merge

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp

@@ -139,8 +139,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
     {
@@ -315,8 +315,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
             c_grid_desc_m_n);
     }
     using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 CGridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
@@ -634,10 +634,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
                 Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
                 FloatCShuffle,              // typename SrcData,
                 FloatC,                     // typename DstData,
-                decltype(
-                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
                 Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
                 5,                                          // index_t VectorDim,
                 CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp

@@ -142,8 +142,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
     {
@@ -323,13 +323,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
     }
 
     using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 CGridDesc_M_N{}))>;
 
     using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 C0GridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
@@ -654,12 +654,9 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
                 FloatC,                     // typename Src0Data,
                 FloatC,                     // typename Src1Data,
                 FloatC,                     // typename DstData,
-                decltype(
-                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
                 Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
                 5,                                          // index_t VectorDim,
                 CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,

include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp

@@ -151,8 +151,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
 
     using ThisThreadBlock = ThisThreadBlock<BlockSize>;
 
-    using GridwiseGemmPipe = remove_cvref_t<decltype(
-        GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
+    using GridwiseGemmPipe = remove_cvref_t<
+        decltype(GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
 
     __host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
     {
@@ -331,18 +331,18 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
             c_grid_desc_m_n);
     }
     using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 CGridDesc_M_N{}))>;
 
     using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 C0GridDesc_M_N{}))>;
 
     using C1GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
-        remove_cvref_t<decltype(
-            MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
+        remove_cvref_t<
+            decltype(MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
                 C1GridDesc_M_N{}))>;
 
     using DefaultBlock2CTileMap =
@@ -674,14 +674,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
                 FloatC,                     // typename Src1Data,
                 FloatC,                     // typename Src2Data,
                 FloatC,                     // typename DstData,
-                decltype(
-                    c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
-                decltype(
-                    c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
+                decltype(c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
                 Sequence<0, 1, 2, 3, 4, 5>,                 // typename DimAccessOrder,
                 5,                                          // index_t VectorDim,
                 CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,

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

@@ -78,8 +78,8 @@ struct GridwiseSparseEmbeddingsForwardLayernorm
     using ThreadwiseWolfordDesc2D = decltype(make_naive_tensor_descriptor_packed(make_tuple(
         Number<DimSubBlocks * DimThreadSize>{}, Number<RowSubBlocks * RowVectorSize>{})));
 
-    using ThreadwiseWolfordDescReduce = decltype(
-        make_naive_tensor_descriptor_packed(make_tuple(Number<DimSubBlocks * DimThreadSize>{})));
+    using ThreadwiseWolfordDescReduce = decltype(make_naive_tensor_descriptor_packed(
+        make_tuple(Number<DimSubBlocks * DimThreadSize>{})));
 
     using ThreadwiseWelford =
         ThreadwiseWelford<AccDataType, ThreadwiseWolfordDesc2D, ThreadwiseWolfordDescReduce>;

include/ck/tensor_operation/gpu/grid/normalization/gridwise_normalization_splitk_1st.hpp

@@ -87,9 +87,9 @@ struct GridwiseNormalizationSplitK1st
             int left_kPerBlock  = math::integer_divide_ceil(k, kGridSize);
             int kRightmostBlock = kRaw - left_kPerBlock * (kGridSize - 1);
             int kPerThread      = kRightmostBlock < K_BlockTileSize
-                                 ? 0
-                                 : KThreadSliceSize * (kRightmostBlock / K_BlockTileSize);
-            int kPerBlockTail = kRightmostBlock - kPerThread * KThreadClusterSize;
+                                      ? 0
+                                      : KThreadSliceSize * (kRightmostBlock / K_BlockTileSize);
+            int kPerBlockTail   = kRightmostBlock - kPerThread * KThreadClusterSize;
 
             if(kPerBlockTail > 0)
             {

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp

@@ -129,6 +129,9 @@ struct ThreadwiseTensorSliceTransfer_v3r1
 
         constexpr auto src_access_lengths = SliceLengths{} / src_scalar_per_access;
 
+        static_assert(SliceLengths::At(SrcVectorDim) % SrcScalarPerVector == 0,
+                      "SliceLengths[SrcVectorDim] must be divisible by SrcScalarPerVector");
+
         constexpr auto src_dim_access_order = SrcDimAccessOrder{};
 
         constexpr auto ordered_src_access_lengths =

include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp

@@ -104,13 +104,13 @@ struct ThreadwiseTensorSliceTransfer_v6r1
 
             // apply pointwise operation
             static_for<0, ScalarPerVector, 1>{}([&](auto i) {
-                SrcData v;
+                DstData v;
 
                 // apply element-wise operation
                 element_op_(v, src_vector_container.template AsType<SrcData>()[i]);
 
                 // apply type convert
-                dst_vector_container.template AsType<DstData>()(i) = type_convert<DstData>(v);
+                dst_vector_container.template AsType<DstData>()(i) = v;
             });
 
             const bool is_dst_valid =

include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp

@@ -236,8 +236,6 @@ struct TransformConvBwdDataToGemm_v1
         const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
         const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
-        const index_t AK0 = K / AK1;
-
         // n_do_ho_wo_k for 3d or n_ho_wo_k for 2d
         const auto out_grid_desc =
             make_out_grid_desc<NDimSpatial, ALayout, ConvBwdDataSpecialization>(
@@ -247,6 +245,8 @@ struct TransformConvBwdDataToGemm_v1
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
                          Filter1x1Stride1Pad0)
         {
+            const index_t AK0 = math::integer_divide_ceil(K, AK1);
+
             // A: output tensor
             const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
                 out_grid_desc,
@@ -308,6 +308,9 @@ struct TransformConvBwdDataToGemm_v1
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
+            const index_t AK0 =
+                math::integer_divide_ceil(ZDotSlice * YDotSlice * XDotSlice * K, AK1);
+
             if constexpr(NDimSpatial == 2)
             {
                 // A: output tensor
@@ -332,7 +335,7 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                     make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
-                const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
+                const auto out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc =
                     transform_tensor_descriptor(
                         out_n_ydot_htilde_xdot_wtilde_k_grid_desc,
                         make_tuple(make_pass_through_transform(N),
@@ -340,7 +343,7 @@ struct TransformConvBwdDataToGemm_v1
                                    make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
                                    make_slice_transform(XDot, I0, XDotSlice),
                                    make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                                   make_unmerge_transform(make_tuple(AK0, AK1))),
+                                   make_pass_through_transform(K)),
                         make_tuple(Sequence<0>{},
                                    Sequence<1>{},
                                    Sequence<2>{},
@@ -352,21 +355,28 @@ struct TransformConvBwdDataToGemm_v1
                                    Sequence<2>{},
                                    Sequence<3>{},
                                    Sequence<4>{},
-                                   Sequence<5, 6>{}));
+                                   Sequence<5>{}));
 
-                const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
-                    out_n_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
-                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, AK0)),
-                               make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice)),
-                               make_pass_through_transform(AK1)),
-                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}, Sequence<6>{}),
-                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+                const auto out_gemmk_gemmmraw_grid_desc = transform_tensor_descriptor(
+                    out_n_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K)),
+                               make_merge_transform(make_tuple(N, HTildeSlice, WTildeSlice))),
+                    make_tuple(Sequence<1, 3, 5>{}, Sequence<0, 2, 4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-                const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                const auto out_gemmk_gemmm_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
-                        out_gemmak0_gemmmraw_gemmak1_grid_desc,
-                        make_tuple(AK0, GemmMPerBlock, AK1),
-                        Sequence<false, DoPadGemmM, false>{});
+                        out_gemmk_gemmmraw_grid_desc,
+                        make_tuple(AK1, GemmMPerBlock),
+                        Sequence<true, DoPadGemmM>{});
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_gemmk_gemmm_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                               make_pass_through_transform(
+                                   out_gemmk_gemmm_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
                 return out_gemmak0_gemmm_gemmak1_grid_desc;
             }
@@ -411,7 +421,7 @@ struct TransformConvBwdDataToGemm_v1
                                    Sequence<7>{}));
 
                 const auto
-                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc =
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc =
                         transform_tensor_descriptor(
                             out_n_zdot_dtilde_ydot_htilde_xdot_wtilde_k_grid_desc,
                             make_tuple(make_pass_through_transform(N),
@@ -421,7 +431,7 @@ struct TransformConvBwdDataToGemm_v1
                                        make_slice_transform(HTilde, IHTildeSliceBegin, HTildeSlice),
                                        make_slice_transform(XDot, I0, XDotSlice),
                                        make_slice_transform(WTilde, IWTildeSliceBegin, WTildeSlice),
-                                       make_unmerge_transform(make_tuple(AK0, AK1))),
+                                       make_pass_through_transform(K)),
                             make_tuple(Sequence<0>{},
                                        Sequence<1>{},
                                        Sequence<2>{},
@@ -437,22 +447,29 @@ struct TransformConvBwdDataToGemm_v1
                                        Sequence<4>{},
                                        Sequence<5>{},
                                        Sequence<6>{},
-                                       Sequence<7, 8>{}));
+                                       Sequence<7>{}));
 
-                const auto out_gemmak0_gemmmraw_gemmak1_grid_desc = transform_tensor_descriptor(
-                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_ak0_ak1_grid_desc,
+                const auto out_gemmk_gemmmraw_grid_desc = transform_tensor_descriptor(
+                    out_n_zdotslice_dtildeslice_ydotslice_htildeslice_xdotslice_wtildeslice_k_grid_desc,
                     make_tuple(
-                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, AK0)),
-                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice)),
-                        make_pass_through_transform(AK1)),
-                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}, Sequence<8>{}),
-                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K)),
+                        make_merge_transform(make_tuple(N, DTildeSlice, HTildeSlice, WTildeSlice))),
+                    make_tuple(Sequence<1, 3, 5, 7>{}, Sequence<0, 2, 4, 6>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-                const auto out_gemmak0_gemmm_gemmak1_grid_desc =
+                const auto out_gemmk_gemmm_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
-                        out_gemmak0_gemmmraw_gemmak1_grid_desc,
-                        make_tuple(AK0, GemmMPerBlock, AK1),
-                        Sequence<false, DoPadGemmM, false>{});
+                        out_gemmk_gemmmraw_grid_desc,
+                        make_tuple(AK1, GemmMPerBlock),
+                        Sequence<true, DoPadGemmM>{});
+
+                const auto out_gemmak0_gemmm_gemmak1_grid_desc = transform_tensor_descriptor(
+                    out_gemmk_gemmm_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
+                               make_pass_through_transform(
+                                   out_gemmk_gemmm_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
                 return out_gemmak0_gemmm_gemmak1_grid_desc;
             }
@@ -505,8 +522,6 @@ struct TransformConvBwdDataToGemm_v1
         const index_t ConvDilationH = conv_filter_dilations[HIdx - NonSpatialDimsNum];
         const index_t ConvDilationW = conv_filter_dilations[WIdx - NonSpatialDimsNum];
 
-        const index_t BK0 = K / BK1;
-
         // assume packed
         // k_y_x_c for 2d or k_z_y_x_c for 3d
         const auto wei_grid_desc = make_wei_grid_desc<BLayout>(K, Z, Y, X, C);
@@ -515,6 +530,8 @@ struct TransformConvBwdDataToGemm_v1
                      ck::tensor_operation::device::ConvolutionBackwardDataSpecialization::
                          Filter1x1Stride1Pad0)
         {
+            const index_t BK0 = math::integer_divide_ceil(K, BK1);
+
             // B: weight tensor
             const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc =
                 transform_tensor_descriptor(make_naive_tensor_descriptor_packed(make_tuple(K, C)),
@@ -551,6 +568,9 @@ struct TransformConvBwdDataToGemm_v1
             const auto YDotSlice = math::integer_divide_ceil(Y - i_ytilde, YTilde);
             const auto XDotSlice = math::integer_divide_ceil(X - i_xtilde, XTilde);
 
+            const index_t BK0 =
+                math::integer_divide_ceil(ZDotSlice * YDotSlice * XDotSlice * K, BK1);
+
             // B weight tensor
             if constexpr(NDimSpatial == 2)
             {
@@ -566,43 +586,47 @@ struct TransformConvBwdDataToGemm_v1
                     make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
                     make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3, 4>{}, Sequence<5>{}));
 
-                const auto wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc =
-                    transform_tensor_descriptor(
-                        wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
-                        make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
-                                   make_slice_transform(YDot, I0, YDotSlice),
-                                   make_slice_transform(XDot, I0, XDotSlice),
-                                   make_freeze_transform(i_ytilde),
-                                   make_freeze_transform(i_xtilde),
-                                   make_pass_through_transform(C)),
-                        make_tuple(Sequence<0>{},
-                                   Sequence<1>{},
-                                   Sequence<3>{},
-                                   Sequence<2>{},
-                                   Sequence<4>{},
-                                   Sequence<5>{}),
-                        make_tuple(Sequence<0, 1>{},
-                                   Sequence<2>{},
-                                   Sequence<3>{},
-                                   Sequence<>{},
-                                   Sequence<>{},
-                                   Sequence<4>{}));
+                const auto wei_k_ydotslice_xdotslice_c_grid_desc = transform_tensor_descriptor(
+                    wei_k_ydot_ytilde_xdot_xtilde_c_grid_desc,
+                    make_tuple(make_pass_through_transform(K),
+                               make_slice_transform(YDot, I0, YDotSlice),
+                               make_slice_transform(XDot, I0, XDotSlice),
+                               make_freeze_transform(i_ytilde),
+                               make_freeze_transform(i_xtilde),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<3>{},
+                               Sequence<2>{},
+                               Sequence<4>{},
+                               Sequence<5>{}),
+                    make_tuple(Sequence<0>{},
+                               Sequence<1>{},
+                               Sequence<2>{},
+                               Sequence<>{},
+                               Sequence<>{},
+                               Sequence<3>{}));
 
-                const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
-                    wei_bk0_bk1_ydotslice_xdotslice_c_grid_desc,
-                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, BK0)),
-                               make_pass_through_transform(C),
-                               make_pass_through_transform(BK1)),
-                    make_tuple(Sequence<2, 3, 0>{}, Sequence<4>{}, Sequence<1>{}),
-                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+                const auto wei_gemmk_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    wei_k_ydotslice_xdotslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(YDotSlice, XDotSlice, K)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1, 2, 0>{}, Sequence<3>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                const auto wei_gemmk_gemmn_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
-                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
-                        make_tuple(wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0),
-                                   GemmNPerBlock,
-                                   BK1),
-                        Sequence<false, DoPadGemmN, false>{});
+                        wei_gemmk_gemmnraw_grid_desc,
+                        make_tuple(BK1, GemmNPerBlock),
+                        Sequence<true, DoPadGemmN>{});
+
+                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_gemmk_gemmn_padded_grid_desc,
+                    make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                               make_pass_through_transform(
+                                   wei_gemmk_gemmn_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
 
                 return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
             }
@@ -631,10 +655,10 @@ struct TransformConvBwdDataToGemm_v1
                                    Sequence<5, 6>{},
                                    Sequence<7>{}));
 
-                const auto wei_bk0_bk1_zdotslice_ydotslice_xdotslice_c_grid_desc =
+                const auto wei_gemmk_zdotslice_ydotslice_xdotslice_c_grid_desc =
                     transform_tensor_descriptor(
                         wei_k_zdot_ztilde_ydot_ytilde_xdot_xtilde_c_grid_desc,
-                        make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
+                        make_tuple(make_pass_through_transform(K),
                                    make_slice_transform(ZDot, I0, ZDotSlice),
                                    make_slice_transform(YDot, I0, YDotSlice),
                                    make_slice_transform(XDot, I0, XDotSlice),
@@ -650,33 +674,37 @@ struct TransformConvBwdDataToGemm_v1
                                    Sequence<4>{},
                                    Sequence<6>{},
                                    Sequence<7>{}),
-                        make_tuple(Sequence<0, 1>{},
+                        make_tuple(Sequence<0>{},
+                                   Sequence<1>{},
                                    Sequence<2>{},
                                    Sequence<3>{},
-                                   Sequence<4>{},
                                    Sequence<>{},
                                    Sequence<>{},
                                    Sequence<>{},
-                                   Sequence<5>{}));
+                                   Sequence<4>{}));
 
-                const auto wei_gemmbk0_gemmnraw_gemmbk1_grid_desc = transform_tensor_descriptor(
-                    wei_bk0_bk1_zdotslice_ydotslice_xdotslice_c_grid_desc,
-                    make_tuple(
-                        make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, BK0)),
-                        make_pass_through_transform(C),
-                        make_pass_through_transform(BK1)),
-                    make_tuple(Sequence<2, 3, 4, 0>{}, Sequence<5>{}, Sequence<1>{}),
-                    make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
+                const auto wei_gemmk_gemmnraw_grid_desc = transform_tensor_descriptor(
+                    wei_gemmk_zdotslice_ydotslice_xdotslice_c_grid_desc,
+                    make_tuple(make_merge_transform(make_tuple(ZDotSlice, YDotSlice, XDotSlice, K)),
+                               make_pass_through_transform(C)),
+                    make_tuple(Sequence<1, 2, 3, 0>{}, Sequence<4>{}),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}));
 
-                const auto wei_gemmbk0_gemmn_gemmbk1_grid_desc =
+                const auto wei_gemmk_gemm_padded_grid_desc =
                     ck::tensor_operation::device::PadTensorDescriptor(
-                        wei_gemmbk0_gemmnraw_gemmbk1_grid_desc,
-                        make_tuple(wei_gemmbk0_gemmnraw_gemmbk1_grid_desc.GetLength(I0),
-                                   GemmNPerBlock,
-                                   BK1),
-                        Sequence<false, DoPadGemmN, false>{});
+                        wei_gemmk_gemmnraw_grid_desc,
+                        make_tuple(BK1, GemmNPerBlock),
+                        Sequence<true, DoPadGemmN>{});
 
-                return wei_gemmbk0_gemmn_gemmbk1_grid_desc;
+                const auto wei_gemmbk0_gemm_gemmbk1_grid_desc = transform_tensor_descriptor(
+                    wei_gemmk_gemm_padded_grid_desc,
+                    make_tuple(
+                        make_unmerge_transform(make_tuple(BK0, BK1)),
+                        make_pass_through_transform(wei_gemmk_gemm_padded_grid_desc.GetLength(I1))),
+                    make_tuple(Sequence<0>{}, Sequence<1>{}),
+                    make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
+
+                return wei_gemmbk0_gemm_gemmbk1_grid_desc;
             }
             else
             {

include/ck/utility/inner_product.hpp

@@ -94,8 +94,8 @@ __device__ void inner_product<half2_t, half2_t, float>(const half2_t& a, const h
     const vector_type<half_t, 2> b_vector{b};
 
     static_for<0, 2, 1>{}([&](auto i) {
-        c += type_convert<int32_t>(a_vector.AsType<half_t>()[i]) *
-             type_convert<int32_t>(b_vector.AsType<half_t>()[i]);
+        c += type_convert<float>(a_vector.AsType<half_t>()[i]) *
+             type_convert<float>(b_vector.AsType<half_t>()[i]);
     });
 #endif
 }

include/ck/version.h.in

+/*******************************************************************************
+ *
+ * MIT License
+ *
+ * Copyright (c) 2023 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.
+ *
+ *******************************************************************************/
+
+/* the configured version and settings for miopen- Composable Kernel */
+
+#ifndef CK_VERSION_H_
+#define CK_VERSION_H_
+
+// clang-format off
+#define CK_VERSION @CMAKE_PROJECT_VERSION@
+#define CK_VERSION_MAJOR @CMAKE_PROJECT_VERSION_MAJOR@
+#define CK_VERSION_MINOR @CMAKE_PROJECT_VERSION_MINOR@
+#define CK_VERSION_PATCH @CMAKE_PROJECT_VERSION_PATCH@
+#define CK_COMMIT_ID @COMMIT_ID@
+// clang-format on
+
+#endif

library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp

@@ -92,11 +92,11 @@ struct ReferenceGemm : public device::BaseOperator
                         ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
                 }
 
-                AccDataType v_c;
+                CDataType v_c;
 
                 arg.c_element_op_(v_c, v_acc);
 
-                arg.c_m_n_(m, n) = ck::type_convert<CDataType>(v_c);
+                arg.c_m_n_(m, n) = v_c;
             };
 
             make_ParallelTensorFunctor(

library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp

@@ -32,6 +32,8 @@ using F32_Tuple     = ck::Tuple<F32>;
 using I32_Tuple     = ck::Tuple<I32>;
 using I32_F32_Tuple = ck::Tuple<I32, F32>;
 
+using F32_F32_Tuple = ck::Tuple<F32, F32>;
+
 // GEMM layout
 using Row = ck::tensor_layout::gemm::RowMajor;
 using Col = ck::tensor_layout::gemm::ColumnMajor;
@@ -95,10 +97,11 @@ using AddFastGelu    = ck::tensor_operation::element_wise::AddFastGelu;
 using AddReluAdd     = ck::tensor_operation::element_wise::AddReluAdd;
 using FastGelu       = ck::tensor_operation::element_wise::FastGelu;
 using AddMultiply    = ck::tensor_operation::element_wise::AddMultiply;
+using MultiplyAdd    = ck::tensor_operation::element_wise::MultiplyAdd;
 using ScaleAdd       = ck::tensor_operation::element_wise::ScaleAdd;
 using Gelu           = ck::tensor_operation::element_wise::Gelu;
 using Swish          = ck::tensor_operation::element_wise::Swish;
-using AddBias        = ck::tensor_operation::element_wise::AddBias;
+using Add            = ck::tensor_operation::element_wise::Add;
 
 template <typename Activation>
 using Activation_Mul_Clamp = ck::tensor_operation::element_wise::Activation_Mul_Clamp<Activation>;

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
 void add_device_batched_gemm_xdl_bf16_bf16_bf16_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Col, Row, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
@@ -37,7 +37,7 @@ void add_device_batched_gemm_xdl_bf16_bf16_bf16_gmk_gnk_gmn_instances(
         DeviceBatchedGemm<Row, Col, Row, BF16, BF16, BF16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_batched_gemm_xdl_f16_f16_f16_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Col, Row, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
@@ -58,7 +58,7 @@ void add_device_batched_gemm_xdl_f16_f16_f16_gmk_gnk_gmn_instances(
         DeviceBatchedGemm<Row, Col, Row, F16, F16, F16, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
 void add_device_batched_gemm_xdl_f32_f32_f32_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemm<Col, Row, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
@@ -79,7 +79,7 @@ void add_device_batched_gemm_xdl_f32_f32_f32_gmk_gnk_gmn_instances(
         DeviceBatchedGemm<Row, Col, Row, F32, F32, F32, PassThrough, PassThrough, PassThrough>>>&
         instances);
 #endif
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 void add_device_batched_gemm_xdl_int8_int8_int8_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<DeviceBatchedGemm<Col,
                                                   Row,
@@ -154,7 +154,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-#ifdef __fp32__
+#ifdef CK_ENABLE_FP32
         if constexpr(is_same_v<ADataType, float> && is_same_v<BDataType, float> &&
                      is_same_v<CDataType, float>)
         {
@@ -180,7 +180,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
             }
         }
 #endif
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
                      is_same_v<CDataType, half_t>)
         {
@@ -206,7 +206,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
             }
         }
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
         if constexpr(is_same_v<ADataType, bhalf_t> && is_same_v<BDataType, bhalf_t> &&
                      is_same_v<CDataType, bhalf_t>)
         {
@@ -232,7 +232,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
             }
         }
 #endif
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
         if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
                      is_same_v<CDataType, int8_t>)
         {

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

@@ -14,7 +14,7 @@
 
 using CDE0ElementOp = ck::tensor_operation::element_wise::AddRelu;
 using CDE1ElementOp = ck::tensor_operation::element_wise::Add;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -13,7 +13,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemmSoftmaxGemmPermute<2,
@@ -59,7 +59,7 @@ void add_device_batched_gemm_bias_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_
                                                             MaskingSpecialization::MaskDisabled>>>&
         instances);
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
 void add_device_batched_gemm_bias_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemmSoftmaxGemmPermute<2,
@@ -148,7 +148,7 @@ struct DeviceOperationInstanceFactory<
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
                      is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t> &&
                      Acc0BiasDataType::Size() == 1 &&
@@ -166,7 +166,7 @@ struct DeviceOperationInstanceFactory<
             }
         }
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
         else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
                           is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16> &&
                           Acc0BiasDataType::Size() == 1 &&

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_batched_gemm_gemm_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instance(
     std::vector<std::unique_ptr<DeviceBatchedGemmGemm<Row,
                                                       Col,

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

@@ -19,7 +19,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
                                                         Row,
@@ -124,7 +124,7 @@ void add_device_batched_gemm_multi_d_dl_f16_f16_f16_gmk_gnk_gmn_irregular_instan
                                                         PassThrough,
                                                         PassThrough>>>& instances);
 #endif
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
 void add_device_batched_gemm_multi_d_dl_i8_i8_i8_gkm_gkn_gmn_instances(
     std::vector<std::unique_ptr<DeviceBatchedGemmMultiD<Col,
                                                         Row,
@@ -263,7 +263,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
                      is_same_v<EDataType, half_t>)
         {
@@ -297,7 +297,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceBatche
             }
         }
 #endif
-#ifdef __int8__
+#ifdef CK_ENABLE_INT8
         else if constexpr(is_same_v<ADataType, int8_t> && is_same_v<BDataType, int8_t> &&
                           is_same_v<EDataType, int8_t>)
         {

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

@@ -11,7 +11,7 @@
 #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
 
 #include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 namespace ck {
 namespace tensor_operation {
 namespace device {

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

@@ -16,7 +16,7 @@ namespace ck {
 namespace tensor_operation {
 namespace device {
 namespace instance {
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
 void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_gmk_gnk_gno_gmo_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemmSoftmaxGemmPermute<2,
@@ -59,7 +59,7 @@ void add_device_batched_gemm_softmax_gemm_permute_xdl_cshuffle_f16_f16_f16_f16_g
                                                             MaskingSpecialization::MaskDisabled>>>&
         instances);
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
 void add_device_batched_gemm_masking_softmax_gemm_permute_xdl_cshuffle_bf16_bf16_bf16_bf16_gmk_gnk_gno_gmo_instances(
     std::vector<std::unique_ptr<
         DeviceBatchedGemmSoftmaxGemmPermute<2,
@@ -148,7 +148,7 @@ struct DeviceOperationInstanceFactory<
     static auto GetInstances()
     {
         std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
-#ifdef __fp16__
+#ifdef CK_ENABLE_FP16
         if constexpr(is_same_v<ADataType, half_t> && is_same_v<B0DataType, half_t> &&
                      is_same_v<B1DataType, half_t> && is_same_v<CDataType, half_t>)
         {
@@ -164,7 +164,7 @@ struct DeviceOperationInstanceFactory<
             }
         }
 #endif
-#ifdef __bf16__
+#ifdef CK_ENABLE_BF16
         else if constexpr(is_same_v<ADataType, BF16> && is_same_v<B0DataType, BF16> &&
                           is_same_v<B1DataType, BF16> && is_same_v<CDataType, BF16>)
         {