Skip to content

Preserve integer dtype and precision in abs - #586

Open
AryanRG wants to merge 1 commit into
pydata:masterfrom
AryanRG:fix-integer-absolute
Open

AryanRG wants to merge 1 commit into
pydata:masterfrom
AryanRG:fix-integer-absolute

Conversation

@AryanRG

@AryanRG AryanRG commented Nov 16, 2026

Copy link
Copy Markdown

abs() currently converts integer inputs to float64 despite the documented integer result. This loses precision for values such as -(2**53 + 1).

The integer-preserving dispatch checks for abs, but the wrapped NumPy function is named absolute. Correcting that check enables the existing integer kernels. The int64 helpers and function pointers now use long long, matching the VM's storage on Windows as well. Minimum signed integers retain NumPy's wraparound result through an explicit guard that avoids signed-overflow undefined behavior.

Extended the existing integer-function test with dtype, exact-value and minimum-integer checks. The regression fails on the original build. The rebuilt final source passes all 6042 tests with thread-count environment overrides cleared, and independent sanitizer checks pass. The default guarded environment produces three matching baseline threading-configuration failures.

Windows/MSVC and actual MKL builds were not run locally. The portability checks here were compile-only, and MKL integer helpers were tested with a type/prototype stub.

Prepared with AI assistance.

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.

1 participant