Some question about nvcc compile options

[MARD1NO]

Recently I want to integrate some cutlass operators into PyTorch, and I found in CUDAExtension we can pass some compile options. In other project it usually set:

"--ptxas-options=-O2",
"--ptxas-options=-allow-expensive-optimizations=true",

In my GEMM test, I found turn off these 2 options will get better performance, but there is no documentations about ptx. I just want to know what these options means, and hope cutlass develop team will give me some suggestions! :D

[hwu36]

when you compile cutlass, just use cutlass's nvcc flags. that is what nvcc team uses to optimize cutlass.