[PATCH] D47370: AMDGPU: Round up kernel argument allocation size

[Matt Arsenault via Phabricator via llvm-commits]

arsenm added inline comments.


================
Comment at: lib/Target/AMDGPU/AMDGPUSubtarget.cpp:426
+  // Being able to dereference past the end is useful for emitting scalar loads.
+  return alignTo(TotalSize, 4);
 }
----------------
t-tye wrote:
> I believe you can align this to 16. See HSA spec at www.hsafoundation.com/html_spec111/HSA_Library.htm#PRM/Topics/04_SyntaxSemantics/kernarg_segment.htm which says:
> 
> "The alignment of the base address of the kernel's kernarg segment variables is the larger of 16 bytes and the maximum alignment of the kernel's kernarg segment variables."
> 
> I suspect that the OpenCL runtime simply aligns all kernarg allocations to 256 but not sure of other languages.
I think we really only need to pad up to 4 at the end to avoid vmem extloads. Wider scalar loads at the end may be useful in some cases, but we do so badly at this now I wouldn't worry about it yet


https://reviews.llvm.org/D47370