yuxiao

I was run example 14_ampere_tf32_tensorop_gemm, but data init method is custom. when stages is 4, I set K=128, 256, 512, 1024, the result from gpu have higher accurate compare with...

I used cuda11.3 in A100. @hwu36

I just did gemm(MxK KxN) TF32 relative error compare with FP64. when K increase, relative errors are decrease. It is unreasonable.

I use win10 system and python 3.6 .when I use command "pip install rasterio==1.0a12" ,it have same error message.but when I use "pip install rasterio" can install success。But The version...

I have try all the solution but do not solve the question.