lucas-rami wrote: > A 1024-lane workgroup is 16 waves, and all waves of a workgroup must be resident on one CU at the same time (barriers and LDS depend on it). That needs 4 waves per SIMD, so at most 128 VGPRs on gfx942.
If that's true then I was misunderstanding what the ABI actually is. I understood the current implementation which sets the minimum waves/EU as a "good in general" default but that suggests it is actually an ABI requirement that all waves of a workgroup fit concurrently on a CU. In which case this PR indeed makes no sense. If we want to give users more control, the only thing we could/should is to stop rejecting the whole "amdgpu-waves-per-eu" attribute when the minimum provided value is below the default one, allowing the maximum to still apply. For example, for a default range [4,10] and with "amdgpu-waves-per-eu" set to [1,6] we would get a final range [4,6] (currently we would get [4,10], the default, because 1 < 4). https://github.com/llvm/llvm-project/pull/138284 _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
