diff --git a/include/ck/ck.hpp b/include/ck/ck.hpp index 0e7e690421..734fe549db 100644 --- a/include/ck/ck.hpp +++ b/include/ck/ck.hpp @@ -200,9 +200,6 @@ // workaround: compiler issue on gfx908 #define CK_WORKAROUND_SWDEV_388832 1 -// workaround: Grouped Conv2d_bwd_data fails for already implemented instance -#define CK_WORKAROUND_GITHUB_ISSUE_824 1 - // flag to enable (1) or disable (0) the debugging output in some kernels #define DEBUG_LOG 0 diff --git a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp index ec1cc53991..5437379018 100644 --- a/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp +++ b/include/ck/tensor_operation/gpu/grid/gridwise_gemm_multiple_d_xdl_cshuffle.hpp @@ -266,12 +266,13 @@ struct GridwiseGemmMultipleD_xdl_cshuffle (NPerBlock % (NXdlPerWave * NPerXdl)) == 0, "Invalid tuning param!"); - const auto M = a_grid_desc_m_k.GetLength(I0); - const auto N = b_grid_desc_n_k.GetLength(I0); - const auto K = a_grid_desc_m_k.GetLength(I1); + const auto M = a_grid_desc_m_k.GetLength(I0); + const auto N = b_grid_desc_n_k.GetLength(I0); + const auto AK = a_grid_desc_m_k.GetLength(I1); + const auto BK = b_grid_desc_n_k.GetLength(I1); // check consistency of desc - if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1))) + if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1) && AK == BK)) { return false; } @@ -289,13 +290,13 @@ struct GridwiseGemmMultipleD_xdl_cshuffle } // check tile size - if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0)) + if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && AK % KPerBlock == 0)) { return false; } // check gridwise gemm pipeline - const auto num_k_loop = K / KPerBlock; + const auto num_k_loop = AK / KPerBlock; if(!GridwiseGemmPipe::IsSupported(num_k_loop)) { diff --git a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp b/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp index 6665d765f8..32ea8ae39c 100644 --- a/include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp +++ b/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 = diff --git a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp b/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp index 98a46ef450..b1b203b43f 100644 --- a/include/ck/tensor_operation/operator_transform/transform_conv_bwd_data_to_gemm_v1.hpp +++ b/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( @@ -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{}); + out_gemmk_gemmmraw_grid_desc, + make_tuple(AK1, GemmMPerBlock), + Sequence{}); + + 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{}); + out_gemmk_gemmmraw_grid_desc, + make_tuple(AK1, GemmMPerBlock), + Sequence{}); + + 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(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{}); + wei_gemmk_gemmnraw_grid_desc, + make_tuple(BK1, GemmNPerBlock), + Sequence{}); + + 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{}); + wei_gemmk_gemmnraw_grid_desc, + make_tuple(BK1, GemmNPerBlock), + Sequence{}); - 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 { diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp index d05e8e3a87..26acf4f5f7 100644 --- a/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_conv_bwd_data/device_grouped_conv_bwd_data_xdl_instance.hpp @@ -47,24 +47,26 @@ using device_grouped_conv_bwd_data_xdl_f16_instances = // ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector| // ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| // ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8> - -#ifndef CK_WORKAROUND_GITHUB_ISSUE_824 - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - // TODO: After enable, add instance for small conv.K and conv.C -#endif + // generic instance + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, + + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; @@ -75,33 +77,34 @@ template -using device_grouped_conv_bwd_data_xdl_bf16_instances = - std::tuple< - // clang-format off - // ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer| - // ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector| - // ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| - // ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8> - -#ifndef CK_WORKAROUND_GITHUB_ISSUE_824 - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> - // TODO: After enable, add instance for small conv.K and conv.C -#endif - // clang-format on - >; +using device_grouped_conv_bwd_data_xdl_bf16_instances = std::tuple< + // clang-format off + // ##############################################| NDim| ALayout| BLayout| DsLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| AElementwise| BElementwise| CDEElementwise| ConvolutionBackward| DoPad| DoPad| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffleMXdl| CShuffleNXdl| CDEBlockTransfer| CDEBlockTransfer| + // ##############################################| Spatial| | | | | Type| Type| Type| DataType| Type| Type| Operation| Operation| Operation| DataSpecialization| GemmM| GemmN| PrefetchStage| Size| Block| Block| Block| | | XDL| XDL| PerWave| PerWave| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| PerWave| PerWave| _MBlock_MPerBlock| ScalarPerVector| + // ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | Lengths_AK0_M_AK1| ArrangeOrder| | | PerVector| PerVector_AK1| | Lengths_BK0_N_BK1| ArrangeOrder| | | PerVector| PerVector_BK1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| + // ##############################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + // generic instance + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 16, 1, 4>, 1>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 4>, 1>, + + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 4, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> + // clang-format on + >; // f32_f32_f32_f32 template , S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4> + // generic instance + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 16, 1, 4>, 1>, + // instances for small conv.K and conv.C + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 1, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 4, 1, 1, 1, S<1, 32, 1, 4>, 1>, -#ifndef CK_WORKAROUND_GITHUB_ISSUE_824 - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4>, - DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>, - // TODO: After enable, add instance for small conv.K and conv.C -#endif + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 32, 1, 4>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 8>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 4, 1, 1, 1, S<1, 16, 1, 4>, 4>, + DeviceGroupedConvBwdDataMultipleD_Xdl_CShuffle_v1< NDimSpatial, ALayout, BLayout, DsLayout, ELayout, F32, F32, F32, F32, Empty_Tuple, F32, PassThrough, PassThrough, PassThrough, ConvSpec, true, true, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 8, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8> // clang-format on >; diff --git a/test/contraction/test_contraction_interface.cpp b/test/contraction/test_contraction_interface.cpp index c9e720c597..12a307f5da 100644 --- a/test/contraction/test_contraction_interface.cpp +++ b/test/contraction/test_contraction_interface.cpp @@ -38,7 +38,7 @@ class ContractionInstanceWrapper //#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector| //#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl| //#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | - DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>; + DeviceContractionMultipleD_Xdl_CShuffle< NumDim, NumDim, NumDim, F32, F32, F32, F32, ck::Tuple, F32, Pass, Pass, Bilinear, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, ABlockTransferSrcVectorDim, 4, 4, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, BBlockTransferSrcVectorDim, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, CDEBlockTransferScalarPerVector>; // clang-format on bool isSupported(std::vector& ADims, diff --git a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp index c2aab749d3..bdad37fe14 100644 --- a/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp +++ b/test/grouped_convnd_bwd_data/test_grouped_convnd_bwd_data.cpp @@ -87,6 +87,9 @@ TYPED_TEST(TestGroupedConvndBwdData2d, Test2D) {2, 2, 128, 128, 256, {1, 1}, {7, 7}, {2, 2}, {1, 1}, {0, 0}, {0, 0}}); this->conv_params.push_back( {2, 2, 128, 128, 256, {1, 1}, {3, 3}, {1, 1}, {1, 1}, {0, 0}, {0, 0}}); + this->conv_params.push_back({2, 1, 1, 1, 32, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); + this->conv_params.push_back({2, 1, 1, 64, 3, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); + this->conv_params.push_back({2, 1, 1, 1, 1, {8, 8}, {32, 32}, {1, 1}, {1, 1}, {1, 1}, {1, 1}}); this->template Run<2>(); } @@ -99,5 +102,11 @@ TYPED_TEST(TestGroupedConvndBwdData3d, Test3D) {3, 2, 2, 128, 256, {3, 3, 3}, {14, 14, 3}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); this->conv_params.push_back( {3, 2, 32, 128, 256, {1, 1, 1}, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}, {0, 0, 0}, {0, 0, 0}}); + this->conv_params.push_back( + {3, 1, 1, 1, 32, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); + this->conv_params.push_back( + {3, 1, 1, 64, 3, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); + this->conv_params.push_back( + {3, 1, 1, 1, 1, {3, 3, 3}, {32, 32, 32}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}, {1, 1, 1}}); this->template Run<3>(); } diff --git a/test/grouped_gemm/test_grouped_gemm_util.hpp b/test/grouped_gemm/test_grouped_gemm_util.hpp index b61118b512..04b31dcc91 100644 --- a/test/grouped_gemm/test_grouped_gemm_util.hpp +++ b/test/grouped_gemm/test_grouped_gemm_util.hpp @@ -147,14 +147,14 @@ struct DeviceGroupedGemmSplitkInstanceWrapper 32, 4, 2, - S<1, 4, 32, 1>, + S<1, 4, 16, 1>, ABlockTransferThreadClusterArrageOrder, ABlockTransferSrcAccessOrder, ABlockTransferSrcVectorDim::value, ABlockTransferSrcScalarPerVector, ABlockTransferDstScalarPerVector_K1::value, ABlockLdsAddExtraM::value, - S<1, 4, 32, 1>, + S<1, 4, 16, 1>, BBlockTransferThreadClusterArrageOrder, BBlockTransferSrcAccessOrder, BBlockTransferSrcVectorDim::value,