fast_cgemm.cu

Nividia提供的CUDA的BLAS库源码

#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            0
#define TRANSB            0
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_2_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*conj(transpose(A))*B + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            1
#define TRANSB            0
#define CONJGA            1
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_3_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*transpose(A)*B + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            1
#define TRANSB            0
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_4_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*A*conj(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            0
#define TRANSB            1
#define CONJGA            0
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_5_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*A*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            0
#define TRANSB            1
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_6_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*conjg(transpose(A))*conjg(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            1
#define TRANSB            1
#define CONJGA            1
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_7_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*conjg(transpose(A))*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            1
#define TRANSB            1
#define CONJGA            1
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_8_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /*  C = alpha*transpose(A)*conjg(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            1
#define TRANSB            1
#define CONJGA            0
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_9_main_hw_tex_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*transpose(A)*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           1
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 0
#define TRANSA            1
#define TRANSB            1
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_1_main_sw_gld (struct cublasCgemmParams parms) 
{
    /* C = alpha*A*B + beta*C. */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            0
#define TRANSB            0
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_2_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*conj(transpose(A))*B + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            0
#define CONJGA            1
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_3_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*transpose(A)*B + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            0
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_4_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*A*conj(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            0
#define TRANSB            1
#define CONJGA            0
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_5_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*A*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            0
#define TRANSB            1
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_6_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*conjg(transpose(A))*conjg(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            1
#define CONJGA            1
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_7_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*conjg(transpose(A))*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            1
#define CONJGA            1
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_8_main_sw_gld (struct cublasCgemmParams parms)
{
    /*  C = alpha*transpose(A)*conjg(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            1
#define CONJGA            0
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_9_main_sw_gld (struct cublasCgemmParams parms)
{
    /* C = alpha*transpose(A)*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   0
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            1
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_1_main_sw_gld_fulltile (struct cublasCgemmParams parms) 
{
    /* C = alpha*A*B + beta*C. */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 1
#define TRANSA            0
#define TRANSB            0
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_2_main_sw_gld_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*conj(transpose(A))*B + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            0
#define CONJGA            1
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_3_main_sw_gld_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*transpose(A)*B + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 1
#define TRANSA            1
#define TRANSB            0
#define CONJGA            0
#define CONJGB            0
#include "cgemm.h"
}

__global__ void fast_cgemm_4_main_sw_gld_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*A*conj(transpose(B)) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 1
#define TRANSA            0
#define TRANSB            1
#define CONJGA            0
#define CONJGB            1
#include "cgemm.h"
}

__global__ void fast_cgemm_5_main_sw_gld_fulltile (struct cublasCgemmParams parms)
{
    /* C = alpha*A*transpose(B) + beta*C */
#undef  USE_TEX
#undef  FAST_IMUL
#undef  FULL_TILES_ONLY
#undef  USE_MIXED_STEPPER
#undef  TRANSA
#undef  TRANSB
#undef  CONJGA
#undef  CONJGB
#define USE_TEX           0
#define FAST_IMUL         1
#define FULL_TILES_ONLY   1
#define USE_MIXED_STEPPER 1