From 1f0cd618247971f0803aff94ddf795d9e3d19428 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Sun, 29 Jul 2018 15:07:51 +0200 Subject: Added first version of a tuner for the ConvGemm direct kernel --- src/tuning/kernels/xconvgemm.cpp | 38 ++++++++ src/tuning/kernels/xconvgemm.hpp | 186 +++++++++++++++++++++++++++++++++++++++ src/tuning/tuning.cpp | 10 ++- src/tuning/tuning.hpp | 7 ++ 4 files changed, 239 insertions(+), 2 deletions(-) create mode 100644 src/tuning/kernels/xconvgemm.cpp create mode 100644 src/tuning/kernels/xconvgemm.hpp (limited to 'src') diff --git a/src/tuning/kernels/xconvgemm.cpp b/src/tuning/kernels/xconvgemm.cpp new file mode 100644 index 00000000..15dfe829 --- /dev/null +++ b/src/tuning/kernels/xconvgemm.cpp @@ -0,0 +1,38 @@ + +// ================================================================================================= +// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This +// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- +// width of 100 characters per line. +// +// Author(s): +// Cedric Nugteren +// +// This file uses the auto-tuner to tune the convgemm kernels. +// +// ================================================================================================= + +#include "tuning/kernels/xconvgemm.hpp" + +// Shortcuts to the clblast namespace +using half = clblast::half; +using float2 = clblast::float2; +using double2 = clblast::double2; + +// Function to tune a specific variation V (not within the clblast namespace) +template +void StartVariation(int argc, char *argv[]) { + const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv); + switch(clblast::GetPrecision(command_line_args)) { + case clblast::Precision::kHalf: clblast::Tuner(argc, argv, V, clblast::XConvGemmGetTunerDefaults, clblast::XConvGemmGetTunerSettings, clblast::XConvGemmTestValidArguments, clblast::XConvGemmSetConstraints, clblast::XConvGemmComputeLocalMemSize, clblast::XConvGemmSetArguments); break; + case clblast::Precision::kSingle: clblast::Tuner(argc, argv, V, clblast::XConvGemmGetTunerDefaults, clblast::XConvGemmGetTunerSettings, clblast::XConvGemmTestValidArguments, clblast::XConvGemmSetConstraints, clblast::XConvGemmComputeLocalMemSize, clblast::XConvGemmSetArguments); break; + case clblast::Precision::kDouble: clblast::Tuner(argc, argv, V, clblast::XConvGemmGetTunerDefaults, clblast::XConvGemmGetTunerSettings, clblast::XConvGemmTestValidArguments, clblast::XConvGemmSetConstraints, clblast::XConvGemmComputeLocalMemSize, clblast::XConvGemmSetArguments); break; + } +} + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + StartVariation<1>(argc, argv); + return 0; +} + +// ================================================================================================= diff --git a/src/tuning/kernels/xconvgemm.hpp b/src/tuning/kernels/xconvgemm.hpp new file mode 100644 index 00000000..83c93860 --- /dev/null +++ b/src/tuning/kernels/xconvgemm.hpp @@ -0,0 +1,186 @@ + +// ================================================================================================= +// This file is part of the CLBlast project. The project is licensed under Apache Version 2.0. This +// project loosely follows the Google C++ styleguide and uses a tab-size of two spaces and a max- +// width of 100 characters per line. +// +// Author(s): +// Cedric Nugteren +// +// This file uses the auto-tuner to tune the ConvGemm kernels. These kernels are based on the GEMM +// direct kernel and will use those parameters, this tuner is just optional to use for advanced +// users. +// +// ================================================================================================= + +#include +#include + +#include "utilities/utilities.hpp" +#include "tuning/tuning.hpp" + +namespace clblast { +// ================================================================================================= + +// Helper functions +template +size_t OutputHeight(const Arguments &args) { + const auto size = args.height + 2 * args.pad_h; + const auto padding = args.dilation_h * (args.kernel_h - 1) + 1; + if (size >= padding) { return (size - padding) / args.stride_h + 1; } + return 1; +} +template +size_t OutputWidth(const Arguments &args) { + const auto size = args.width + 2 * args.pad_w; + const auto padding = args.dilation_w * (args.kernel_w - 1) + 1; + if (size >= padding) { return (size - padding) / args.stride_w + 1; } + return 1; +} + +// Settings for this kernel (default command-line arguments) +TunerDefaults XConvGemmGetTunerDefaults(const int) { + auto settings = TunerDefaults(); + settings.options = {kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, + kArgNumKernels, kArgBatchCount, kArgFraction}; + settings.channels = 32; + settings.height = 64; + settings.width = 64; + settings.kernel_h = 3; + settings.kernel_w = 3; + settings.num_kernels = 32; + settings.default_batch_count = 16; + settings.default_fraction = 1.0; + settings.default_num_runs = 2; + return settings; +} + +// Settings for this kernel (general) +template +TunerSettings XConvGemmGetTunerSettings(const int, const Arguments &args) { + auto settings = TunerSettings(); + + // Identification of the kernel + settings.kernel_family = "xconvgemm"; + settings.kernel_name = "Xconvgemm"; + settings.sources = +"#define ROUTINE_CONVGEMM" +#include "../src/kernels/level3/xgemm_direct_part1.opencl" +#include "../src/kernels/level3/xgemm_direct_part2.opencl" +#include "../src/kernels/level3/xgemm_direct_part3.opencl" +#include "../src/kernels/levelx/xconvgemm_part1.opencl" +#include "../src/kernels/levelx/xconvgemm_part2.opencl" + ; + + // Helper variables + const auto patch_size = args.kernel_h * args.kernel_w * args.channels; + const auto num_patches = OutputHeight(args) * OutputWidth(args); + + // Buffer sizes + settings.size_a = args.batch_count * args.channels * args.height * args.width; + settings.size_b = args.num_kernels * args.channels * args.kernel_h * args.kernel_w; + settings.size_a = args.batch_count * args.num_kernels * OutputHeight(args) * OutputWidth(args); + + // Inputs and outputs IDs (X:0, Y:1, A:2, B:3, C:4, temp:5) + settings.inputs = {2, 3, 4}; + settings.outputs = {4}; + + // Sets the base thread configuration + settings.global_size = {num_patches, args.num_kernels}; + settings.global_size_ref = settings.global_size; + settings.local_size = {1, 1}; + settings.local_size_ref = {8, 8}; + + // Transforms the thread configuration based on the parameters + settings.mul_local = {{"MDIMCD", "NDIMCD"}}; + settings.mul_global = {{"MDIMCD", "NDIMCD"}}; + settings.div_global = {{"WGD", "WGD"}}; + + // Sets the tuning parameters and their possible values + settings.parameters = { + {"WGD", {8, 16, 32}}, + {"MDIMCD", {8, 16, 32}}, + {"NDIMCD", {8, 16, 32}}, + {"MDIMAD", {8, 16, 32}}, + {"NDIMBD", {8, 16, 32}}, + {"KWID", {1}}, + {"VWMD", {1, 2, 4, 8}}, + {"VWND", {1, 2, 4, 8}}, + {"PADA", {0}}, + {"PADB", {0}}, + }; + + // Describes how to compute the performance metrics + settings.metric_amount = args.batch_count * 2 * num_patches * args.num_kernels * patch_size; + settings.performance_unit = "GFLOPS"; + + return settings; +} + +// Tests for valid arguments +template +void XConvGemmTestValidArguments(const int, const Arguments &) { } +std::vector XConvGemmSetConstraints(const int) { + auto constraints = std::vector(); + auto MultipleOfX = [] (std::vector v) { return IsMultiple(v[0], v[1]); }; + auto MultipleOfXMulY = [] (std::vector v) { return IsMultiple(v[0], v[1]*v[2]); }; + auto MultipleOfXMulYDivZ = [] (std::vector v) { return IsMultiple(v[0], (v[1]*v[2])/v[3]); }; + // Requirement for unrolling the WGD loop + constraints.push_back({MultipleOfX, {"WGD", "KWID"}}); + // Required for integer MWID and NWID + constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMCD", "VWMD"}}); + constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMCD", "VWND"}}); + // Required for integer MWIAD and NWIBD + constraints.push_back({MultipleOfXMulY, {"WGD", "MDIMAD", "VWMD"}}); + constraints.push_back({MultipleOfXMulY, {"WGD", "NDIMBD", "VWND"}}); + // WGD has to be a multiple of KDIMAD = ((MDIMCD*NDIMCD)/(MDIMAD)) and KDIMBD = (...) + constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "MDIMAD"}}); + constraints.push_back({MultipleOfXMulYDivZ, {"WGD", "MDIMCD", "NDIMCD", "NDIMBD"}}); + + return constraints; +} +template +LocalMemSizeInfo XConvGemmComputeLocalMemSize(const int) { + return { + [] (std::vector v) -> size_t { + return GetBytes(PrecisionValue()) * ((v[0]*(v[0] + v[1]) + v[0]*(v[0] + v[2]))); + }, + {"WGD", "PADA", "PADB"} + }; +} + +// Sets the kernel's arguments +template +void XConvGemmSetArguments(const int, Kernel &kernel, const Arguments &args, std::vector>& buffers) { + const auto output_h = OutputHeight(args); + const auto output_w = OutputWidth(args); + const auto patch_size = args.kernel_h * args.kernel_w * args.channels; + const auto num_patches = output_h * output_w; + const auto result_stride = args.num_kernels * output_h * output_w; + kernel.SetArgument(0, static_cast(num_patches)); + kernel.SetArgument(1, static_cast(args.num_kernels)); + kernel.SetArgument(2, static_cast(patch_size)); + kernel.SetArgument(3, buffers[3]()); // 3 == B matrix ==> kernel buffer + kernel.SetArgument(4, 0); // c_offset + kernel.SetArgument(5, buffers[4]()); // 4 == C matrix ==> result buffer + kernel.SetArgument(6, 0); // c_offset + kernel.SetArgument(7, static_cast(result_stride)); + kernel.SetArgument(8, buffers[2]()); // 2 == A matrix ==> image buffer + kernel.SetArgument(9, 0); // c_offset + kernel.SetArgument(10, static_cast(args.height)); + kernel.SetArgument(11, static_cast(args.width)); + kernel.SetArgument(12, static_cast(args.channels)); + kernel.SetArgument(13, static_cast(args.kernel_h)); + kernel.SetArgument(14, static_cast(args.kernel_w)); + kernel.SetArgument(15, 0); // pad_h + kernel.SetArgument(16, 0); // pad_w + kernel.SetArgument(17, 1); // stride_h + kernel.SetArgument(18, 1); // stride_w + kernel.SetArgument(19, 1); // dilation_h + kernel.SetArgument(20, 1); // dilation_w + kernel.SetArgument(21, static_cast(output_h)); + kernel.SetArgument(22, static_cast(output_w)); +} + +// ================================================================================================= +} // namespace clblast diff --git a/src/tuning/tuning.cpp b/src/tuning/tuning.cpp index d382fb18..f76af774 100644 --- a/src/tuning/tuning.cpp +++ b/src/tuning/tuning.cpp @@ -122,8 +122,14 @@ void Tuner(int argc, char* argv[], const int V, if (o == kArgM) { args.m = GetArgument(command_line_args, help, kArgM, defaults.default_m); } if (o == kArgN) { args.n = GetArgument(command_line_args, help, kArgN, defaults.default_n); } if (o == kArgK) { args.k = GetArgument(command_line_args, help, kArgK, defaults.default_k); } - if (o == kArgAlpha) { args.alpha = GetArgument(command_line_args, help, kArgAlpha, GetScalar()); } - if (o == kArgBeta) { args.beta = GetArgument(command_line_args, help, kArgBeta, GetScalar()); } + if (o == kArgChannels) { args.channels = GetArgument(command_line_args, help, kArgChannels, defaults.channels); } + if (o == kArgHeight) { args.height = GetArgument(command_line_args, help, kArgHeight, defaults.height); } + if (o == kArgWidth) { args.width = GetArgument(command_line_args, help, kArgWidth, defaults.width); } + if (o == kArgKernelH) { args.kernel_h = GetArgument(command_line_args, help, kArgKernelH, defaults.kernel_h); } + if (o == kArgKernelW) { args.kernel_w = GetArgument(command_line_args, help, kArgKernelW, defaults.kernel_w); } + if (o == kArgNumKernels) { args.num_kernels = GetArgument(command_line_args, help, kArgNumKernels, defaults.num_kernels); } + if (o == kArgAlpha) { args.alpha = GetArgument(command_line_args, help, kArgAlpha, GetScalar()); } + if (o == kArgBeta) { args.beta = GetArgument(command_line_args, help, kArgBeta, GetScalar()); } if (o == kArgBatchCount) { args.batch_count = GetArgument(command_line_args, help, kArgBatchCount, defaults.default_batch_count); } } args.fraction = GetArgument(command_line_args, help, kArgFraction, defaults.default_fraction); diff --git a/src/tuning/tuning.hpp b/src/tuning/tuning.hpp index 37a042ff..61cc3bda 100644 --- a/src/tuning/tuning.hpp +++ b/src/tuning/tuning.hpp @@ -41,6 +41,13 @@ struct TunerDefaults { size_t default_m = 1; size_t default_n = 1; size_t default_k = 1; + size_t channels = 1; + size_t height = 1; + size_t width = 1; + size_t kernel_h = 3; + size_t kernel_w = 3; + size_t num_kernels = 1; + size_t batch_count = 1; // Other defaults size_t default_batch_count = 1; -- cgit v1.2.3