strmm.cu

Nividia提供的CUDA的BLAS库源码

#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              0
#define LOWER             1
#define TRANS             0
#include "strmm_l.h"
}

__global__ void strmm_l_up_tr_main_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0 
#define ALPHA0            1
#define UNIT              0
#define LOWER             0
#define TRANS             1
#include "strmm_l.h"
}

__global__ void strmm_l_lo_tr_main_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              0
#define LOWER             1
#define TRANS             1
#include "strmm_l.h"
}

__global__ void strmm_r_up_nt_main_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              0
#define LOWER             0
#define TRANS             0
#include "strmm_r.h"
}

__global__ void strmm_r_lo_nt_main_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              0
#define LOWER             1
#define TRANS             0
#include "strmm_r.h"
}

__global__ void strmm_r_up_tr_main_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              0
#define LOWER             0
#define TRANS             1
#include "strmm_r.h"
}

__global__ void strmm_r_lo_tr_main_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              0
#define LOWER             1
#define TRANS             1
#include "strmm_r.h"
}

__global__ void strmm_l_up_nt_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             0
#define TRANS             0
#include "strmm_l.h"
}

__global__ void strmm_l_lo_nt_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             1
#define TRANS             0
#include "strmm_l.h"
}

__global__ void strmm_l_up_tr_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             0
#define TRANS             1
#include "strmm_l.h"
}

__global__ void strmm_l_lo_tr_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             1
#define TRANS             1
#include "strmm_l.h"
}

__global__ void strmm_r_up_nt_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             0
#define TRANS             0
#include "strmm_r.h"
}

__global__ void strmm_r_lo_nt_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             1
#define TRANS             0
#include "strmm_r.h"
}

__global__ void strmm_r_up_tr_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             0
#define TRANS             1
#include "strmm_r.h"
}

__global__ void strmm_r_lo_tr_main_unit_alpha0_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         0
#define ALPHA0            1
#define UNIT              1
#define LOWER             1
#define TRANS             1
#include "strmm_r.h"
}

__global__ void fast_strmm_l_up_nt_main_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         1
#define ALHPA0            0
#define UNIT              0
#define LOWER             0
#define TRANS             0
#include "strmm_l.h"
}

__global__ void fast_strmm_l_lo_nt_main_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         1
#define ALPHA0            0
#define UNIT              0
#define LOWER             1
#define TRANS             0
#include "strmm_l.h"
}

__global__ void fast_strmm_l_up_tr_main_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         1
#define ALPHA0            0
#define UNIT              0
#define LOWER             0
#define TRANS             1
#include "strmm_l.h"
}

__global__ void fast_strmm_l_lo_tr_main_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER
#undef  TRANS
#define USE_MIXED_STEPPER 1
#define FULL_TILES_ONLY   0
#define FAST_IMUL         1
#define ALPHA0            0
#define UNIT              0
#define LOWER             1
#define TRANS             1
#include "strmm_l.h"
}

__global__ void fast_strmm_r_up_nt_main_sw (struct cublasStrmmParams parms) 
{
#undef  USE_MIXED_STEPPER
#undef  FULL_TILES_ONLY
#undef  FAST_IMUL
#undef  ALPHA0
#undef  UNIT
#undef  LOWER