Skip to content

Fix catastrophic cancellation in f32 and f16 vcos microkernels near x = pi + 2*pi*n by replacing floating-point theta offsets with integer quadrant range reduction and symmetry/sign folding, ensuring exact numerical equivalence with YNNPACK.#10810

Open
copybara-service[bot] wants to merge 1 commit into
masterfrom
test_951618805

Conversation

@copybara-service

@copybara-service copybara-service Bot commented Jul 21, 2026

Copy link
Copy Markdown
Contributor

Fix catastrophic cancellation in f32 and f16 vcos microkernels near x = pi + 2pin by replacing floating-point theta offsets with integer quadrant range reduction and symmetry/sign folding, ensuring exact numerical equivalence with YNNPACK.

@copybara-service copybara-service Bot changed the title [XNNPACK] Add poly-5 and poly-3 minimax polynomial microkernels for f32/f16 vsin and vcos. Implement branchless minimax polynomial approximation for f32 and f16 vsin/vcos, including RVV templates. Jul 21, 2026
@ken-unger

Copy link
Copy Markdown
Contributor

The rvv cmake and bzl files would need to be updated.

@copybara-service copybara-service Bot changed the title Implement branchless minimax polynomial approximation for f32 and f16 vsin/vcos, including RVV templates. Fix catastrophic cancellation in f32 and f16 vcos microkernels near x = pi + 2*pi*n by replacing floating-point theta offsets with integer quadrant range reduction and symmetry/sign folding, ensuring exact numerical equivalence with YNNPACK. Jul 22, 2026
@copybara-service
copybara-service Bot force-pushed the test_951618805 branch 2 times, most recently from 3345a6d to 2f9bf21 Compare July 23, 2026 18:03
… = pi + 2*pi*n by replacing floating-point theta offsets with integer quadrant range reduction and symmetry/sign folding, ensuring exact numerical equivalence with YNNPACK.

PiperOrigin-RevId: 951618805
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.

2 participants