The GitHub Actions job "Lint" on tvm.git/fix/cuda-min-max-nan has succeeded. Run started by GitHub user LngelKyo (triggered by LngelKyo).
Head commit for run: b629c8ee765cbf4a12d9db8ac72bda9b85d218ba / LngelKyo <[email protected]> [Fix][CUDA] Fix constant-operand min/max codegen and test coverage Fixes the four defects found in review of the previous commit: - rhs constant NaN printed `min(va)` / `max(va)` with unbalanced parentheses (three opens, two closes). CUDA has no unary min/max, so the generated kernel did not compile; the surviving else-branch operand must be the constant itself. - lhs constant non-NaN printed the ternary without an enclosing pair of parentheses, so nesting it in a compound expression parsed as `(x + (c cmp vb)) ? c : vb`. Wrapped the whole ternary. - dropped the redundant const_cast: PrintConst takes a const FloatImmNode*. - the rhs constant branch of the float_imm test used the lhs-constant form for min (T.min(const, B) instead of T.min(B, const)), so case 4 was never exercised for min and the expectation still assumed a rhs constant, producing a false failure on the -0.0 lane. Adds test_min_max_float_imm_operand_nested_cuda: a constant non-NaN lhs inside `C[i] = max(0.0, B[i]) + 1.0`, which fails without the enclosing parentheses. All four template formulas were checked lane-by-lane against np.where(cmp | isnan(lhs), lhs, rhs) for every op x const_side x const combination before pushing. Report URL: https://github.com/apache/tvm/actions/runs/35670355145 With regards, GitHub Actions via GitBox --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
