optimalbyteeandoptimalbytee
50c4f72af7
* Speed up large-radius Unsharp Mask with guarded FFT convolution * Address Unsharp Mask review feedback --------- Co-authored-by: optimalbytee <[email protected]>
1.4 KiB
Unsharp Mask performance
Large-radius native Unsharp Mask uses separable FFT convolution with the original
discrete Gaussian coefficients and the existing rustfft dependency. A streamed
separable path preserves the original tap order for exact processing.
FFT is eligible for radii 64–1000 with at least 65,536 output pixels, a nonzero amount and a zero or negative threshold. Positive thresholds, unstable samples, unsupported geometry and insufficient scratch use the exact path. WebAssembly retains exact processing. Alpha, selections, edges and cancellation are preserved. Smart Sharpen, High Pass and shared blur math are unchanged.
FFT changes summation order and permits small rounding differences. The measured maximum was one 8-bit storage level, approximately one 16-bit level, and 8.762e-6 for the HDR fixture. Exact-path comparisons were byte-identical. Transform axes are capped at 32,768; FFT scratch grouping has a 1 GiB budget, excluding document tiles. This is not a whole-process memory limit.
Before/after numbers, benchmark methodology, numerical tolerances and preview measurement limits are recorded in PR #2987. The comparisons use synthetic images and measure CPU computation, not slider event-to-screen latency. No published performance budget or rendering-parity floor is raised by these measurements.