Repository navigation
Conversation
davidrohr
left a comment
davidrohr
left a comment
There was a problem hiding this comment.
Why did you change all the GPUd() to GPUhd(). GPUhd() means it is executed by the host code compiled with CUDA, so that should be only the code that drives the devices?
Does that code really need the GPUCommonMath functions? IMHO in that case one can use std:: directly.
Named double functions for GPUCommonMath::Sqrt,Sin,Cos,Tan,ASin,ACos,ATan,ATan2,Log,Copysign. Clz64, Ctz64, Popcount64: 64-bit counterparts of Clz, Ctz and Popcount on the same builtins and CUDA/HIP intrinsics. Named rather than uint64_t overloads, which would change existing calls with 64-bit arguments and make calls with int arguments ambiguous. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Signed-off-by: Felix Schlepper <felix.schlepper@cern.ch>
|
Hi @davidrohr, you right of course. I had declared functors with GPUhd() in code linked to O2 using CAMath. My compilation then failed without declaring the functions also GPUhd(), so I had 'bug' in my code... sorry. I think I never understood the difference between GPUhd() and GPUd(). Also I had the question about the deterministic mode, currently it forces some computations to be made in double precision and then downcasts to float, but this trick does not work if the computation is supposed to be in double see for example |
This branch has not been deployed
2 participants
…, host-device calls
Named double functions for GPUCommonMath::Sqrt,Sin,Cos,Tan,ASin,ACos,ATan,ATan2,Log,Copysign.
Clz64, Ctz64, Popcount64: 64-bit counterparts of Clz, Ctz and Popcount on the same builtins and CUDA/HIP intrinsics.
And AtomicLoad/Store.