From 4ec3c47404d050473ff0063c7fa56f0959f29b68 Mon Sep 17 00:00:00 2001 From: ltqin Date: Tue, 1 Nov 2022 00:24:25 +0800 Subject: [PATCH] Add Conv Forward on Navi21 for ResNet50 (#490) * add device of dl * fix k1 of GridwiseGemmDl_km_kn_mn_v1r3 * init version for dl conv * add example(init) * result right * disable elementwise operation * check parameters * add fp32,int8 example and change check code * change deive file and class name * add check vector access of C * add instance * add to ckProfiler * add Filter1x1Pad0 instances * fix ignore error * fix for CI Co-authored-by: letaoqin [ROCm/composable_kernel commit: 8ee36118be9b19b15c2471bffeeeb624afb14044] --- example/09_convnd_fwd/CMakeLists.txt | 4 + .../09_convnd_fwd/convnd_fwd_dl_common.hpp | 170 ++++ example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp | 39 + example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp | 39 + example/09_convnd_fwd/convnd_fwd_dl_int8.cpp | 39 + .../run_convnd_fwd_dl_example.inc | 97 ++ .../gpu/device/device_grouped_conv_fwd.hpp | 56 ++ ...ice_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp | 837 ++++++++++++++++++ .../gpu/device/impl/device_gemm_dl.hpp | 1 + .../gpu/grid/gridwise_gemm_dl_v1r3.hpp | 3 +- .../gpu/grouped_convolution_forward_dl.hpp | 118 +++ .../gpu/grouped_conv2d_fwd/CMakeLists.txt | 4 + ..._fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp | 101 +++ ..._fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp | 102 +++ ...fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp | 102 +++ .../include/profile_grouped_conv_fwd_impl.hpp | 134 ++- 16 files changed, 1803 insertions(+), 43 deletions(-) create mode 100644 example/09_convnd_fwd/convnd_fwd_dl_common.hpp create mode 100644 example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp create mode 100644 example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp create mode 100644 example/09_convnd_fwd/convnd_fwd_dl_int8.cpp create mode 100644 example/09_convnd_fwd/run_convnd_fwd_dl_example.inc create mode 100644 include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp create mode 100644 include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp create mode 100644 library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp create mode 100644 library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp diff --git a/example/09_convnd_fwd/CMakeLists.txt b/example/09_convnd_fwd/CMakeLists.txt index b373d1d6c0..1428f49fc1 100644 --- a/example/09_convnd_fwd/CMakeLists.txt +++ b/example/09_convnd_fwd/CMakeLists.txt @@ -4,3 +4,7 @@ add_example_executable(example_convnd_fwd_xdl_bf16 convnd_fwd_xdl_bf16.cpp) add_example_executable(example_convnd_fwd_xdl_int8 convnd_fwd_xdl_int8.cpp) # FIXME: re-enable this exampe as test when SWDEV-335738 is fixed add_example_executable_no_testing(example_convnd_fwd_xdl_fp64 convnd_fwd_xdl_fp64.cpp) + +add_example_executable(example_convnd_fwd_dl_fp16 convnd_fwd_dl_fp16.cpp) +add_example_executable(example_convnd_fwd_dl_fp32 convnd_fwd_dl_fp32.cpp) +add_example_executable(example_convnd_fwd_dl_int8 convnd_fwd_dl_int8.cpp) diff --git a/example/09_convnd_fwd/convnd_fwd_dl_common.hpp b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp new file mode 100644 index 0000000000..9fe09a1f2e --- /dev/null +++ b/example/09_convnd_fwd/convnd_fwd_dl_common.hpp @@ -0,0 +1,170 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include +#include +#include +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/utility/check_err.hpp" +#include "ck/library/utility/device_memory.hpp" +#include "ck/library/utility/host_tensor.hpp" +#include "ck/library/utility/host_tensor_generator.hpp" +#include "ck/library/utility/convolution_parameter.hpp" +#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" +#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp" + +void print_helper_msg() +{ + std::cout << "arg1: verification (0=no, 1=yes)\n" + << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n" + << "arg3: time kernel (0=no, 1=yes)\n" + << ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl; +} + +template +bool run_grouped_conv_fwd_dl(bool do_verification, + int init_method, + bool time_kernel, + const ck::utils::conv::ConvParam& conv_param, + const HostTensorDescriptor& in_g_n_c_wis_desc, + const HostTensorDescriptor& wei_g_k_c_xs_desc, + const HostTensorDescriptor& out_g_n_k_wos_desc, + const InElementOp& in_element_op, + const WeiElementOp& wei_element_op, + const OutElementOp& out_element_op) +{ + Tensor in(in_g_n_c_wis_desc); + Tensor wei(wei_g_k_c_xs_desc); + Tensor out_host(out_g_n_k_wos_desc); + Tensor out_device(out_g_n_k_wos_desc); + + std::cout << "in: " << in.mDesc << std::endl; + std::cout << "wei: " << wei.mDesc << std::endl; + std::cout << "out: " << out_host.mDesc << std::endl; + + switch(init_method) + { + case 0: break; + case 1: + in.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + wei.GenerateTensorValue(GeneratorTensor_2{-5, 5}); + break; + case 2: + in.GenerateTensorValue(GeneratorTensor_3{0.0, 1.0}); + wei.GenerateTensorValue(GeneratorTensor_3{-0.5, 0.5}); + break; + default: + in.GenerateTensorValue(GeneratorTensor_1{1}); + wei.GenerateTensorValue(GeneratorTensor_1{1}); + } + + DeviceMem in_device_buf(sizeof(InDataType) * in.mDesc.GetElementSpaceSize()); + DeviceMem wei_device_buf(sizeof(WeiDataType) * wei.mDesc.GetElementSpaceSize()); + DeviceMem out_device_buf(sizeof(OutDataType) * out_device.mDesc.GetElementSpaceSize()); + + in_device_buf.ToDevice(in.mData.data()); + wei_device_buf.ToDevice(wei.mData.data()); + + std::array a_g_n_c_wis_lengths{}; + std::array a_g_n_c_wis_strides{}; + std::array b_g_k_c_xs_lengths{}; + std::array b_g_k_c_xs_strides{}; + std::array c_g_n_k_wos_lengths{}; + std::array c_g_n_k_wos_strides{}; + std::array conv_filter_strides{}; + std::array conv_filter_dilations{}; + std::array input_left_pads{}; + std::array input_right_pads{}; + + auto copy = [](auto& x, auto& y) { std::copy(x.begin(), x.end(), y.begin()); }; + + copy(in_g_n_c_wis_desc.GetLengths(), a_g_n_c_wis_lengths); + copy(in_g_n_c_wis_desc.GetStrides(), a_g_n_c_wis_strides); + copy(wei_g_k_c_xs_desc.GetLengths(), b_g_k_c_xs_lengths); + copy(wei_g_k_c_xs_desc.GetStrides(), b_g_k_c_xs_strides); + copy(out_g_n_k_wos_desc.GetLengths(), c_g_n_k_wos_lengths); + copy(out_g_n_k_wos_desc.GetStrides(), c_g_n_k_wos_strides); + copy(conv_param.conv_filter_strides_, conv_filter_strides); + copy(conv_param.conv_filter_dilations_, conv_filter_dilations); + copy(conv_param.input_left_pads_, input_left_pads); + copy(conv_param.input_right_pads_, input_right_pads); + + // do Conv + auto conv = DeviceConvNDFwdInstance{}; + auto invoker = conv.MakeInvoker(); + auto argument = conv.MakeArgument(in_device_buf.GetDeviceBuffer(), + wei_device_buf.GetDeviceBuffer(), + out_device_buf.GetDeviceBuffer(), + a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + c_g_n_k_wos_lengths, + c_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads, + in_element_op, + wei_element_op, + out_element_op); + + if(!conv.IsSupportedArgument(argument)) + { + return true; + } + + float avg_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel}); + + std::size_t flop = conv_param.GetFlops(); + std::size_t num_btype = conv_param.GetByte(); + + float tflops = static_cast(flop) / 1.E9 / avg_time; + float gb_per_sec = num_btype / 1.E6 / avg_time; + std::cout << "Perf: " << avg_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, " + << conv.GetTypeString() << std::endl; + + if(do_verification) + { + auto ref_conv = ck::tensor_operation::host::ReferenceConvFwd(); + + auto ref_invoker = ref_conv.MakeInvoker(); + auto ref_argument = ref_conv.MakeArgument(in, + wei, + out_host, + conv_param.conv_filter_strides_, + conv_param.conv_filter_dilations_, + conv_param.input_left_pads_, + conv_param.input_right_pads_, + in_element_op, + wei_element_op, + out_element_op); + + ref_invoker.Run(ref_argument); + + out_device_buf.FromDevice(out_device.mData.data()); + + return ck::utils::check_err( + out_device.mData, out_host.mData, "Error: incorrect results!", 1e-5f, 1e-4f); + } + + return true; +} diff --git a/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp new file mode 100644 index 0000000000..9f3b049741 --- /dev/null +++ b/example/09_convnd_fwd/convnd_fwd_dl_fp16.cpp @@ -0,0 +1,39 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include "convnd_fwd_dl_common.hpp" + +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp" + +#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" + +using InDataType = ck::half_t; +using WeiDataType = ck::half_t; +using AccDataType = float; +using OutDataType = ck::half_t; + +template +using S = ck::Sequence; + +using InElementOp = ck::tensor_operation::element_wise::PassThrough; +using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; +using OutElementOp = ck::tensor_operation::element_wise::PassThrough; + +static constexpr auto ConvSpec = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; + +static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +template +// clang-format off +using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK +// ######| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| +// ######| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| +// ######| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | +// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + < NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>; +// clang-format on + +#include "run_convnd_fwd_dl_example.inc" + +int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; } diff --git a/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp new file mode 100644 index 0000000000..bfb241d768 --- /dev/null +++ b/example/09_convnd_fwd/convnd_fwd_dl_fp32.cpp @@ -0,0 +1,39 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include "convnd_fwd_dl_common.hpp" + +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp" + +#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" + +using InDataType = float; +using WeiDataType = float; +using AccDataType = float; +using OutDataType = float; + +template +using S = ck::Sequence; + +using InElementOp = ck::tensor_operation::element_wise::PassThrough; +using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; +using OutElementOp = ck::tensor_operation::element_wise::PassThrough; + +static constexpr auto ConvSpec = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; + +static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +template +// clang-format off +using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK +// ######| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| +// ######| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| +// ######| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | +// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + < NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>; +// clang-format on + +#include "run_convnd_fwd_dl_example.inc" + +int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; } diff --git a/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp new file mode 100644 index 0000000000..142e360baa --- /dev/null +++ b/example/09_convnd_fwd/convnd_fwd_dl_int8.cpp @@ -0,0 +1,39 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include "convnd_fwd_dl_common.hpp" + +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp" + +#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" + +using InDataType = int8_t; +using WeiDataType = int8_t; +using AccDataType = int32_t; +using OutDataType = int8_t; + +template +using S = ck::Sequence; + +using InElementOp = ck::tensor_operation::element_wise::PassThrough; +using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; +using OutElementOp = ck::tensor_operation::element_wise::PassThrough; + +static constexpr auto ConvSpec = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; + +static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +template +// clang-format off +using DeviceGroupedConvNDFwdInstance = ck::tensor_operation::device::DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK +// ######| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| +// ######| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| +// ######| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | +// ######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + < NDimSpatial, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4>; +// clang-format on + +#include "run_convnd_fwd_dl_example.inc" + +int main(int argc, char* argv[]) { return run_convnd_fwd_dl_example(argc, argv) ? 0 : 1; } diff --git a/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc new file mode 100644 index 0000000000..52a285b566 --- /dev/null +++ b/example/09_convnd_fwd/run_convnd_fwd_dl_example.inc @@ -0,0 +1,97 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +bool run_convnd_fwd_dl_example(int argc, char* argv[]) +{ + print_helper_msg(); + bool do_verification = true; + int init_method = 1; + bool time_kernel = false; + + ck::utils::conv::ConvParam conv_param{ + 2, 1, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}; + + if(argc == 1) + { + // use default + } + else if(argc == 4) + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + } + else + { + do_verification = std::stoi(argv[1]); + init_method = std::stoi(argv[2]); + time_kernel = std::stoi(argv[3]); + const ck::index_t num_dim_spatial = std::stoi(argv[4]); + + conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv); + } + + const auto in_element_op = InElementOp{}; + const auto wei_element_op = WeiElementOp{}; + const auto out_element_op = OutElementOp{}; + + const auto run = [&](auto ndim_spatial, auto in_layout, auto wei_layout, auto out_layout) { + constexpr ck::index_t ndim_spatial_value = ndim_spatial.value; + std::cout << "ndim_spatial_value: " << ndim_spatial_value << std::endl; + + using InLayout = decltype(in_layout); + using WeiLayout = decltype(wei_layout); + using OutLayout = decltype(out_layout); + + const auto in_g_n_c_wis_desc = + ck::utils::conv::make_input_host_tensor_descriptor_g_n_c_wis_packed( + conv_param); + + const auto wei_g_k_c_xs_desc = + ck::utils::conv::make_weight_host_tensor_descriptor_g_k_c_xs_packed( + conv_param); + + const auto out_g_n_k_wos_desc = + ck::utils::conv::make_output_host_tensor_descriptor_g_n_k_wos_packed( + conv_param); + + return run_grouped_conv_fwd_dl< + ndim_spatial_value, + InDataType, + WeiDataType, + OutDataType, + InElementOp, + WeiElementOp, + OutElementOp, + DeviceGroupedConvNDFwdInstance>( + do_verification, + init_method, + time_kernel, + conv_param, + in_g_n_c_wis_desc, + wei_g_k_c_xs_desc, + out_g_n_k_wos_desc, + in_element_op, + wei_element_op, + out_element_op); + }; + + namespace ctc = ck::tensor_layout::convolution; + + if(conv_param.num_dim_spatial_ == 1) + { + return run(ck::Number<1>{}, ctc::GNWC{}, ctc::GKXC{}, ctc::GNWK{}); + } + else if(conv_param.num_dim_spatial_ == 2) + { + return run(ck::Number<2>{}, ctc::GNHWC{}, ctc::GKYXC{}, ctc::GNHWK{}); + } + else if(conv_param.num_dim_spatial_ == 3) + { + return run(ck::Number<3>{}, ctc::GNDHWC{}, ctc::GKZYXC{}, ctc::GNDHWK{}); + } + + return true; +} diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp new file mode 100644 index 0000000000..481e2e6aee --- /dev/null +++ b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp @@ -0,0 +1,56 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include + +#include "ck/tensor_operation/gpu/device/device_base.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +// Convolution Forward: +// input : input image A[G, N, C, Hi, Wi], +// input : weight B[G, K, C, Y, X], +// input : D0[G, N, K, Ho, Wo], D1[G, N, K, Ho, Wo], ... +// output : output image E[G, N, K, Ho, Wo] +// C = a_op(A) * b_op(B) +// E = cde_op(C, D0, D1, ...) +template +struct DeviceGroupedConvFwd : public BaseOperator +{ + virtual std::unique_ptr + MakeArgumentPointer(const void* p_a, // input image + const void* p_b, // weight + void* p_c, // output image + const std::array& a_g_n_c_wis_lengths, + const std::array& a_g_n_c_wis_strides, + const std::array& b_g_k_c_xs_lengths, + const std::array& b_g_k_c_xs_strides, + const std::array& c_g_n_k_wos_lengths, + const std::array& c_g_n_k_wos_strides, + const std::array& conv_filter_strides, + const std::array& conv_filter_dilations, + const std::array& input_left_pads, + const std::array& input_right_pads, + const AElementwiseOperation& a_element_op, + const BElementwiseOperation& b_element_op, + const CElementwiseOperation& c_element_op) = 0; + + virtual std::unique_ptr MakeInvokerPointer() = 0; +}; + +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp new file mode 100644 index 0000000000..47c821171b --- /dev/null +++ b/include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp @@ -0,0 +1,837 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include +#include +#include +#include +#include + +#include "ck/utility/common_header.hpp" +#include "ck/tensor_description/tensor_descriptor.hpp" +#include "ck/tensor_description/tensor_descriptor_helper.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" +#include "ck/tensor_operation/operator_transform/transform_conv_fwd_to_gemm.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/device/matrix_padder.hpp" +#include "ck/tensor_operation/gpu/grid/gridwise_gemm_dl_v1r3.hpp" +#include "ck/host_utility/device_prop.hpp" +#include "ck/host_utility/kernel_launch.hpp" +#include "ck/host_utility/io.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { + +namespace { + +struct ComputePtrOffsetOfStridedBatch +{ + ComputePtrOffsetOfStridedBatch(index_t BatchStrideA, index_t BatchStrideB, index_t BatchStrideC) + : BatchStrideA_(BatchStrideA), BatchStrideB_(BatchStrideB), BatchStrideC_(BatchStrideC) + { + } + + __host__ __device__ constexpr long_index_t GetAPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideA_); + } + + __host__ __device__ constexpr long_index_t GetBPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideB_); + } + + __host__ __device__ constexpr long_index_t GetCPtrOffset(index_t g_idx) const + { + return g_idx * static_cast(BatchStrideC_); + } + + index_t BatchStrideA_; + index_t BatchStrideB_; + index_t BatchStrideC_; +}; + +/* + * \brief Wrapper function of GridwiseGemm::Run to realize BatchedGEMM. + * + * \tparam ComputePtrOffsetOfBatch Class that computes the base pointer offsets of A, B, C matrix + * given the batch. For example, ComputePtrOffsetOfStridedBatch() computes the offsets of evenly + * strided batched, but we can easily extend to other layouts. The returned offset can be either \p + * index_t or \p long_index_t. If it returns \p long_index_t, we are not subject to the 2GB + * limitations. + * + * \tparam Block2ETileMap Block2ETileMap::CalculateBottomIndex() takes in id of a workgroup and + * returns the 2D index of the tile that it computes. \see + * GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3::Run(). + * + * \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2 + * tiles from different matrices. Keep in mind that these 2 matrices can share the same grid + * descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link + * device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link + * DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of + * pointer offset into \p ComputePtrOffsetOfStridedBatch. + * + * \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes. + * Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to + * realize BatchedGemm and GroupedGemm (and the corresponding GEMM fusion). + * + */ +template +__global__ void +#if CK_USE_LAUNCH_BOUNDS + __launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU) +#endif + kernel_grouped_conv_fwd_dl( + const ABDataType* __restrict__ p_a_grid, + const ABDataType* __restrict__ p_b_grid, + CDataType* __restrict__ p_c_grid, + const index_t batch_count, + const AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1, + const BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1, + const CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11, + const Block2CTileMap block_2_ctile_map, + const ComputePtrOffsetOfBatch compute_ptr_offset_of_batch) +{ +#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx906__) || defined(__gfx1030__)) + // offset base pointer for each work-group + const index_t num_blocks_per_batch = + __builtin_amdgcn_readfirstlane(get_grid_size() / batch_count); + const index_t g_idx = __builtin_amdgcn_readfirstlane(get_block_1d_id() / num_blocks_per_batch); + + const long_index_t a_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetAPtrOffset(g_idx))); + const long_index_t b_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetBPtrOffset(g_idx))); + const long_index_t c_batch_offset = __builtin_amdgcn_readfirstlane( + static_cast(compute_ptr_offset_of_batch.GetCPtrOffset(g_idx))); + + constexpr index_t shared_block_size = + GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(ABDataType); + + __shared__ ABDataType p_shared[shared_block_size]; + + GridwiseGemm::Run(p_a_grid + a_batch_offset, + p_b_grid + b_batch_offset, + p_c_grid + c_batch_offset, + p_shared, + a_grid_desc_k0_m0_m1_k1, + b_grid_desc_k0_n0_n1_k1, + c_grid_desc_m0_m10_m11_n0_n10_n11, + block_2_ctile_map, + integral_constant{}, + integral_constant{}); +#else + ignore = p_a_grid; + ignore = p_b_grid; + ignore = p_c_grid; + ignore = batch_count; + ignore = a_grid_desc_k0_m0_m1_k1; + ignore = b_grid_desc_k0_n0_n1_k1; + ignore = c_grid_desc_m0_m10_m11_n0_n10_n11; + ignore = compute_ptr_offset_of_batch; + ignore = block_2_ctile_map; + + compute_ptr_offset_of_batch.GetAPtrOffset(0); + compute_ptr_offset_of_batch.GetBPtrOffset(0); + compute_ptr_offset_of_batch.GetCPtrOffset(0); +#endif +} + +} // namespace + +// +// @brief Device Convolution operation. +// +// Supports: +// @li Forward convolution with up to 3 spatial dimentions +// @li Input tensor in GNWC data format +// @li Weight tensor in GKXC data format +// @li Output tensor in GNWK data format +// +// 1D: +// out[N, Wo, K] = in[N, Wi, C] * wei[K, X, C] +// 2D: +// out[N, Ho, Wo, K] = in[N, Hi, Wi, C] * wei[K, Y, X, C] +// 3D: +// out[N, Do, Ho, Wo, K] = in[N, Di, Hi, Wi, C] * wei[K, Z, Y, X, C] +// +template < + index_t NDimSpatial, + typename ADataType, + typename BDataType, + typename CDataType, + typename AccDataType, + typename ALayout, + typename BLayout, + typename CLayout, + typename AElementwiseOperation, + typename BElementwiseOperation, + typename CElementwiseOperation, + ConvolutionForwardSpecialization ConvForwardSpecialization, + GemmSpecialization GemmSpec, + index_t BlockSize, + index_t MPerBlock, + index_t NPerBlock, + index_t K0PerBlock, + index_t K1, + index_t M1PerThread, + index_t N1PerThread, + index_t KPerThread, + typename M1N1ThreadClusterM1Xs, + typename M1N1ThreadClusterN1Xs, + typename ABlockTransferThreadSliceLengths_K0_M0_M1_K1, + typename ABlockTransferThreadClusterLengths_K0_M0_M1_K1, + typename ABlockTransferThreadClusterArrangeOrder, + typename ABlockTransferSrcAccessOrder, + typename ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1, + typename ABlockTransferSrcVectorTensorContiguousDimOrder, + typename ABlockTransferDstVectorTensorLengths_K0_M0_M1_K1, + typename BBlockTransferThreadSliceLengths_K0_N0_N1_K1, + typename BBlockTransferThreadClusterLengths_K0_N0_N1_K1, + typename BBlockTransferThreadClusterArrangeOrder, + typename BBlockTransferSrcAccessOrder, + typename BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1, + typename BBlockTransferSrcVectorTensorContiguousDimOrder, + typename BBlockTransferDstVectorTensorLengths_K0_N0_N1_K1, + typename CThreadTransferSrcDstAccessOrder, + index_t CThreadTransferSrcDstVectorDim, + index_t CThreadTransferDstScalarPerVector, + enable_if_t< + is_same_v && + is_same_v && + is_same_v, + bool> = false> +struct DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK : public DeviceGroupedConvFwd +{ + using DeviceOp = DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK; + + static constexpr auto I0 = Number<0>{}; + static constexpr auto I1 = Number<1>{}; + static constexpr auto I2 = Number<2>{}; + static constexpr auto I3 = Number<3>{}; + + static constexpr auto conv_to_gemm_transformer = + TransformConvFwdToGemm{}; + + static constexpr auto matrix_padder = + MatrixPadder{MPerBlock, NPerBlock, K0PerBlock}; + + template + static auto + MakeAGridDescriptor_AK0_M_AK1(const std::array& a_g_n_c_wis_lengths, + const std::array& a_g_n_c_wis_strides, + const std::array& b_g_k_c_xs_lengths, + const std::array& b_g_k_c_xs_strides, + const std::array& c_g_n_k_wos_lengths, + const std::array& c_g_n_k_wos_strides, + const std::array& conv_filter_strides, + const std::array& conv_filter_dilations, + const std::array& input_left_pads, + const std::array& input_right_pads) + { + const auto in_gemmmraw_gemmkraw_desc = + conv_to_gemm_transformer.template MakeADescriptor_M_K(a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + c_g_n_k_wos_lengths, + c_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads); + + const auto in_gemmm_gemmk_desc = + matrix_padder.PadADescriptor_M_K(in_gemmmraw_gemmkraw_desc); + + const auto M = in_gemmm_gemmk_desc.GetLength(I0); + const auto K = in_gemmm_gemmk_desc.GetLength(I1); + + const auto AK0 = K / K1; + + return transform_tensor_descriptor( + in_gemmm_gemmk_desc, + make_tuple(make_unmerge_transform(make_tuple(AK0, K1)), make_pass_through_transform(M)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + + template + static auto + MakeBGridDescriptor_BK0_N_BK1(const std::array& b_g_k_c_xs_lengths, + const std::array& b_g_k_c_xs_strides) + { + const auto wei_gemmnraw_gemmkraw_desc = + conv_to_gemm_transformer.template MakeBDescriptor_N_K(b_g_k_c_xs_lengths, + b_g_k_c_xs_strides); + + const auto wei_gemmn_gemmk_desc = + matrix_padder.PadBDescriptor_N_K(wei_gemmnraw_gemmkraw_desc); + + const auto N = wei_gemmn_gemmk_desc.GetLength(I0); + const auto K = wei_gemmn_gemmk_desc.GetLength(I1); + + const auto BK0 = K / K1; + + return transform_tensor_descriptor( + wei_gemmn_gemmk_desc, + make_tuple(make_unmerge_transform(make_tuple(BK0, K1)), make_pass_through_transform(N)), + make_tuple(Sequence<1>{}, Sequence<0>{}), + make_tuple(Sequence<0, 2>{}, Sequence<1>{})); + } + + template + static auto + MakeCGridDescriptor_M_N(const std::array& c_g_n_k_wos_lengths, + const std::array& c_g_n_k_wos_strides) + { + const auto out_gemmmraw_gemmnraw_desc = + conv_to_gemm_transformer.template MakeCDescriptor_M_N(c_g_n_k_wos_lengths, + c_g_n_k_wos_strides); + + const auto out_gemmm_gemmn_desc = + matrix_padder.PadCDescriptor_M_N(out_gemmmraw_gemmnraw_desc); + + return out_gemmm_gemmn_desc; + } + + // desc for problem definition + using AGridDesc_AK0_M_AK1 = remove_cvref_t({}, {}, {}, {}, {}, {}, {}, {}, {}, {}))>; + using BGridDesc_BK0_N_BK1 = + remove_cvref_t({}, {}))>; + using CGridDesc_M_N = remove_cvref_t({}, {}))>; + + // GridwiseGemm + using GridwiseGemm = + GridwiseGemmDl_km_kn_mn_v1r3; + + using AGridDesc_K0_M0_M1_K1 = + decltype(GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(AGridDesc_AK0_M_AK1{})); + using BGridDesc_K0_N0_N1_K1 = + decltype(GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(BGridDesc_BK0_N_BK1{})); + using CGridDesc_M0_M10_M11_N0_N10_N11 = + decltype(GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(CGridDesc_M_N{})); + using DefaultBlock2CTileMap = + decltype(GridwiseGemm::MakeDefaultBlock2CTileMap(CGridDesc_M_N{})); + + // Argument + struct Argument : public BaseArgument + { + Argument(const void* p_a, + const void* p_b, + void* p_c, + const std::array& a_g_n_c_wis_lengths, + const std::array& a_g_n_c_wis_strides, + const std::array& b_g_k_c_xs_lengths, + const std::array& b_g_k_c_xs_strides, + const std::array& c_g_n_k_wos_lengths, + const std::array& c_g_n_k_wos_strides, + const std::array& conv_filter_strides, + const std::array& conv_filter_dilations, + const std::array& input_left_pads, + const std::array& input_right_pads, + const AElementwiseOperation& a_element_op, + const BElementwiseOperation& b_element_op, + const CElementwiseOperation& c_element_op) + : p_a_grid_{static_cast(p_a)}, + p_b_grid_{static_cast(p_b)}, + p_c_grid_{static_cast(p_c)}, + num_group_{a_g_n_c_wis_lengths[0]}, + a_grid_desc_ak0_m_ak1_{ + DeviceOp::MakeAGridDescriptor_AK0_M_AK1(a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + c_g_n_k_wos_lengths, + c_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads)}, + b_grid_desc_bk0_n_bk1_{DeviceOp::MakeBGridDescriptor_BK0_N_BK1( + b_g_k_c_xs_lengths, b_g_k_c_xs_strides)}, + c_grid_desc_m_n_{DeviceOp::MakeCGridDescriptor_M_N(c_g_n_k_wos_lengths, + c_g_n_k_wos_strides)}, + a_grid_desc_k0_m0_m1_k1_{}, + b_grid_desc_k0_n0_n1_k1_{}, + c_grid_desc_m0_m10_m11_n0_n10_n11_{}, + block_2_ctile_map_{}, + compute_ptr_offset_of_batch_{ + a_g_n_c_wis_strides[0], b_g_k_c_xs_strides[0], c_g_n_k_wos_strides[0]}, + a_element_op_{a_element_op}, + b_element_op_{b_element_op}, + c_element_op_{c_element_op}, + a_g_n_c_wis_lengths_{a_g_n_c_wis_lengths}, + a_g_n_c_wis_strides_{a_g_n_c_wis_strides}, + b_g_k_c_xs_lengths_{b_g_k_c_xs_lengths}, + b_g_k_c_xs_strides_{b_g_k_c_xs_strides}, + c_g_n_k_wos_lengths_{c_g_n_k_wos_lengths}, + c_g_n_k_wos_strides_{c_g_n_k_wos_strides}, + conv_filter_strides_{conv_filter_strides}, + conv_filter_dilations_{conv_filter_dilations}, + input_left_pads_{input_left_pads}, + input_right_pads_{input_right_pads} + { + // A/B/E Batch Stride + compute_ptr_offset_of_batch_.BatchStrideA_ = a_g_n_c_wis_strides[0]; + compute_ptr_offset_of_batch_.BatchStrideB_ = b_g_k_c_xs_strides[0]; + compute_ptr_offset_of_batch_.BatchStrideC_ = c_g_n_k_wos_strides[0]; + + // populate desc for Ds/E + if(GridwiseGemm::CheckValidity( + a_grid_desc_ak0_m_ak1_, b_grid_desc_bk0_n_bk1_, c_grid_desc_m_n_)) + { + + a_grid_desc_k0_m0_m1_k1_ = + GridwiseGemm::MakeAGridDescriptor_K0_M0_M1_K1(a_grid_desc_ak0_m_ak1_); + b_grid_desc_k0_n0_n1_k1_ = + GridwiseGemm::MakeBGridDescriptor_K0_N0_N1_K1(b_grid_desc_bk0_n_bk1_); + c_grid_desc_m0_m10_m11_n0_n10_n11_ = + GridwiseGemm::MakeCGridDescriptor_M0_M10_M11_N0_N10_N11(c_grid_desc_m_n_); + + block_2_ctile_map_ = GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_); + } + } + + void Print() const + { + std::cout << "A[K0, M, K1]: " << a_grid_desc_ak0_m_ak1_ << std::endl; + std::cout << "B[K0, N, K1]: " << b_grid_desc_bk0_n_bk1_ << std::endl; + std::cout << "C[M, N]: " << c_grid_desc_m_n_ << std::endl; + std::cout << "num_group: " << num_group_ << std::endl; + + std::cout << "A[k0, m0, m1, k1]: " << a_grid_desc_k0_m0_m1_k1_ << std::endl; + std::cout << "B[k0, n0, n1, k1]: " << b_grid_desc_k0_n0_n1_k1_ << std::endl; + std::cout << "A[m0, m10, m11, n0, n10, n11]: " << c_grid_desc_m0_m10_m11_n0_n10_n11_ + << std::endl; + } + + // private: + // pointers + const ADataType* p_a_grid_; + const BDataType* p_b_grid_; + CDataType* p_c_grid_; + + // tensor descriptors for problem definiton + index_t num_group_; + AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_; + BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_; + CGridDesc_M_N c_grid_desc_m_n_; + + // tensor descriptors for block/thread-wise copy + AGridDesc_K0_M0_M1_K1 a_grid_desc_k0_m0_m1_k1_; + BGridDesc_K0_N0_N1_K1 b_grid_desc_k0_n0_n1_k1_; + CGridDesc_M0_M10_M11_N0_N10_N11 c_grid_desc_m0_m10_m11_n0_n10_n11_; + + // block-to-e-tile map + DefaultBlock2CTileMap block_2_ctile_map_; + + // for computing batch offset + ComputePtrOffsetOfStridedBatch compute_ptr_offset_of_batch_; + + // element-wise op + AElementwiseOperation a_element_op_; + BElementwiseOperation b_element_op_; + CElementwiseOperation c_element_op_; + + // for checking IsSupportedArgument() + std::array a_g_n_c_wis_lengths_; + std::array a_g_n_c_wis_strides_; + std::array b_g_k_c_xs_lengths_; + std::array b_g_k_c_xs_strides_; + std::array c_g_n_k_wos_lengths_; + std::array c_g_n_k_wos_strides_; + std::array conv_filter_strides_; + std::array conv_filter_dilations_; + std::array input_left_pads_; + std::array input_right_pads_; + }; + + // Invoker + struct Invoker : public BaseInvoker + { + using Argument = DeviceOp::Argument; + + float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{}) + { + // if(stream_config.log_level_ > 0) + { + arg.Print(); + } + + if(!GridwiseGemm::CheckValidity( + arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_)) + { + throw std::runtime_error( + "wrong! DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK has invalid setting"); + } + + const index_t grid_size = + GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_.GetLength(I0), + arg.c_grid_desc_m_n_.GetLength(I1)) * + arg.num_group_; + + auto launch_kernel = [&](auto has_main_k_block_loop, + auto has_double_tail_k_block_loop) { + constexpr bool has_main_loop = has_main_k_block_loop.value; + constexpr bool has_double_loop = has_double_tail_k_block_loop; + + const auto kernel = + kernel_grouped_conv_fwd_dl; + + return launch_and_time_kernel(stream_config, + kernel, + dim3(grid_size), + dim3(BlockSize), + 0, + arg.p_a_grid_, + arg.p_b_grid_, + arg.p_c_grid_, + arg.a_g_n_c_wis_lengths_[0], // Group count + arg.a_grid_desc_k0_m0_m1_k1_, + arg.b_grid_desc_k0_n0_n1_k1_, + arg.c_grid_desc_m0_m10_m11_n0_n10_n11_, + arg.block_2_ctile_map_, + arg.compute_ptr_offset_of_batch_); + }; + + const auto K0 = arg.a_grid_desc_k0_m0_m1_k1_.GetLength(I0); + const bool has_main_k_block_loop = GridwiseGemm::CalculateHasMainKBlockLoop(K0); + const bool has_double_tail_k_block_loop = + GridwiseGemm::CalculateHasDoubleTailKBlockLoop(K0); + + if(has_main_k_block_loop && has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else if(has_main_k_block_loop && !has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else if(!has_main_k_block_loop && has_double_tail_k_block_loop) + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + else + { + return launch_kernel(integral_constant{}, + integral_constant{}); + } + } + + float Run(const BaseArgument* p_arg, + const StreamConfig& stream_config = StreamConfig{}) override + { + return Run(*dynamic_cast(p_arg), stream_config); + } + }; + + static bool IsSupportedArgument(const Argument& arg) + { + namespace ctc = tensor_layout::convolution; + + // check device + if(!(ck::get_device_name() == "gfx906" || ck::get_device_name() == "gfx1030")) + { + return false; + } + + // check ConvolutionForwardSpecialization + if constexpr(ConvForwardSpecialization == + ConvolutionForwardSpecialization::Filter1x1Stride1Pad0) + { + // check if it's 1x1, stride=1 conv + for(index_t i = 0; i < NDimSpatial; ++i) + { + const index_t X = arg.b_g_k_c_xs_lengths_[i + 3]; + const index_t ConvStride = arg.conv_filter_strides_[i]; + const index_t LeftPad = arg.input_left_pads_[i]; + const index_t RightPad = arg.input_right_pads_[i]; + + if(!(X == 1 && ConvStride == 1 && LeftPad == 0 && RightPad == 0)) + { + std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X + << " ConvStride = " << ConvStride << " LeftPad = " << LeftPad + << " RightPad = " << RightPad << std::endl; + return false; + } + } + } + else if constexpr(ConvForwardSpecialization == + ConvolutionForwardSpecialization::Filter1x1Pad0) + { + // check if it's 1x1 conv + for(index_t i = 0; i < NDimSpatial; ++i) + { + const index_t X = arg.b_g_k_c_xs_lengths_[i + 3]; + const index_t LeftPad = arg.input_left_pads_[i]; + const index_t RightPad = arg.input_right_pads_[i]; + + if(!(X == 1 && LeftPad == 0 && RightPad == 0)) + { + std::cout << "Filter1x1Stride1Pad0 check: i = " << i << " X = " << X + << " LeftPad = " << LeftPad << " RightPad = " << RightPad + << std::endl; + return false; + } + } + } + + // check vector access of A + // FIXME: layout + if constexpr(is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v) + { + auto srcVectorLengths = ABlockTransferSrcVectorTensorLengths_K0_M0_M1_K1{}; + if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1) + { + return false; + } + if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0) + { + return false; + } + + const index_t C = arg.a_g_n_c_wis_lengths_[2]; + + if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0) + { + return false; + } + } + else + { + return false; + } + + // check vector access of B + // FIXME: layout + if constexpr(is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v) + + { + auto srcVectorLengths = BBlockTransferSrcVectorTensorLengths_K0_N0_N1_K1{}; + if(srcVectorLengths[I1] != 1 || srcVectorLengths[I2] != 1) + { + return false; + } + if(K1 % srcVectorLengths[I3] != 0 || K0PerBlock % srcVectorLengths[I0] != 0) + { + return false; + } + + const index_t C = arg.b_g_k_c_xs_lengths_[2]; + + if(C % (srcVectorLengths[I0] * srcVectorLengths[I3]) != 0) + { + return false; + } + } + else + { + return false; + } + + // check vector access of C + if constexpr(is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v || is_same_v || + is_same_v) + { + const index_t K = arg.c_g_n_k_wos_lengths_[2]; + + if(!(K % CThreadTransferDstScalarPerVector == 0 && CThreadTransferSrcDstVectorDim == 5)) + { + return false; + } + } + else + { + return false; + } + // check Gridwise GEMM + return GridwiseGemm::CheckValidity( + arg.a_grid_desc_ak0_m_ak1_, arg.b_grid_desc_bk0_n_bk1_, arg.c_grid_desc_m_n_); + } + + bool IsSupportedArgument(const BaseArgument* p_arg) override + { + return IsSupportedArgument(*dynamic_cast(p_arg)); + } + + static auto MakeArgument(const void* p_a, + const void* p_b, + void* p_c, + const std::array& a_g_n_c_wis_lengths, + const std::array& a_g_n_c_wis_strides, + const std::array& b_g_k_c_xs_lengths, + const std::array& b_g_k_c_xs_strides, + const std::array& c_g_n_k_wos_lengths, + const std::array& c_g_n_k_wos_strides, + const std::array& conv_filter_strides, + const std::array& conv_filter_dilations, + const std::array& input_left_pads, + const std::array& input_right_pads, + const AElementwiseOperation& a_element_op, + const BElementwiseOperation& b_element_op, + const CElementwiseOperation& c_element_op) + { + return Argument{p_a, + p_b, + p_c, + a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + c_g_n_k_wos_lengths, + c_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads, + a_element_op, + b_element_op, + c_element_op}; + } + + static auto MakeInvoker() { return Invoker{}; } + + std::unique_ptr + MakeArgumentPointer(const void* p_a, + const void* p_b, + void* p_c, + const std::array& a_g_n_c_wis_lengths, + const std::array& a_g_n_c_wis_strides, + const std::array& b_g_k_c_xs_lengths, + const std::array& b_g_k_c_xs_strides, + const std::array& c_g_n_k_wos_lengths, + const std::array& c_g_n_k_wos_strides, + const std::array& conv_filter_strides, + const std::array& conv_filter_dilations, + const std::array& input_left_pads, + const std::array& input_right_pads, + const AElementwiseOperation& a_element_op, + const BElementwiseOperation& b_element_op, + const CElementwiseOperation& c_element_op) override + { + return std::make_unique(p_a, + p_b, + p_c, + a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + c_g_n_k_wos_lengths, + c_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads, + a_element_op, + b_element_op, + c_element_op); + } + + std::unique_ptr MakeInvokerPointer() override + { + return std::make_unique(Invoker{}); + } + + std::string GetTypeString() const override + { + auto str = std::stringstream(); + + // clang-format off + str << "DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK" + << "<" + << BlockSize << ", " + << MPerBlock << ", " + << NPerBlock << ", " + << K0PerBlock << ", " + << getConvForwardSpecializationString(ConvForwardSpecialization) + << ">"; + // clang-format on + + return str.str(); + } +}; + +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp index 0d93e06418..7dc542abb9 100644 --- a/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp +++ b/include/ck/tensor_operation/gpu/device/impl/device_gemm_dl.hpp @@ -214,6 +214,7 @@ struct DeviceGemmDl : public DeviceGemm{}; // K1 should be Number<...> - static constexpr auto K1 = AGridDesc_K0_M_K1{}.GetLength(I2); + static constexpr auto K1 = Number{}; __host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte() { diff --git a/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp new file mode 100644 index 0000000000..c8ce393179 --- /dev/null +++ b/library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp @@ -0,0 +1,118 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#pragma once + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_conv_fwd.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +// grouped conv2d forward, GNHWC/GKYXC/GNHWK +void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances( + std::vector>>& instances); + +void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances( + std::vector>>& instances); + +void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances( + std::vector>>& instances); + +template +struct DeviceOperationInstanceFactory> +{ + using DeviceOp = DeviceGroupedConvFwd; + + static auto GetInstances() + { + std::vector> op_ptrs; + + if constexpr(NumDimSpatial == 2 && is_same_v && + is_same_v && is_same_v) + { + if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(op_ptrs); + } + else if constexpr(is_same_v && is_same_v && + is_same_v) + { + add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs); + } + } + + return op_ptrs; + } +}; + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt index 0d2d7f846a..5ef1b6866e 100644 --- a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/CMakeLists.txt @@ -6,4 +6,8 @@ add_instance_library(device_grouped_conv2d_fwd_instance device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instance.cpp # NHWGC, GKYXC, NHWGK device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instance.cpp + #dl + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp ) diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp new file mode 100644 index 0000000000..15ec1b1316 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instance.cpp @@ -0,0 +1,101 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using InDataType = ck::half_t; +using WeiDataType = ck::half_t; +using AccDataType = float; +using OutDataType = ck::half_t; + +template +using S = ck::Sequence; + +using InElementOp = ck::tensor_operation::element_wise::PassThrough; +using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; +using OutElementOp = ck::tensor_operation::element_wise::PassThrough; + +using InLayout = ck::tensor_layout::convolution::GNHWC; +using WeiLayout = ck::tensor_layout::convolution::GKYXC; +using OutLayout = ck::tensor_layout::convolution::GNHWK; + +static constexpr auto ConvSpec = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; +static constexpr auto Filter1x1Pad0 = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0; +static constexpr auto Filter1x1Stride1Pad0 = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0; + +static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances = std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Pad0_instances = std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Pad0, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Stride1Pad0_instances = + std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Stride1Pad0, GemmPadingSpec, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances{}); + + add_device_operation_instances( + instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Pad0_instances{}); + + add_device_operation_instances( + instances, + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_Filter1x1Stride1Pad0_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp new file mode 100644 index 0000000000..2ed4f18ff5 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instance.cpp @@ -0,0 +1,102 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using InDataType = float; +using WeiDataType = float; +using AccDataType = float; +using OutDataType = float; + +template +using S = ck::Sequence; + +using InElementOp = ck::tensor_operation::element_wise::PassThrough; +using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; +using OutElementOp = ck::tensor_operation::element_wise::PassThrough; + +using InLayout = ck::tensor_layout::convolution::GNHWC; +using WeiLayout = ck::tensor_layout::convolution::GKYXC; +using OutLayout = ck::tensor_layout::convolution::GNHWK; + +static constexpr auto ConvSpec = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; +static constexpr auto Filter1x1Pad0 = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0; +static constexpr auto Filter1x1Stride1Pad0 = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0; + +static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances = std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Pad0_instances = std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Pad0, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Stride1Pad0_instances = + std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Stride1Pad0, GemmPadingSpec, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances{}); + + add_device_operation_instances( + instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Pad0_instances{}); + + add_device_operation_instances( + instances, + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_Filter1x1Stride1Pad0_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp new file mode 100644 index 0000000000..e92d9e3cc0 --- /dev/null +++ b/library/src/tensor_operation_instance/gpu/grouped_conv2d_fwd/device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instance.cpp @@ -0,0 +1,102 @@ +// SPDX-License-Identifier: MIT +// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. + +#include + +#include "ck/ck.hpp" +#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_dl_nhwc_kyxc_nhwk.hpp" +#include "ck/library/tensor_operation_instance/add_device_operation_instance.hpp" +#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp" +#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" +#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" + +namespace ck { +namespace tensor_operation { +namespace device { +namespace instance { + +using InDataType = int8_t; +using WeiDataType = int8_t; +using AccDataType = int32_t; +using OutDataType = int8_t; + +template +using S = ck::Sequence; + +using InElementOp = ck::tensor_operation::element_wise::PassThrough; +using WeiElementOp = ck::tensor_operation::element_wise::PassThrough; +using OutElementOp = ck::tensor_operation::element_wise::PassThrough; + +using InLayout = ck::tensor_layout::convolution::GNHWC; +using WeiLayout = ck::tensor_layout::convolution::GKYXC; +using OutLayout = ck::tensor_layout::convolution::GNHWK; + +static constexpr auto ConvSpec = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; +static constexpr auto Filter1x1Pad0 = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Pad0; +static constexpr auto Filter1x1Stride1Pad0 = + ck::tensor_operation::device::ConvolutionForwardSpecialization::Filter1x1Stride1Pad0; + +static constexpr auto GemmPadingSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances = std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, ConvSpec, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Pad0_instances = std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Pad0, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +using device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Stride1Pad0_instances = + std::tuple< + // clang-format off + // ###############################| NDim| InData| WeiData| OutData| AccData| InLayout| WeiLayout| OutLayout| In| Wei| Out| Convolution| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer| + // ###############################| Spatial| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Forward| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector| + // ###############################| | | | | | | | | Operation| Operation| Operation| Specialization| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | | + // ###############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | + DeviceGroupedConvFwdDl_NHWC_KYXC_NHWK< 2, InDataType, WeiDataType, OutDataType, AccDataType, InLayout, WeiLayout, OutLayout, InElementOp, WeiElementOp, OutElementOp, Filter1x1Stride1Pad0, GemmPadingSpec, 256, 128, 128, 16, 4, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<8, 1, 1, 4>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 4>, S<1, 2, 0, 3>, S<1, 1, 1, 4>, S<0, 1, 2, 3, 4, 5>, 5, 4> + // clang-format on + >; + +void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances( + std::vector>>& instances) +{ + add_device_operation_instances(instances, + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances{}); + + add_device_operation_instances( + instances, device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Pad0_instances{}); + + add_device_operation_instances( + instances, + device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_Filter1x1Stride1Pad0_instances{}); +} + +} // namespace instance +} // namespace device +} // namespace tensor_operation +} // namespace ck diff --git a/profiler/include/profile_grouped_conv_fwd_impl.hpp b/profiler/include/profile_grouped_conv_fwd_impl.hpp index 8d7ebe0465..32bded1bd4 100644 --- a/profiler/include/profile_grouped_conv_fwd_impl.hpp +++ b/profiler/include/profile_grouped_conv_fwd_impl.hpp @@ -14,6 +14,9 @@ #include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp" +#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd.hpp" +#include "ck/library/tensor_operation_instance/gpu/grouped_convolution_forward_dl.hpp" + #include "ck/library/utility/check_err.hpp" #include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/host_tensor.hpp" @@ -136,25 +139,6 @@ bool profile_grouped_conv_fwd_impl(int do_verification, ref_invoker.Run(ref_argument); } - using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD, - OutLayout, - InDataType, - WeiDataType, - ck::Tuple<>, - OutDataType, - InElementOp, - WeiElementOp, - OutElementOp>; - - // get device op instances - const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< - DeviceOp>::GetInstances(); - - std::cout << "found " << op_ptrs.size() << " instances" << std::endl; - std::string best_op_name; float best_avg_time = 0; float best_tflops = 0; @@ -163,29 +147,7 @@ bool profile_grouped_conv_fwd_impl(int do_verification, // profile device op instances bool pass = true; - for(auto& op_ptr : op_ptrs) - { - auto argument_ptr = - op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(), - wei_device_buf.GetDeviceBuffer(), - std::array{}, - out_device_buf.GetDeviceBuffer(), - a_g_n_c_wis_lengths, - a_g_n_c_wis_strides, - b_g_k_c_xs_lengths, - b_g_k_c_xs_strides, - std::array, 0>{{}}, - std::array, 0>{{}}, - e_g_n_k_wos_lengths, - e_g_n_k_wos_strides, - conv_filter_strides, - conv_filter_dilations, - input_left_pads, - input_right_pads, - in_element_op, - wei_element_op, - out_element_op); - + auto run_impl = [&](auto& op_ptr, auto& argument_ptr) { if(op_ptr->IsSupportedArgument(argument_ptr.get())) { // re-init output to zero before profiling next kernel @@ -237,6 +199,94 @@ bool profile_grouped_conv_fwd_impl(int do_verification, { std::cout << op_ptr->GetTypeString() << " does not support this problem" << std::endl; } + }; + + // xdl + { + using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD, + OutLayout, + InDataType, + WeiDataType, + ck::Tuple<>, + OutDataType, + InElementOp, + WeiElementOp, + OutElementOp>; + + // get device op instances + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + + std::cout << "xdl found " << op_ptrs.size() << " instances" << std::endl; + + for(auto& op_ptr : op_ptrs) + { + auto argument_ptr = op_ptr->MakeArgumentPointer( + in_device_buf.GetDeviceBuffer(), + wei_device_buf.GetDeviceBuffer(), + std::array{}, + out_device_buf.GetDeviceBuffer(), + a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + std::array, 0>{{}}, + std::array, 0>{{}}, + e_g_n_k_wos_lengths, + e_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads, + in_element_op, + wei_element_op, + out_element_op); + + run_impl(op_ptr, argument_ptr); + } + } + + // dl + { + using DeviceOp = ck::tensor_operation::device::DeviceGroupedConvFwd; + + const auto op_ptrs = ck::tensor_operation::device::instance::DeviceOperationInstanceFactory< + DeviceOp>::GetInstances(); + std::cout << "dl found " << op_ptrs.size() << " instances" << std::endl; + + for(auto& op_ptr : op_ptrs) + { + auto argument_ptr = op_ptr->MakeArgumentPointer(in_device_buf.GetDeviceBuffer(), + wei_device_buf.GetDeviceBuffer(), + out_device_buf.GetDeviceBuffer(), + a_g_n_c_wis_lengths, + a_g_n_c_wis_strides, + b_g_k_c_xs_lengths, + b_g_k_c_xs_strides, + e_g_n_k_wos_lengths, + e_g_n_k_wos_strides, + conv_filter_strides, + conv_filter_dilations, + input_left_pads, + input_right_pads, + in_element_op, + wei_element_op, + out_element_op); + + run_impl(op_ptr, argument_ptr); + } } std::cout << "Best configuration parameters:"