From d45911b61dedafcbd74f65df263b4197697d6a81 Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Tue, 23 Oct 2018 20:52:25 +0200 Subject: Added groundwork for col2im algorithm plus first non-working version of kernel and test --- test/correctness/routines/levelx/xcol2im.cpp | 26 ++++ test/performance/routines/levelx/xcol2im.cpp | 33 +++++ test/routines/levelx/xcol2im.hpp | 195 +++++++++++++++++++++++++++ 3 files changed, 254 insertions(+) create mode 100644 test/correctness/routines/levelx/xcol2im.cpp create mode 100644 test/performance/routines/levelx/xcol2im.cpp create mode 100644 test/routines/levelx/xcol2im.hpp (limited to 'test') diff --git a/test/correctness/routines/levelx/xcol2im.cpp b/test/correctness/routines/levelx/xcol2im.cpp new file mode 100644 index 00000000..306d6fcc --- /dev/null +++ b/test/correctness/routines/levelx/xcol2im.cpp @@ -0,0 +1,26 @@ + +// ================================================================================================= +// 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 +// +// ================================================================================================= + +#include "test/correctness/testblas.hpp" +#include "test/routines/levelx/xcol2im.hpp" + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + auto errors = size_t{0}; + errors += clblast::RunTests, float, float>(argc, argv, false, "SCOL2IM"); + errors += clblast::RunTests, double, double>(argc, argv, true, "DCOL2IM"); + errors += clblast::RunTests, clblast::float2, clblast::float2>(argc, argv, true, "CCOL2IM"); + errors += clblast::RunTests, clblast::double2, clblast::double2>(argc, argv, true, "ZCOL2IM"); + errors += clblast::RunTests, clblast::half, clblast::half>(argc, argv, true, "HCOL2IM"); + if (errors > 0) { return 1; } else { return 0; } +} + +// ================================================================================================= diff --git a/test/performance/routines/levelx/xcol2im.cpp b/test/performance/routines/levelx/xcol2im.cpp new file mode 100644 index 00000000..76a5be30 --- /dev/null +++ b/test/performance/routines/levelx/xcol2im.cpp @@ -0,0 +1,33 @@ + +// ================================================================================================= +// 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 +// +// ================================================================================================= + +#include "test/performance/client.hpp" +#include "test/routines/levelx/xcol2im.hpp" + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + const auto command_line_args = clblast::RetrieveCommandLineArguments(argc, argv); + switch(clblast::GetPrecision(command_line_args, clblast::Precision::kSingle)) { + case clblast::Precision::kHalf: + clblast::RunClient, clblast::half, clblast::half>(argc, argv); break; + case clblast::Precision::kSingle: + clblast::RunClient, float, float>(argc, argv); break; + case clblast::Precision::kDouble: + clblast::RunClient, double, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: + clblast::RunClient, clblast::float2, clblast::float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient, clblast::double2, clblast::double2>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/levelx/xcol2im.hpp b/test/routines/levelx/xcol2im.hpp new file mode 100644 index 00000000..7393c432 --- /dev/null +++ b/test/routines/levelx/xcol2im.hpp @@ -0,0 +1,195 @@ + +// ================================================================================================= +// 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 implements a class with static methods to describe the Xcol2im routine. Examples of +// such 'descriptions' are how to calculate the size a of buffer or how to run the routine. These +// static methods are used by the correctness tester and the performance tester. +// +// ================================================================================================= + +#ifndef CLBLAST_TEST_ROUTINES_XCOL2IM_H_ +#define CLBLAST_TEST_ROUTINES_XCOL2IM_H_ + +#include "test/routines/common.hpp" + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template +class TestXcol2im { +public: + + // The BLAS level: 4 for the extra routines + static size_t BLASLevel() { return 4; } + + // The list of arguments relevant for this routine + static std::vector GetOptions() { + return {kArgChannels, kArgHeight, kArgWidth, kArgKernelH, kArgKernelW, kArgPadH, kArgPadW, + kArgStrideH, kArgStrideW, kArgDilationH, kArgDilationW, + kArgAOffset, kArgBOffset}; + } + static std::vector BuffersIn() { return {kBufMatA, kBufMatB}; } // b = col + static std::vector BuffersOut() { return {kBufMatA}; } // a = im + + // Describes how to obtain the sizes of the buffers + static size_t ColHeight(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; + } + static size_t ColWidth(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; + } + static size_t NumPatches(const Arguments &args) { + return ColHeight(args) * ColWidth(args) * args.channels; + } + static size_t GetSizeA(const Arguments &args) { + return args.height * args.width * args.channels + args.a_offset; + } + static size_t GetSizeB(const Arguments &args) { + return args.kernel_w * args.kernel_h * NumPatches(args) + args.b_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments &args, Queue&) { + args.a_size = GetSizeA(args); // im + args.b_size = GetSizeB(args); // col + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments &) { return 1; } // N/A for this routine + static size_t DefaultLDB(const Arguments &) { return 1; } // N/A for this routine + static size_t DefaultLDC(const Arguments &) { return 1; } // N/A for this routine + + // Describes which transpose options are relevant for this routine + using Transposes = std::vector; + static Transposes GetATransposes(const Transposes &) { return {}; } // N/A for this routine + static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + + // Describes how to prepare the input data + static void PrepareData(const Arguments&, Queue&, const int, std::vector&, + std::vector&, std::vector&, std::vector&, std::vector&, + std::vector&, std::vector&) {} // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments &args, Buffers &buffers, Queue &queue) { + #ifdef OPENCL_API + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Col2im(args.channels, args.height, args.width, + args.kernel_h, args.kernel_w, + args.pad_h, args.pad_w, + args.stride_h, args.stride_w, + args.dilation_h, args.dilation_w, + buffers.b_mat(), args.b_offset, // col + buffers.a_mat(), args.a_offset, // im + &queue_plain, &event); + if (status == StatusCode::kSuccess) { clWaitForEvents(1, &event); clReleaseEvent(event); } + #elif CUDA_API + auto status = Col2im(args.channels, args.height, args.width, + args.kernel_h, args.kernel_w, + args.pad_h, args.pad_w, + args.stride_h, args.stride_w, + args.dilation_h, args.dilation_w, + buffers.b_mat(), args.b_offset, // col + buffers.a_mat(), args.a_offset, // im + queue.GetContext()(), queue.GetDevice()()); + cuStreamSynchronize(queue()); + #endif + return status; + } + + // Describes how to run a naive version of the routine (for correctness/performance comparison). + // Note that a proper clBLAS or CPU BLAS comparison is not available for non-BLAS routines. + static StatusCode RunReference1(const Arguments &args, Buffers &buffers, Queue &queue) { + auto buffers_host = BuffersHost(); + DeviceToHost(args, buffers, buffers_host, queue, BuffersIn()); + const auto status = RunReference(args, buffers_host); + HostToDevice(args, buffers, buffers_host, queue, BuffersOut()); + return status; + } + + static StatusCode RunReference2(const Arguments &args, BuffersHost &buffers_host, Queue&) { + return RunReference(args, buffers_host); + } + static StatusCode RunReference3(const Arguments &, BuffersCUDA &, Queue &) { + return StatusCode::kUnknownError; + } + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector DownloadResult(const Arguments &args, Buffers &buffers, Queue &queue) { + std::vector result(args.a_size, static_cast(0)); + buffers.a_mat.Read(queue, args.a_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments &args) { return args.height * args.width; } + static size_t ResultID2(const Arguments &args) { return args.channels; } + static size_t GetResultIndex(const Arguments &args, const size_t id1, const size_t id2) { + return id1 + args.height * args.width * id2 + args.a_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments &) { + return 1; + } + static size_t GetBytes(const Arguments &args) { + const auto im = args.channels * args.width * args.height; // possibly less with striding + const auto col = args.kernel_h * args.kernel_w * NumPatches(args); + return (im + col) * sizeof(T); + } +}; + +// ================================================================================================= + +template +StatusCode RunReference(const Arguments &args, BuffersHost &buffers_host) { + // Reference taken from im2col but swapped the input/output + const auto col_h = TestXcol2im::ColHeight(args); + const auto col_w = TestXcol2im::ColWidth(args); + for (auto c_id = size_t{0}; c_id < args.channels; ++c_id) { // image channels + for (auto kh_id = size_t{0}; kh_id < args.kernel_h; ++kh_id) { // kernel height + for (auto kw_id = size_t{0}; kw_id < args.kernel_w; ++kw_id) { // kernel width + for (auto h_id = size_t{0}; h_id < col_h; ++h_id) { // image height + for (auto w_id = size_t{0}; w_id < col_w; ++w_id) { // image width + + // Reads the input value + const auto kernel_index = kw_id + args.kernel_w * kh_id; + const auto patch_index = w_id + col_w * h_id; + const auto col_index = patch_index + kernel_index * col_w * col_h + + c_id * col_w * col_h * args.kernel_h * args.kernel_w; + const auto val = buffers_host.b_mat[col_index + args.b_offset]; + + // Sets the output value + const auto h_index = kh_id * args.dilation_h + args.stride_h * h_id - args.pad_h; + const auto w_index = kw_id * args.dilation_w + args.stride_w * w_id - args.pad_w; + if (h_index >= 0 && h_index < args.height && + w_index >= 0 && w_index < args.width) { + const auto im_index = w_index + args.width * (h_index + args.height * c_id); + buffers_host.a_mat[im_index + args.a_offset] = val; + } + } + } + } + } + } + return StatusCode::kSuccess; +} + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XCOL2IM_H_ +#endif -- cgit v1.2.3