cublas_impl.h

Go to the documentation of this file.

/*
 *  SPDX-FileCopyrightText: Copyright 2021, Siavash Ameli <sameli@berkeley.edu>
 *  SPDX-License-Identifier: BSD-3-Clause
 *  SPDX-FileType: SOURCE
 *
 *  This program is free software: you can redistribute it and/or modify it
 *  under the terms of the license found in the LICENSE.txt file in the root
 *  directory of this source tree.
 */
 
#ifndef _CU_BASIC_ALGEBRA_CUBLAS_IMPL_H_
#define _CU_BASIC_ALGEBRA_CUBLAS_IMPL_H_
 
// =======
// Headers
// =======
 
// Avoid CUBLAS numeration value not handled in switch [-Wswitch-enum] warning
#ifdef _MSC_VER
    #pragma warning(push, 0)  // Suppress all warnings from the followings
    #include <cublas_v2.h>
    #pragma warning(pop)  // Restore previous warning level
#elif defined(__INTEL_LLVM_COMPILER) || defined(__INTEL_COMPILER)
    #pragma warning(push, 0)
    #include <cublas_v2.h>
    #pragma warning(pop)
#elif defined(__GNUC__) || defined(__clang__)
    #pragma GCC diagnostic push
    #pragma GCC diagnostic ignored "-Wswitch-enum"
    #include <cublas_v2.h>
    #pragma GCC diagnostic pop
#else
    #include <cublas_v2.h>  // cublasOperation_t, CUBLAS_OP_N, CUBLAS_OP_T,
                            // cudaError_t
#endif
 
// Restrict qualifier
// Note: generally, the restrict qualifier is only needed to be be added to the
// function signature in the implementation (*.cu files) and not the
// declarations (in *h files). However, CUDA seems to have a bug that the
// restrict qualifier should be added to BOTH declaration and implementation,
// otherwise, it corrupts the array pointers.
#if defined(_MSC_VER)
    #define RESTRICT __restrict
#elif defined(__INTEL_COMPILER)
    #define RESTRICT __restrict
#elif defined(__CUDA__) || defined(__GNUC__) || defined(__clang__)
    #define RESTRICT __restrict__
#else
    #define RESTRICT
#endif
 
 
// ===========
// cublas impl
// ===========
 
 
namespace cublas_impl
{
    // cublasTgemv 
    template <typename DataType, typename ComputeType>
    cudaError_t cublasTgemv(
            cublasOperation_t trans,
            int m,
            int n,
            const DataType* RESTRICT alpha,
            const DataType* RESTRICT A,
            int lda,
            const DataType* RESTRICT x,
            int incx,
            const DataType* RESTRICT beta,
            DataType* RESTRICT y,
            int incy);
 
    // cublasTcopy
    template <typename DataType>
    cudaError_t cublasTcopy(
            int n,
            const DataType* RESTRICT x,
            int incx,
            DataType* RESTRICT y,
            int incy);
 
    // cublasTaxpy
    template <typename DataType>
    cudaError_t cublasTaxpy(
            int n,
            const DataType* RESTRICT alpha,
            const DataType* RESTRICT x,
            int incx,
            DataType* RESTRICT y,
            int incy);
 
    // cublasTdot
    template <typename DataType, typename ComputeType>
    cudaError_t cublasTdot(
            int n,
            const DataType* RESTRICT x,
            int incx,
            const DataType* RESTRICT y,
            int incy,
            DataType* RESTRICT result);
 
    // cublasTnrm2
    template <typename DataType, typename ComputeType>
    cudaError_t cublasTnrm2(
            int n,
            const DataType* RESTRICT x,
            int incx,
            DataType* RESTRICT result);
 
    // cublasTscal
    template <typename DataType>
    cudaError_t cublasTscal(
            int n,
            const DataType* RESTRICT alpha,
            DataType* RESTRICT x,
            int incx);
}
 
#endif  // _CU_BASIC_ALGEBRA_CUBLAS_IMPL_H_