Haicheng Wu comments

Results 323 comments of


                                            Haicheng Wu

[QST] make error caused by glibc 2.32 in ubuntu 18.04 "//usr/local/lib/libpthread.so.0: undefined reference to"

@d-k-b , any idea?

Does Conv2d Kernels Support Float32 on SM75?

Cutlass has simt conv kernel for sm75

Fix typos in cute docs

@ccecka , @thakkarV , more work needed in this pr?

* Use Cutlass ops when possible to +15% speed

@danthe3rd

Fix typos in convolution tests

Found some compiler bugs when working on yours. Addressing these bugs now.

Fix typos in convolution tests

Sorry. This one is good. Will merge it around the time tagging 3.5.1 which is this weekend or next Monday. I thought you were pinging me for your mixed input...

[QST] StreamK ReductionStrategy: "Atomic" or "Mixed"

the output is not deterministic if you use atomic with float point. it is hard to debug the numeric issues. atomic is usually faster though.

recursion error reached hipifying cutlass portion of xformers

FYI, it is waste of time to hipify cutlass since cutlass uses lots of inline ptx. You could rewrite these inline ptx in cuda, but it is not cutlass anymore.

[BUG] Stride is ignored for dst tensor of a Conv2dFprop

conv epilogue reuses gemm epilogue which assumes the output a row major dense matrix. the column number is `K` which is 1 in your case, and the row number is...

[BUG] Stride is ignored for dst tensor of a Conv2dFprop

i would say that you would have to hack what ever that is not right. the code is like ```for (row) for (col) *memory_ptr(row, col) = ``` you essentially needs...