Merge pull request #43 from ROCmSoftwarePlatform/develop

Merge develop into master

composable_kernel/include/problem_transform/transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk.hpp

@@ -21,8 +21,8 @@ template <typename... Wei,
           typename ConvDilations,
           typename InLeftPads,
           typename InRightPads,
-          index_t IYTildaValue,
-          index_t IXTildaValue,
+          typename IYTilda,
+          typename IXTilda,
           index_t GemmK1Value>
 __host__ __device__ constexpr auto
 transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
@@ -33,8 +33,8 @@ transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
     const ConvDilations& conv_dilations,
     const InLeftPads& in_left_pads,
     const InRightPads& in_right_pads,
-    Number<IYTildaValue>,
-    Number<IXTildaValue>,
+    IYTilda i_ytilda,
+    IXTilda i_xtilda,
     Number<GemmK1Value>)
 {
     constexpr auto I0 = Number<0>{};
@@ -42,9 +42,7 @@ transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
     constexpr auto I2 = Number<2>{};
     constexpr auto I3 = Number<3>{};
 
-    constexpr auto GemmK1  = Number<GemmK1Value>{};
-    constexpr auto IYTilda = Number<IYTildaValue>{};
-    constexpr auto IXTilda = Number<IXTildaValue>{};
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
 
     const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
     const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
@@ -98,8 +96,8 @@ transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
     const auto WTildaSlice = IWTildaSliceEnd - IWTildaSliceBegin;
 
     // GemmK is different for each GEMM
-    const auto YDotSlice = math::integer_divide_ceil(Y - IYTilda, YTilda);
-    const auto XDotSlice = math::integer_divide_ceil(X - IXTilda, XTilda);
+    const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilda, YTilda);
+    const auto XDotSlice = math::integer_divide_ceil(X - i_xtilda, XTilda);
 
     const auto K1 = GemmK1;
     const auto K0 = K / K1;
@@ -183,8 +181,8 @@ transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
                                     make_tuple(make_unmerge_transform(make_tuple(K0, K1)),
                                                make_slice_transform(YDot, I0, YDotSlice),
                                                make_slice_transform(XDot, I0, XDotSlice),
-                                               make_freeze_transform(IYTilda),
-                                               make_freeze_transform(IXTilda),
+                                               make_freeze_transform(i_ytilda),
+                                               make_freeze_transform(i_xtilda),
                                                make_pass_through_transform(C)),
                                     make_tuple(Sequence<0>{},
                                                Sequence<1>{},
@@ -241,9 +239,9 @@ transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
     const auto in_n_htildaslice_wtildaslice_c_grid_desc = transform_tensor_descriptor(
         in_n_ytilda_htilda_xtilda_wtilda_c_grid_desc,
         make_tuple(make_pass_through_transform(N),
-                   make_freeze_transform(IYTilda),
+                   make_freeze_transform(i_ytilda),
                    make_slice_transform(HTilda, IHTildaSliceBegin, HTildaSlice),
-                   make_freeze_transform(IXTilda),
+                   make_freeze_transform(i_xtilda),
                    make_slice_transform(WTilda, IWTildaSliceBegin, WTildaSlice),
                    make_pass_through_transform(C)),
         make_tuple(Sequence<0>{},
@@ -271,5 +269,84 @@ transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk(
                       in_gemmm_gemmn_grid_desc);
 }
 
+// A: out
+// B: wei
+// C: in
+// Number of GEMMs = 1
+// GemmM = N * Ho * Wo
+// GemmN = C
+// GemmK = K
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_data_convolution_into_gemm_v4r1r2_nhwc_kyxc_nhwk_1x1(
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const TensorDescriptor<Wei...>& /* wei_k_y_x_c_grid_desc */,
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const ConvStrides& conv_strides,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto K1 = GemmK1;
+    const auto K0 = K / K1;
+
+    // A: output tensor
+    const auto out_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+                                    make_tuple(make_pass_through_transform(N * Ho * Wo),
+                                               make_unmerge_transform(make_tuple(K0, K1))),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0, 2>{}));
+
+    // B: weight tensor
+    const auto wei_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C)),
+        make_tuple(make_unmerge_transform(make_tuple(K0, K1)), make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // 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_gemmm_gemmn_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>{}));
+
+    return make_tuple(out_gemmk0_gemmm_gemmk1_grid_desc,
+                      wei_gemmk0_gemmn_gemmk1_grid_desc,
+                      in_gemmm_gemmn_grid_desc);
+}
+
 } // namespace ck
 #endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R2_ATOMIC_NCHW_KCYX_NKHW_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// GemmM = K
+// GemmK = N * Ho * Wo
+// GemmN = C * Y * X
+template <typename... Wei,
+          typename... In,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r2_atomic_nchw_kcyx_nkhw_pad(
+    const TensorDescriptor<Wei...>& wei_k_c_y_x_grid_desc,
+    const TensorDescriptor<In...>& in_n_c_hi_wi_grid_desc,
+    const TensorDescriptor<Out...>& out_n_k_ho_wo_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_c_hi_wi_grid_desc.GetLength(I0);
+    const auto C = in_n_c_hi_wi_grid_desc.GetLength(I1);
+    const auto K = out_n_k_ho_wo_grid_desc.GetLength(I1);
+
+    const auto Hi = in_n_c_hi_wi_grid_desc.GetLength(I2);
+    const auto Wi = in_n_c_hi_wi_grid_desc.GetLength(I3);
+
+    const auto Ho = out_n_k_ho_wo_grid_desc.GetLength(I2);
+    const auto Wo = out_n_k_ho_wo_grid_desc.GetLength(I3);
+
+    const auto Y = wei_k_c_y_x_grid_desc.GetLength(I2);
+    const auto X = wei_k_c_y_x_grid_desc.GetLength(I3);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = K;
+    const auto GemmN      = C * Y * X;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: output tensor
+    const auto out_gemmktotal_gemmm_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N, K, Ho * Wo)),
+        make_tuple(make_pass_through_transform(K), make_merge_transform(make_tuple(N, Ho * Wo))),
+        make_tuple(Sequence<1>{}, Sequence<0, 2>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: input tensor
+    const auto in_n_c_hip_wip_grid_desc = transform_tensor_descriptor(
+        in_n_c_hi_wi_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_pad_transform(Hi, InLeftPadH, InRightPadH),
+                   make_pad_transform(Wi, InLeftPadW, InRightPadW)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
+
+    const auto in_n_c_y_ho_x_wo_grid_desc = transform_tensor_descriptor(
+        in_n_c_hip_wip_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_pass_through_transform(C),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW))),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2, 3>{}, Sequence<4, 5>{}));
+
+    const auto in_gemmktotal_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_c_y_ho_x_wo_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(C, Y, X)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 2, 4>{}, Sequence<0, 3, 5>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, C * Y * X)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(C * Y * X)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_ATOMIC_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r4_atomic_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = Y * X * C;
+    const auto GemmN      = K;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: 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_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmktotal_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: output tensor
+    const auto out_gemmktotal_gemmn_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+    const auto out_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(in_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R4_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: in
+// B: wei
+// C: out
+// GemmM = N * Ho * Wo
+// GemmN = K
+// GemmK = Y * X * C
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r4_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM  = Y * X * C;
+    const auto GemmN  = K;
+    const auto GemmK  = N * Ho * Wo;
+    const auto GemmK0 = GemmK / GemmK1;
+
+    // A: 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_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmk_gemmm_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmk0_gemmm_gemmk1_grid_desc =
+        transform_tensor_descriptor(in_gemmk_gemmm_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmM)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // B: output tensor
+    const auto out_gemmk_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K)),
+        make_tuple(make_pass_through_transform(N * Ho * Wo), make_pass_through_transform(K)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmk0_gemmn_gemmk1_grid_desc =
+        transform_tensor_descriptor(out_gemmk_gemmn_grid_desc,
+                                    make_tuple(make_unmerge_transform(make_tuple(GemmK0, GemmK1)),
+                                               make_pass_through_transform(GemmN)),
+                                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc = transform_tensor_descriptor(
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C)),
+        make_tuple(make_pass_through_transform(K), make_pass_through_transform(Y * X * C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    return make_tuple(in_gemmk0_gemmm_gemmk1_grid_desc,
+                      out_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/problem_transform/transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk.hpp

+#ifndef CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
+#define CK_TRANSFORM_BACKWARD_WEIGHT_CONVOLUTION_INTO_GEMM_V4R4R5_NHWC_KYXC_NHWK_HPP
+
+#include "common_header.hpp"
+#include "tensor_descriptor.hpp"
+#include "tensor_descriptor_helper.hpp"
+
+namespace ck {
+
+// A: out
+// B: in
+// C: wei
+// GemmM = K
+// GemmN = Y * X * C
+// GemmKTotal = N * Ho * Wo
+template <typename... In,
+          typename... Wei,
+          typename... Out,
+          typename ConvStrides,
+          typename ConvDilations,
+          typename InLeftPads,
+          typename InRightPads,
+          index_t GemmK1Value,
+          typename GemmKBatchType,
+          typename GemmKPadType>
+__host__ __device__ constexpr auto
+transform_backward_weight_convolution_into_gemm_v4r4r5_nhwc_kyxc_nhwk_pad(
+    const TensorDescriptor<In...>& in_n_hi_wi_c_grid_desc,
+    const TensorDescriptor<Wei...>& wei_k_y_x_c_grid_desc,
+    const TensorDescriptor<Out...>& out_n_ho_wo_k_grid_desc,
+    const ConvStrides& conv_strides,
+    const ConvDilations& conv_dilations,
+    const InLeftPads& in_left_pads,
+    const InRightPads& in_right_pads,
+    Number<GemmK1Value>,
+    GemmKBatchType GemmKBatch,
+    GemmKPadType GemmKPad)
+{
+    constexpr auto I0 = Number<0>{};
+    constexpr auto I1 = Number<1>{};
+    constexpr auto I2 = Number<2>{};
+    constexpr auto I3 = Number<3>{};
+
+    constexpr auto GemmK1 = Number<GemmK1Value>{};
+
+    const auto N = in_n_hi_wi_c_grid_desc.GetLength(I0);
+    const auto C = in_n_hi_wi_c_grid_desc.GetLength(I3);
+    const auto K = out_n_ho_wo_k_grid_desc.GetLength(I3);
+
+    const auto Hi = in_n_hi_wi_c_grid_desc.GetLength(I1);
+    const auto Wi = in_n_hi_wi_c_grid_desc.GetLength(I2);
+
+    const auto Ho = out_n_ho_wo_k_grid_desc.GetLength(I1);
+    const auto Wo = out_n_ho_wo_k_grid_desc.GetLength(I2);
+
+    const auto Y = wei_k_y_x_c_grid_desc.GetLength(I1);
+    const auto X = wei_k_y_x_c_grid_desc.GetLength(I2);
+
+    const auto ConvStrideH = conv_strides[I0];
+    const auto ConvStrideW = conv_strides[I1];
+
+    const auto ConvDilationH = conv_dilations[I0];
+    const auto ConvDilationW = conv_dilations[I1];
+
+    const auto InLeftPadH = in_left_pads[I0];
+    const auto InLeftPadW = in_left_pads[I1];
+
+    const auto InRightPadH = in_right_pads[I0];
+    const auto InRightPadW = in_right_pads[I1];
+
+    const auto GemmM      = K;
+    const auto GemmN      = Y * X * C;
+    const auto GemmKTotal = N * Ho * Wo;
+    const index_t GemmK0  = GemmKPad / (GemmKBatch * GemmK1);
+
+    // A: output tensor
+    const auto out_gemmktotal_gemmm_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(N * Ho * Wo, K));
+
+    const auto out_gemmkpad_gemmm_grid_desc = transform_tensor_descriptor(
+        out_gemmktotal_gemmm_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc = transform_tensor_descriptor(
+        out_gemmkpad_gemmm_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmM)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // B: 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_y_ho_x_wo_c_grid_desc = transform_tensor_descriptor(
+        in_n_hip_wip_c_grid_desc,
+        make_tuple(make_pass_through_transform(N),
+                   make_embed_transform(make_tuple(Y, Ho), make_tuple(ConvDilationH, ConvStrideH)),
+                   make_embed_transform(make_tuple(X, Wo), make_tuple(ConvDilationW, ConvStrideW)),
+                   make_pass_through_transform(C)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
+        make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
+
+    const auto in_gemmktotal_gemmn_grid_desc =
+        transform_tensor_descriptor(in_n_y_ho_x_wo_c_grid_desc,
+                                    make_tuple(make_merge_transform(make_tuple(Y, X, C)),
+                                               make_merge_transform(make_tuple(N, Ho, Wo))),
+                                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                                    make_tuple(Sequence<1>{}, Sequence<0>{}));
+
+    const auto in_gemmkpad_gemmn_grid_desc = transform_tensor_descriptor(
+        in_gemmktotal_gemmn_grid_desc,
+        make_tuple(make_right_pad_transform(GemmKTotal, GemmKPad - GemmKTotal),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0>{}, Sequence<1>{}));
+
+    const auto in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc = transform_tensor_descriptor(
+        in_gemmkpad_gemmn_grid_desc,
+        make_tuple(make_unmerge_transform(make_tuple(GemmKBatch, GemmK0, GemmK1)),
+                   make_pass_through_transform(GemmN)),
+        make_tuple(Sequence<0>{}, Sequence<1>{}),
+        make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
+
+    // C: weight tensor
+    const auto wei_gemmm_gemmn_grid_desc =
+        make_naive_tensor_descriptor_packed(make_tuple(K, Y * X * C));
+
+    return make_tuple(out_gemmkbatch_gemmk0_gemmm_gemmk1_grid_desc,
+                      in_gemmkbatch_gemmk0_gemmn_gemmk1_grid_desc,
+                      wei_gemmm_gemmn_grid_desc);
+}
+
+} // namespace ck
+#endif

composable_kernel/include/tensor_description/multi_index_transform_helper.hpp

@@ -31,7 +31,7 @@ __host__ __device__ constexpr auto make_left_pad_transform(
     return LeftPad<LowLength, LeftPadLength, SkipIsValidCheck>{low_length, left_pad};
 }
 
-template <typename LowLength, typename RightPadLength, bool SkipIsValidCheck>
+template <typename LowLength, typename RightPadLength, bool SkipIsValidCheck = false>
 __host__ __device__ constexpr auto make_right_pad_transform(
     const LowLength& low_length,
     const RightPadLength& right_pad,

composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp

@@ -10,6 +10,7 @@ namespace ck {
 
 template <index_t BlockSize,
           typename FloatAB,
+          typename FloatAcc,
           typename AK0MK1BlockDesc,
           typename BK0NK1BlockDesc,
           index_t MPerXDL,
@@ -29,14 +30,18 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     static constexpr index_t MPerBlock = AK0MK1BlockDesc{}.GetLength(I1);
     static constexpr index_t NPerBlock = BK0NK1BlockDesc{}.GetLength(I1);
 
-    static constexpr index_t K0        = BK0NK1BlockDesc{}.GetLength(I0);
-    static constexpr index_t KPerBlock = K0;
+    static constexpr index_t K0 = BK0NK1BlockDesc{}.GetLength(I0);
 
     static constexpr auto xdlops_gemm = XdlopsGemm<FloatAB, MPerXDL, NPerXDL, K1>{};
 
     static constexpr index_t MWaves = MPerBlock / (MRepeat * MPerXDL);
     static constexpr index_t NWaves = NPerBlock / (NRepeat * NPerXDL);
 
+    StaticBufferV2<AddressSpaceEnum_t::Vgpr, vector_type<FloatAcc, 16>, MRepeat * NRepeat, true>
+        c_thread_buf_;
+
+    __host__ __device__ constexpr auto& GetCThreadBuffer() { return c_thread_buf_; }
+
     __device__ static auto GetWaveIdx()
     {
         const index_t thread_id = get_thread_local_1d_id();
@@ -162,7 +167,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     {
         return transform_tensor_descriptor(
             AK0MK1BlockDesc{},
-            make_tuple(make_pass_through_transform(Number<KPerBlock>{}),
+            make_tuple(make_pass_through_transform(Number<K0>{}),
                        make_unmerge_transform(
                            make_tuple(Number<MRepeat>{}, Number<MWaves>{}, Number<MPerXDL>{})),
                        make_pass_through_transform(Number<K1>{})),
@@ -174,7 +179,7 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
     {
         return transform_tensor_descriptor(
             BK0NK1BlockDesc{},
-            make_tuple(make_pass_through_transform(Number<KPerBlock>{}),
+            make_tuple(make_pass_through_transform(Number<K0>{}),
                        make_unmerge_transform(
                            make_tuple(Number<NRepeat>{}, Number<NWaves>{}, Number<NPerXDL>{})),
                        make_pass_through_transform(Number<K1>{})),
@@ -195,48 +200,43 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
         auto b_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAB>(
             b_thread_desc_.GetElementSpaceSize());
 
-        vector_type<FloatAB, K1> a_thread_vec;
-
-        vector_type<FloatAB, K1> b_thread_vec;
-
-        static_for<0, KPerBlock, xdlops_gemm.KPerXdlops / xdlops_gemm.KPerThread>{}([&](auto k0) {
+        static_for<0, MRepeat, 1>{}([&](auto m0) {
             // read A
             a_thread_copy_.Run(a_k0_m0_m1_m2_k1_block_desc,
-                               make_tuple(k0, I0, I0, I0, I0),
+                               make_tuple(I0, m0, I0, I0, I0),
                                a_block_buf,
                                a_thread_desc_,
                                make_tuple(I0, I0, I0, I0, I0),
                                a_thread_buf);
 
-            // read B
-            b_thread_copy_.Run(b_k0_n0_n1_n2_k1_block_desc,
-                               make_tuple(k0, I0, I0, I0, I0),
-                               b_block_buf,
-                               b_thread_desc_,
-                               make_tuple(I0, I0, I0, I0, I0),
-                               b_thread_buf);
+            static_for<0, NRepeat, 1>{}([&](auto n0) {
+                // read B
+                b_thread_copy_.Run(b_k0_n0_n1_n2_k1_block_desc,
+                                   make_tuple(I0, n0, I0, I0, I0),
+                                   b_block_buf,
+                                   b_thread_desc_,
+                                   make_tuple(I0, I0, I0, I0, I0),
+                                   b_thread_buf);
 
-            using mfma_input_type = typename vector_type<FloatAB, xdlops_gemm.KPerThread>::type;
+                static_for<0, K0, xdlops_gemm.K0PerXdlops>{}([&](auto k0) {
+                    vector_type<FloatAB, K1> a_thread_vec;
+                    vector_type<FloatAB, K1> b_thread_vec;
 
-            static_for<0, MRepeat, 1>{}([&](auto m0) {
-                static_for<0, NRepeat, 1>{}([&](auto n0) {
                     static_for<0, K1, 1>{}([&](auto i) {
                         a_thread_vec.template AsType<FloatAB>()(i) = a_thread_buf
-                            [Number<a_thread_desc_.CalculateOffset(make_tuple(0, m0, 0, 0, i))>{}];
-                    });
-
-                    static_for<0, K1, 1>{}([&](auto i) {
+                            [Number<a_thread_desc_.CalculateOffset(make_tuple(k0, 0, 0, 0, i))>{}];
                         b_thread_vec.template AsType<FloatAB>()(i) = b_thread_buf
-                            [Number<b_thread_desc_.CalculateOffset(make_tuple(0, n0, 0, 0, i))>{}];
+                            [Number<b_thread_desc_.CalculateOffset(make_tuple(k0, 0, 0, 0, i))>{}];
                     });
 
-                    constexpr index_t c_offset =
-                        c_thread_desc_.CalculateOffset(make_tuple(m0, n0, 0));
+                    using mfma_input_type =
+                        typename vector_type<FloatAB, xdlops_gemm.K1PerXdlops>::type;
+
+                    constexpr index_t c_offset = c_thread_desc_.CalculateOffset(make_tuple(m0, n0));
 
-                    xdlops_gemm.template Run<c_offset>(
-                        a_thread_vec.template AsType<mfma_input_type>(),
-                        b_thread_vec.template AsType<mfma_input_type>(),
-                        c_thread_buf);
+                    xdlops_gemm.template Run(a_thread_vec.template AsType<mfma_input_type>(),
+                                             b_thread_vec.template AsType<mfma_input_type>(),
+                                             c_thread_buf.GetVector(Number<c_offset>{}));
                 });
             });
         });
@@ -244,35 +244,35 @@ struct BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1
 
     private:
     // A[K, M]
-    static constexpr auto a_thread_desc_ = make_naive_tensor_descriptor_packed(
-        make_tuple(I1, Number<MRepeat>{}, I1, I1, Number<K1>{}));
+    static constexpr auto a_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{}));
 
     // B[K, N]
-    static constexpr auto b_thread_desc_ = make_naive_tensor_descriptor_packed(
-        make_tuple(I1, Number<NRepeat>{}, I1, I1, Number<K1>{}));
+    static constexpr auto b_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<K0>{}, I1, I1, I1, Number<K1>{}));
 
-    static constexpr auto c_thread_desc_ = make_naive_tensor_descriptor_packed(
-        make_tuple(Number<MRepeat>{}, Number<NRepeat>{}, Number<xdlops_gemm.GetNumXdlops()>{}));
+    static constexpr auto c_thread_desc_ =
+        make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
 
     using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
                                                          FloatAB,
                                                          decltype(a_k0_m0_m1_m2_k1_block_desc),
                                                          decltype(a_thread_desc_),
-                                                         Sequence<1, MRepeat, 1, 1, K1>,
+                                                         Sequence<K0, 1, 1, 1, K1>,
                                                          Sequence<0, 1, 2, 3, 4>,
                                                          4,
                                                          K1,
-                                                         1>;
+                                                         K1>;
 
     using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
                                                          FloatAB,
                                                          decltype(b_k0_n0_n1_n2_k1_block_desc),
                                                          decltype(b_thread_desc_),
-                                                         Sequence<1, NRepeat, 1, 1, K1>,
+                                                         Sequence<K0, 1, 1, 1, K1>,
                                                          Sequence<0, 1, 2, 3, 4>,
                                                          4,
                                                          K1,
-                                                         1>;
+                                                         K1>;
 
     AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
     BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};

composable_kernel/include/tensor_operation/gridwise_generic_2d_reduction_multiblock.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#ifndef CK_GRIDWISE_GENERIC_2D_REDUCTION_MULTIBLOCK_HPP
+#define CK_GRIDWISE_GENERIC_2D_REDUCTION_MULTIBLOCK_HPP
+
+#include "reduction_common.hpp"
+#include "reduction_operator.hpp"
+#include "reduction_functions_blockwise.hpp"
+
+#include "blockwise_tensor_slice_transfer.hpp"
+
+namespace ck {
+
+template <index_t BlockSize,
+          typename srcDataType,
+          typename dstDataType, // not used together with the beta input
+          typename compType,
+          typename src2dDescType,
+          typename dst1dDescType,
+          ReduceTensorOp_t op,
+          NanPropagation_t nanPropaOpt,
+          ReduceTensorIndices_t reduceIndicesOpt,
+          index_t GredAccessesPerThreadInBlock>
+struct GridwiseReduction_xy_to_x_multiblock
+{
+    using opReduce       = typename reduce_binary_operator<compType, op>::opType;
+    using preUnaryOpType = typename reduce_unary_operator<compType, op, true, false>::preUnaryOp;
+    using posUnaryOpType = typename reduce_unary_operator<compType, op, true, false>::posUnaryOp;
+
+    static constexpr auto buffer2dDesc = make_naive_tensor_descriptor_packed(
+        make_tuple(Number<GredAccessesPerThreadInBlock>{}, Number<BlockSize>{}));
+    using blockwise_reduce =
+        BlockwiseReduction_2d_block_buffer<decltype(buffer2dDesc), true, opReduce, nanPropaOpt>;
+
+    static constexpr index_t BlockBufferSize = buffer2dDesc.GetElementSize();
+
+    static constexpr auto I0 = Number<0>{};
+
+    template <int RunId>
+    __device__ static void Run(const src2dDescType& src2dDesc,
+                               const dst1dDescType& dst1dDesc,
+                               int origReduceLen,
+                               int BlkGroupSize,
+                               srcDataType alpha,
+                               const srcDataType* const __restrict__ p_src_global,
+                               dstDataType beta,
+                               srcDataType* const __restrict__ ws_values_global,
+                               int* const __restrict__ ws_indices_global);
+
+    template <>
+    __device__ static void Run<1>(const src2dDescType& src2dDesc,
+                                  const dst1dDescType& dst1dDesc,
+                                  int origReduceLen,
+                                  int BlkGroupSize,
+                                  srcDataType alpha,
+                                  const srcDataType* const __restrict__ p_src_global,
+                                  dstDataType beta,
+                                  srcDataType* const __restrict__ ws_values_global,
+                                  int* const __restrict__ ws_indices_global)
+    {
+        (void)ws_indices_global;
+
+        (void)alpha; // unused
+        (void)beta;  // unused
+
+        const auto zeroVal = opReduce::GetReductionZeroVal();
+
+        // LDS
+        __shared__ compType p_in_block_buffer[BlockBufferSize];
+
+        const auto src_global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_src_global, src2dDesc.GetElementSpaceSize(), type_convert<srcDataType>{}(zeroVal));
+        auto workspace_global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            ws_values_global, dst1dDesc.GetLength(I0) * BlkGroupSize);
+
+        auto in_block_buf =
+            make_dynamic_buffer<AddressSpaceEnum_t::Lds>(p_in_block_buffer, BlockBufferSize);
+        StaticBuffer<AddressSpaceEnum_t::Vgpr, compType, 1, true> accuValue_buf;
+
+        accuValue_buf(I0) = zeroVal;
+
+        const auto toReduceLength = src2dDesc.GetLength(Number<1>{});
+        const int divider         = origReduceLen;
+
+        const preUnaryOpType preUnaryOp(divider);
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / BlkGroupSize;
+        const index_t block_local_id  = block_global_id % BlkGroupSize;
+
+        const index_t reduceSizePerBlock =
+            (((toReduceLength + BlkGroupSize - 1) / BlkGroupSize + BlockBufferSize - 1) /
+             BlockBufferSize) *
+            BlockBufferSize;
+
+        constexpr auto in_block_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<1>{}, Number<BlockSize * GredAccessesPerThreadInBlock>{}));
+
+        using ThreadSliceLengths   = Sequence<1, GredAccessesPerThreadInBlock>;
+        using ThreadClusterLengths = Sequence<1, BlockSize>;
+
+        auto blockwise_src_load = BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                                                  InMemoryDataOperationEnum_t::Set,
+                                                                  Sequence<1, BlockBufferSize>,
+                                                                  ThreadSliceLengths,
+                                                                  ThreadClusterLengths,
+                                                                  Sequence<0, 1>,
+                                                                  srcDataType,
+                                                                  compType,
+                                                                  src2dDescType,
+                                                                  decltype(in_block_desc),
+                                                                  Sequence<0, 1>,
+                                                                  Sequence<0, 1>,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  false,
+                                                                  true>(
+            src2dDesc,
+            make_multi_index(blkgroup_id, block_local_id * reduceSizePerBlock),
+            in_block_desc,
+            make_multi_index(0, 0));
+
+        constexpr auto in_block_copy_step = make_multi_index(0, BlockBufferSize);
+
+        const index_t toReduceBlocks = (reduceSizePerBlock + BlockSize - 1) / BlockSize;
+
+        for(index_t reducedBlocks = 0; reducedBlocks < toReduceBlocks;
+            reducedBlocks += GredAccessesPerThreadInBlock)
+        {
+            blockwise_src_load.RunRead(src2dDesc, src_global_buf);
+            blockwise_src_load.RunWrite(in_block_desc, in_block_buf);
+            __syncthreads();
+
+            // do element-wise pre-reduction operation
+            blockwise_reduce::operate_on_elements(preUnaryOp, in_block_buf);
+
+            index_t BlocksInOneOp = (reducedBlocks < toReduceBlocks - GredAccessesPerThreadInBlock)
+                                        ? GredAccessesPerThreadInBlock
+                                        : toReduceBlocks - reducedBlocks;
+            blockwise_reduce::Reduce(in_block_buf, BlocksInOneOp, accuValue_buf(I0));
+
+            blockwise_src_load.MoveSrcSliceWindow(src2dDesc, in_block_copy_step);
+        }
+
+        constexpr auto ReducedDataDesc =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
+
+        const auto workspace_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(dst1dDesc.GetLength(I0) * BlkGroupSize));
+
+        // The first thread in the block stores the reduced result to the global location
+        // representing the block
+        if(thread_local_id == 0)
+        {
+            auto threadwise_workspace_store =
+                ThreadwiseTensorSliceTransfer_v1r3<compType,
+                                                   srcDataType,
+                                                   decltype(ReducedDataDesc),
+                                                   decltype(workspace_desc),
+                                                   Sequence<1>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   1,
+                                                   InMemoryDataOperationEnum_t::Set,
+                                                   1,
+                                                   true>(workspace_desc,
+                                                         make_multi_index(block_global_id));
+
+            threadwise_workspace_store.Run(ReducedDataDesc,
+                                           make_tuple(I0),
+                                           accuValue_buf,
+                                           workspace_desc,
+                                           workspace_global_buf);
+        }
+    };
+
+    template <>
+    __device__ static void Run<2>(const src2dDescType& src2dDesc,
+                                  const dst1dDescType& dst1dDesc,
+                                  int origReduceLen,
+                                  int BlkGroupSize,
+                                  srcDataType alpha,
+                                  const srcDataType* const __restrict__ p_src_global,
+                                  dstDataType beta,
+                                  srcDataType* const __restrict__ ws_values_global,
+                                  int* const __restrict__ ws_indices_global)
+    {
+        (void)alpha; // unused
+        (void)beta;  // unused
+
+        const auto zeroVal = opReduce::GetReductionZeroVal();
+
+        // LDS
+        __shared__ compType p_in_block_values_buffer[BlockBufferSize];
+        __shared__ int p_in_block_indices_buffer[BlockBufferSize];
+
+        const auto src_global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            p_src_global, src2dDesc.GetElementSpaceSize(), type_convert<srcDataType>{}(zeroVal));
+        auto workspace_global_val_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            ws_values_global, dst1dDesc.GetLength(I0) * BlkGroupSize);
+        auto workspace_global_idx_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
+            ws_indices_global, dst1dDesc.GetLength(I0) * BlkGroupSize);
+
+        auto in_block_val_buf =
+            make_dynamic_buffer<AddressSpaceEnum_t::Lds>(p_in_block_values_buffer, BlockBufferSize);
+        auto in_block_idx_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
+            p_in_block_indices_buffer, BlockBufferSize);
+        StaticBuffer<AddressSpaceEnum_t::Vgpr, compType, 1, true> accuValue_buf;
+        StaticBuffer<AddressSpaceEnum_t::Vgpr, int, 1, true> accuIndex_buf;
+
+        accuValue_buf(I0) = zeroVal;
+        accuIndex_buf(I0) = 0;
+
+        const auto toReduceLength = src2dDesc.GetLength(Number<1>{});
+        const int divider         = origReduceLen;
+
+        const preUnaryOpType preUnaryOp(divider);
+
+        const index_t thread_local_id = get_thread_local_1d_id();
+        const index_t block_global_id = get_block_1d_id();
+        const index_t blkgroup_id     = block_global_id / BlkGroupSize;
+        const index_t block_local_id  = block_global_id % BlkGroupSize;
+
+        const index_t reduceSizePerBlock =
+            (((toReduceLength + BlkGroupSize - 1) / BlkGroupSize + BlockBufferSize - 1) /
+             BlockBufferSize) *
+            BlockBufferSize;
+
+        constexpr auto in_block_desc = make_naive_tensor_descriptor_packed(
+            make_tuple(Number<1>{}, Number<BlockSize * GredAccessesPerThreadInBlock>{}));
+
+        using ThreadSliceLengths   = Sequence<1, GredAccessesPerThreadInBlock>;
+        using ThreadClusterLengths = Sequence<1, BlockSize>;
+
+        auto blockwise_src_load = BlockwiseTensorSliceTransfer_v4<BlockSize,
+                                                                  InMemoryDataOperationEnum_t::Set,
+                                                                  Sequence<1, BlockBufferSize>,
+                                                                  ThreadSliceLengths,
+                                                                  ThreadClusterLengths,
+                                                                  Sequence<0, 1>,
+                                                                  srcDataType,
+                                                                  compType,
+                                                                  src2dDescType,
+                                                                  decltype(in_block_desc),
+                                                                  Sequence<0, 1>,
+                                                                  Sequence<0, 1>,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  1,
+                                                                  false,
+                                                                  true>(
+            src2dDesc,
+            make_multi_index(blkgroup_id, block_local_id * reduceSizePerBlock),
+            in_block_desc,
+            make_multi_index(0, 0));
+
+        constexpr auto in_block_copy_step = make_multi_index(0, BlockBufferSize);
+
+        const index_t toReduceBlocks = (reduceSizePerBlock + BlockSize - 1) / BlockSize;
+
+        int indexOffset = block_local_id * reduceSizePerBlock;
+
+        for(index_t reducedBlocks = 0; reducedBlocks < toReduceBlocks;
+            reducedBlocks += GredAccessesPerThreadInBlock)
+        {
+            blockwise_reduce::init_buffer_indices(in_block_idx_buf, indexOffset);
+
+            blockwise_src_load.RunRead(src2dDesc, src_global_buf);
+            blockwise_src_load.RunWrite(in_block_desc, in_block_val_buf);
+
+            __syncthreads();
+
+            // unary operation before reducing, needed by AMAX; For MIN/MAX, nothing is actually
+            // done here
+            blockwise_reduce::operate_on_elements(preUnaryOp, in_block_val_buf);
+
+            index_t BlocksInOneOp = (reducedBlocks < toReduceBlocks - GredAccessesPerThreadInBlock)
+                                        ? GredAccessesPerThreadInBlock
+                                        : toReduceBlocks - reducedBlocks;
+
+            blockwise_reduce::Reduce2(in_block_val_buf,
+                                      in_block_idx_buf,
+                                      BlocksInOneOp,
+                                      accuValue_buf(I0),
+                                      accuIndex_buf(I0));
+
+            indexOffset += BlockBufferSize;
+
+            blockwise_src_load.MoveSrcSliceWindow(src2dDesc, in_block_copy_step);
+        }
+
+        constexpr auto ReducedDataDesc =
+            make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
+
+        const auto workspace_desc =
+            make_naive_tensor_descriptor_packed(make_tuple(dst1dDesc.GetLength(I0) * BlkGroupSize));
+
+        // The first thread in the block stores the reduced result to the global location
+        // representing the block
+        if(thread_local_id == 0)
+        {
+            auto threadwise_workspace_val_store =
+                ThreadwiseTensorSliceTransfer_v1r3<compType,
+                                                   srcDataType,
+                                                   decltype(ReducedDataDesc),
+                                                   decltype(workspace_desc),
+                                                   Sequence<1>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   1,
+                                                   InMemoryDataOperationEnum_t::Set,
+                                                   1,
+                                                   true>(workspace_desc,
+                                                         make_multi_index(block_global_id));
+
+            auto threadwise_workspace_idx_store =
+                ThreadwiseTensorSliceTransfer_v1r3<int,
+                                                   int,
+                                                   decltype(ReducedDataDesc),
+                                                   decltype(workspace_desc),
+                                                   Sequence<1>,
+                                                   Sequence<0>,
+                                                   0,
+                                                   1,
+                                                   InMemoryDataOperationEnum_t::Set,
+                                                   1,
+                                                   true>(workspace_desc,
+                                                         make_multi_index(block_global_id));
+
+            threadwise_workspace_val_store.Run(ReducedDataDesc,
+                                               make_tuple(I0),
+                                               accuValue_buf,
+                                               workspace_desc,
+                                               workspace_global_val_buf);
+            threadwise_workspace_idx_store.Run(ReducedDataDesc,
+                                               make_tuple(I0),
+                                               accuIndex_buf,
+                                               workspace_desc,
+                                               workspace_global_idx_buf);
+        }
+    };
+};
+
+} // namespace ck
+#endif

composable_kernel/include/tensor_operation/reduction_functions_blockwise.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#ifndef CK_REDUCTION_FUNCTIONS_BLOCKWISE_HPP
+#define CK_REDUCTION_FUNCTIONS_BLOCKWISE_HPP
+
+#include "data_type.hpp"
+
+#include "reduction_common.hpp"
+#include "reduction_operator.hpp"
+#include "reduction_functions_binop.hpp"
+
+namespace ck {
+
+template <typename buffer2dDescType,
+          bool blockIsOneRow,
+          typename opReduce,
+          NanPropagation_t nanPropaOpt>
+struct BlockwiseReduction_2d_block_buffer
+{
+    using compType = typename opReduce::dataType;
+
+    static constexpr auto buffer2dDesc = buffer2dDescType{};
+
+    static constexpr index_t BlockSize =
+        blockIsOneRow ? buffer2dDesc.GetLength(Number<1>{}) : buffer2dDesc.GetLength(Number<0>{});
+    static constexpr index_t NumBlocks =
+        blockIsOneRow ? buffer2dDesc.GetLength(Number<0>{}) : buffer2dDesc.GetLength(Number<1>{});
+    using binop = detail::binop_with_nan_check<nanPropaOpt, opReduce, compType>;
+
+    // This interface does not accumulate on indices
+    template <typename BufferType>
+    __device__ static void
+    Reduce(BufferType& block_buffer, index_t toReduceBlocks, compType& accuData)
+    {
+        const index_t thread_local_id = get_thread_local_1d_id();
+        compType lAccuData            = opReduce::GetReductionZeroVal();
+
+        index_t offset;
+        for(index_t otherDimInd = 0; otherDimInd < toReduceBlocks; otherDimInd++)
+        {
+            offset = blockIsOneRow
+                         ? buffer2dDesc.CalculateOffset(make_tuple(otherDimInd, thread_local_id))
+                         : buffer2dDesc.CalculateOffset(make_tuple(thread_local_id, otherDimInd));
+            compType opData = type_convert<compType>{}(block_buffer[offset]);
+
+            binop::calculate(lAccuData, opData);
+        }
+
+        offset = blockIsOneRow ? buffer2dDesc.CalculateOffset(make_tuple(0, thread_local_id))
+                               : buffer2dDesc.CalculateOffset(make_tuple(thread_local_id, 0));
+
+        block_buffer(offset) = lAccuData;
+
+        __syncthreads();
+
+        for(index_t indOffset = BlockSize / 2; indOffset > 0; indOffset /= 2)
+        {
+            if(thread_local_id < indOffset)
+            {
+                index_t offset1 =
+                    blockIsOneRow ? buffer2dDesc.CalculateOffset(make_tuple(0, thread_local_id))
+                                  : buffer2dDesc.CalculateOffset(make_tuple(thread_local_id, 0));
+
+                index_t offset2 =
+                    blockIsOneRow
+                        ? buffer2dDesc.CalculateOffset(make_tuple(0, thread_local_id + indOffset))
+                        : buffer2dDesc.CalculateOffset(make_tuple(thread_local_id + indOffset, 0));
+
+                compType opData1 = type_convert<compType>{}(block_buffer[offset1]);
+                compType opData2 = type_convert<compType>{}(block_buffer[offset2]);
+                binop::calculate(opData1, opData2);
+                block_buffer(offset1) = type_convert<compType>{}(opData1);
+            }
+
+            __syncthreads();
+        }
+
+        if(thread_local_id == 0)
+        {
+            compType tmpVal = type_convert<compType>{}(block_buffer[0]);
+
+            binop::calculate(accuData, tmpVal);
+        }
+    };
+
+    // This interface accumulates on both data values and indices
+    template <typename BufferType, typename IdxBufferType>
+    __device__ static void Reduce2(BufferType& block_buffer,
+                                   IdxBufferType& block_indices_buffer,
+                                   index_t toReduceBlocks,
+                                   compType& accuData,
+                                   int& accuIndex)
+    {
+        const index_t thread_local_id = get_thread_local_1d_id();
+        compType lAccuData            = opReduce::GetReductionZeroVal();
+        int lAccuIndex                = 0;
+
+        if constexpr(blockIsOneRow)
+        {
+            for(index_t otherDimInd = 0; otherDimInd < toReduceBlocks; otherDimInd++)
+            {
+                for(index_t indOffset = 1; indOffset < BlockSize; indOffset *= 2)
+                {
+                    if(thread_local_id % (indOffset * 2) == 0)
+                    {
+                        index_t offset1 =
+                            buffer2dDesc.CalculateOffset(make_tuple(otherDimInd, thread_local_id));
+                        index_t offset2 = buffer2dDesc.CalculateOffset(
+                            make_tuple(otherDimInd, thread_local_id + indOffset));
+
+                        compType currVal1 = type_convert<compType>{}(block_buffer[offset1]);
+                        compType currVal2 = type_convert<compType>{}(block_buffer[offset2]);
+                        int currIndex1    = block_indices_buffer[offset1];
+                        int currIndex2    = block_indices_buffer[offset2];
+
+                        binop::calculate(currVal1, currVal2, currIndex1, currIndex2);
+                        block_buffer(offset1)         = type_convert<compType>{}(currVal1);
+                        block_indices_buffer(offset1) = currIndex1;
+                    }
+                    __syncthreads();
+                }
+            }
+
+            if(thread_local_id == 0)
+            {
+                for(index_t otherDimInd = 0; otherDimInd < toReduceBlocks; otherDimInd++)
+                {
+                    index_t offset = buffer2dDesc.CalculateOffset(make_tuple(otherDimInd, 0));
+
+                    compType tmpVal = type_convert<compType>{}(block_buffer[offset]);
+                    int tmpIndex    = block_indices_buffer[offset];
+
+                    binop::calculate(lAccuData, tmpVal, lAccuIndex, tmpIndex);
+                }
+
+                binop::calculate(accuData, lAccuData, accuIndex, lAccuIndex);
+            }
+        }
+        else
+        {
+            index_t offset;
+
+            for(index_t otherDimInd = 0; otherDimInd < toReduceBlocks; otherDimInd++)
+            {
+                offset = buffer2dDesc.CalculateOffset(make_tuple(thread_local_id, otherDimInd));
+                compType currVal = type_convert<compType>{}(block_buffer[offset]);
+                int currIndex    = block_indices_buffer[offset];
+
+                binop::calculate(lAccuData, currVal, lAccuIndex, currIndex);
+            }
+
+            offset = buffer2dDesc.CalculateOffset(make_tuple(thread_local_id, 0));
+
+            block_buffer(offset)         = lAccuData;
+            block_indices_buffer(offset) = lAccuIndex;
+
+            __syncthreads();
+
+            for(index_t indOffset = 1; indOffset < BlockSize; indOffset *= 2)
+            {
+                if(thread_local_id % (indOffset * 2) == 0)
+                {
+                    index_t offset1 = buffer2dDesc.CalculateOffset(make_tuple(thread_local_id, 0));
+                    index_t offset2 =
+                        buffer2dDesc.CalculateOffset(make_tuple(thread_local_id + indOffset, 0));
+
+                    compType currVal1 = type_convert<compType>{}(block_buffer[offset1]);
+                    compType currVal2 = type_convert<compType>{}(block_buffer[offset2]);
+                    int currIndex1    = block_indices_buffer[offset1];
+                    int currIndex2    = block_indices_buffer[offset2];
+
+                    binop::calculate(currVal1, currVal2, currIndex1, currIndex2);
+                    block_buffer(offset1)         = type_convert<compType>{}(currVal1);
+                    block_indices_buffer(offset1) = currIndex1;
+                }
+
+                __syncthreads();
+            }
+
+            if(thread_local_id == 0)
+            {
+                compType tmpVal = type_convert<compType>{}(block_buffer[0]);
+                int tmpIndex    = block_indices_buffer[0];
+
+                binop::calculate(accuData, tmpVal, accuIndex, tmpIndex);
+            }
+        }
+    };
+
+    template <typename BufferType>
+    __device__ static void set_buffer_value(BufferType& block_buffer, compType value)
+    {
+        index_t thread_id = get_thread_local_1d_id();
+
+        for(index_t otherDimInd = 0; otherDimInd < NumBlocks; otherDimInd++)
+        {
+            index_t offset = blockIsOneRow
+                                 ? buffer2dDesc.CalculateOffset(make_tuple(otherDimInd, thread_id))
+                                 : buffer2dDesc.CalculateOffset(make_tuple(thread_id, otherDimInd));
+
+            block_buffer(offset) = value;
+
+            __syncthreads();
+        }
+    };
+
+    // Initialize the block-wise indices buffer, the index for each element in the block-wise data
+    // buffer
+    // is calculated according to its position in the buffer and the global starting index
+    template <typename IdxBufferType>
+    __device__ static void init_buffer_indices(IdxBufferType& block_indices_buffer, int indexStart)
+    {
+        index_t thread_id = get_thread_local_1d_id();
+
+        for(index_t otherDimInd = 0; otherDimInd < NumBlocks; otherDimInd++)
+        {
+            index_t offset = blockIsOneRow
+                                 ? buffer2dDesc.CalculateOffset(make_tuple(otherDimInd, thread_id))
+                                 : buffer2dDesc.CalculateOffset(make_tuple(thread_id, otherDimInd));
+
+            block_indices_buffer(offset) = offset + indexStart;
+
+            __syncthreads();
+        }
+    };
+
+    // Execute unary operation on the block buffer elements
+    template <typename unary_op_type, typename BufferType>
+    __device__ static void operate_on_elements(unary_op_type& unary_op, BufferType& block_buffer)
+    {
+        index_t thread_id = get_thread_local_1d_id();
+
+        for(index_t otherDimInd = 0; otherDimInd < NumBlocks; otherDimInd++)
+        {
+            index_t offset = blockIsOneRow
+                                 ? buffer2dDesc.CalculateOffset(make_tuple(otherDimInd, thread_id))
+                                 : buffer2dDesc.CalculateOffset(make_tuple(thread_id, otherDimInd));
+
+            block_buffer(offset) = unary_op(block_buffer[offset]);
+
+            __syncthreads();
+        }
+    };
+};
+
+}; // end of namespace ck
+
+#endif

composable_kernel/include/tensor_operation/reduction_functions_threadwise.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#ifndef CK_REDUCTION_FUNCTIONS_THREADWISE_HPP
+#define CK_REDUCTION_FUNCTIONS_THREADWISE_HPP
+
+#include "data_type.hpp"
+
+#include "reduction_common.hpp"
+#include "reduction_operator.hpp"
+#include "reduction_functions_binop.hpp"
+
+namespace ck {
+
+template <typename BufferType, typename opReduce, NanPropagation_t nanPropaOpt>
+struct ThreadReduce
+{
+    using compType = typename opReduce::dataType;
+
+    static_assert(BufferType::IsStaticBuffer(), "Thread-wise reduction needs use StaticBuffer!");
+
+    static_assert(
+        std::is_same<typename BufferType::type, compType>::value,
+        "Data type of StaticBuffer for Thread-wise reduction should be same as the compType!");
+
+    static constexpr index_t ThreadBufferLen = BufferType::Size();
+
+    using binop = detail::binop_with_nan_check<nanPropaOpt, opReduce, compType>;
+
+    // This interface does not accumulate on indices
+    __device__ static void Reduce(const BufferType& thread_buffer, compType& accuData)
+    {
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { binop::calculate(accuData, thread_buffer[I]); });
+    };
+
+    // This interface accumulates on both data values and indices and
+    // is called by Direct_ThreadWise reduction method at first-time reduction
+    __device__ static void
+    Reduce2(const BufferType& thread_buffer, compType& accuData, int& accuIndex, int indexStart)
+    {
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) {
+            int currIndex = I + indexStart;
+            binop::calculate(accuData, thread_buffer[I], accuIndex, currIndex);
+        });
+    };
+
+    // Set the elements in the per-thread buffer to a specific value
+    // cppcheck-suppress constParameter
+    __device__ static void set_buffer_value(BufferType& thread_buffer, compType value)
+    {
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) { thread_buffer(I) = value; });
+    };
+
+    // Execute unary operation on the per-thread buffer elements
+    template <typename unary_op_type>
+    __device__ static void operate_on_elements(unary_op_type& unary_op, BufferType& thread_buffer)
+    {
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { thread_buffer(I) = unary_op(thread_buffer[I]); });
+    };
+};
+
+template <typename BufferType,
+          typename IdxBufferType,
+          typename opReduce,
+          NanPropagation_t nanPropaOpt>
+struct ThreadReduceWithIndicesInput
+{
+    using compType = typename opReduce::dataType;
+
+    static_assert(BufferType::IsStaticBuffer(), "Thread-wise reduction needs use StaticBuffer!");
+    static_assert(IdxBufferType::IsStaticBuffer(),
+                  "Thread-wise reduction needs use StaticBuffer for indices!");
+
+    static_assert(
+        std::is_same<typename BufferType::type, compType>::value,
+        "Data type of StaticBuffer for Thread-wise reduction should be same as the compType!");
+    static_assert(std::is_same<typename IdxBufferType::type, index_t>::value,
+                  "Indices type of StaticBuffer for Thread-wise reduction should be index_t!");
+
+    static_assert(BufferType::Size() == IdxBufferType::Size(),
+                  "StaticBuffers for data and indices should have the same sizes!");
+
+    static constexpr index_t ThreadBufferLen = BufferType::Size();
+
+    using binop = detail::binop_with_nan_check<nanPropaOpt, opReduce, compType>;
+
+    // This interface accumulates on both data values and indices and
+    // is called by Direct_ThreadWise reduction method at second-time reduction
+    __device__ static void Reduce(const BufferType& thread_buffer,
+                                  const IdxBufferType& thread_indices_buffer,
+                                  compType& accuData,
+                                  int& accuIndex)
+    {
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) {
+            binop::calculate(accuData, thread_buffer[I], accuIndex, thread_indices_buffer[I]);
+        });
+    };
+
+    // Set the elements in the per-thread buffer to a specific value
+    // cppcheck-suppress constParameter
+    __device__ static void set_buffer_value(BufferType& thread_buffer, compType value)
+    {
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) { thread_buffer(I) = value; });
+    };
+
+    // Execute unary operation on the per-thread buffer elements
+    template <typename unary_op_type>
+    __device__ static void operate_on_elements(unary_op_type& unary_op, BufferType& thread_buffer)
+    {
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { thread_buffer(I) = unary_op(thread_buffer[I]); });
+    };
+};
+
+}; // end of namespace ck
+
+#endif

composable_kernel/include/tensor_operation/reduction_functions_warpwise.hpp

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2020 Advanced Micro Devices, Inc.
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to deal
+ * in the Software without restriction, including without limitation the rights
+ * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+ * copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ *
+ *******************************************************************************/
+#ifndef CK_REDUCTION_FUNCTIONS_WARPWISE_HPP
+#define CK_REDUCTION_FUNCTIONS_WARPWISE_HPP
+
+#include "data_type.hpp"
+
+#include "reduction_common.hpp"
+#include "reduction_operator.hpp"
+#include "reduction_functions_binop.hpp"
+
+namespace ck {
+
+template <typename BufferType, index_t BlockSize, typename opReduce, NanPropagation_t nanPropaOpt>
+struct WarpReduce
+{
+    using compType = typename opReduce::dataType;
+    using binop    = detail::binop_with_nan_check<nanPropaOpt, opReduce, compType>;
+
+    static_assert(BufferType::IsStaticBuffer(),
+                  "Per-thread buffer for WarpWise reduction should be StaticBuffer!");
+    static_assert(std::is_same<typename BufferType::type, compType>::value,
+                  "Data type of per-thread StaticBuffer for WarpWise reduction should be same as "
+                  "the compType!");
+
+    static constexpr index_t ThreadBufferLen = BufferType::Size();
+    static constexpr bool have_builtin_shuffle =
+        std::is_same<compType, float>::value || std::is_same<compType, double>::value;
+
+    // This interface does not accumulate on indices
+    __device__ static void Reduce(const BufferType& thread_buffer, compType& accuData)
+    {
+        if constexpr(have_builtin_shuffle)
+            ReduceImpl1(thread_buffer, accuData);
+        else
+            ReduceImpl2(thread_buffer, accuData);
+    };
+
+    // This interface implementation uses HIP built-in device shuffling functions
+    __device__ static void ReduceImpl1(const BufferType& thread_buffer, compType& accuData)
+    {
+        compType lAccuData = opReduce::GetReductionZeroVal();
+
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { binop::calculate(lAccuData, thread_buffer[I]); });
+
+        // synchronize among all threads in this warp
+        __all(1);
+
+        for(index_t stride = warpSize / 2; stride > 0; stride /= 2)
+        {
+            compType tmpVal = __shfl_down(lAccuData, stride, warpSize);
+            binop::calculate(lAccuData, tmpVal);
+            __all(1);
+        }
+
+        binop::calculate(accuData, lAccuData);
+    };
+
+    // This interface implementation does not use HIP built-in device shuffling functions
+    // since for fp16, built-in shuffling functions is not provided by HIP
+    __device__ static void ReduceImpl2(const BufferType& thread_buffer, compType& accuData)
+    {
+        compType lAccuData = opReduce::GetReductionZeroVal();
+
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { binop::calculate(lAccuData, thread_buffer[I]); });
+
+        __syncthreads();
+
+        index_t thread_id        = get_thread_local_1d_id();
+        index_t warpId           = thread_id / warpSize;
+        index_t thread_inwarp_id = thread_id % warpSize;
+
+        __shared__ compType shuffle_buffer[BlockSize];
+
+        compType* myBuffer = &shuffle_buffer[warpId * warpSize];
+
+        myBuffer[thread_inwarp_id] = lAccuData;
+
+        __syncthreads();
+
+        for(index_t stride = warpSize / 2; stride > 0; stride /= 2)
+        {
+            if(thread_inwarp_id < stride)
+            {
+                compType currVal1 = myBuffer[thread_inwarp_id];
+                compType currVal2 = myBuffer[thread_inwarp_id + stride];
+
+                binop::calculate(currVal1, currVal2);
+
+                myBuffer[thread_inwarp_id] = currVal1;
+            }
+
+            __syncthreads();
+        }
+        if(thread_inwarp_id == 0)
+            binop::calculate(accuData, myBuffer[0]);
+    };
+
+    // This interface accumulates on both data values and indices and is called by Direct_WarpWise
+    // reduction method at first-time reduction
+    __device__ static void
+    Reduce2(const BufferType& thread_buffer, compType& accuData, int& accuIndex, int indexStart)
+    {
+        if constexpr(have_builtin_shuffle)
+            Reduce2Impl1(thread_buffer, accuData, accuIndex, indexStart);
+        else
+            Reduce2Impl2(thread_buffer, accuData, accuIndex, indexStart);
+    };
+
+    // This interface implementation uses HIP built-in device shuffling functions
+    __device__ static void Reduce2Impl1(const BufferType& thread_buffer,
+                                        compType& accuData,
+                                        int& accuIndex,
+                                        int indexStart)
+    {
+        compType lAccuData       = opReduce::GetReductionZeroVal();
+        int lAccuIndex           = 0;
+        index_t thread_inwarp_id = get_thread_local_1d_id() % warpSize;
+
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) {
+            int currIndex = thread_inwarp_id * ThreadBufferLen + I + indexStart;
+            binop::calculate(lAccuData, thread_buffer[I], lAccuIndex, currIndex);
+        });
+
+        // synchronize among all threads in this warp
+        __all(1);
+
+        for(index_t stride = 1; stride < warpSize; stride *= 2)
+        {
+            compType tmpVal = __shfl_down(lAccuData, stride, warpSize);
+            int tmpIndex    = __shfl_down(lAccuIndex, stride, warpSize);
+
+            binop::calculate(lAccuData, tmpVal, lAccuIndex, tmpIndex);
+            __all(1);
+        }
+
+        if(thread_inwarp_id == 0)
+            binop::calculate(accuData, lAccuData, accuIndex, lAccuIndex);
+    };
+
+    // This interface implementation does not use HIP built-in device shuffling functions since for
+    // fp16, built-in shuffling functions is not provided by HIP
+    __device__ static void Reduce2Impl2(const BufferType& thread_buffer,
+                                        compType& accuData,
+                                        int& accuIndex,
+                                        int indexStart)
+    {
+        compType lAccuData       = opReduce::GetReductionZeroVal();
+        int lAccuIndex           = 0;
+        index_t thread_id        = get_thread_local_1d_id();
+        index_t warpId           = thread_id / warpSize;
+        index_t thread_inwarp_id = thread_id % warpSize;
+
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) {
+            int currIndex = thread_inwarp_id * ThreadBufferLen + I + indexStart;
+            binop::calculate(lAccuData, thread_buffer[I], lAccuIndex, currIndex);
+        });
+
+        __shared__ compType shuffle_data_buffer[BlockSize];
+        __shared__ int shuffle_indices_buffer[BlockSize];
+
+        compType* myDataBuffer = &shuffle_data_buffer[warpId * warpSize];
+        int* myIndicesBuffer   = &shuffle_indices_buffer[warpId * warpSize];
+
+        myDataBuffer[thread_inwarp_id]    = lAccuData;
+        myIndicesBuffer[thread_inwarp_id] = lAccuIndex;
+
+        __syncthreads();
+
+        for(index_t stride = 1; stride < warpSize; stride *= 2)
+        {
+            compType currVal1 = myDataBuffer[thread_inwarp_id];
+            compType currVal2 = myDataBuffer[thread_inwarp_id + stride];
+            int currIndex1    = myIndicesBuffer[thread_inwarp_id];
+            int currIndex2    = myIndicesBuffer[thread_inwarp_id + stride];
+
+            binop::calculate(currVal1, currVal2, currIndex1, currIndex2);
+
+            myDataBuffer[thread_inwarp_id]    = currVal1;
+            myIndicesBuffer[thread_inwarp_id] = currIndex1;
+
+            __syncthreads();
+        }
+
+        if(thread_inwarp_id == 0)
+            binop::calculate(accuData, myDataBuffer[0], accuIndex, myIndicesBuffer[0]);
+    };
+
+    // cppcheck-suppress constParameter
+    __device__ static void set_buffer_value(BufferType& thread_buffer, compType value)
+    {
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) { thread_buffer(I) = value; });
+
+        __all(1);
+    };
+
+    // Execute unary operation on the per-thread buffer elements
+    template <typename unary_op_type>
+    __device__ static void operate_on_elements(unary_op_type& unary_op, BufferType& thread_buffer)
+    {
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { thread_buffer(I) = unary_op(thread_buffer[I]); });
+
+        __all(1);
+    };
+};
+
+template <typename BufferType,
+          typename IdxBufferType,
+          index_t BlockSize,
+          typename opReduce,
+          NanPropagation_t nanPropaOpt>
+struct WarpReduceWithIndicesInput
+{
+    using compType = typename opReduce::dataType;
+    using binop    = detail::binop_with_nan_check<nanPropaOpt, opReduce, compType>;
+
+    static_assert(BufferType::IsStaticBuffer(),
+                  "Per-thread buffer for WarpWise reduction should be StaticBuffer!");
+    static_assert(IdxBufferType::IsStaticBuffer(),
+                  "Per-thread buffer for WarpWise reduction should be StaticBuffer for indices!");
+
+    static_assert(std::is_same<typename BufferType::type, compType>::value,
+                  "Data type of per-thread StaticBuffer for WarpWise reduction should be same as "
+                  "the compType!");
+    static_assert(
+        std::is_same<typename IdxBufferType::type, index_t>::value,
+        "Indices type per-thread of StaticBuffer for WarpWise reduction should be index_t!");
+
+    static_assert(BufferType::Size() == IdxBufferType::Size(),
+                  "StaticBuffers for data and indices should have the same sizes!");
+
+    static constexpr index_t ThreadBufferLen = BufferType::Size();
+    static constexpr bool have_builtin_shuffle =
+        std::is_same<compType, float>::value || std::is_same<compType, double>::value;
+
+    // This interface accumulates on both data values and indices and is called by Direct_WarpWise
+    // reduction method at second-time reduction
+    __device__ static void Reduce(const BufferType& thread_buffer,
+                                  const IdxBufferType& thread_indices_buffer,
+                                  compType& accuData,
+                                  int& accuIndex)
+    {
+        if constexpr(have_builtin_shuffle)
+            ReduceImpl1(thread_buffer, thread_indices_buffer, accuData, accuIndex);
+        else
+            ReduceImpl2(thread_buffer, thread_indices_buffer, accuData, accuIndex);
+    };
+
+    // This interface implementation uses HIP built-in device shuffling functions
+    __device__ static void ReduceImpl1(const BufferType& thread_buffer,
+                                       const IdxBufferType& thread_indices_buffer,
+                                       compType& accuData,
+                                       int& accuIndex)
+    {
+        compType lAccuData = opReduce::GetReductionZeroVal();
+        int lAccuIndex     = 0;
+
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) {
+            binop::calculate(lAccuData, thread_buffer[I], lAccuIndex, thread_indices_buffer[I]);
+        });
+
+        // synchronize among all threads in this warp
+        __all(1);
+
+        for(index_t stride = 1; stride < warpSize; stride *= 2)
+        {
+            compType tmpVal = __shfl_down(lAccuData, stride, warpSize);
+            int tmpIndex    = __shfl_down(lAccuIndex, stride, warpSize);
+
+            binop::calculate(lAccuData, tmpVal, lAccuIndex, tmpIndex);
+            __all(1);
+        }
+
+        binop::calculate(accuData, lAccuData, accuIndex, lAccuIndex);
+    };
+
+    // This interface implementation does not use HIP built-in device shuffling functions
+    // since for fp16, built-in shuffling functions is not provided by HIP
+    __device__ static void ReduceImpl2(const BufferType& thread_buffer,
+                                       const IdxBufferType& thread_indices_buffer,
+                                       compType& accuData,
+                                       int& accuIndex)
+    {
+        compType lAccuData       = opReduce::GetReductionZeroVal();
+        int lAccuIndex           = 0;
+        index_t thread_id        = get_thread_local_1d_id();
+        index_t warpId           = thread_id / warpSize;
+        index_t thread_inwarp_id = thread_id % warpSize;
+
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) {
+            binop::calculate(lAccuData, thread_buffer[I], lAccuIndex, thread_indices_buffer[I]);
+        });
+
+        __shared__ compType shuffle_data_buffer[BlockSize];
+        __shared__ int shuffle_indices_buffer[BlockSize];
+
+        compType* myDataBuffer = &shuffle_data_buffer[warpId * warpSize];
+        int* myIndicesBuffer   = &shuffle_indices_buffer[warpId * warpSize];
+
+        myDataBuffer[thread_inwarp_id]    = lAccuData;
+        myIndicesBuffer[thread_inwarp_id] = lAccuIndex;
+
+        __syncthreads();
+
+        for(index_t stride = 1; stride < warpSize; stride *= 2)
+        {
+            compType currVal1 = myDataBuffer[thread_inwarp_id];
+            compType currVal2 = myDataBuffer[thread_inwarp_id + stride];
+            int currIndex1    = myIndicesBuffer[thread_inwarp_id];
+            int currIndex2    = myIndicesBuffer[thread_inwarp_id + stride];
+
+            binop::calculate(currVal1, currVal2, currIndex1, currIndex2);
+
+            myDataBuffer[thread_inwarp_id]    = currVal1;
+            myIndicesBuffer[thread_inwarp_id] = currIndex1;
+
+            __syncthreads();
+        }
+
+        if(thread_inwarp_id == 0)
+            binop::calculate(accuData, myDataBuffer[0], accuIndex, myIndicesBuffer[0]);
+    };
+
+    // cppcheck-suppress constParameter
+    __device__ static void set_buffer_value(BufferType& thread_buffer, compType value)
+    {
+        static_for<0, ThreadBufferLen, 1>{}([&](auto I) { thread_buffer(I) = value; });
+
+        __all(1);
+    };
+
+    // Execute unary operation on the per-thread buffer elements
+    template <typename unary_op_type>
+    __device__ static void operate_on_elements(unary_op_type& unary_op, BufferType& thread_buffer)
+    {
+        static_for<0, ThreadBufferLen, 1>{}(
+            [&](auto I) { thread_buffer(I) = unary_op(thread_buffer[I]); });
+
+        __all(1);
+    };
+};
+
+}; // end of namespace ck
+
+#endif

composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer.hpp

@@ -397,7 +397,7 @@ struct ThreadwiseTensorSliceTransfer_v2
                       "wrong! SrcDesc need to known at compile-time");
     }
 
-    __device__ void SetDstSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
+    __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
     {
         src_coord_ = make_tensor_coordinate(src_desc, src_slice_origin_idx);
     }
@@ -713,9 +713,6 @@ struct ThreadwiseTensorSliceTransfer_v3
         : src_coord_(make_tensor_coordinate(src_desc, src_slice_origin)),
           dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin))
     {
-        // TODO: fix this
-        static_assert(is_same<SrcData, DstData>::value,
-                      "wrong! current implementation assume SrcData and DstData are same type");
     }
 
     __device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
@@ -985,7 +982,8 @@ struct ThreadwiseTensorSliceTransfer_v3
                 constexpr index_t buffer_offset =
                     buffer_desc_.CalculateOffset(dst_data_idx + i * dst_scalar_step_in_vector);
 
-                dst_tmp_vector.template AsType<DstData>()(i) = buffer_[Number<buffer_offset>{}];
+                dst_tmp_vector.template AsType<DstData>()(i) =
+                    type_convert<DstData>{}(buffer_[Number<buffer_offset>{}]);
             });
 
             using dst_vector_t = typename decltype(dst_tmp_vector)::type;

composable_kernel/include/utility/config.hpp

@@ -90,8 +90,8 @@
 #endif
 
 // pass tensor descriptor by value or void*
-#define CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE 0
-#define CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER 1
+#define CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VALUE 1
+#define CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER 0
 
 // merge transformation use magic number division
 #define CK_EXPERIMENTAL_MERGE_USE_MAGIC_DIVISION 0