cgemm.cu

来自「Nividia提供的CUDA的BLAS库源码」· CU 代码 · 共 2,129 行 · 第 1/4 页

CU
2,129 行
字号
1234
}__global__ void cgemm_4_main_hw_tex (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         0#define FULL_TILES_ONLY   0#define USE_MIXED_STEPPER 0#define TRANSA            0#define TRANSB            1#define CONJGA            0#define CONJGB            1#include "cgemm.h"}__global__ void cgemm_5_main_hw_tex (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         0#define FULL_TILES_ONLY   0#define USE_MIXED_STEPPER 0#define TRANSA            0#define TRANSB            1#define CONJGA            0#define CONJGB            0#include "cgemm.h"}__global__ void cgemm_6_main_hw_tex (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         0#define FULL_TILES_ONLY   0#define USE_MIXED_STEPPER 0#define TRANSA            1#define TRANSB            1#define CONJGA            1#define CONJGB            1#include "cgemm.h"}__global__ void cgemm_7_main_hw_tex (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         0#define FULL_TILES_ONLY   0#define USE_MIXED_STEPPER 0#define TRANSA            1#define TRANSB            1#define CONJGA            1#define CONJGB            0#include "cgemm.h"}__global__ void cgemm_8_main_hw_tex (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         0#define FULL_TILES_ONLY   0#define USE_MIXED_STEPPER 0#define TRANSA            1#define TRANSB            1#define CONJGA            0#define CONJGB            1#include "cgemm.h"}__global__ void cgemm_9_main_hw_tex (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         0#define FULL_TILES_ONLY   0#define USE_MIXED_STEPPER 0#define TRANSA            1#define TRANSB            1#define CONJGA            0#define CONJGB            0#include "cgemm.h"}__global__ void cgemm_1_main_sw_tex_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           1#define FAST_IMUL         0#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 cgemm_2_main_sw_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         0#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 cgemm_3_main_sw_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         0#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 cgemm_4_main_sw_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         0#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 cgemm_5_main_sw_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         0#define FULL_TILES_ONLY   1#define USE_MIXED_STEPPER 1#define TRANSA            0#define TRANSB            1#define CONJGA            0#define CONJGB            0#include "cgemm.h"}__global__ void cgemm_6_main_sw_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         0#define FULL_TILES_ONLY   1#define USE_MIXED_STEPPER 1#define TRANSA            1#define TRANSB            1#define CONJGA            1#define CONJGB            1#include "cgemm.h"}__global__ void cgemm_7_main_sw_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         0#define FULL_TILES_ONLY   1#define USE_MIXED_STEPPER 1#define TRANSA            1#define TRANSB            1#define CONJGA            1#define CONJGB            0#include "cgemm.h"}__global__ void cgemm_8_main_sw_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         0#define FULL_TILES_ONLY   1#define USE_MIXED_STEPPER 1#define TRANSA            1#define TRANSB            1#define CONJGA            0#define CONJGB            1#include "cgemm.h"}__global__ void cgemm_9_main_sw_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         0#define FULL_TILES_ONLY   1#define USE_MIXED_STEPPER 1#define TRANSA            1#define TRANSB            1#define CONJGA            0#define CONJGB            0#include "cgemm.h"}__global__ void cgemm_1_main_hw_tex_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           1#define FAST_IMUL         0#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 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         0#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 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         0#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 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         0#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 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         0#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 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         0#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 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         0#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 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         0#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 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         0#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"}
1234

cgemm.cu - 源码说明

本页面展示了「Nividia提供的CUDA的BLAS库源码」中的 cgemm.cu 源码文件，采用 CU 编程语言编写，共 2,129 行代码。您可以在线阅读完整代码内容，也可以返回资源详情页下载完整源码包进行本地学习和开发。

虫虫开发者社区收录了大量与CUDA相关的技术资源，包括源代码、技术文档、电路图等，是电子工程师和嵌入式开发者的专业学习平台。

⌨️ 快捷键说明

复制代码Ctrl + C
搜索代码Ctrl + F
全屏模式F11
增大字号Ctrl + =
减小字号Ctrl + -
显示快捷键?