Skip to content

tpoisonooo/chgemm

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

13 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

chgemm


License build

chgemm is an symmetric int8 project, which is slightly different from BLAS sgemm:

  1. when you input an int8_t type of matrix [-127,+127], you will get an int32_t one. PS: pay attention to the overflow;
  2. considering the application scene of the deeep learning, the packAB interface is open and can be adjusted;
  3. the common design plan is alpha*A*B+beta*C=C, but mine is C=A*B, because they have no utility in deep learning inference;
  4. row major;
  5. the speed of this project is quicker than any other projects'.

chgemm 是一个 int8 gemm 工程,与 BLAS gemm 不完全相同:

  1. 输入为 [-127, +127] 范围内的 int8_t 类型矩阵,输出为 int32_t 矩阵。需注意溢出;
  2. 更多地为深度学习应用场景考虑,packAB 接口暴露出来可以调整;
  3. 实现为 C = A * B。alpha 和 beta 在深度学习推理中无实用意义;
  4. 行主序实现,放弃远古 fortran 时代的列主序;
  5. 不低于其他项目的 symmint8 gemm 速度。

test result

Compiled on RK3399 with -O3 flag. The current peek can be 18.6 gflops, and the orange line is the single-core fp32 limit(14.3 gflops).

速度

-O3 编译,目前在 rk3399 单核结果。目前极限可以到 18.6 gflops,橙线是 rk3399 单核 fp32 极限。 在 aws A72 单核测试约 23 gflops,是此实现方法的极限(发挥 100% 性能)。

尺寸和gflops结果


使用方式

  1. 修改makefile中的OLDNEW挑选不同实现方式。首次运行需要OLDNEW是同一个
  2. make run 即输出速度结果
  3. parameters.h可修改测试参数

集成方式

参照 MMult_4x8_21.c 调用矩阵乘法,将代码嵌入到自己的项目中。可根据推理库的实现做相应修改。

application with chgemm inside

chgemm is pleased to support ncnn available, check gemm_symm_int8.h.

About

symmetric int8 gemm

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published