diff options
Diffstat (limited to 'test/routines/levelx/xgemmbatched.hpp')
| -rw-r--r-- | test/routines/levelx/xgemmbatched.hpp | 27 |
1 files changed, 18 insertions, 9 deletions
diff --git a/test/routines/levelx/xgemmbatched.hpp b/test/routines/levelx/xgemmbatched.hpp index ab5f20c5..ebfd8b19 100644 --- a/test/routines/levelx/xgemmbatched.hpp +++ b/test/routines/levelx/xgemmbatched.hpp @@ -16,15 +16,7 @@ #ifndef CLBLAST_TEST_ROUTINES_XGEMMBATCHED_H_ #define CLBLAST_TEST_ROUTINES_XGEMMBATCHED_H_ -#include <vector> -#include <string> - -#ifdef CLBLAST_REF_CLBLAS - #include "test/wrapper_clblas.hpp" -#endif -#ifdef CLBLAST_REF_CBLAS - #include "test/wrapper_cblas.hpp" -#endif +#include "test/routines/common.hpp" namespace clblast { // ================================================================================================= @@ -168,6 +160,23 @@ class TestXgemmBatched { } #endif + // Describes how to run the cuBLAS routine (for correctness/performance comparison) + #ifdef CLBLAST_REF_CUBLAS + static StatusCode RunReference3(const Arguments<T> &args, BuffersCUDA<T> &buffers, Queue &) { + for (auto batch = size_t{0}; batch < args.batch_count; ++batch) { + auto status = cublasXgemm(reinterpret_cast<cublasHandle_t>(args.cublas_handle), args.layout, + convertToCUBLAS(args.a_transpose), + convertToCUBLAS(args.b_transpose), + args.m, args.n, args.k, args.alphas[batch], + buffers.a_mat, args.a_offsets[batch], args.a_ld, + buffers.b_mat, args.b_offsets[batch], args.b_ld, args.betas[batch], + buffers.c_mat, args.c_offsets[batch], args.c_ld); + if (status != CUBLAS_STATUS_SUCCESS) { return StatusCode::kUnknownError; } + } + 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.c_size, static_cast<T>(0)); |