/* ************************************************************************ * Copyright 2018-2019 Advanced Micro Devices, Inc. * * ************************************************************************ */ #ifndef _ROCBLAS_HPP_ #define _ROCBLAS_HPP_ /* library headers */ #include "rocblas.h" /*!\file * This file exposes C++ templated BLAS interface with only the precision templated. */ /* * =========================================================================== * README: Please follow the naming convention * Big case for matrix, e.g. matrix A, B, C GEMM (C = A*B) * Lower case for vector, e.g. vector x, y GEMV (y = A*x) * =========================================================================== */ /* * =========================================================================== * level 1 BLAS * =========================================================================== */ // scal template rocblas_status (*rocblas_scal)( rocblas_handle handle, rocblas_int n, const U* alpha, T* x, rocblas_int incx); template <> static constexpr auto rocblas_scal = rocblas_sscal; template <> static constexpr auto rocblas_scal = rocblas_dscal; template <> static constexpr auto rocblas_scal = rocblas_cscal; template <> static constexpr auto rocblas_scal = rocblas_zscal; template <> static constexpr auto rocblas_scal = rocblas_csscal; template <> static constexpr auto rocblas_scal = rocblas_zdscal; // scal_batched template rocblas_status (*rocblas_scal_batched)(rocblas_handle handle, rocblas_int n, const U* alpha, T* const x[], rocblas_int incx, rocblas_int batch_count); template <> static constexpr auto rocblas_scal_batched = rocblas_sscal_batched; template <> static constexpr auto rocblas_scal_batched = rocblas_dscal_batched; template <> static constexpr auto rocblas_scal_batched = rocblas_cscal_batched; template <> static constexpr auto rocblas_scal_batched = rocblas_zscal_batched; template <> static constexpr auto rocblas_scal_batched = rocblas_csscal_batched; template <> static constexpr auto rocblas_scal_batched = rocblas_zdscal_batched; // scal_strided_batched template rocblas_status (*rocblas_scal_strided_batched)(rocblas_handle handle, rocblas_int n, const U* alpha, T* x, rocblas_int incx, rocblas_stride stride_x, rocblas_int batch_count); template <> static constexpr auto rocblas_scal_strided_batched = rocblas_sscal_strided_batched; template <> static constexpr auto rocblas_scal_strided_batched = rocblas_dscal_strided_batched; template <> static constexpr auto rocblas_scal_strided_batched = rocblas_cscal_strided_batched; template <> static constexpr auto rocblas_scal_strided_batched = rocblas_zscal_strided_batched; template <> static constexpr auto rocblas_scal_strided_batched = rocblas_csscal_strided_batched; template <> static constexpr auto rocblas_scal_strided_batched = rocblas_zdscal_strided_batched; // copy template rocblas_status (*rocblas_copy)( rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, T* y, rocblas_int incy); template <> static constexpr auto rocblas_copy = rocblas_scopy; template <> static constexpr auto rocblas_copy = rocblas_dcopy; template <> static constexpr auto rocblas_copy = rocblas_ccopy; template <> static constexpr auto rocblas_copy = rocblas_zcopy; template rocblas_status (*rocblas_copy_batched)(rocblas_handle handle, rocblas_int n, const T* const x[], rocblas_int incx, T* const y[], rocblas_int incy, rocblas_int batch_count); template <> static constexpr auto rocblas_copy_batched = rocblas_scopy_batched; template <> static constexpr auto rocblas_copy_batched = rocblas_dcopy_batched; template <> static constexpr auto rocblas_copy_batched = rocblas_ccopy_batched; template <> static constexpr auto rocblas_copy_batched = rocblas_zcopy_batched; template rocblas_status (*rocblas_copy_strided_batched)(rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, rocblas_stride stridex, T* y, rocblas_int incy, rocblas_stride stridey, rocblas_int batch_count); template <> static constexpr auto rocblas_copy_strided_batched = rocblas_scopy_strided_batched; template <> static constexpr auto rocblas_copy_strided_batched = rocblas_dcopy_strided_batched; template <> static constexpr auto rocblas_copy_strided_batched = rocblas_ccopy_strided_batched; template <> static constexpr auto rocblas_copy_strided_batched = rocblas_zcopy_strided_batched; // swap template rocblas_status (*rocblas_swap)( rocblas_handle handle, rocblas_int n, T* x, rocblas_int incx, T* y, rocblas_int incy); template <> static constexpr auto rocblas_swap = rocblas_sswap; template <> static constexpr auto rocblas_swap = rocblas_dswap; template <> static constexpr auto rocblas_swap = rocblas_cswap; template <> static constexpr auto rocblas_swap = rocblas_zswap; // swap_batched template rocblas_status (*rocblas_swap_batched)(rocblas_handle handle, rocblas_int n, T* x[], rocblas_int incx, T* y[], rocblas_int incy, rocblas_int batch_count); template <> static constexpr auto rocblas_swap_batched = rocblas_sswap_batched; template <> static constexpr auto rocblas_swap_batched = rocblas_dswap_batched; template <> static constexpr auto rocblas_swap_batched = rocblas_cswap_batched; template <> static constexpr auto rocblas_swap_batched = rocblas_zswap_batched; // swap_strided_batched template rocblas_status (*rocblas_swap_strided_batched)(rocblas_handle handle, rocblas_int n, T* x, rocblas_int incx, rocblas_stride stridex, T* y, rocblas_int incy, rocblas_stride stridey, rocblas_int batch_count); template <> static constexpr auto rocblas_swap_strided_batched = rocblas_sswap_strided_batched; template <> static constexpr auto rocblas_swap_strided_batched = rocblas_dswap_strided_batched; template <> static constexpr auto rocblas_swap_strided_batched = rocblas_cswap_strided_batched; template <> static constexpr auto rocblas_swap_strided_batched = rocblas_zswap_strided_batched; // dot template rocblas_status (*rocblas_dot)(rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, const T* y, rocblas_int incy, T* result); template <> static constexpr auto rocblas_dot = rocblas_sdot; template <> static constexpr auto rocblas_dot = rocblas_ddot; template <> static constexpr auto rocblas_dot = rocblas_hdot; template <> static constexpr auto rocblas_dot = rocblas_bfdot; template <> static constexpr auto rocblas_dot = rocblas_cdotu; template <> static constexpr auto rocblas_dot = rocblas_zdotu; // dotc template rocblas_status (*rocblas_dotc)(rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, const T* y, rocblas_int incy, T* result); template <> static constexpr auto rocblas_dotc = rocblas_cdotc; template <> static constexpr auto rocblas_dotc = rocblas_zdotc; // dot_batched template rocblas_status (*rocblas_dot_batched)(rocblas_handle handle, rocblas_int n, const T* const x[], rocblas_int incx, const T* const y[], rocblas_int incy, rocblas_int batch_count, T* result); template <> static constexpr auto rocblas_dot_batched = rocblas_sdot_batched; template <> static constexpr auto rocblas_dot_batched = rocblas_ddot_batched; template <> static constexpr auto rocblas_dot_batched = rocblas_hdot_batched; template <> static constexpr auto rocblas_dot_batched = rocblas_bfdot_batched; template <> static constexpr auto rocblas_dot_batched = rocblas_cdotu_batched; template <> static constexpr auto rocblas_dot_batched = rocblas_zdotu_batched; // dotc_batched template rocblas_status (*rocblas_dotc_batched)(rocblas_handle handle, rocblas_int n, const T* const x[], rocblas_int incx, const T* const y[], rocblas_int incy, rocblas_int batch_count, T* result); template <> static constexpr auto rocblas_dotc_batched = rocblas_cdotc_batched; template <> static constexpr auto rocblas_dotc_batched = rocblas_zdotc_batched; // dot_strided_batched template rocblas_status (*rocblas_dot_strided_batched)(rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, rocblas_stride stridex, const T* y, rocblas_int incy, rocblas_stride stridey, rocblas_int batch_count, T* result); template <> static constexpr auto rocblas_dot_strided_batched = rocblas_sdot_strided_batched; template <> static constexpr auto rocblas_dot_strided_batched = rocblas_ddot_strided_batched; template <> static constexpr auto rocblas_dot_strided_batched = rocblas_hdot_strided_batched; template <> static constexpr auto rocblas_dot_strided_batched = rocblas_bfdot_strided_batched; template <> static constexpr auto rocblas_dot_strided_batched = rocblas_cdotu_strided_batched; template <> static constexpr auto rocblas_dot_strided_batched = rocblas_zdotu_strided_batched; // dotc template rocblas_status (*rocblas_dotc_strided_batched)(rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, rocblas_stride stridex, const T* y, rocblas_int incy, rocblas_stride stridey, rocblas_int batch_count, T* result); template <> static constexpr auto rocblas_dotc_strided_batched = rocblas_cdotc_strided_batched; template <> static constexpr auto rocblas_dotc_strided_batched = rocblas_zdotc_strided_batched; // asum template rocblas_status (*rocblas_asum)( rocblas_handle handle, rocblas_int n, const T1* x, rocblas_int incx, T2* result); template <> static constexpr auto rocblas_asum = rocblas_sasum; template <> static constexpr auto rocblas_asum = rocblas_dasum; template <> static constexpr auto rocblas_asum = rocblas_scasum; template <> static constexpr auto rocblas_asum = rocblas_dzasum; // asum_batched template rocblas_status (*rocblas_asum_batched)(rocblas_handle handle, rocblas_int n, const T1* const x[], rocblas_int incx, rocblas_int batch_count, T2* result); template <> static constexpr auto rocblas_asum_batched = rocblas_sasum_batched; template <> static constexpr auto rocblas_asum_batched = rocblas_dasum_batched; template <> static constexpr auto rocblas_asum_batched = rocblas_scasum_batched; template <> static constexpr auto rocblas_asum_batched = rocblas_dzasum_batched; // asum_strided_batched template rocblas_status (*rocblas_asum_strided_batched)(rocblas_handle handle, rocblas_int n, const T1* x, rocblas_int incx, rocblas_stride stridex, rocblas_int batch_count, T2* result); template <> static constexpr auto rocblas_asum_strided_batched = rocblas_sasum_strided_batched; template <> static constexpr auto rocblas_asum_strided_batched = rocblas_dasum_strided_batched; template <> static constexpr auto rocblas_asum_strided_batched = rocblas_scasum_strided_batched; template <> static constexpr auto rocblas_asum_strided_batched = rocblas_dzasum_strided_batched; // nrm2 template rocblas_status (*rocblas_nrm2)( rocblas_handle handle, rocblas_int n, const T1* x, rocblas_int incx, T2* result); template <> static constexpr auto rocblas_nrm2 = rocblas_snrm2; template <> static constexpr auto rocblas_nrm2 = rocblas_dnrm2; template <> static constexpr auto rocblas_nrm2 = rocblas_scnrm2; template <> static constexpr auto rocblas_nrm2 = rocblas_dznrm2; // nrm2_batched template rocblas_status (*rocblas_nrm2_batched)(rocblas_handle handle, rocblas_int n, const T1* const x[], rocblas_int incx, rocblas_int batch_count, T2* results); template <> static constexpr auto rocblas_nrm2_batched = rocblas_snrm2_batched; template <> static constexpr auto rocblas_nrm2_batched = rocblas_dnrm2_batched; template <> static constexpr auto rocblas_nrm2_batched = rocblas_scnrm2_batched; template <> static constexpr auto rocblas_nrm2_batched = rocblas_dznrm2_batched; // nrm2_strided_batched template rocblas_status (*rocblas_nrm2_strided_batched)(rocblas_handle handle, rocblas_int n, const T1* x, rocblas_int incx, rocblas_stride stridex, rocblas_int batch_count, T2* results); template <> static constexpr auto rocblas_nrm2_strided_batched = rocblas_snrm2_strided_batched; template <> static constexpr auto rocblas_nrm2_strided_batched = rocblas_dnrm2_strided_batched; template <> static constexpr auto rocblas_nrm2_strided_batched = rocblas_scnrm2_strided_batched; template <> static constexpr auto rocblas_nrm2_strided_batched = rocblas_dznrm2_strided_batched; // iamax and iamin need to be full functions rather than references, in order // to allow them to be passed as template arguments // // iamax and iamin. // // // Define the signature type. // template using rocblas_iamax_iamin_t = rocblas_status (*)( rocblas_handle handle, rocblas_int n, const T* x, rocblas_int incx, rocblas_int* result); // // iamax // template rocblas_iamax_iamin_t rocblas_iamax; template <> static constexpr auto rocblas_iamax = rocblas_isamax; template <> static constexpr auto rocblas_iamax = rocblas_idamax; template <> static constexpr auto rocblas_iamax = rocblas_icamax; template <> static constexpr auto rocblas_iamax = rocblas_izamax; // // iamin // template rocblas_iamax_iamin_t rocblas_iamin; template <> static constexpr auto rocblas_iamin = rocblas_isamin; template <> static constexpr auto rocblas_iamin = rocblas_idamin; template <> static constexpr auto rocblas_iamin = rocblas_icamin; template <> static constexpr auto rocblas_iamin = rocblas_izamin; // axpy template rocblas_status (*rocblas_axpy)(rocblas_handle handle, rocblas_int n, const T* alpha, const T* x, rocblas_int incx, T* y, rocblas_int incy); template <> static constexpr auto rocblas_axpy = rocblas_haxpy; template <> static constexpr auto rocblas_axpy = rocblas_saxpy; template <> static constexpr auto rocblas_axpy = rocblas_daxpy; template <> static constexpr auto rocblas_axpy = rocblas_caxpy; template <> static constexpr auto rocblas_axpy = rocblas_zaxpy; // rot template rocblas_status (*rocblas_rot)(rocblas_handle handle, rocblas_int n, T* x, rocblas_int incx, T* y, rocblas_int incy, const U* c, const V* s); template <> static constexpr auto rocblas_rot = rocblas_srot; template <> static constexpr auto rocblas_rot = rocblas_drot; template <> static constexpr auto rocblas_rot = rocblas_crot; template <> static constexpr auto rocblas_rot = rocblas_csrot; template <> static constexpr auto rocblas_rot = rocblas_zrot; template <> static constexpr auto rocblas_rot = rocblas_zdrot; // rot_batched template rocblas_status (*rocblas_rot_batched)(rocblas_handle handle, rocblas_int n, T* const x[], rocblas_int incx, T* const y[], rocblas_int incy, const U* c, const V* s, rocblas_int batch_count); template <> static constexpr auto rocblas_rot_batched = rocblas_srot_batched; template <> static constexpr auto rocblas_rot_batched = rocblas_drot_batched; template <> static constexpr auto rocblas_rot_batched = rocblas_crot_batched; template <> static constexpr auto rocblas_rot_batched = rocblas_csrot_batched; template <> static constexpr auto rocblas_rot_batched = rocblas_zrot_batched; template <> static constexpr auto rocblas_rot_batched = rocblas_zdrot_batched; // rot_strided_batched template rocblas_status (*rocblas_rot_strided_batched)(rocblas_handle handle, rocblas_int n, T* x, rocblas_int incx, rocblas_stride stride_x, T* y, rocblas_int incy, rocblas_stride stride_y, const U* c, const V* s, rocblas_int batch_count); template <> static constexpr auto rocblas_rot_strided_batched = rocblas_srot_strided_batched; template <> static constexpr auto rocblas_rot_strided_batched = rocblas_drot_strided_batched; template <> static constexpr auto rocblas_rot_strided_batched = rocblas_crot_strided_batched; template <> static constexpr auto rocblas_rot_strided_batched = rocblas_csrot_strided_batched; template <> static constexpr auto rocblas_rot_strided_batched = rocblas_zrot_strided_batched; template <> static constexpr auto rocblas_rot_strided_batched = rocblas_zdrot_strided_batched; // rotg template rocblas_status (*rocblas_rotg)(rocblas_handle handle, T* a, T* b, U* c, T* s); template <> static constexpr auto rocblas_rotg = rocblas_srotg; template <> static constexpr auto rocblas_rotg = rocblas_drotg; template <> static constexpr auto rocblas_rotg = rocblas_crotg; template <> static constexpr auto rocblas_rotg = rocblas_zrotg; // rotg_batched template rocblas_status (*rocblas_rotg_batched)(rocblas_handle handle, T* const a[], T* const b[], U* const c[], T* const s[], rocblas_int batch_count); template <> static constexpr auto rocblas_rotg_batched = rocblas_srotg_batched; template <> static constexpr auto rocblas_rotg_batched = rocblas_drotg_batched; template <> static constexpr auto rocblas_rotg_batched = rocblas_crotg_batched; template <> static constexpr auto rocblas_rotg_batched = rocblas_zrotg_batched; //rotg_strided_batched template rocblas_status (*rocblas_rotg_strided_batched)(rocblas_handle handle, T* a, rocblas_stride stride_a, T* b, rocblas_stride stride_b, U* c, rocblas_stride stride_c, T* s, rocblas_stride stride_s, rocblas_int batch_count); template <> static constexpr auto rocblas_rotg_strided_batched = rocblas_srotg_strided_batched; template <> static constexpr auto rocblas_rotg_strided_batched = rocblas_drotg_strided_batched; template <> static constexpr auto rocblas_rotg_strided_batched = rocblas_crotg_strided_batched; template <> static constexpr auto rocblas_rotg_strided_batched = rocblas_zrotg_strided_batched; //rotm template rocblas_status (*rocblas_rotm)(rocblas_handle handle, rocblas_int n, T* x, rocblas_int incx, T* y, rocblas_int incy, const T* param); template <> static constexpr auto rocblas_rotm = rocblas_srotm; template <> static constexpr auto rocblas_rotm = rocblas_drotm; // rotm_batched template rocblas_status (*rocblas_rotm_batched)(rocblas_handle handle, rocblas_int n, T* const x[], rocblas_int incx, T* const y[], rocblas_int incy, const T* const param[], rocblas_int batch_count); template <> static constexpr auto rocblas_rotm_batched = rocblas_srotm_batched; template <> static constexpr auto rocblas_rotm_batched = rocblas_drotm_batched; // rotm_strided_batched template rocblas_status (*rocblas_rotm_strided_batched)(rocblas_handle handle, rocblas_int n, T* x, rocblas_int incx, rocblas_stride stride_x, T* y, rocblas_int incy, rocblas_stride stride_y, const T* param, rocblas_stride stride_param, rocblas_int batch_count); template <> static constexpr auto rocblas_rotm_strided_batched = rocblas_srotm_strided_batched; template <> static constexpr auto rocblas_rotm_strided_batched = rocblas_drotm_strided_batched; //rotmg template rocblas_status (*rocblas_rotmg)(rocblas_handle handle, T* d1, T* d2, T* x1, const T* y1, T* param); template <> static constexpr auto rocblas_rotmg = rocblas_srotmg; template <> static constexpr auto rocblas_rotmg = rocblas_drotmg; //rotmg_batched template rocblas_status (*rocblas_rotmg_batched)(rocblas_handle handle, T* const d1[], T* const d2[], T* const x1[], const T* const y1[], T* const param[], rocblas_int batch_count); template <> static constexpr auto rocblas_rotmg_batched = rocblas_srotmg_batched; template <> static constexpr auto rocblas_rotmg_batched = rocblas_drotmg_batched; //rotmg_strided_batched template rocblas_status (*rocblas_rotmg_strided_batched)(rocblas_handle handle, T* d1, rocblas_stride stride_d1, T* d2, rocblas_stride stride_d2, T* x1, rocblas_stride stride_x1, const T* y1, rocblas_stride stride_y1, T* param, rocblas_stride stride_param, rocblas_int batch_count); template <> static constexpr auto rocblas_rotmg_strided_batched = rocblas_srotmg_strided_batched; template <> static constexpr auto rocblas_rotmg_strided_batched = rocblas_drotmg_strided_batched; /* * =========================================================================== * level 2 BLAS * =========================================================================== */ // ger template rocblas_status (*rocblas_ger)(rocblas_handle handle, rocblas_int m, rocblas_int n, const T* alpha, const T* x, rocblas_int incx, const T* y, rocblas_int incy, T* A, rocblas_int lda); template <> static constexpr auto rocblas_ger = rocblas_sger; template <> static constexpr auto rocblas_ger = rocblas_dger; template rocblas_status (*rocblas_ger_batched)(rocblas_handle handle, rocblas_int m, rocblas_int n, const T* alpha, const T* const x[], rocblas_int incx, const T* const y[], rocblas_int incy, T* const A[], rocblas_int lda, rocblas_int batch_count); template <> static constexpr auto rocblas_ger_batched = rocblas_sger_batched; template <> static constexpr auto rocblas_ger_batched = rocblas_dger_batched; template rocblas_status (*rocblas_ger_strided_batched)(rocblas_handle handle, rocblas_int m, rocblas_int n, const T* alpha, const T* x, rocblas_int incx, rocblas_int stride_x, const T* y, rocblas_int incy, rocblas_int stride_y, T* A, rocblas_int lda, rocblas_int stride_a, rocblas_int batch_count); template <> static constexpr auto rocblas_ger_strided_batched = rocblas_sger_strided_batched; template <> static constexpr auto rocblas_ger_strided_batched = rocblas_dger_strided_batched; // syr template rocblas_status (*rocblas_syr)(rocblas_handle handle, rocblas_fill uplo, rocblas_int n, const T* alpha, const T* x, rocblas_int incx, T* A, rocblas_int lda); template <> static constexpr auto rocblas_syr = rocblas_ssyr; template <> static constexpr auto rocblas_syr = rocblas_dsyr; // syr strided batched template rocblas_status (*rocblas_syr_strided_batched)(rocblas_handle handle, rocblas_fill uplo, rocblas_int n, const T* alpha, const T* x, rocblas_int incx, rocblas_stride stridex, T* A, rocblas_int lda, rocblas_stride strideA, rocblas_int batch_count); template <> static constexpr auto rocblas_syr_strided_batched = rocblas_ssyr_strided_batched; template <> static constexpr auto rocblas_syr_strided_batched = rocblas_dsyr_strided_batched; // syr batched template rocblas_status (*rocblas_syr_batched)(rocblas_handle handle, rocblas_fill uplo, rocblas_int n, const T* alpha, const T* const x[], rocblas_int incx, T* A[], rocblas_int lda, rocblas_int batch_count); template <> static constexpr auto rocblas_syr_batched = rocblas_ssyr_batched; template <> static constexpr auto rocblas_syr_batched = rocblas_dsyr_batched; // gemv template rocblas_status (*rocblas_gemv)(rocblas_handle handle, rocblas_operation transA, rocblas_int m, rocblas_int n, const T* alpha, const T* A, rocblas_int lda, const T* x, rocblas_int incx, const T* beta, T* y, rocblas_int incy); template <> static constexpr auto rocblas_gemv = rocblas_sgemv; template <> static constexpr auto rocblas_gemv = rocblas_dgemv; template <> static constexpr auto rocblas_gemv = rocblas_cgemv; template <> static constexpr auto rocblas_gemv = rocblas_zgemv; // gemv_strided_batched template rocblas_status (*rocblas_gemv_strided_batched)(rocblas_handle handle, rocblas_operation transA, rocblas_int m, rocblas_int n, const T* alpha, const T* A, rocblas_int lda, rocblas_int stride_a, const T* x, rocblas_int incx, rocblas_int stride_x, const T* beta, T* y, rocblas_int incy, rocblas_int stride_y, rocblas_int batch_count); template <> static constexpr auto rocblas_gemv_strided_batched = rocblas_sgemv_strided_batched; template <> static constexpr auto rocblas_gemv_strided_batched = rocblas_dgemv_strided_batched; template <> static constexpr auto rocblas_gemv_strided_batched = rocblas_cgemv_strided_batched; template <> static constexpr auto rocblas_gemv_strided_batched = rocblas_zgemv_strided_batched; // gemv_batched template rocblas_status (*rocblas_gemv_batched)(rocblas_handle handle, rocblas_operation transA, rocblas_int m, rocblas_int n, const T* alpha, const T* const A[], rocblas_int lda, const T* const x[], rocblas_int incx, const T* beta, T* const y[], rocblas_int incy, rocblas_int batch_count); template <> static constexpr auto rocblas_gemv_batched = rocblas_sgemv_batched; template <> static constexpr auto rocblas_gemv_batched = rocblas_dgemv_batched; template <> static constexpr auto rocblas_gemv_batched = rocblas_cgemv_batched; template <> static constexpr auto rocblas_gemv_batched = rocblas_zgemv_batched; // trsv template rocblas_status (*rocblas_trsv)(rocblas_handle handle, rocblas_fill uplo, rocblas_operation transA, rocblas_diagonal diag, rocblas_int m, const T* A, rocblas_int lda, T* x, rocblas_int incx); template <> static constexpr auto rocblas_trsv = rocblas_strsv; template <> static constexpr auto rocblas_trsv = rocblas_dtrsv; // symv template rocblas_status (*rocblas_symv)(rocblas_handle handle, rocblas_fill uplo, rocblas_int n, const T* alpha, const T* A, rocblas_int lda, const T* x, rocblas_int incx, const T* beta, T* y, rocblas_int incy); /* not implemented template <> static constexpr auto rocblas_symv = rocblas_ssymv; template <> static constexpr auto rocblas_symv = rocblas_dsymv; */ /* * =========================================================================== * level 3 BLAS * =========================================================================== */ // geam template rocblas_status (*rocblas_geam)(rocblas_handle handle, rocblas_operation transA, rocblas_operation transB, rocblas_int m, rocblas_int n, const T* alpha, const T* A, rocblas_int lda, const T* beta, const T* B, rocblas_int ldb, T* C, rocblas_int ldc); template <> static constexpr auto rocblas_geam = rocblas_sgeam; template <> static constexpr auto rocblas_geam = rocblas_dgeam; // gemm template rocblas_status (*rocblas_gemm)(rocblas_handle handle, rocblas_operation transA, rocblas_operation transB, rocblas_int m, rocblas_int n, rocblas_int k, const T* alpha, const T* A, rocblas_int lda, const T* B, rocblas_int ldb, const T* beta, T* C, rocblas_int ldc); template <> static constexpr auto rocblas_gemm = rocblas_hgemm; template <> static constexpr auto rocblas_gemm = rocblas_sgemm; template <> static constexpr auto rocblas_gemm = rocblas_dgemm; template <> static constexpr auto rocblas_gemm = rocblas_cgemm; template <> static constexpr auto rocblas_gemm = rocblas_zgemm; // gemm_batched template rocblas_status (*rocblas_gemm_batched)(rocblas_handle handle, rocblas_operation transA, rocblas_operation transB, rocblas_int m, rocblas_int n, rocblas_int k, const T* alpha, const T* const A[], rocblas_int lda, const T* const B[], rocblas_int ldb, const T* beta, T* C[], rocblas_int ldc, rocblas_int batch_count); template <> static constexpr auto rocblas_gemm_batched = rocblas_hgemm_batched; template <> static constexpr auto rocblas_gemm_batched = rocblas_sgemm_batched; template <> static constexpr auto rocblas_gemm_batched = rocblas_dgemm_batched; template <> static constexpr auto rocblas_gemm_batched = rocblas_cgemm_batched; template <> static constexpr auto rocblas_gemm_batched = rocblas_zgemm_batched; // gemm_strided_batched template rocblas_status (*rocblas_gemm_strided_batched)(rocblas_handle handle, rocblas_operation transA, rocblas_operation transB, rocblas_int m, rocblas_int n, rocblas_int k, const T* alpha, const T* A, rocblas_int lda, rocblas_stride bsa, const T* B, rocblas_int ldb, rocblas_int bsb, const T* beta, T* C, rocblas_int ldc, rocblas_stride bsc, rocblas_int batch_count); template <> static constexpr auto rocblas_gemm_strided_batched = rocblas_hgemm_strided_batched; template <> static constexpr auto rocblas_gemm_strided_batched = rocblas_sgemm_strided_batched; template <> static constexpr auto rocblas_gemm_strided_batched = rocblas_dgemm_strided_batched; template <> static constexpr auto rocblas_gemm_strided_batched = rocblas_cgemm_strided_batched; template <> static constexpr auto rocblas_gemm_strided_batched = rocblas_zgemm_strided_batched; #if 0 // trmm template rocblas_status (*rocblas_trmm)(rocblas_handle handle, rocblas_side side, rocblas_fill uplo, rocblas_operation transA, rocblas_diagonal diag, rocblas_int m, rocblas_int n, const T* alpha, T* A, rocblas_int lda, T* B, rocblas_int ldb); template <> static constexpr auto rocblas_trmm = rocblas_strmm; template <> static constexpr auto rocblas_trmm = rocblas_dtrmm; #endif // trsm template rocblas_status (*rocblas_trsm)(rocblas_handle handle, rocblas_side side, rocblas_fill uplo, rocblas_operation transA, rocblas_diagonal diag, rocblas_int m, rocblas_int n, const T* alpha, T* A, rocblas_int lda, T* B, rocblas_int ldb); template <> static constexpr auto rocblas_trsm = rocblas_strsm; template <> static constexpr auto rocblas_trsm = rocblas_dtrsm; // trtri template rocblas_status (*rocblas_trtri)(rocblas_handle handle, rocblas_fill uplo, rocblas_diagonal diag, rocblas_int n, T* A, rocblas_int lda, T* invA, rocblas_int ldinvA); template <> static constexpr auto rocblas_trtri = rocblas_strtri; template <> static constexpr auto rocblas_trtri = rocblas_dtrtri; // trtri_batched template rocblas_status (*rocblas_trtri_batched)(rocblas_handle handle, rocblas_fill uplo, rocblas_diagonal diag, rocblas_int n, T* A[], rocblas_int lda, T* invA[], rocblas_int ldinvA, rocblas_int batch_count); template <> static constexpr auto rocblas_trtri_batched = rocblas_strtri_batched; template <> static constexpr auto rocblas_trtri_batched = rocblas_dtrtri_batched; // trtri_strided_batched template rocblas_status (*rocblas_trtri_strided_batched)(rocblas_handle handle, rocblas_fill uplo, rocblas_diagonal diag, rocblas_int n, T* A, rocblas_int lda, rocblas_stride bsa, T* invA, rocblas_int ldinvA, rocblas_stride bsinvA, rocblas_int batch_count); template <> static constexpr auto rocblas_trtri_strided_batched = rocblas_strtri_strided_batched; template <> static constexpr auto rocblas_trtri_strided_batched = rocblas_dtrtri_strided_batched; #endif // _ROCBLAS_HPP_