strsm.cu

Nividia提供的CUDA的BLAS库源码

}

__global__ void fast_strsm_l_lo_nt_main_hw (struct cublasStrsmParams parms) 
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             1
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_tr_main_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             1
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_nt_main_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             0
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_tr_main_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             0
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_nt_main_fulltile_hw (struct cublasStrsmParams parms) 
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY 
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   1
#define LOWER             1
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_tr_main_fulltile_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   1
#define LOWER             1
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_nt_main_fulltile_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   1
#define LOWER             0
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_tr_main_fulltile_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   1
#define LOWER             0
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_nt_main_alpha0_hw (struct cublasStrsmParams parms) 
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   0
#define LOWER             1
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_tr_main_alpha0_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   0
#define LOWER             1
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_nt_main_alpha0_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   0
#define LOWER             0
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_tr_main_alpha0_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   0
#define LOWER             0
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_nt_main_fulltile_alpha0_hw (struct cublasStrsmParams parms) 
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY 
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   1
#define LOWER             1
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_lo_tr_main_fulltile_alpha0_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   1
#define LOWER             1
#define TRANS             1
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_nt_main_fulltile_alpha0_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   1
#define LOWER             0
#define TRANS             0
#include "strsm_l.h"
}

__global__ void fast_strsm_l_up_tr_main_fulltile_alpha0_hw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 0
#define FAST_IMUL         1
#define ALPHA_IS_ZERO     1
#define FULL_TILES_ONLY   1
#define LOWER             0
#define TRANS             1
#include "strsm_l.h"
}

__global__ void strsm_l_lo_nt_main_sw (struct cublasStrsmParams parms) 
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 1
#define FAST_IMUL         0
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             1
#define TRANS             0
#include "strsm_l.h"
}

__global__ void strsm_l_lo_tr_main_sw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 1
#define FAST_IMUL         0
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             1
#define TRANS             1
#include "strsm_l.h"
}

__global__ void strsm_l_up_nt_main_sw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER
#undef FAST_IMUL
#undef ALPHA_IS_ZERO
#undef FULL_TILES_ONLY
#undef LOWER
#undef TRANS
#define NOUNIT            0
#define USE_MIXED_STEPPER 1
#define FAST_IMUL         0
#define ALPHA_IS_ZERO     0
#define FULL_TILES_ONLY   0
#define LOWER             0
#define TRANS             0
#include "strsm_l.h"
}

__global__ void strsm_l_up_tr_main_sw (struct cublasStrsmParams parms)
{
#undef NOUNIT
#undef USE_MIXED_STEPPER