Haicheng Wu

Results 323 comments of Haicheng Wu

Sorry, I don't know. I checked your code and I cannot find anything obvious. Maybe you can change this https://github.com/NVIDIA/cutlass/blob/master/examples/16_ampere_tensorop_conv2dfprop/ampere_tensorop_conv2dfprop.cu#L167 to `kAnalytic` first. `kOptimized` uses complex `params` to compute things...

Sorry, I haven't looked into your case again. I know nvrtc + conv can work since many users make it work. We are working on enable cutlass conv with python....

> waiting for https://github.com/NVIDIA/cutlass/issues/474? yes if you still cannot figure out.

pycutlass implemented in 2.10 supports to use nvrtc to build conv.

it is supported in 2.10

Would you please show me the kernel name of gemm and conv now?

Yes, after the toolkit supports it.

Thanks, would you please list the new kernel name of fprop, wgrad, and dgrad?

The conv kernel name is still the old one. Is it your intention to keep conv kernel name unchanged?

Thank you. I will let different stake holders to review this. Since cutlass is used everywhere, it will take a while.