diff options
Diffstat (limited to 'test')
-rw-r--r-- | test/correctness/routines/levelx/xomatcopy.cc | 30 | ||||
-rw-r--r-- | test/correctness/testblas.cc | 4 | ||||
-rw-r--r-- | test/correctness/testblas.h | 16 | ||||
-rw-r--r-- | test/performance/client.cc | 17 | ||||
-rw-r--r-- | test/performance/client.h | 17 | ||||
-rw-r--r-- | test/performance/routines/levelx/xomatcopy.cc | 36 | ||||
-rw-r--r-- | test/routines/levelx/xomatcopy.h | 164 |
7 files changed, 270 insertions, 14 deletions
diff --git a/test/correctness/routines/levelx/xomatcopy.cc b/test/correctness/routines/levelx/xomatcopy.cc new file mode 100644 index 00000000..43021bc9 --- /dev/null +++ b/test/correctness/routines/levelx/xomatcopy.cc @@ -0,0 +1,30 @@ + +// ================================================================================================= +// 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 <www.cedricnugteren.nl> +// +// ================================================================================================= + +#include "correctness/testblas.h" +#include "routines/levelx/xomatcopy.h" + +// Shortcuts to the clblast namespace +using float2 = clblast::float2; +using double2 = clblast::double2; + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + auto errors = size_t{0}; + errors += clblast::RunTests<clblast::TestXomatcopy<float>, float, float>(argc, argv, false, "SOMATCOPY"); + errors += clblast::RunTests<clblast::TestXomatcopy<double>, double, double>(argc, argv, true, "DOMATCOPY"); + errors += clblast::RunTests<clblast::TestXomatcopy<float2>, float2, float2>(argc, argv, true, "COMATCOPY"); + errors += clblast::RunTests<clblast::TestXomatcopy<double2>, double2, double2>(argc, argv, true, "ZOMATCOPY"); + errors += clblast::RunTests<clblast::TestXomatcopy<half>, half, half>(argc, argv, true, "HOMATCOPY"); + if (errors > 0) { return 1; } else { return 0; } +} + +// ================================================================================================= diff --git a/test/correctness/testblas.cc b/test/correctness/testblas.cc index e5addc48..9f842d91 100644 --- a/test/correctness/testblas.cc +++ b/test/correctness/testblas.cc @@ -174,8 +174,8 @@ void TestBlas<T,U>::TestRegular(std::vector<Arguments<U>> &test_vector, const st template <typename T, typename U> void TestBlas<T,U>::TestInvalid(std::vector<Arguments<U>> &test_vector, const std::string &name) { if (!PrecisionSupported<T>(device_)) { return; } - if (!compare_clblas_) { return; } - if (std::is_same<T, half>::value) { return; } + if (!compare_clblas_) { return; } // not supported for CPU BLAS routines + if (std::is_same<T, half>::value) { return; } // not supported for half-precision TestStart("invalid buffer sizes", name); // Iterates over all the to-be-tested combinations of arguments diff --git a/test/correctness/testblas.h b/test/correctness/testblas.h index cce10751..e849466a 100644 --- a/test/correctness/testblas.h +++ b/test/correctness/testblas.h @@ -129,6 +129,13 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na const auto reference_routine2 = C::RunReference2; // CBLAS #endif + // Non-BLAS routines cannot be fully tested + if (!silent && C::BLASLevel() == 4) { + fprintf(stdout, "\n* NOTE: This non-BLAS routine is tested against a custom implementation,\n"); + fprintf(stdout, " not against clBLAS or a CPU BLAS library. Thus, the arguments '-clblas'\n"); + fprintf(stdout, " and '-cblas' have no effect.\n"); + } + // Creates a tester auto options = C::GetOptions(); TestBlas<T,U> tester{argc, argv, silent, name, options, @@ -176,8 +183,9 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na auto ap_sizes = std::vector<size_t>{args.ap_size}; // Sets the dimensions of the matrices or vectors depending on the BLAS level - auto dimensions = (C::BLASLevel() == 3) ? tester.kMatrixDims : - (C::BLASLevel() == 2) ? tester.kMatrixVectorDims : + auto dimensions = (C::BLASLevel() == 4) ? tester.kMatrixDims : // non-BLAS extra routines + (C::BLASLevel() == 3) ? tester.kMatrixDims : // level 3 + (C::BLASLevel() == 2) ? tester.kMatrixVectorDims : // level 2 tester.kVectorDims; // else: level 1 // For the options relevant to this routine, sets the vectors to proper values @@ -318,7 +326,9 @@ size_t RunTests(int argc, char *argv[], const bool silent, const std::string &na // Runs the tests tester.TestRegular(regular_test_vector, case_name); #ifdef CLBLAST_REF_CLBLAS - tester.TestInvalid(invalid_test_vector, case_name); + if (C::BLASLevel() != 4) { + tester.TestInvalid(invalid_test_vector, case_name); + } #endif } } diff --git a/test/performance/client.cc b/test/performance/client.cc index 5a7226df..4c0c79a6 100644 --- a/test/performance/client.cc +++ b/test/performance/client.cc @@ -42,8 +42,10 @@ Client<T,U>::Client(const Routine run_routine, // applicable, but are searched for anyway to be able to create one common argument parser. All // arguments have a default value in case they are not found. template <typename T, typename U> -Arguments<U> Client<T,U>::ParseArguments(int argc, char *argv[], const GetMetric default_a_ld, - const GetMetric default_b_ld, const GetMetric default_c_ld) { +Arguments<U> Client<T,U>::ParseArguments(int argc, char *argv[], const size_t level, + const GetMetric default_a_ld, + const GetMetric default_b_ld, + const GetMetric default_c_ld) { auto args = Arguments<U>{}; auto help = std::string{"\n* Options given/available:\n"}; @@ -116,6 +118,17 @@ Arguments<U> Client<T,U>::ParseArguments(int argc, char *argv[], const GetMetric // which is thus always displayed (unless silence is specified). if (!args.silent) { fprintf(stdout, "%s\n", help.c_str()); } + // Comparison against a non-BLAS routine is not supported + if (level == 4) { // level-4 == level-X + if (args.compare_clblas != 0 || args.compare_cblas != 0) { + if (!args.silent) { + fprintf(stdout, "* Disabling clBLAS and CPU BLAS comparisons for this non-BLAS routine\n\n"); + } + } + args.compare_clblas = 0; + args.compare_cblas = 0; + } + // Comparison against clBLAS or a CPU BLAS library is not supported in case of half-precision if (args.precision == Precision::kHalf) { if (args.compare_clblas != 0 || args.compare_cblas != 0) { diff --git a/test/performance/client.h b/test/performance/client.h index 8d0597d7..493a7aed 100644 --- a/test/performance/client.h +++ b/test/performance/client.h @@ -53,8 +53,10 @@ class Client { // Parses all command-line arguments, filling in the arguments structure. If no command-line // argument is given for a particular argument, it is filled in with a default value. - Arguments<U> ParseArguments(int argc, char *argv[], const GetMetric default_a_ld, - const GetMetric default_b_ld, const GetMetric default_c_ld); + Arguments<U> ParseArguments(int argc, char *argv[], const size_t level, + const GetMetric default_a_ld, + const GetMetric default_b_ld, + const GetMetric default_c_ld); // The main client function, setting-up arguments, matrices, OpenCL buffers, etc. After set-up, it // calls the client routines. @@ -97,14 +99,14 @@ void RunClient(int argc, char *argv[]) { // Sets the reference to test against #ifdef CLBLAST_REF_CLBLAS - const auto reference1 = C::RunReference1; // clBLAS when available + auto reference1 = C::RunReference1; // clBLAS when available #else - const auto reference1 = ReferenceNotAvailable<T,U>; + auto reference1 = ReferenceNotAvailable<T,U>; #endif #ifdef CLBLAST_REF_CBLAS - const auto reference2 = C::RunReference2; // CBLAS when available + auto reference2 = C::RunReference2; // CBLAS when available #else - const auto reference2 = ReferenceNotAvailable<T,U>; + auto reference2 = ReferenceNotAvailable<T,U>; #endif // Creates a new client @@ -112,7 +114,8 @@ void RunClient(int argc, char *argv[]) { C::GetFlops, C::GetBytes); // Simple command line argument parser with defaults - auto args = client.ParseArguments(argc, argv, C::DefaultLDA, C::DefaultLDB, C::DefaultLDC); + auto args = client.ParseArguments(argc, argv, C::BLASLevel(), + C::DefaultLDA, C::DefaultLDB, C::DefaultLDC); if (args.print_help) { return; } // Runs the client diff --git a/test/performance/routines/levelx/xomatcopy.cc b/test/performance/routines/levelx/xomatcopy.cc new file mode 100644 index 00000000..851f6ee1 --- /dev/null +++ b/test/performance/routines/levelx/xomatcopy.cc @@ -0,0 +1,36 @@ + +// ================================================================================================= +// 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 <www.cedricnugteren.nl> +// +// ================================================================================================= + +#include "performance/client.h" +#include "routines/levelx/xomatcopy.h" + +// Shortcuts to the clblast namespace +using float2 = clblast::float2; +using double2 = clblast::double2; + +// Main function (not within the clblast namespace) +int main(int argc, char *argv[]) { + switch(clblast::GetPrecision(argc, argv, clblast::Precision::kSingle)) { + case clblast::Precision::kHalf: + clblast::RunClient<clblast::TestXomatcopy<half>, half, half>(argc, argv); break; + case clblast::Precision::kSingle: + clblast::RunClient<clblast::TestXomatcopy<float>, float, float>(argc, argv); break; + case clblast::Precision::kDouble: + clblast::RunClient<clblast::TestXomatcopy<double>, double, double>(argc, argv); break; + case clblast::Precision::kComplexSingle: + clblast::RunClient<clblast::TestXomatcopy<float2>, float2, float2>(argc, argv); break; + case clblast::Precision::kComplexDouble: + clblast::RunClient<clblast::TestXomatcopy<double2>, double2, double2>(argc, argv); break; + } + return 0; +} + +// ================================================================================================= diff --git a/test/routines/levelx/xomatcopy.h b/test/routines/levelx/xomatcopy.h new file mode 100644 index 00000000..c499ba75 --- /dev/null +++ b/test/routines/levelx/xomatcopy.h @@ -0,0 +1,164 @@ + +// ================================================================================================= +// 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 <www.cedricnugteren.nl> +// +// This file implements a class with static methods to describe the Xomatcopy 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_XOMATCOPY_H_ +#define CLBLAST_TEST_ROUTINES_XOMATCOPY_H_ + +#include <vector> +#include <string> + +namespace clblast { +// ================================================================================================= + +// See comment at top of file for a description of the class +template <typename T> +class TestXomatcopy { + 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<std::string> GetOptions() { + return {kArgM, kArgN, + kArgLayout, kArgATransp, + kArgALeadDim, kArgBLeadDim, + kArgAOffset, kArgBOffset, + kArgAlpha}; + } + + // Describes how to obtain the sizes of the buffers + static size_t GetSizeA(const Arguments<T> &args) { + const auto a_rotated = (args.layout == Layout::kRowMajor); + const auto a_two = (a_rotated) ? args.m : args.n; + return a_two * args.a_ld + args.a_offset; + } + static size_t GetSizeB(const Arguments<T> &args) { + const auto b_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + const auto b_two = (b_rotated) ? args.n : args.m; + return b_two * args.b_ld + args.b_offset; + } + + // Describes how to set the sizes of all the buffers + static void SetSizes(Arguments<T> &args) { + args.a_size = GetSizeA(args); + args.b_size = GetSizeB(args); + } + + // Describes what the default values of the leading dimensions of the matrices are + static size_t DefaultLDA(const Arguments<T> &args) { return args.n; } + static size_t DefaultLDB(const Arguments<T> &args) { return args.m; } + static size_t DefaultLDC(const Arguments<T> &) { return 1; } // N/A for this routine + + // Describes which omatcopyose options are relevant for this routine + using Transposes = std::vector<Transpose>; + static Transposes GetATransposes(const Transposes &all) { return all; } + static Transposes GetBTransposes(const Transposes &) { return {}; } // N/A for this routine + + // Describes how to run the CLBlast routine + static StatusCode RunRoutine(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + auto queue_plain = queue(); + auto event = cl_event{}; + auto status = Omatcopy<T>(args.layout, args.a_transpose, + args.m, args.n, args.alpha, + buffers.a_mat(), args.a_offset, args.a_ld, + buffers.b_mat(), args.b_offset, args.b_ld, + &queue_plain, &event); + clWaitForEvents(1, &event); + 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. + + #ifdef CLBLAST_REF_CLBLAS + static StatusCode RunReference1(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + return RunReference2(args, buffers, queue); + } + #endif + + #ifdef CLBLAST_REF_CBLAS + static StatusCode RunReference2(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + + // Data transfer from OpenCL to std::vector + std::vector<T> a_mat_cpu(args.a_size, static_cast<T>(0)); + std::vector<T> b_mat_cpu(args.b_size, static_cast<T>(0)); + buffers.a_mat.Read(queue, args.a_size, a_mat_cpu); + buffers.b_mat.Read(queue, args.b_size, b_mat_cpu); + + // Checking for invalid arguments + const auto a_rotated = (args.layout == Layout::kRowMajor); + const auto b_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + const auto a_base = (a_rotated) ? args.a_ld*(args.m-1) + args.n : args.a_ld*(args.n-1) + args.m; + const auto b_base = (b_rotated) ? args.b_ld*(args.m-1) + args.n : args.b_ld*(args.n-1) + args.m; + if ((args.m == 0) || (args.n == 0)) { return StatusCode::kInvalidDimension; } + if ((args.a_ld < args.m && !a_rotated) || (args.a_ld < args.n && a_rotated)) { return StatusCode::kInvalidLeadDimA; } + if ((args.b_ld < args.m && !b_rotated) || (args.b_ld < args.n && b_rotated)) { return StatusCode::kInvalidLeadDimB; } + if (buffers.a_mat.GetSize() < (a_base + args.a_offset) * sizeof(T)) { return StatusCode::kInsufficientMemoryA; } + if (buffers.b_mat.GetSize() < (b_base + args.b_offset) * sizeof(T)) { return StatusCode::kInsufficientMemoryB; } + + // Matrix copy, scaling, and/or transpose + for (auto id1 = size_t{0}; id1 < args.m; ++id1) { + for (auto id2 = size_t{0}; id2 < args.n; ++id2) { + const auto a_one = (a_rotated) ? id2 : id1; + const auto a_two = (a_rotated) ? id1 : id2; + const auto b_one = (b_rotated) ? id2 : id1; + const auto b_two = (b_rotated) ? id1 : id2; + const auto a_index = a_two * args.a_ld + a_one + args.a_offset; + const auto b_index = b_two * args.b_ld + b_one + args.b_offset; + b_mat_cpu[b_index] = args.alpha * a_mat_cpu[a_index]; + } + } + + // Data transfer back to OpenCL + buffers.b_mat.Write(queue, args.b_size, b_mat_cpu); + return StatusCode::kSuccess; + } + #endif + + // Describes how to download the results of the computation (more importantly: which buffer) + static std::vector<T> DownloadResult(const Arguments<T> &args, Buffers<T> &buffers, Queue &queue) { + std::vector<T> result(args.b_size, static_cast<T>(0)); + buffers.b_mat.Read(queue, args.b_size, result); + return result; + } + + // Describes how to compute the indices of the result buffer + static size_t ResultID1(const Arguments<T> &args) { return args.m; } + static size_t ResultID2(const Arguments<T> &args) { return args.n; } + static size_t GetResultIndex(const Arguments<T> &args, const size_t id1, const size_t id2) { + const auto b_rotated = (args.layout == Layout::kColMajor && args.a_transpose != Transpose::kNo) || + (args.layout == Layout::kRowMajor && args.a_transpose == Transpose::kNo); + const auto b_one = (b_rotated) ? id2 : id1; + const auto b_two = (b_rotated) ? id1 : id2; + return b_two * args.b_ld + b_one + args.b_offset; + } + + // Describes how to compute performance metrics + static size_t GetFlops(const Arguments<T> &args) { + return args.m*args.n; + } + static size_t GetBytes(const Arguments<T> &args) { + return (2*args.m*args.n) * sizeof(T); + } +}; + +// ================================================================================================= +} // namespace clblast + +// CLBLAST_TEST_ROUTINES_XOMATCOPY_H_ +#endif |