[PATCH] D159459: [AMDGPU] Add ASM and MC updates for preloading kernargs

Tue Sep 12 15:20:08 PDT 2023

kerbowa added a comment.

In D159459#4644061 <https://reviews.llvm.org/D159459#4644061>, @arsenm wrote:

>> Alignment of the arguments in SGPRs is equivalent to the kernarg segment when
>
> accessed via the kernarg_segment_ptr.
>
> The alignment should not be relevant to the registers. The registers should always be packed

I don't think they are with the current FW/runtime. The placement of the arguments is exactly the same as the kernarg segment. In the future, the runtime could fix the alignment so that the arguments are packed but Jack said he didn't want to pursue this yet.

================
Comment at: llvm/test/MC/AMDGPU/user-sgpr-count-diag.s:6
   .amdhsa_user_sgpr_queue_ptr 1
+  .amdhsa_accum_offset 4
   .amdhsa_next_free_vgpr 32
----------------
arsenm wrote:
> Why amdhsa_accum_offset? I don't recognize this one
It errors out if it's not there.

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D159459/new/

https://reviews.llvm.org/D159459