[SSCP] Add atomic support #941
Conversation
illuhad
changed the title
| // TODO: It seems that at least on gfx90a, there are additional unsafe atomic | ||
| // operations that can yield a performance improvement. Should we expose those? |
There was a problem hiding this comment.
For reference I think this GROAMCS issue is related. I think we should definitely expose them, however would it be possible to use the __builtin_amdgcn_ds_atomic_fadd_f32 https://gitlab.com/gromacs/gromacs/-/merge_requests/3493
There was a problem hiding this comment.
Usually the set of enabled clang builtins depends on the frontend. It's likely that amdgcn builtins won't work by default with SSCP because the clang frontend does not know that it will run on anything else but the host CPU.
We could just create a simple __hip/opensycl_sscp_amdgcn_ds_atomic_fadd_f32() declaration and implement this inside amdgpu/atomic.cpp with a simple call to the amdgcn builtin. As long as this function is only invoked when JITing for amdgpu (and a GPU that supports it), it should be fine with SSCP.
Co-authored-by: sbalint98 <balint@streamhpc.com>
|
This PR now provides complete implementations of atomic operations for SPIR-V, PTX, and amdgcn. |
|
Passes all current atomic tests on all backends. CC @RaulPPelaez |
WIP. Adds support for atomics to the SSCP compiler.