From f7f8ec644f51d16f888b6a7086009b79c0beef8f Mon Sep 17 00:00:00 2001 From: Cedric Nugteren Date: Thu, 13 Apr 2017 21:31:27 +0200 Subject: Fixed CUDA malloc and cuBLAS handles: cuBLAS as a performance-reference now works --- test/wrapper_cuda.hpp | 96 +++++++++++++++++++++++++++++++++++---------------- 1 file changed, 67 insertions(+), 29 deletions(-) (limited to 'test/wrapper_cuda.hpp') diff --git a/test/wrapper_cuda.hpp b/test/wrapper_cuda.hpp index 509de9d1..51f897c4 100644 --- a/test/wrapper_cuda.hpp +++ b/test/wrapper_cuda.hpp @@ -29,17 +29,47 @@ namespace clblast { // ================================================================================================= +#ifdef CLBLAST_REF_CUBLAS + template + void cublasSetup(Arguments &args) { + cudaSetDevice(static_cast(args.device_id)); + auto status = cublasCreate(reinterpret_cast(&args.cublas_handle)); + if (status != CUBLAS_STATUS_SUCCESS) { + throw std::runtime_error("CUDA cublasCreate error"); + } + } +#endif + +#ifdef CLBLAST_REF_CUBLAS + template + void cublasTeardown(Arguments &args) { + auto status = cublasDestroy(reinterpret_cast(args.cublas_handle)); + if (status != CUBLAS_STATUS_SUCCESS) { + throw std::runtime_error("CUDA cublasDestroy error"); + } + } +#endif + +// ================================================================================================= + // Copies data from the CUDA device to the host and frees-up the CUDA memory afterwards #ifdef CLBLAST_REF_CUBLAS template - void CUDAToHost(T* buffer_cuda, std::vector &buffer_host, const size_t size) { - cudaMemcpy( + void CUDAToHost(T** buffer_cuda, std::vector &buffer_host, const size_t size) { + auto status1 = cudaMemcpy( reinterpret_cast(buffer_host.data()), - reinterpret_cast(buffer_cuda), + reinterpret_cast(*buffer_cuda), size*sizeof(T), cudaMemcpyDeviceToHost ); - cudaFree(buffer_cuda); + if (status1 != cudaSuccess) { + throw std::runtime_error("CUDA cudaMemcpy error with status: "+ToString(static_cast(status1))); + } + auto status2 = cudaFree(*buffer_cuda); + if (status2 != cudaSuccess) { + throw std::runtime_error("CUDA cudaFree error with status: "+ToString(static_cast(status2))); + } + *buffer_cuda = nullptr; } #else template void CUDAToHost(T*, const std::vector&, const size_t) { } @@ -48,14 +78,22 @@ namespace clblast { // Allocates space on the CUDA device and copies in data from the host #ifdef CLBLAST_REF_CUBLAS template - void HostToCUDA(T* buffer_cuda, std::vector &buffer_host, const size_t size) { - cudaMalloc(reinterpret_cast(&buffer_cuda), size*sizeof(T)); - cudaMemcpy( - reinterpret_cast(buffer_cuda), + void HostToCUDA(T** buffer_cuda, std::vector &buffer_host, const size_t size) { + if (*buffer_cuda == nullptr) { + auto status1 = cudaMalloc(reinterpret_cast(buffer_cuda), size*sizeof(T)); + if (status1 != cudaSuccess) { + throw std::runtime_error("CUDA cudaMalloc error with status: "+ToString(static_cast(status1))); + } + } + auto status2 = cudaMemcpy( + reinterpret_cast(*buffer_cuda), reinterpret_cast(buffer_host.data()), size*sizeof(T), cudaMemcpyHostToDevice ); + if (status2 != cudaSuccess) { + throw std::runtime_error("CUDA cudaMemcpy error with status: "+ToString(static_cast(status2))); + } } #else template void HostToCUDA(T*, const std::vector&, const size_t) { } @@ -65,26 +103,26 @@ namespace clblast { template struct BuffersCUDA { - T* x_vec; - T* y_vec; - T* a_mat; - T* b_mat; - T* c_mat; - T* ap_mat; - T* scalar; + T* x_vec = nullptr; + T* y_vec = nullptr; + T* a_mat = nullptr; + T* b_mat = nullptr; + T* c_mat = nullptr; + T* ap_mat = nullptr; + T* scalar = nullptr; }; template void CUDAToHost(const Arguments &args, BuffersCUDA &buffers, BuffersHost &buffers_host, const std::vector &names) { for (auto &name: names) { - if (name == kBufVecX) { buffers_host.x_vec = std::vector(args.x_size, static_cast(0)); CUDAToHost(buffers.x_vec, buffers_host.x_vec, args.x_size); } - else if (name == kBufVecY) { buffers_host.y_vec = std::vector(args.y_size, static_cast(0)); CUDAToHost(buffers.y_vec, buffers_host.y_vec, args.y_size); } - else if (name == kBufMatA) { buffers_host.a_mat = std::vector(args.a_size, static_cast(0)); CUDAToHost(buffers.a_mat, buffers_host.a_mat, args.a_size); } - else if (name == kBufMatB) { buffers_host.b_mat = std::vector(args.b_size, static_cast(0)); CUDAToHost(buffers.b_mat, buffers_host.b_mat, args.b_size); } - else if (name == kBufMatC) { buffers_host.c_mat = std::vector(args.c_size, static_cast(0)); CUDAToHost(buffers.c_mat, buffers_host.c_mat, args.c_size); } - else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector(args.ap_size, static_cast(0)); CUDAToHost(buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } - else if (name == kBufScalar) { buffers_host.scalar = std::vector(args.scalar_size, static_cast(0)); CUDAToHost(buffers.scalar, buffers_host.scalar, args.scalar_size); } + if (name == kBufVecX) { buffers_host.x_vec = std::vector(args.x_size, static_cast(0)); CUDAToHost(&buffers.x_vec, buffers_host.x_vec, args.x_size); } + else if (name == kBufVecY) { buffers_host.y_vec = std::vector(args.y_size, static_cast(0)); CUDAToHost(&buffers.y_vec, buffers_host.y_vec, args.y_size); } + else if (name == kBufMatA) { buffers_host.a_mat = std::vector(args.a_size, static_cast(0)); CUDAToHost(&buffers.a_mat, buffers_host.a_mat, args.a_size); } + else if (name == kBufMatB) { buffers_host.b_mat = std::vector(args.b_size, static_cast(0)); CUDAToHost(&buffers.b_mat, buffers_host.b_mat, args.b_size); } + else if (name == kBufMatC) { buffers_host.c_mat = std::vector(args.c_size, static_cast(0)); CUDAToHost(&buffers.c_mat, buffers_host.c_mat, args.c_size); } + else if (name == kBufMatAP) { buffers_host.ap_mat = std::vector(args.ap_size, static_cast(0)); CUDAToHost(&buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } + else if (name == kBufScalar) { buffers_host.scalar = std::vector(args.scalar_size, static_cast(0)); CUDAToHost(&buffers.scalar, buffers_host.scalar, args.scalar_size); } else { throw std::runtime_error("Invalid buffer name"); } } } @@ -93,13 +131,13 @@ template void HostToCUDA(const Arguments &args, BuffersCUDA &buffers, BuffersHost &buffers_host, const std::vector &names) { for (auto &name: names) { - if (name == kBufVecX) { HostToCUDA(buffers.x_vec, buffers_host.x_vec, args.x_size); } - else if (name == kBufVecY) { HostToCUDA(buffers.y_vec, buffers_host.y_vec, args.y_size); } - else if (name == kBufMatA) { HostToCUDA(buffers.a_mat, buffers_host.a_mat, args.a_size); } - else if (name == kBufMatB) { HostToCUDA(buffers.b_mat, buffers_host.b_mat, args.b_size); } - else if (name == kBufMatC) { HostToCUDA(buffers.c_mat, buffers_host.c_mat, args.c_size); } - else if (name == kBufMatAP) { HostToCUDA(buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } - else if (name == kBufScalar) { HostToCUDA(buffers.scalar, buffers_host.scalar, args.scalar_size); } + if (name == kBufVecX) { HostToCUDA(&buffers.x_vec, buffers_host.x_vec, args.x_size); } + else if (name == kBufVecY) { HostToCUDA(&buffers.y_vec, buffers_host.y_vec, args.y_size); } + else if (name == kBufMatA) { HostToCUDA(&buffers.a_mat, buffers_host.a_mat, args.a_size); } + else if (name == kBufMatB) { HostToCUDA(&buffers.b_mat, buffers_host.b_mat, args.b_size); } + else if (name == kBufMatC) { HostToCUDA(&buffers.c_mat, buffers_host.c_mat, args.c_size); } + else if (name == kBufMatAP) { HostToCUDA(&buffers.ap_mat, buffers_host.ap_mat, args.ap_size); } + else if (name == kBufScalar) { HostToCUDA(&buffers.scalar, buffers_host.scalar, args.scalar_size); } else { throw std::runtime_error("Invalid buffer name"); } } } -- cgit v1.2.3