_xs_lengths,
-                            b_g_k_c_xs_strides,
-                            e_g_n_c_wis_lengths,
-                            e_g_n_c_wis_strides,
-                            conv_filter_strides,
-                            conv_filter_dilations,
-                            input_left_pads,
-                            input_right_pads,
-                            {i_ytilde, i_xtilde});
-
-                    // desc for problem definition
-                    const auto a_grid_desc_m_k = transform_k0_m_k1_to_m_k(a_grid_desc_ak0_m_ak1);
-                    const auto b_grid_desc_n_k = transform_k0_m_k1_to_m_k(b_grid_desc_bk0_n_bk1);
-
-                    a_grid_desc_m_k_container_.push_back(a_grid_desc_m_k);
-                    b_grid_desc_n_k_container_.push_back(b_grid_desc_n_k);
-                    ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n);
-                    e_grid_desc_m_n_container_.push_back(e_grid_desc_m_n);
-
-                    // desc for blockwise copy
-                    a_grid_desc_ak0_m_ak1_container_.push_back(a_grid_desc_ak0_m_ak1);
-                    b_grid_desc_bk0_n_bk1_container_.push_back(b_grid_desc_bk0_n_bk1);
-
-                    // block-to-e-tile-map
-                    auto block_2_etile_map =
-                        GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
-
-                    block_2_etile_map_container_.push_back(block_2_etile_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))
-                    {
-                        ds_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
-                            GridwiseGemm::MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                                ds_grid_desc_m_n));
+                                tildes);
 
-                        e_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
-                            GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
-                                e_grid_desc_m_n));
+                        const auto b_grid_desc_bk0_n_bk1 =
+                            transform_conv_to_gemm.template MakeBDescriptor_BK0_N_BK1<BLayout>(
+                                a_g_n_k_wos_lengths,
+                                a_g_n_k_wos_strides,
+                                b_g_k_c_xs_lengths,
+                                b_g_k_c_xs_strides,
+                                e_g_n_c_wis_lengths,
+                                e_g_n_c_wis_strides,
+                                conv_filter_strides,
+                                conv_filter_dilations,
+                                input_left_pads,
+                                input_right_pads,
+                                tildes);
+
+                        DsGridDesc_M_N ds_grid_desc_m_n;
+
+                        // populate Ds desc
+                        static_for<0, NumDTensor, 1>{}([&](auto i) {
+                            using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
+
+                            ds_grid_desc_m_n(i) =
+                                transform_conv_to_gemm.template MakeCDescriptor_M_N<DLayout>(
+                                    a_g_n_k_wos_lengths,
+                                    a_g_n_k_wos_strides,
+                                    b_g_k_c_xs_lengths,
+                                    b_g_k_c_xs_strides,
+                                    ds_g_n_c_wis_lengths[i],
+                                    ds_g_n_c_wis_strides[i],
+                                    conv_filter_strides,
+                                    conv_filter_dilations,
+                                    input_left_pads,
+                                    input_right_pads,
+                                    tildes);
+                        });
+
+                        const auto e_grid_desc_m_n =
+                            transform_conv_to_gemm.template MakeCDescriptor_M_N<ELayout>(
+                                a_g_n_k_wos_lengths,
+                                a_g_n_k_wos_strides,
+                                b_g_k_c_xs_lengths,
+                                b_g_k_c_xs_strides,
+                                e_g_n_c_wis_lengths,
+                                e_g_n_c_wis_strides,
+                                conv_filter_strides,
+                                conv_filter_dilations,
+                                input_left_pads,
+                                input_right_pads,
+                                tildes);
+
+                        // desc for problem definition
+                        const auto a_grid_desc_m_k =
+                            transform_k0_m_k1_to_m_k(a_grid_desc_ak0_m_ak1);
+                        const auto b_grid_desc_n_k =
+                            transform_k0_m_k1_to_m_k(b_grid_desc_bk0_n_bk1);
+
+                        a_grid_desc_m_k_container_.push_back(a_grid_desc_m_k);
+                        b_grid_desc_n_k_container_.push_back(b_grid_desc_n_k);
+                        ds_grid_desc_m_n_container_.push_back(ds_grid_desc_m_n);
+                        e_grid_desc_m_n_container_.push_back(e_grid_desc_m_n);
+
+                        // desc for blockwise copy
+                        a_grid_desc_ak0_m_ak1_container_.push_back(a_grid_desc_ak0_m_ak1);
+                        b_grid_desc_bk0_n_bk1_container_.push_back(b_grid_desc_bk0_n_bk1);
+
+                        // block-to-e-tile-map
+                        auto block_2_etile_map =
+                            GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n);
+
+                        block_2_etile_map_container_.push_back(block_2_etile_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))
+                        {
+                            ds_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
+                                GridwiseGemm::
+                                    MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                        ds_grid_desc_m_n));
+
+                            e_grid_desc_mblock_mperblock_nblock_nperblock_container_.push_back(
+                                GridwiseGemm::MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
+                                    e_grid_desc_m_n));
+                        }
                     }
                 }
             }
@@ -803,7 +846,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
         // vector load for A matrix from global memory to LDS
         if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
-                     is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
+                     is_same_v<ALayout, tensor_layout::convolution::GNDHWK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::NHWGK> ||
+                     is_same_v<ALayout, tensor_layout::convolution::NDHWGK>)
         {
             if(!(ABlockTransferSrcVectorDim == 2 && ConvK % ABlockTransferSrcScalarPerVector == 0))
             {
@@ -816,7 +861,8 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
         }
 
         // vector load for B matrix from global memory to LDS
-        if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC>)
+        if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                     is_same_v<BLayout, tensor_layout::convolution::GKZYXC>)
         {
             if(!(BBlockTransferSrcVectorDim == 1 && ConvC % BBlockTransferSrcScalarPerVector == 0))
             {
@@ -835,7 +881,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
             using DLayout = remove_cvref_t<tuple_element_t<i.value, DsLayout>>;
 
             if constexpr(is_same_v<DLayout, tensor_layout::convolution::GNHWC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::GNDHWC> ||
                          is_same_v<DLayout, tensor_layout::convolution::NHWGC> ||
+                         is_same_v<DLayout, tensor_layout::convolution::NDHWGC> ||
                          is_same_v<DLayout, tensor_layout::convolution::G_NHW_C> ||
                          is_same_v<DLayout, tensor_layout::convolution::GC> ||
                          is_same_v<DLayout, tensor_layout::convolution::G_C>)
@@ -859,7 +907,9 @@ struct DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1
 
         // vector store for E
         if constexpr(is_same_v<ELayout, tensor_layout::convolution::GNHWC> ||
-                     is_same_v<ELayout, tensor_layout::convolution::NHWGC>)
+                     is_same_v<ELayout, tensor_layout::convolution::GNDHWC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::NHWGC> ||
+                     is_same_v<ELayout, tensor_layout::convolution::NDHWGC>)
         {
             // vector store C matrix into global memory
             if(!(ConvC % CDEBlockTransferScalarPerVector_NPerBlock == 0))

include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp

@@ -18,32 +18,53 @@ template <
     index_t NDimSpatial,
     typename ALayout,
     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization ConvBwdDataSpecialization>
-constexpr auto
-make_out_n_ho_wo_k_grid_desc(const index_t N,
-                             const index_t Ho,
-                             const index_t Wo,
-                             const index_t K,
-                             const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides)
+constexpr auto make_out_grid_desc(const index_t N,
+                                  const index_t Do,
+                                  const index_t Ho,
+                                  const index_t Wo,
+                                  const index_t K,
+                                  const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_strides)
 {
+    const auto KStride = Number<1>{};
 
     if constexpr(is_same_v<ALayout, tensor_layout::convolution::NHWGK>)
     {
         const index_t NStride  = out_g_n_k_wos_strides[1];
         const index_t HiStride = out_g_n_k_wos_strides[3];
         const index_t WiStride = out_g_n_k_wos_strides[4];
-        const auto CStride     = Number<1>{};
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
                          Filter1x1Stride1Pad0)
         {
 
             return make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, K),
-                                                make_tuple(WiStride, CStride));
+                                                make_tuple(WiStride, KStride));
         }
         else
         {
             return make_naive_tensor_descriptor(make_tuple(N, Ho, Wo, K),
-                                                make_tuple(NStride, HiStride, WiStride, CStride));
+                                                make_tuple(NStride, HiStride, WiStride, KStride));
+        }
+    }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::NDHWGK>)
+    {
+        const index_t NStride  = out_g_n_k_wos_strides[1];
+        const index_t DoStride = out_g_n_k_wos_strides[3];
+        const index_t HoStride = out_g_n_k_wos_strides[4];
+        const index_t WoStride = out_g_n_k_wos_strides[5];
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+
+            return make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, K),
+                                                make_tuple(WoStride, KStride));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor(
+                make_tuple(N, Do, Ho, Wo, K),
+                make_tuple(NStride, DoStride, HoStride, WoStride, KStride));
         }
     }
     else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNHWK>)
@@ -60,12 +81,80 @@ make_out_n_ho_wo_k_grid_desc(const index_t N,
             return make_naive_tensor_descriptor_packed(make_tuple(N, Ho, Wo, K));
         }
     }
+    else if constexpr(is_same_v<ALayout, tensor_layout::convolution::GNDHWK>)
+    {
+        // assume packed
+        if constexpr(ConvBwdDataSpecialization ==
+                     ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
+                         Filter1x1Stride1Pad0)
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N * Do * Ho * Wo, K));
+        }
+        else
+        {
+            return make_naive_tensor_descriptor_packed(make_tuple(N, Do, Ho, Wo, K));
+        }
+    }
     else
     {
         throw std::runtime_error("wrong! unsupported layout: " + ALayout::name());
     }
 }
 
+template <typename BLayout>
+constexpr auto make_wei_grid_desc(
+    const index_t K, const index_t Z, const index_t Y, const index_t X, const index_t C)
+{
+
+    if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKYXC>)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+    }
+    else if constexpr(is_same_v<BLayout, tensor_layout::convolution::GKZYXC>)
+    {
+        return make_naive_tensor_descriptor_packed(make_tuple(K, Z, Y, X, C));
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + BLayout::name());
+    }
+}
+
+template <index_t NDimSpatial, typename CLayout>
+constexpr auto make_in_grid_desc(const index_t N,
+                                 const index_t Di,
+                                 const index_t Hi,
+                                 const index_t Wi,
+                                 const index_t C,
+                                 const std::array<index_t, NDimSpatial + 3>& in_g_n_c_wis_strides)
+{
+
+    if constexpr(is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                 is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                 is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>)
+    {
+        return make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
+                                            make_tuple(in_g_n_c_wis_strides[1],
+                                                       in_g_n_c_wis_strides[3],
+                                                       in_g_n_c_wis_strides[4],
+                                                       in_g_n_c_wis_strides[2]));
+    }
+    else if constexpr(is_same_v<CLayout, tensor_layout::convolution::GNDHWC> ||
+                      is_same_v<CLayout, tensor_layout::convolution::NDHWGC>)
+    {
+        return make_naive_tensor_descriptor(make_tuple(N, Di, Hi, Wi, C),
+                                            make_tuple(in_g_n_c_wis_strides[1],
+                                                       in_g_n_c_wis_strides[3],
+                                                       in_g_n_c_wis_strides[4],
+                                                       in_g_n_c_wis_strides[5],
+                                                       in_g_n_c_wis_strides[2]));
+    }
+    else
+    {
+        throw std::runtime_error("wrong! unsupported layout: " + CLayout::name());
+    }
+}
+
 } // namespace
 
 template <
@@ -82,10 +171,26 @@ struct TransformConvBwdDataToGemm_v1
     static constexpr auto I0 = Number<0>{};
     static constexpr auto I1 = Number<1>{};
 
+    static constexpr auto NonSpatialDimsNum = Number<3>{};
+
+    static constexpr auto DIdx = Number<NonSpatialDimsNum>{};
+    static constexpr auto HIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+    static constexpr auto WIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{} : Number<NonSpatialDimsNum + 2>{};
+
+    static constexpr auto ZIdx = Number<NonSpatialDimsNum>{};
+    static constexpr auto YIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum>{} : Number<NonSpatialDimsNum + 1>{};
+    static constexpr auto XIdx =
+        NDimSpatial == 2 ? Number<NonSpatialDimsNum + 1>{} : Number<NonSpatialDimsNum + 2>{};
+
     template <typename ALayout,
-              typename std::enable_if<NDimSpatial == 2 &&
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
                                           (is_same_v<ALayout, tensor_layout::convolution::GNHWK> ||
-                                           is_same_v<ALayout, tensor_layout::convolution::NHWGK>),
+                                           is_same_v<ALayout, tensor_layout::convolution::GNDHWK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NHWGK> ||
+                                           is_same_v<ALayout, tensor_layout::convolution::NDHWGK>),
                                       bool>::type = false>
     static auto MakeADescriptor_AK0_M_AK1(
         const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
@@ -100,35 +205,43 @@ struct TransformConvBwdDataToGemm_v1
         const std::array<index_t, NDimSpatial>& /* input_right_pads */,
         const std::array<index_t, NDimSpatial>& tildes)
     {
-        index_t i_ytilde = tildes[0];
-        index_t i_xtilde = tildes[1];
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
 
         const index_t N = in_g_n_c_wis_lengths[1];
         const index_t K = wei_g_k_c_xs_lengths[1];
 
-        const index_t Hi = in_g_n_c_wis_lengths[3];
-        const index_t Wi = in_g_n_c_wis_lengths[4];
+        const index_t Di = NDimSpatial == 3 ? in_g_n_c_wis_lengths[DIdx] : 1;
+        const index_t Hi = in_g_n_c_wis_lengths[HIdx];
+        const index_t Wi = in_g_n_c_wis_lengths[WIdx];
 
-        const index_t Ho = out_g_n_k_wos_lengths[3];
-        const index_t Wo = out_g_n_k_wos_lengths[4];
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
 
-        const index_t Y = wei_g_k_c_xs_lengths[3];
-        const index_t X = wei_g_k_c_xs_lengths[4];
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
 
-        const index_t InLeftPadH = input_left_pads[0];
-        const index_t InLeftPadW = input_left_pads[1];
+        const index_t InLeftPadD = input_left_pads[DIdx - NonSpatialDimsNum];
+        const index_t InLeftPadH = input_left_pads[HIdx - NonSpatialDimsNum];
+        const index_t InLeftPadW = input_left_pads[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
         const index_t AK0 = K / AK1;
 
-        const auto out_n_ho_wo_k_grid_desc =
-            make_out_n_ho_wo_k_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
-                N, Ho, Wo, K, out_g_n_k_wos_strides);
+        // n_do_ho_wo_k for 3d or n_ho_wo_k for 2d
+        const auto out_grid_desc =
+            make_out_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
+                N, Do, Ho, Wo, K, out_g_n_k_wos_strides);
 
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
@@ -136,8 +249,8 @@ struct TransformConvBwdDataToGemm_v1
         {
             // A: output tensor
             const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
-                out_n_ho_wo_k_grid_desc,
-                make_tuple(make_pass_through_transform(N * Ho * Wo),
+                out_grid_desc,
+                make_tuple(make_pass_through_transform(N * Do * Ho * Wo),
                            make_unmerge_transform(make_tuple(AK0, AK1))),
                 make_tuple(Sequence<0>{}, Sequence<1>{}),
                 make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
@@ -152,103 +265,208 @@ struct TransformConvBwdDataToGemm_v1
         }
         else
         {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
             const auto YDot = math::integer_divide_ceil(Y, YTilde);
             const auto XDot = math::integer_divide_ceil(X, XTilde);
 
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
             const auto HTilde =
                 Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
             const auto WTilde =
                 Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
 
             // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
             const auto IHTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
             const auto IWTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
 
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
             const auto IHTildeSliceEnd = math::min(
                 HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
             const auto IWTildeSliceEnd = math::min(
                 WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
 
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
             const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
             const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
 
             // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
-            // A: output tensor
-            const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
-                out_n_ho_wo_k_grid_desc,
-                make_tuple(make_pass_through_transform(N),
-                           make_pad_transform(Ho, I0, I0),
-                           make_pad_transform(Wo, I0, I0),
-                           make_pass_through_transform(K)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
-
-            const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
-                out_n_hop_wop_k_grid_desc,
-                make_tuple(
-                    make_pass_through_transform(N),
-                    make_embed_transform(make_tuple(YDot, HTilde),
-                                         make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
-                    make_embed_transform(make_tuple(XDot, WTilde),
-                                         make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
-                    make_pass_through_transform(K)),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
-
-            const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
-                transform_tensor_descriptor(
-                    out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+            if constexpr(NDimSpatial == 2)
+            {
+                // A: output tensor
+                const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                    out_grid_desc,
                     make_tuple(make_pass_through_transform(N),
-                               make_slice_transform(YDot, I0, YDotSlice),
-                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
-                               make_slice_transform(XDot, I0, XDotSlice),
-                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                               make_unmerge_transform(make_tuple(AK0, AK1))),
-                    make_tuple(Sequence<0>{},
-                               Sequence<1>{},
-                               Sequence<2>{},
-                               Sequence<3>{},
-                               Sequence<4>{},
-                               Sequence<5>{}),
-                    make_tuple(Sequence<0>{},
-                               Sequence<1>{},
-                               Sequence<2>{},
-                               Sequence<3>{},
-                               Sequence<4>{},
-                               Sequence<5, 6>{}));
-
-            const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
-                out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
-                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, AK0)),
-                           make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                           make_pass_through_transform(AK1)),
-                make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-
-            const auto out_gemmak0_gemmm_gemmak1_grid_desc =
-                ck::tensor_operation::device::PadTensorDescriptor(
-                    out_gemmak0_gemmmraw_gemmak1_grid_desc,
-                    make_tuple(AK0, GemmMPerBlock, AK1),
-                    Sequence<false, DoPadGemmM, false>{});
-
-            return out_gemmak0_gemmm_gemmak1_grid_desc;
+                               make_pad_transform(Ho, I0, I0),
+                               make_pad_transform(Wo, I0, I0),
+                               make_pass_through_transform(K)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+                const auto out_n_ydot_htilde_xdot_wtilde_k_grid_desc = transform_tensor_descriptor(
+                    out_n_hop_wop_k_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(YDot, HTilde),
+                                             make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, WTilde),
+                                             make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                        make_pass_through_transform(K)),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+                const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_unmerge_transform(make_tuple(AK0, AK1))),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5, 6>{}));
+
+                const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, AK0)),
+                               make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                               make_pass_through_transform(AK1)),
+                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        out_gemmak0_gemmmraw_gemmak1_grid_desc,
+                        make_tuple(AK0, GemmMPerBlock, AK1),
+                        Sequence<false, DoPadGemmM, false>{});
+
+                return out_gemmak0_gemmm_gemmak1_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                // A: output tensor
+                const auto out_n_hop_wop_k_grid_desc = transform_tensor_descriptor(
+                    out_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(Do, I0, I0),
+                               make_pad_transform(Ho, I0, I0),
+                               make_pad_transform(Wo, I0, I0),
+                               make_pass_through_transform(K)),
+                    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 out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc =
+                    transform_tensor_descriptor(
+                        out_n_hop_wop_k_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(
+                                       make_tuple(ZDot, DTilde),
+                                       make_tuple(-ConvDilationD / GcdStrideDilationD, I1)),
+                                   make_embed_transform(
+                                       make_tuple(YDot, HTilde),
+                                       make_tuple(-ConvDilationH / GcdStrideDilationH, I1)),
+                                   make_embed_transform(
+                                       make_tuple(XDot, WTilde),
+                                       make_tuple(-ConvDilationW / GcdStrideDilationW, I1)),
+                                   make_pass_through_transform(K)),
+                        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
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
+                        transform_tensor_descriptor(
+                            out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
+                            make_tuple(make_pass_through_transform(N),
+                                       make_slice_transform(ZDot, I0, ZDotSlice),
+                                       make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                       make_slice_transform(YDot, I0, YDotSlice),
+                                       make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                       make_slice_transform(XDot, I0, XDotSlice),
+                                       make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                       make_unmerge_transform(make_tuple(AK0, AK1))),
+                            make_tuple(Sequence<0>{},
+                                       Sequence<1>{},
+                                       Sequence<2>{},
+                                       Sequence<3>{},
+                                       Sequence<4>{},
+                                       Sequence<5>{},
+                                       Sequence<6>{},
+                                       Sequence<7>{}),
+                            make_tuple(Sequence<0>{},
+                                       Sequence<1>{},
+                                       Sequence<2>{},
+                                       Sequence<3>{},
+                                       Sequence<4>{},
+                                       Sequence<5>{},
+                                       Sequence<6>{},
+                                       Sequence<7, 8>{}));
+
+                const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, AK0)),
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                        make_pass_through_transform(AK1)),
+                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        out_gemmak0_gemmmraw_gemmak1_grid_desc,
+                        make_tuple(AK0, GemmMPerBlock, AK1),
+                        Sequence<false, DoPadGemmM, false>{});
+
+                return out_gemmak0_gemmm_gemmak1_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
         }
     }
 
     template <typename BLayout,
-              typename std::enable_if<NDimSpatial == 2 &&
-                                          is_same_v<BLayout, tensor_layout::convolution::GKYXC>,
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
+                                          (is_same_v<BLayout, tensor_layout::convolution::GKYXC> ||
+                                           is_same_v<BLayout, tensor_layout::convolution::GKZYXC>),
                                       bool>::type = false>
     static auto MakeBDescriptor_BK0_N_BK1(
         const std::array<index_t, NDimSpatial + 3>& out_g_n_k_wos_lengths,
@@ -263,30 +481,35 @@ struct TransformConvBwdDataToGemm_v1
         const std::array<index_t, NDimSpatial>& /* input_right_pads */,
         const std::array<index_t, NDimSpatial>& tildes)
     {
-        index_t i_ytilde = tildes[0];
-        index_t i_xtilde = tildes[1];
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
 
         const index_t N = in_g_n_c_wis_lengths[1];
         const index_t K = wei_g_k_c_xs_lengths[1];
         const index_t C = wei_g_k_c_xs_lengths[2];
 
-        const index_t Ho = out_g_n_k_wos_lengths[3];
-        const index_t Wo = out_g_n_k_wos_lengths[4];
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
 
-        const index_t Y = wei_g_k_c_xs_lengths[3];
-        const index_t X = wei_g_k_c_xs_lengths[4];
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
 
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
         const index_t BK0 = K / BK1;
 
         // assume packed
-        const auto wei_k_y_x_c_grid_desc =
-            make_naive_tensor_descriptor_packed(make_tuple(K, Y, X, C));
+        // k_y_x_c for 2d or k_z_y_x_c for 3d
+        const auto wei_grid_desc = make_wei_grid_desc<BLayout>(K, Z, Y, X, C);
 
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
@@ -299,7 +522,7 @@ struct TransformConvBwdDataToGemm_v1
                                                        make_pass_through_transform(C)),
                                             make_tuple(Sequence<0>{}, Sequence<1>{}),
                                             make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
-            make_naive_tensor_descriptor(make_tuple(N * Ho * Wo, C), make_tuple(I0, I1));
+            make_naive_tensor_descriptor(make_tuple(N * Do * Ho * Wo, C), make_tuple(I0, I1));
 
             const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
                 ck::tensor_operation::device::PadTensorDescriptor(
@@ -311,75 +534,163 @@ struct TransformConvBwdDataToGemm_v1
         }
         else
         {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            const auto ZDot = math::integer_divide_ceil(Z, ZTilde);
             const auto YDot = math::integer_divide_ceil(Y, YTilde);
             const auto XDot = math::integer_divide_ceil(X, XTilde);
 
             // GemmK is different for each GEMM
+            const auto ZDotSlice = math::integer_divide_ceil(Z - i_ztilde, ZTilde);
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
             // B weight tensor
-            const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
-                wei_k_y_x_c_grid_desc,
-                make_tuple(make_pass_through_transform(K),
-                           make_embed_transform(make_tuple(YDot, YTilde),
-                                                make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
-                           make_embed_transform(make_tuple(XDot, XTilde),
-                                                make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
-                           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 wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc =
-                transform_tensor_descriptor(wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
-                                            make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                                       make_slice_transform(YDot, I0, YDotSlice),
-                                                       make_slice_transform(XDot, I0, XDotSlice),
-                                                       make_freeze_transform(i_ytilde),
-                                                       make_freeze_transform(i_xtilde),
-                                                       make_pass_through_transform(C)),
-                                            make_tuple(Sequence<0>{},
-                                                       Sequence<1>{},
-                                                       Sequence<3>{},
-                                                       Sequence<2>{},
-                                                       Sequence<4>{},
-                                                       Sequence<5>{}),
-                                            make_tuple(Sequence<0, 1>{},
-                                                       Sequence<2>{},
-                                                       Sequence<3>{},
-                                                       Sequence<>{},
-                                                       Sequence<>{},
-                                                       Sequence<4>{}));
-
-            const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
-                wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc,
-                make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, BK0)),
-                           make_pass_through_transform(C),
-                           make_pass_through_transform(BK1)),
-                make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
-
-            const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
-                ck::tensor_operation::device::PadTensorDescriptor(
-                    wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+            if constexpr(NDimSpatial == 2)
+            {
+                const auto wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc = transform_tensor_descriptor(
+                    wei_grid_desc,
                     make_tuple(
-                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0), GemmNPerBlock, BK1),
-                    Sequence<false, DoPadGemmN, false>{});
-
-            return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+                        make_pass_through_transform(K),
+                        make_embed_transform(make_tuple(YDot, YTilde),
+                                             make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                        make_embed_transform(make_tuple(XDot, XTilde),
+                                             make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                        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 wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                        make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_freeze_transform(i_ytilde),
+                                   make_freeze_transform(i_xtilde),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<3>{},
+                                   Sequence<2>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{}),
+                        make_tuple(Sequence<0, 1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<4>{}));
+
+                const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, BK0)),
+                               make_pass_through_transform(C),
+                               make_pass_through_transform(BK1)),
+                    make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                        make_tuple(wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0),
+                                   GemmNPerBlock,
+                                   BK1),
+                        Sequence<false, DoPadGemmN, false>{});
+
+                return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+                const auto wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_grid_desc,
+                        make_tuple(
+                            make_pass_through_transform(K),
+                            make_embed_transform(make_tuple(ZDot, ZTilde),
+                                                 make_tuple(ConvStrideD / GcdStrideDilationD, I1)),
+                            make_embed_transform(make_tuple(YDot, YTilde),
+                                                 make_tuple(ConvStrideH / GcdStrideDilationH, I1)),
+                            make_embed_transform(make_tuple(XDot, XTilde),
+                                                 make_tuple(ConvStrideW / GcdStrideDilationW, I1)),
+                            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 wei_bk0_bk1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                        make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                                   make_slice_transform(ZDot, I0, ZDotSlice),
+                                   make_slice_transform(YDot, I0, YDotSlice),
+                                   make_slice_transform(XDot, I0, XDotSlice),
+                                   make_freeze_transform(i_ztilde),
+                                   make_freeze_transform(i_ytilde),
+                                   make_freeze_transform(i_xtilde),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<3>{},
+                                   Sequence<5>{},
+                                   Sequence<2>{},
+                                   Sequence<4>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0, 1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<>{},
+                                   Sequence<5>{}));
+
+                const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_bk0_bk1_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, BK0)),
+                        make_pass_through_transform(C),
+                        make_pass_through_transform(BK1)),
+                    make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+
+                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
+                        make_tuple(wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0),
+                                   GemmNPerBlock,
+                                   BK1),
+                        Sequence<false, DoPadGemmN, false>{});
+
+                return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
         }
     }
 
     template <typename CLayout,
-              typename std::enable_if<NDimSpatial == 2 &&
+              typename std::enable_if<(NDimSpatial == 2 || NDimSpatial == 3) &&
                                           (is_same_v<CLayout, tensor_layout::convolution::GNHWC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::GNDHWC> ||
                                            is_same_v<CLayout, tensor_layout::convolution::NHWGC> ||
+                                           is_same_v<CLayout, tensor_layout::convolution::NDHWGC> ||
                                            is_same_v<CLayout, tensor_layout::convolution::G_NHW_C>),
                                       bool>::type = false>
     static auto
@@ -395,153 +706,309 @@ struct TransformConvBwdDataToGemm_v1
                         const std::array<index_t, NDimSpatial>& input_right_pads,
                         const std::array<index_t, NDimSpatial>& tildes)
     {
-        index_t i_ytilde = tildes[0];
-        index_t i_xtilde = tildes[1];
+        index_t i_ztilde = tildes[ZIdx - NonSpatialDimsNum];
+        index_t i_ytilde = tildes[YIdx - NonSpatialDimsNum];
+        index_t i_xtilde = tildes[XIdx - NonSpatialDimsNum];
 
         const index_t N = in_g_n_c_wis_lengths[1];
         const index_t C = wei_g_k_c_xs_lengths[2];
 
-        const index_t Hi = in_g_n_c_wis_lengths[3];
-        const index_t Wi = in_g_n_c_wis_lengths[4];
+        const index_t Di = NDimSpatial == 3 ? in_g_n_c_wis_lengths[DIdx] : 1;
+        const index_t Hi = in_g_n_c_wis_lengths[HIdx];
+        const index_t Wi = in_g_n_c_wis_lengths[WIdx];
 
-        const index_t Ho = out_g_n_k_wos_lengths[3];
-        const index_t Wo = out_g_n_k_wos_lengths[4];
+        const index_t Do = NDimSpatial == 3 ? out_g_n_k_wos_lengths[DIdx] : 1;
+        const index_t Ho = out_g_n_k_wos_lengths[HIdx];
+        const index_t Wo = out_g_n_k_wos_lengths[WIdx];
 
-        const index_t Y = wei_g_k_c_xs_lengths[3];
-        const index_t X = wei_g_k_c_xs_lengths[4];
+        const index_t Z = NDimSpatial == 3 ? wei_g_k_c_xs_lengths[ZIdx] : 1;
+        const index_t Y = wei_g_k_c_xs_lengths[YIdx];
+        const index_t X = wei_g_k_c_xs_lengths[XIdx];
 
-        const index_t InLeftPadH = input_left_pads[0];
-        const index_t InLeftPadW = input_left_pads[1];
+        const index_t InLeftPadD = input_left_pads[DIdx - NonSpatialDimsNum];
+        const index_t InLeftPadH = input_left_pads[HIdx - NonSpatialDimsNum];
+        const index_t InLeftPadW = input_left_pads[WIdx - NonSpatialDimsNum];
 
-        const index_t InRightPadH = input_right_pads[0];
-        const index_t InRightPadW = input_right_pads[1];
+        const index_t InRightPadD = input_right_pads[DIdx - NonSpatialDimsNum];
+        const index_t InRightPadH = input_right_pads[HIdx - NonSpatialDimsNum];
+        const index_t InRightPadW = input_right_pads[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvStrideH = conv_filter_strides[0];
-        const index_t ConvStrideW = conv_filter_strides[1];
+        const index_t ConvStrideD = conv_filter_strides[DIdx - NonSpatialDimsNum];
+        const index_t ConvStrideH = conv_filter_strides[HIdx - NonSpatialDimsNum];
+        const index_t ConvStrideW = conv_filter_strides[WIdx - NonSpatialDimsNum];
 
-        const index_t ConvDilationH = conv_filter_dilations[0];
-        const index_t ConvDilationW = conv_filter_dilations[1];
+        const index_t ConvDilationD = conv_filter_dilations[DIdx - NonSpatialDimsNum];
+        const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
+        const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
         // assume strided
-        const auto in_n_hi_wi_c_grid_desc =
-            make_naive_tensor_descriptor(make_tuple(N, Hi, Wi, C),
-                                         make_tuple(in_g_n_c_wis_strides[1],
-                                                    in_g_n_c_wis_strides[3],
-                                                    in_g_n_c_wis_strides[4],
-                                                    in_g_n_c_wis_strides[2]));
+        // n_hi_wi_c for 2d n_di_hi_wi_c for 3d
+        const auto in_grid_desc =
+            make_in_grid_desc<NDimSpatial, CLayout>(N, Di, Hi, Wi, C, in_g_n_c_wis_strides);
 
         if constexpr(ConvBwdDataSpecialization ==
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
                          Filter1x1Stride1Pad0)
         {
             // C: input tensor
-            const auto in_n_y_ho_x_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(I1, Ho), make_tuple(I1, ConvStrideH)),
-                           make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, 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_gemmnraw_grid_desc = transform_tensor_descriptor(
-                in_n_y_ho_x_wo_c_grid_desc,
-                make_tuple(make_freeze_transform(I0),
-                           make_freeze_transform(I0),
-                           make_merge_transform(make_tuple(N, Ho, Wo)),
-                           make_pass_through_transform(C)),
-                make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
-                make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
-
-            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
-                in_gemmmraw_gemmnraw_grid_desc,
-                make_tuple(GemmMPerBlock, GemmNPerBlock),
-                Sequence<DoPadGemmM, DoPadGemmN>{});
-
-            return in_gemmm_gemmn_grid_desc;
+            if constexpr(NDimSpatial == 2)
+            {
+                const auto in_n_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                        make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, 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_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_y_ho_x_wo_c_grid_desc,
+                    make_tuple(make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_merge_transform(make_tuple(N, Ho, Wo)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1>{}, Sequence<3>{}, Sequence<0, 2, 4>{}, Sequence<5>{}),
+                    make_tuple(Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else if constexpr(NDimSpatial == 3)
+            {
+
+                // C: input tensor
+                const auto in_n_x_do_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(
+                        make_pass_through_transform(N),
+                        make_embed_transform(make_tuple(I1, Do), make_tuple(I1, ConvStrideD)),
+                        make_embed_transform(make_tuple(I1, Ho), make_tuple(I1, ConvStrideH)),
+                        make_embed_transform(make_tuple(I1, Wo), make_tuple(I1, 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_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_x_do_y_ho_x_wo_c_grid_desc,
+                    make_tuple(make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_freeze_transform(I0),
+                               make_merge_transform(make_tuple(N, Do, Ho, Wo)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1>{},
+                               Sequence<3>{},
+                               Sequence<5>{},
+                               Sequence<0, 2, 4, 6>{},
+                               Sequence<7>{}),
+                    make_tuple(
+                        Sequence<>{}, Sequence<>{}, Sequence<>{}, Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
         }
         else
         {
+            const auto GcdStrideDilationD = math::gcd(ConvStrideD, ConvDilationD);
             const auto GcdStrideDilationH = math::gcd(ConvStrideH, ConvDilationH);
             const auto GcdStrideDilationW = math::gcd(ConvStrideW, ConvDilationW);
 
+            const auto ZTilde = ConvStrideD / GcdStrideDilationD;
             const auto YTilde = ConvStrideH / GcdStrideDilationH;
             const auto XTilde = ConvStrideW / GcdStrideDilationW;
 
+            const auto DTilde =
+                Do + math::integer_divide_ceil(ConvDilationD * (Z - I1), ConvStrideD);
             const auto HTilde =
                 Ho + math::integer_divide_ceil(ConvDilationH * (Y - I1), ConvStrideH);
             const auto WTilde =
                 Wo + math::integer_divide_ceil(ConvDilationW * (X - I1), ConvStrideW);
 
-            // only work on HTilde and WTilde that contribute to non-padding area of input tensor
+            // only work on DTilde, HTilde and WTilde that contribute to
+            // non-padding area of input tensor
+            const auto IDTildeSliceBegin = math::integer_divide_floor(
+                math::max(I0, InLeftPadD - ConvDilationD * (ZTilde - I1)), ConvStrideD);
             const auto IHTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadH - ConvDilationH * (YTilde - I1)), ConvStrideH);
             const auto IWTildeSliceBegin = math::integer_divide_floor(
                 math::max(I0, InLeftPadW - ConvDilationW * (XTilde - I1)), ConvStrideW);
 
+            const auto IDTildeSliceEnd = math::min(
+                DTilde, math::integer_divide_ceil(InLeftPadD + Di - I1, ConvStrideD) + I1);
             const auto IHTildeSliceEnd = math::min(
                 HTilde, math::integer_divide_ceil(InLeftPadH + Hi - I1, ConvStrideH) + I1);
             const auto IWTildeSliceEnd = math::min(
                 WTilde, math::integer_divide_ceil(InLeftPadW + Wi - I1, ConvStrideW) + I1);
 
+            const auto DTildeSlice = IDTildeSliceEnd - IDTildeSliceBegin;
             const auto HTildeSlice = IHTildeSliceEnd - IHTildeSliceBegin;
             const auto WTildeSlice = IWTildeSliceEnd - IWTildeSliceBegin;
 
             // C: input tensor
-            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_ytilde_htilde_xtilde_wtilde_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(YTilde, HTilde),
-                                                make_tuple(ConvDilationH, ConvStrideH)),
-                           make_embed_transform(make_tuple(XTilde, WTilde),
-                                                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_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
-                in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
-                make_tuple(make_pass_through_transform(N),
-                           make_freeze_transform(i_ytilde),
-                           make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
-                           make_freeze_transform(i_xtilde),
-                           make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                           make_pass_through_transform(C)),
-                make_tuple(Sequence<0>{},
-                           Sequence<1>{},
-                           Sequence<2>{},
-                           Sequence<3>{},
-                           Sequence<4>{},
-                           Sequence<5>{}),
-                make_tuple(Sequence<0>{},
-                           Sequence<>{},
-                           Sequence<1>{},
-                           Sequence<>{},
-                           Sequence<2>{},
-                           Sequence<3>{}));
-
-            const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
-                in_n_htildeslice_wtildeslice_c_grid_desc,
-                make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                           make_pass_through_transform(C)),
-                make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
-                make_tuple(Sequence<0>{}, Sequence<1>{}));
-
-            const auto in_gemmm_gemmn_grid_desc = ck::tensor_operation::device::PadTensorDescriptor(
-                in_gemmmraw_gemmnraw_grid_desc,
-                make_tuple(GemmMPerBlock, GemmNPerBlock),
-                Sequence<DoPadGemmM, DoPadGemmN>{});
-
-            return in_gemmm_gemmn_grid_desc;
+            if constexpr(NDimSpatial == 2)
+            {
+                const auto in_n_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                    in_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_pad_transform(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_ytilde_htilde_xtilde_wtilde_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(YTilde, HTilde),
+                                                        make_tuple(ConvDilationH, ConvStrideH)),
+                                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                                        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_n_htildeslice_wtildeslice_c_grid_desc = transform_tensor_descriptor(
+                    in_n_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(N),
+                               make_freeze_transform(i_ytilde),
+                               make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                               make_freeze_transform(i_xtilde),
+                               make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<3>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<>{},
+                               Sequence<1>{},
+                               Sequence<>{},
+                               Sequence<2>{},
+                               Sequence<3>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_htildeslice_wtildeslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0, 1, 2>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else if(NDimSpatial == 3)
+            {
+                const auto in_n_dip_hip_wip_c_grid_desc = transform_tensor_descriptor(
+                    in_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_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_dip_hip_wip_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_embed_transform(make_tuple(ZTilde, DTilde),
+                                                        make_tuple(ConvDilationD, ConvStrideD)),
+                                   make_embed_transform(make_tuple(YTilde, HTilde),
+                                                        make_tuple(ConvDilationH, ConvStrideH)),
+                                   make_embed_transform(make_tuple(XTilde, WTilde),
+                                                        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_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc =
+                    transform_tensor_descriptor(
+                        in_n_ztilde_dtilde_ytilde_htilde_xtilde_wtilde_c_grid_desc,
+                        make_tuple(make_pass_through_transform(N),
+                                   make_freeze_transform(i_ztilde),
+                                   make_slice_transform(DTilde, IDTildeSliceBegin, DTildeSlice),
+                                   make_freeze_transform(i_ytilde),
+                                   make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
+                                   make_freeze_transform(i_xtilde),
+                                   make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
+                                   make_pass_through_transform(C)),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
+                                   Sequence<2>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{},
+                                   Sequence<5>{},
+                                   Sequence<6>{},
+                                   Sequence<7>{}),
+                        make_tuple(Sequence<0>{},
+                                   Sequence<>{},
+                                   Sequence<1>{},
+                                   Sequence<>{},
+                                   Sequence<2>{},
+                                   Sequence<>{},
+                                   Sequence<3>{},
+                                   Sequence<4>{}));
+
+                const auto in_gemmmraw_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    in_n_dtildeslice_htildeslice_wtildeslice_c_grid_desc,
+                    make_tuple(
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
+                        make_pass_through_transform(C)),
+                    make_tuple(Sequence<0, 1, 2, 3>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+                const auto in_gemmm_gemmn_grid_desc =
+                    ck::tensor_operation::device::PadTensorDescriptor(
+                        in_gemmmraw_gemmnraw_grid_desc,
+                        make_tuple(GemmMPerBlock, GemmNPerBlock),
+                        Sequence<DoPadGemmM, DoPadGemmN>{});
+                return in_gemmm_gemmn_grid_desc;
+            }
+            else
+            {
+                throw std::runtime_error("wrong! only implemented for 2D and 3D now");
+            }
         }
     }
 

library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#pragma once
+
+#include "ck/ck.hpp"
+#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
+#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
+#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+
+using BF16 = ck::bhalf_t;
+using F16  = ck::half_t;
+using F32  = float;
+
+using Empty_Tuple = ck::Tuple<>;
+
+template <ck::index_t... Is>
+using S = ck::Sequence<Is...>;
+
+using namespace ck::tensor_layout::convolution;
+
+using PassThrough = ck::tensor_operation::element_wise::PassThrough;
+
+static constexpr auto ConvBwdDataDefault = ConvolutionBackwardDataSpecialization::Default;
+
+static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
+    ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
+
+// f16_f16_f32_f16
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionBackwardDataSpecialization ConvSpec>
+using device_grouped_conv_bwd_data_xdl_f16_instances =
+    std::tuple<
+        // clang-format off
+        // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+        // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+        // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+        // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
+        
+#ifdef CK_WORKAROUND_SWDEV_3318619
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+#endif
+        // clang-format on
+        >;
+
+// bf16_bf16_f32_bf16
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionBackwardDataSpecialization ConvSpec>
+using device_grouped_conv_bwd_data_xdl_bf16_instances =
+    std::tuple<
+        // clang-format off
+        // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+        // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+        // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+        // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   BF16,   BF16,     F32,      BF16, Empty_Tuple,   BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
+        
+#ifdef CK_WORKAROUND_SWDEV_3318619
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
+#endif
+        // clang-format on
+        >;
+
+// f32_f32_f32_f32
+template <index_t NDimSpatial,
+          typename ALayout,
+          typename BLayout,
+          typename DsLayout,
+          typename ELayout,
+          ConvolutionBackwardDataSpecialization ConvSpec>
+using device_grouped_conv_bwd_data_xdl_f32_instances =
+    std::tuple<
+        // clang-format off
+         // ##############################################|       NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
+         // ##############################################|    Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
+         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
+         // ##############################################|           |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
+         DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,  F32,  F32,      F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                4>
+
+#ifdef CK_WORKAROUND_SWDEV_3318619
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
+        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<NDimSpatial, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
+#endif
+        // clang-format on
+        >;
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

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

@@ -16,7 +16,7 @@ namespace device {
 namespace instance {
 
 // conv2d backward data
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -30,7 +30,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -44,7 +44,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -58,7 +58,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -72,7 +72,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -86,7 +86,7 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -100,6 +100,91 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
                                                                   PassThrough,
                                                                   PassThrough>>>& instances);
 
+// conv3d backward data
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances);
+
 template <ck::index_t NumDimSpatial,
           typename OutLayout,
           typename WeiLayout,
@@ -139,43 +224,96 @@ struct DeviceOperationInstanceFactory<
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-
-        if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
-                     is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, GNHWK>)
+        if constexpr(NumDimSpatial == 2)
         {
-            if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
-                         is_same_v<OutDataType, F16>)
-            {
-                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
-                              is_same_v<OutDataType, F32>)
+
+            if constexpr(is_same_v<InLayout, GNHWC> && is_same_v<WeiLayout, GKYXC> &&
+                         is_same_v<OutLayout, GNHWK>)
             {
-                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs);
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
+                        op_ptrs);
+                }
             }
-            else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
-                              is_same_v<OutDataType, BF16>)
+            else if constexpr(is_same_v<InLayout, NHWGC> && is_same_v<WeiLayout, GKYXC> &&
+                              is_same_v<OutLayout, NHWGK>)
             {
-                add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
+                        op_ptrs);
+                }
             }
         }
-        else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
-                          is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
+        else if constexpr(NumDimSpatial == 3)
         {
-            if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
-                         is_same_v<OutDataType, F16>)
-            {
-                add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(op_ptrs);
-            }
-            else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
-                              is_same_v<OutDataType, F32>)
+
+            if constexpr(is_same_v<InLayout, GNDHWC> && is_same_v<WeiLayout, GKZYXC> &&
+                         is_same_v<OutLayout, GNDHWK>)
             {
-                add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+                        op_ptrs);
+                }
             }
-            else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
-                              is_same_v<OutDataType, BF16>)
+            else if constexpr(is_same_v<InLayout, NDHWGC> && is_same_v<WeiLayout, GKZYXC> &&
+                              is_same_v<OutLayout, NDHWGK>)
             {
-                add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
+                if constexpr(is_same_v<InDataType, F16> && is_same_v<WeiDataType, F16> &&
+                             is_same_v<OutDataType, F16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, F32> && is_same_v<WeiDataType, F32> &&
+                                  is_same_v<OutDataType, F32>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+                        op_ptrs);
+                }
+                else if constexpr(is_same_v<InDataType, BF16> && is_same_v<WeiDataType, BF16> &&
+                                  is_same_v<OutDataType, BF16>)
+                {
+                    add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+                        op_ptrs);
+                }
             }
         }
 

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instance.cpp

@@ -2,14 +2,14 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 // Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -26,19 +26,21 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_bf16_instances<GNHWK,
-                                                          GKYXC,
-                                                          Empty_Tuple,
-                                                          GNHWC,
-                                                          ConvBwdDataDefault>{});
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        GNHWK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        GNHWC,
+                                                        ConvBwdDataDefault>{});
     // 2. Filter1x1Stride1Pad0
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_bf16_instances<GNHWK,
-                                                          GKYXC,
-                                                          Empty_Tuple,
-                                                          GNHWC,
-                                                          ConvBwdDataFilter1x1Stride1Pad0>{});
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        GNHWK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        GNHWC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instance.cpp

@@ -2,14 +2,14 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 // Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -26,19 +26,21 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f16_instances(
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f16_instances<GNHWK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         GNHWC,
-                                                         ConvBwdDataDefault>{});
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataDefault>{});
     // 2. Filter1x1Stride1Pad0
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f16_instances<GNHWK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         GNHWC,
-                                                         ConvBwdDataFilter1x1Stride1Pad0>{});
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instance.cpp

@@ -2,14 +2,14 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 // Compilation parameters for out[g, n, hi, wi, c] * wei[g, k, y, x, c] = in[g, n, ho, wo, k]
-void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_gnhwk_gkyxc_gnhwc_f32_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   GNHWK,
                                                                   GKYXC,
@@ -26,19 +26,21 @@ void add_device_grouped_conv2d_bwd_data_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f32_instances<GNHWK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         GNHWC,
-                                                         ConvBwdDataDefault>{});
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataDefault>{});
     // 2. Filter1x1Stride1Pad0
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f32_instances<GNHWK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         GNHWC,
-                                                         ConvBwdDataFilter1x1Stride1Pad0>{});
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       GNHWK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       GNHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_instance.hpp

-// SPDX-License-Identifier: MIT
-// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
-
-#pragma once
-
-#include "ck/ck.hpp"
-#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
-#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_bwd_data_multiple_d_xdl_cshuffle_v1.hpp"
-#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
-
-#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-
-namespace ck {
-namespace tensor_operation {
-namespace device {
-namespace instance {
-
-using BF16 = ck::bhalf_t;
-using F16  = ck::half_t;
-using F32  = float;
-
-using Empty_Tuple = ck::Tuple<>;
-
-template <ck::index_t... Is>
-using S = ck::Sequence<Is...>;
-
-using NHWGC = ck::tensor_layout::convolution::NHWGC;
-using GNHWC = ck::tensor_layout::convolution::GNHWC;
-
-using GKYXC = ck::tensor_layout::convolution::GKYXC;
-
-using NHWGK = ck::tensor_layout::convolution::NHWGK;
-using GNHWK = ck::tensor_layout::convolution::GNHWK;
-
-using PassThrough = ck::tensor_operation::element_wise::PassThrough;
-
-static constexpr auto ConvBwdDataDefault = ConvolutionBackwardDataSpecialization::Default;
-
-static constexpr auto ConvBwdDataFilter1x1Stride1Pad0 =
-    ConvolutionBackwardDataSpecialization::Filter1x1Stride1Pad0;
-
-// f16_f16_f32_f16
-template <typename ALayout,
-          typename BLayout,
-          typename DsLayout,
-          typename ELayout,
-          ConvolutionBackwardDataSpecialization ConvSpec>
-using device_grouped_conv2d_bwd_data_xdl_f16_instances = std::tuple<
-    // clang-format off
-        // ##############################################|    NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
-        // ##############################################| Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
-        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
-        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
-        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
-        
-#ifdef CK_WORKAROUND_SWDEV_3318619
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2,   GNHWK,   GKYXC, Empty_Tuple,   GNHWC,   F16,   F16,     F32,      F16, Empty_Tuple,   F16,  PassThrough,  PassThrough,    PassThrough,  ConvBwdDataDefault,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-#endif
-    // clang-format on
-    >;
-
-// bf16_bf16_f32_bf16
-template <typename ALayout,
-          typename BLayout,
-          typename DsLayout,
-          typename ELayout,
-          ConvolutionBackwardDataSpecialization ConvSpec>
-using device_grouped_conv2d_bwd_data_xdl_bf16_instances = std::tuple<
-    // clang-format off
-        // ##############################################|    NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
-        // ##############################################| Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
-        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
-        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
-        DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   BF16,   BF16,     F32,      BF16, Empty_Tuple,   BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                8>
-        
-#ifdef CK_WORKAROUND_SWDEV_3318619
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 32, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 8>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-        //     DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  BF16,  BF16,     F32,     BF16, Empty_Tuple,  BF16,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>
-#endif
-    // clang-format on
-    >;
-
-// f32_f32_f32_f32
-template <typename ALayout,
-          typename BLayout,
-          typename DsLayout,
-          typename ELayout,
-          ConvolutionBackwardDataSpecialization ConvSpec>
-using device_grouped_conv2d_bwd_data_xdl_f32_instances = std::tuple<
-    // clang-format off
-        // ##############################################|    NDim| ALayout| BLayout|    DsLayout| ELayout| AData| BData| AccData| CShuffle|      DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad|      NumGemmK| Block|  MPer|  NPer|  KPer| AK1| BK1| MPer| NPer|    MXdl|    NXdl|    ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds|    BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl|   CDEBlockTransfer| CDEBlockTransfer|
-        // ##############################################| Spatial|        |        |            |        |  Type|  Type|    Type| DataType|        Type|  Type|    Operation|    Operation|      Operation|  DataSpecialization| GemmM| GemmN| PrefetchStage|  Size| Block| Block| Block|    |    |  XDL|  XDL| PerWave| PerWave|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraM|     ThreadCluster|  ThreadCluster| SrcAccessOrder|   SrcVectorDim|      SrcScalar|      DstScalar|    ExtraN|      PerWave|      PerWave|  _MBlock_MPerBlock|  ScalarPerVector|
-        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        | Lengths_AK0_M_AK1|   ArrangeOrder|               |               |      PerVector|  PerVector_AK1|          | Lengths_BK0_N_BK1|   ArrangeOrder|               |               |      PerVector|  PerVector_BK1|          |   PerShuffle|   PerShuffle|  _NBlock_NPerBlock|       _NPerBlock|
-        // ##############################################|        |        |        |            |        |      |      |        |         |            |      |             |             |               |                    |      |      |              |      |      |      |      |    |    |     |     |        |        |                  |               |               |               |               |               |          |                  |               |               |               |               |               |          |             |             |                   |                 |
-         DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,  F32,  F32,      F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   2,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<0, 2, 1>,     S<0, 2, 1>,              1,              4,              2,         0,            1,            1,     S<1, 32, 1, 8>,                4>
-
-#ifdef CK_WORKAROUND_SWDEV_3318619
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   256,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   256,    32,   8,   8,   32,   32,       2,       4,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,   128,    32,   8,   8,   32,   32,       4,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    64,   128,    32,   8,   8,   32,   32,       2,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    64,    32,   8,   8,   32,   32,       2,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,   128,    64,    32,   8,   8,   32,   32,       2,       1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   256,    64,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 64, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,   128,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 32, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,   128,    32,   128,    32,   8,   8,   32,   32,       1,       2,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 32, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 8>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    64,    32,    32,   8,   8,   32,   32,       2,       1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              4,              4,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              4,              4,         1,            1,            1,     S<1, 16, 1, 4>,                4>,
-        // DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1<       2, ALayout, BLayout,    DsLayout, ELayout,   F32,   F32,     F32,      F32, Empty_Tuple,   F32,  PassThrough,  PassThrough,    PassThrough,            ConvSpec,  true,  true,             1,    64,    32,    64,    32,   8,   8,   32,   32,       1,       2,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              2,              8,              8,         1,       S<4, 16, 1>,     S<1, 0, 2>,     S<1, 0, 2>,              1,              8,              8,         1,            1,            1,     S<1, 16, 1, 4>,                8>,
-#endif
-    // clang-format on
-    >;
-
-} // namespace instance
-} // namespace device
-} // namespace tensor_operation
-} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instance.cpp

@@ -2,14 +2,14 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 // Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_bf16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -26,19 +26,21 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_bf16_instances<NHWGK,
-                                                          GKYXC,
-                                                          Empty_Tuple,
-                                                          NHWGC,
-                                                          ConvBwdDataDefault>{});
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        NHWGK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        NHWGC,
+                                                        ConvBwdDataDefault>{});
     // 2. Filter1x1Stride1Pad0
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_bf16_instances<NHWGK,
-                                                          GKYXC,
-                                                          Empty_Tuple,
-                                                          NHWGC,
-                                                          ConvBwdDataFilter1x1Stride1Pad0>{});
+        device_grouped_conv_bwd_data_xdl_bf16_instances<2,
+                                                        NHWGK,
+                                                        GKYXC,
+                                                        Empty_Tuple,
+                                                        NHWGC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp

@@ -2,14 +2,14 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 // Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f16_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -26,19 +26,21 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f16_instances<NHWGK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         NHWGC,
-                                                         ConvBwdDataDefault>{});
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataDefault>{});
     // 2. Filter1x1Stride1Pad0
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f16_instances<NHWGK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         NHWGC,
-                                                         ConvBwdDataFilter1x1Stride1Pad0>{});
+        device_grouped_conv_bwd_data_xdl_f16_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv2d_bwd_data/device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instance.cpp

@@ -2,14 +2,14 @@
 // Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
 
 #include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
-#include "device_grouped_conv2d_bwd_data_xdl_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
 
 namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
 // Compilation parameters for out[n, hi, wi, g, c] * wei[g, k, y, x, c] = in[n, ho, wo, g, k]
-void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
+void add_device_grouped_conv2d_bwd_data_xdl_nhwgk_gkyxc_nhwgc_f32_instances(
     std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<2,
                                                                   NHWGK,
                                                                   GKYXC,
@@ -26,19 +26,21 @@ void add_device_grouped_conv2d_bwd_data_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
     // 1. Default
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f32_instances<NHWGK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         NHWGC,
-                                                         ConvBwdDataDefault>{});
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataDefault>{});
     // 2. Filter1x1Stride1Pad0
     add_device_operation_instances(
         instances,
-        device_grouped_conv2d_bwd_data_xdl_f32_instances<NHWGK,
-                                                         GKYXC,
-                                                         Empty_Tuple,
-                                                         NHWGC,
-                                                         ConvBwdDataFilter1x1Stride1Pad0>{});
+        device_grouped_conv_bwd_data_xdl_f32_instances<2,
+                                                       NHWGK,
+                                                       GKYXC,
+                                                       Empty_Tuple,
+                                                       NHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
 }
 
 } // namespace instance

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/CMakeLists.txt

+add_instance_library(device_grouped_conv3d_bwd_data_instance
+   device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp
+   device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp
+)

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = in[g, n, do, ho,
+// wo, k]
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        GNDHWK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        GNDHWC,
+                                                        ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        GNDHWK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        GNDHWC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = in[g, n, do, ho,
+// wo, k]
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_gndhwc_gkzyxc_gndhwk_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[g, n, di, hi, wi, c] * wei[g, k, z, y, x, c] = in[g, n, do, ho,
+// wo, k]
+void add_device_grouped_conv3d_bwd_data_xdl_gndhwk_gkzyxc_gndhwc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  GNDHWK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  GNDHWC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       GNDHWK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       GNDHWC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_bf16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
+// g, k]
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_bf16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  BF16,
+                                                                  BF16,
+                                                                  Empty_Tuple,
+                                                                  BF16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        NDHWGK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        NDHWGC,
+                                                        ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_bf16_instances<3,
+                                                        NDHWGK,
+                                                        GKZYXC,
+                                                        Empty_Tuple,
+                                                        NDHWGC,
+                                                        ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f16_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
+// g, k]
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f16_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F16,
+                                                                  F16,
+                                                                  Empty_Tuple,
+                                                                  F16,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f16_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

library/src/tensor_operation_instance/gpu/grouped_conv3d_bwd_data/device_grouped_conv3d_bwd_data_xdl_ndhwgc_gkzyxc_ndhwgk_f32_instance.cpp

+// SPDX-License-Identifier: MIT
+// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
+
+#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp"
+#include "ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp"
+
+namespace ck {
+namespace tensor_operation {
+namespace device {
+namespace instance {
+// Compilation parameters for out[n, di, hi, wi, g, c] * wei[g, k, z, y, x, c] = in[n, do, ho, wo,
+// g, k]
+void add_device_grouped_conv3d_bwd_data_xdl_ndhwgk_gkzyxc_ndhwgc_f32_instances(
+    std::vector<std::unique_ptr<DeviceGroupedConvBwdDataMultipleD<3,
+                                                                  NDHWGK,
+                                                                  GKZYXC,
+                                                                  Empty_Tuple,
+                                                                  NDHWGC,
+                                                                  F32,
+                                                                  F32,
+                                                                  Empty_Tuple,
+                                                                  F32,
+                                                                  PassThrough,
+                                                                  PassThrough,
+                                                                  PassThrough>>>& instances)
+{
+    // 1. Default
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataDefault>{});
+    // 2. Filter1x1Stride1Pad0
+    add_device_operation_instances(
+        instances,
+        device_grouped_conv_bwd_data_xdl_f32_instances<3,
+                                                       NDHWGK,
+                                                       GKZYXC,
+                                                       Empty_Tuple,
+                                                       NDHWGC,
+                                                       ConvBwdDataFilter1x1Stride1Pad0>{});
+}
+
+} // namespace instance
+} // namespace device
+} // namespace tensor_operation
+} // namespace ck

profiler/src/CMakeLists.txt

@@ -81,4 +81,5 @@ target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_contraction_scale_in
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_pool_fwd_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_batched_gemm_multi_d_instance)
 target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv2d_bwd_data_instance)
+target_link_libraries(${PROFILER_EXECUTABLE} PRIVATE device_grouped_conv3d_bwd_data_instance)
 rocm_install(TARGETS ${PROFILER_EXECUTABLE} COMPONENT profiler)