wehub-resource-sync
158 lines
6.4 KiB
C++
158 lines
6.4 KiB
C++
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License. */
|
|
|
|
#pragma once
|
|
|
|
#include <cublas_v2.h>
|
|
#include <cuda.h>
|
|
#if CUDA_VERSION >= 12030 && defined(__linux__)
|
|
#include <cublas_api.h>
|
|
#endif
|
|
|
|
#include <mutex> // NOLINT
|
|
#include <type_traits>
|
|
|
|
#include "paddle/phi/backends/dynload/dynamic_loader.h"
|
|
#include "paddle/phi/common/port.h"
|
|
|
|
namespace phi {
|
|
namespace dynload {
|
|
|
|
extern std::once_flag cublas_dso_flag;
|
|
extern void *cublas_dso_handle;
|
|
|
|
/**
|
|
* The following macro definition can generate structs
|
|
* (for each function) to dynamic load cublas routine
|
|
* via operator overloading.
|
|
*
|
|
* note: default dynamic linked libs
|
|
*/
|
|
#define DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP(__name) \
|
|
struct DynLoad__##__name { \
|
|
template <typename... Args> \
|
|
inline auto operator()(Args... args) -> DECLARE_TYPE(__name, args...) { \
|
|
using cublas_func = \
|
|
decltype(::__name(std::declval<Args>()...)) (*)(Args...); \
|
|
std::call_once(cublas_dso_flag, []() { \
|
|
cublas_dso_handle = phi::dynload::GetCublasDsoHandle(); \
|
|
}); \
|
|
static void *p_##__name = dlsym(cublas_dso_handle, #__name); \
|
|
return reinterpret_cast<cublas_func>(p_##__name)(args...); \
|
|
} \
|
|
}; \
|
|
extern DynLoad__##__name __name
|
|
|
|
#define CUBLAS_BLAS_ROUTINE_EACH(__macro) \
|
|
__macro(cublasSaxpy_v2); \
|
|
__macro(cublasDaxpy_v2); \
|
|
__macro(cublasCaxpy_v2); \
|
|
__macro(cublasZaxpy_v2); \
|
|
__macro(cublasSscal_v2); \
|
|
__macro(cublasDscal_v2); \
|
|
__macro(cublasScopy_v2); \
|
|
__macro(cublasDcopy_v2); \
|
|
__macro(cublasSgemv_v2); \
|
|
__macro(cublasDgemv_v2); \
|
|
__macro(cublasCgemv_v2); \
|
|
__macro(cublasZgemv_v2); \
|
|
__macro(cublasSgemm_v2); \
|
|
__macro(cublasDgemm_v2); \
|
|
__macro(cublasCgemm_v2); \
|
|
__macro(cublasZgemm_v2); \
|
|
__macro(cublasHgemm); \
|
|
__macro(cublasSgemmEx); \
|
|
__macro(cublasSgeam); \
|
|
__macro(cublasDgeam); \
|
|
__macro(cublasStrsm_v2); \
|
|
__macro(cublasDtrsm_v2); \
|
|
__macro(cublasCtrsm_v2); \
|
|
__macro(cublasZtrsm_v2); \
|
|
__macro(cublasCreate_v2); \
|
|
__macro(cublasDestroy_v2); \
|
|
__macro(cublasSetStream_v2); \
|
|
__macro(cublasSetPointerMode_v2); \
|
|
__macro(cublasGetPointerMode_v2); \
|
|
__macro(cublasSgemmBatched); \
|
|
__macro(cublasDgemmBatched); \
|
|
__macro(cublasCgemmBatched); \
|
|
__macro(cublasZgemmBatched); \
|
|
__macro(cublasStrsmBatched); \
|
|
__macro(cublasDtrsmBatched); \
|
|
__macro(cublasCtrsmBatched); \
|
|
__macro(cublasZtrsmBatched); \
|
|
__macro(cublasSgetrfBatched); \
|
|
__macro(cublasSgetriBatched); \
|
|
__macro(cublasDgetrfBatched); \
|
|
__macro(cublasDgetriBatched); \
|
|
__macro(cublasCgetrfBatched); \
|
|
__macro(cublasCgetriBatched); \
|
|
__macro(cublasZgetrfBatched); \
|
|
__macro(cublasZgetriBatched); \
|
|
__macro(cublasSmatinvBatched); \
|
|
__macro(cublasDmatinvBatched); \
|
|
__macro(cublasCmatinvBatched); \
|
|
__macro(cublasZmatinvBatched); \
|
|
__macro(cublasSgetrsBatched); \
|
|
__macro(cublasDgetrsBatched); \
|
|
__macro(cublasSdot_v2); \
|
|
__macro(cublasDdot_v2); \
|
|
__macro(cublasCdotc_v2); \
|
|
__macro(cublasZdotc_v2); \
|
|
__macro(cublasCdotu_v2); \
|
|
__macro(cublasZdotu_v2); \
|
|
__macro(cublasDotEx); \
|
|
__macro(cublasGemmEx); \
|
|
__macro(cublasSgemmStridedBatched); \
|
|
__macro(cublasDgemmStridedBatched); \
|
|
__macro(cublasCgemmStridedBatched); \
|
|
__macro(cublasZgemmStridedBatched); \
|
|
__macro(cublasHgemmStridedBatched); \
|
|
__macro(cublasSetMathMode); \
|
|
__macro(cublasGetMathMode); \
|
|
__macro(cublasCgeam); \
|
|
__macro(cublasZgeam); \
|
|
__macro(cublasGemmBatchedEx); \
|
|
__macro(cublasGemmStridedBatchedEx);
|
|
|
|
CUBLAS_BLAS_ROUTINE_EACH(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
|
|
|
|
// NVIDIA's cublas_v2.h defines: #define cublasSetWorkspace
|
|
// cublasSetWorkspace_v2 The actual exported symbol in NVIDIA's libcublas.so is
|
|
// cublasSetWorkspace_v2, so dlsym must look up "cublasSetWorkspace_v2" (the _v2
|
|
// name). Non-NVIDIA toolchains (e.g. Iluvatar/COREX) do NOT define this macro
|
|
// and export only "cublasSetWorkspace" in their shared library. We use #ifdef
|
|
// to detect which symbol name dlsym should look up.
|
|
#if !defined(_WIN32)
|
|
#ifdef cublasSetWorkspace // NVIDIA: macro maps to cublasSetWorkspace_v2
|
|
#define CUBLAS_WORKSPACE_ROUTINE(__macro) __macro(cublasSetWorkspace_v2);
|
|
#else // Iluvatar/COREX: only cublasSetWorkspace exists
|
|
#define CUBLAS_WORKSPACE_ROUTINE(__macro) __macro(cublasSetWorkspace);
|
|
#endif
|
|
CUBLAS_WORKSPACE_ROUTINE(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
|
|
#endif
|
|
|
|
#if CUDA_VERSION >= 12030 && defined(__linux__)
|
|
#define CUBLAS_BLAS_ROUTINE_EACH_R5(__macro) \
|
|
__macro(cublasGemmStridedBatchedEx_64); \
|
|
__macro(cublasGemmEx_64); \
|
|
__macro(cublasSgemmEx_64);
|
|
|
|
CUBLAS_BLAS_ROUTINE_EACH_R5(DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP)
|
|
#endif
|
|
|
|
#undef DECLARE_DYNAMIC_LOAD_CUBLAS_WRAP
|
|
} // namespace dynload
|
|
} // namespace phi
|