Closed bjoo closed 4 years ago
Looks good to me.
Dan any comments? @ibaned
Not much! Definitely thank you to Balint for this contribution, and a quick skim looks okay. I don't use storage types in my production code so I just use scalar
instead of the specialized GPU types.
Happy to do it. I am starting on an attempt at permutations in these types today. You can look at the testing code for this pull as well following the link in the comments. Thanks for merging.
Best, B
On May 1, 2020, at 10:58 AM, Dan Ibanez notifications@github.com wrote:
Not much! Definitely thank you to Balint for this contribution, and a quick skim looks okay. I don't use storage types in my production code so I just use scalar instead of the specialized GPU types.
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.
Hi All, Here is my attempt at porting the CUDA simd::abi::cuda_warp to HIP (specifically to hip::simd_abi::hip_wavefront with N <=64.
I wrote a separate testsuite to test specifically the function calls at https://github.com/bjoo/simd-math-testing.git. I am not averse to that being extended for SIMD testing.
Many thanks to @Rombur for pointing out things like threadIdx.x to hipThreadIdx_x and that I need to deal carefully with the math functions (which turned out to be a more tricky slog than I thought -- again I needed to explicitly disambiguate at the lowest level to use ::exp, ::sqrt etc otherwise the compiler got confused.
Hope this is a helpful addition, Best, B