[Openmp-commits] [PATCH] D73077: [libomptarget] Implement wavefront functions for amdgcn

[Matt Arsenault via Phabricator via Openmp-commits]

arsenm added inline comments.


================
Comment at: openmp/libomptarget/deviceRTLs/amdgcn/src/target_impl.hip:21-22
+
+// initialized with a 64-bit mask with bits set in positions less than the
+// thread's lane number in the warp
+DEVICE __kmpc_impl_lanemask_t __kmpc_impl_lanemask_lt() {
----------------
JonChesterfield wrote:
> JonChesterfield wrote:
> > arsenm wrote:
> > > What about wave32?
> > Hi Matt. I didn't think we had any 32 bit warp / wavefront hardware support in tree. If we do there's some changes needed elsewhere in this library as well
> Ah. gfx1010 has some code in clang, e.g. https://reviews.llvm.org/D63209
> 
> I'd still like to land this patch as it progresses support for most amdgcn.
> 
> I'll also start the task of supporting wave32. The complicated part is probably detecting what a given arch uses early enough to pick out the right typedef for lanemask_t.
Putting off wave32 is fine, but might as well fix the hardcoded 63 here anyway


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D73077/new/

https://reviews.llvm.org/D73077