interpolation

Pixel interpolation

‍Verified against 42eca0e8fe on 2026-09-29.

Which resampling kernel the pipeline uses, why, and the one configuration trap that fails quietly.

Each finding below is dated with the commit that established it. A finding is only as good as its hash: before acting on one older than the code you are changing, re-measure it — and re-date it here when you do. Where an earlier version of a claim was wrong, that is recorded rather than quietly corrected: how a claim was wrong is usually the more useful thing to know.

Found 22f623c0be, 2026-06-25.

Mitchell-Netravali (B=C=1/3) is the pipeline interpolator. Lanczos has been removed entirely. Rationale: Lanczos has large negative side-lobes → halos at high-contrast edges and pushes premultiplied alpha out of [0,1]. Mitchell is near-halo-free (~3% residual undershoot), sharp, and a separable partition-of-unity kernel that fits the existing tap machinery for CPU and GPU.

The pipeline's interpolation architecture in src/pixel/interpolation.c is separable — each kernel registers a 1D maketaps, and both dt_interpolation_resample (CPU) and dt_interpolation_resample_cl (GPU) consume the same CPU-computed taps. A new separable kernel is automatically CPU+GPU.

The drawlayer brush matte still forces BILINEAR explicitly — premultiplied alpha wants strictly zero overshoot.

Config option strings in anselconfig.xml.in MUST equal the kernel .name field exactly; USERPREF resolves by strcmp. A mismatch (e.g. "bicubic (Catmull-Rom)" vs "bicubic") silently falls back to default instead of erroring.

Two stale references to Lanczos remain in the tree

Found 42eca0e8fe, 2026-09-29 — during the verification pass that produced this file.

Lanczos is gone from the enum, the kernels and the configuration, but two places still name it, and one of them is a live defect rather than a stale comment:

• **tools/benchmark_darkroom_rc.sh:110-111 writes lanczos2 / lanczos3 into the ANSEL rc** for pixel_interpolator_warp and pixel_interpolator. By the rule stated above — a config string must equal the kernel's .name exactly, or USERPREF silently falls back to the default — the Ansel half of that benchmark is not measuring the interpolator it reports. Lines 139-140 set the darktable rc, where those values are still valid, so only the first pair is wrong.
• src/iop/drawlayer.c:808 justifies forcing bilinear for the brush matte with "unlike the user-pref default (Lanczos)". The default is mitchell. The code is right; only the parenthetical is stale.

Neither is fixed here: this file's commit is documentation only.