Skip to content

Fix dtype validation for conditional SM targets - #106

Open
sylvesterkaczmarek wants to merge 1 commit into
NVIDIA:mainfrom
sylvesterkaczmarek:sylvesterkaczmarek/bug-conditional-sm-parsing
Open

Fix dtype validation for conditional SM targets#106
sylvesterkaczmarek wants to merge 1 commit into
NVIDIA:mainfrom
sylvesterkaczmarek:sylvesterkaczmarek/bug-conditional-sm-parsing

Conversation

@sylvesterkaczmarek

Copy link
Copy Markdown

Description

Closes #105.

Strip one recognized architecture or family suffix before converting the SM number in check_dtype_support. Keep the original target string for diagnostics and leave tileiras target selection unchanged.

The new compiler tests export bytecode without requiring device arguments. They cover plain and suffixed targets, dtype and bytecode version limits, and malformed suffixes.

Validation

  • Nine suffixed-target cases fail on the original parsing code; all 14 regression cases pass with the fix.
  • PYTHONPATH=src python -m pytest --noconftest -o addopts="" test/test_conditional_arch.py test/test_ir_types.py test/test_ast_util.py test/test_float_encoding.py -q: 32 passed.
  • python -m flake8 src/cuda/tile/_passes/check_dtype_support.py test/test_conditional_arch.py: passed.
  • git diff --check: passed.

Tests ran in Linux arm64 with the working-tree Python sources and the cuda-tile 1.6.0 wheel's native extension. GPU fixtures, kernel execution, and a source build of the extension were not run. This fixes the Python validation failure only; the separate tileiras target-support limitation described in the issue remains unchanged.

Checklist

  • I am familiar with the Contributing Guidelines.
  • New or existing tests cover these changes.
  • Documentation reviewed; no API or assembler support changes.

Signed-off-by: Sylvester Kaczmarek <16242628+sylvesterkaczmarek@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

[BUG]: check_dtype_support rejects family-conditional (sm_XXXa) gpu_code targets

1 participant