NVIDIA_SGEMM_PRACTICE icon indicating copy to clipboard operation
NVIDIA_SGEMM_PRACTICE copied to clipboard

Published 20 hours ago • verified publisher icon wangzyon

Results differ from cublas

Open chaoming0625 opened this issue 2 years ago • 2 comments

Dear @wangzyon , when set the size of the matrix to 3, or 9; the mm results are significantly different from cublas. How can we solve it?

chaoming0625 avatar Aug 21 '22 14:08 chaoming0625

Reulsts of kernel 1-5 differ from cublas. I know the reason.

While, result of kenerl 6-7 is tested be same as cubals, however it is different from that of NumPy (in Python). Why?

chaoming0625 avatar Aug 21 '22 14:08 chaoming0625

I generate a 16x16 matrix A, B, and compute it with kernel 7, and numpy.

image image

The following figure is the results calculated through kernel 7.

image

It is significantly different from NumPy.

image

chaoming0625 avatar Aug 21 '22 15:08 chaoming0625