littlewu2508

Results 90 comments of littlewu2508

I found out that bcachefs kernel on current [HEAD](https://github.com/koverstreet/bcachefs/commit/fc5d98c9d4acfd564894bdb39b716f89e4222435) and bcachefs-tools at current [HEAD]( fb2d506) have problems in `bcachefs device add` and `bcachefs data rereplicate`. `bcachefs device add` result in...

> Hi @littlewu2508 Sorry this has taken so long to get to. I updated the PR to address the conflicts with the addition of gfx11 series support. Please look this...

So the question is, what is the preferable configuration for batched sGEMM with one or two dimension = 2? How should I set the search range of `ThreadTile`, `WorkGroup`, `WorkGroupMapping`...

> I'm thrilled to see a community contribution like this one. I have to thank @benjaminulmer who helped me perform the benchmark successfully in https://github.com/ROCmSoftwarePlatform/Tensile/issues/1410. > There's lots of questions...

> I have a technical question about this tuning, though. Have you compared the performance using these gfx1031 kernels on the RX 6700 XT vs. the performance using the existing...

https://github.com/RadeonOpenCompute/rocminfo/issues/42#issuecomment-870846291 may help

> Eclasses are cheap. Write multiple eclasses if it makes sense. Trying to make an all-in-one eclass is only going to cause confusion, mistakes and in the end, someone will...

@heroxbd [rocm-5.1.3-scilibs](https://github.com/littlewu2508/gentoo/tree/rocm-5.1.3-scilibs) is contains ebuilds utilizing rocm.eclass, sucsh as `=sci-libs/rocBLAS-5.1.3`, `=sci-libs/rocThrust-5.1.3`.

My plan is to split the one-size-fits-all `rocm_src_test` into two funcitons, corresponding to two scenarios (cmake test or standalone binary), and let each ebuild decide which to use. This can...

@mgorny I have corrected all the mistakes, and greatly simplified rocm-test. Now I let the ebuild decide which way to choose, by passing `--cmake` or as argument. ebuilds will also...