Repository navigation
[release/2.15] Fix inline asm elementwise tests for real-true16 on gfx1250 (#3862) - #3968
Open
BLOrange-AMD wants to merge 1 commit into
Open
BLOrange-AMD wants to merge 1 commit into
BLOrange-AMD wants to merge 1 commit into
Conversation
…x1250 (#3862) This PR fixes: higher_order_ops/test_inline_asm_elementwise.py::TestInlineAsmElementwise::test_eager_vs_compiled_bitwise_truncate_to_uint16 higher_order_ops/test_inline_asm_elementwise.py::TestInlineAsmElementwise::test_eager_vs_compiled_bitwise_add_fp16_native higher_order_ops/test_inline_asm_elementwise.py::TestInlineAsmElementwise::test_eager_vs_compiled_bitwise_add_bf16_native Through [PR](llvm/llvm-project#190452), gfx1250 enables LLVM's real-true16 feature. - So in PyTorch, need to add gfx1250 to a real-true16 code path, align with Navi3/3.5/4 in TRUE16_ARCH. This fixes add_fp16_native and truncate_to_uint16, whose existing real-true16 sequences assemble unchanged on gfx1250. - add_bf16_native gets its own real-true16 variant. It never needed one before because it is skipped on every previously-listed true16 arch (v_cvt_pk_bf16_f32 requires gfx950+), so gfx1250 is the first real-true16 arch where it actually runs. The new sequence does the add and the pack into physical v0, declared as a clobber, then copies the packed low half into the true16 output with v_mov_b16 $0, v0.l. Writing the packed result into $0 directly does not work either — v_cvt_pk_bf16_f32 rejects a 16-bit destination too. --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com> Co-authored-by: Bo Li <BLOrange-AMD@users.noreply.github.com> Co-authored-by: Prachi Gupta <pracgupt@amd.com>
BLOrange-AMD
changed the title
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR fixes:
higher_order_ops/test_inline_asm_elementwise.py::TestInlineAsmElementwise::test_eager_vs_compiled_bitwise_truncate_to_uint16
higher_order_ops/test_inline_asm_elementwise.py::TestInlineAsmElementwise::test_eager_vs_compiled_bitwise_add_fp16_native
higher_order_ops/test_inline_asm_elementwise.py::TestInlineAsmElementwise::test_eager_vs_compiled_bitwise_add_bf16_native
Through PR, gfx1250 enables LLVM's real-true16 feature.
So in PyTorch, need to add gfx1250 to a real-true16 code path, align with Navi3/3.5/4 in TRUE16_ARCH. This fixes add_fp16_native and truncate_to_uint16, whose existing real-true16 sequences assemble unchanged on gfx1250.
add_bf16_native gets its own real-true16 variant. It never needed one before because it is skipped on every previously-listed true16 arch (v_cvt_pk_bf16_f32 requires gfx950+), so gfx1250 is the first real-true16 arch where it actually runs. The new sequence does the add and the pack into physical v0, declared as a clobber, then copies the packed low half into the true16 output with v_mov_b16 $0, v0.l. Writing the packed result into $0 directly does not work either — v_cvt_pk_bf16_f32 rejects a 16-bit destination too.