Precompute the vision resize tap weights

[namespaceMarcello]

ds4_resize_rgb_bicubic called ds4_cubic() once per contributing source pixel. A tap weight depends only on the output coordinate and on the source sample it lands on, so a 12 MP photo scaled to the DeepSeek canvas evaluated the kernel about 200 million times to obtain a few thousand distinct values.

This builds one weight row per output sample and accumulates exactly as before: same products, same sums, same order. Tabulating the weights also lets the compiler vectorize the inner loop, which the call used to block, and that is where most of the gain comes from.

Preprocessing is synchronous, right before the GPU vision encode, so this is latency removed from every image on every backend.

Ryzen 9 7940HX, default CFLAGS (-O3 -ffast-math -march=native), mean of a timed batch, ms:

image             DeepSeek V4         GLM 5.3
4032x3024      283.00 -> 140.50   456.00 -> 261.00
2560x1440      112.80 ->  63.12    24.19 ->  25.35
1280x720        35.71 ->  18.74     5.62 ->   5.23
640x480         13.50 ->   7.35     1.70 ->   1.66
3200x400        44.33 ->  24.33     7.83 ->   8.49

GLM only pays the resize when the image does not already match its padded canvas; its flat rows are the exact-size copy path, unchanged and within run to run noise.

On output: a differential harness against the previous implementation over 457 geometries reports zero differing floats at -O0, -O2, -O3, -O3 -ffast-math, -O3 -ffast-math -mavx2 -mfma, -O3 -ffast-math -march=x86-64-v4 and -O3 -ffast-math -fno-tree-vectorize. The corpus is built to make rounding decide the result: two-level checkerboards and stripes at exact and near-exact integer ratios, where three quarters of the values land on an exact .5 tie, plus 1x1 sources and targets, decoder-limit dimensions and padded strides.

One combination differs: -ffast-math with -march=native, which resolves to -march=znver4 here. One geometry out of 457 then moves 48 of its 3108 values by one level of 255, a 1036x1 destination from a 16384-wide checkerboard. Neither AVX2 nor AVX-512 alone reproduces it, so it is a tuning-specific reassociation rather than a property of the change.

For scale, on that same geometry the current code already differs from itself between build flags: 1143 of 3108 floats between an -O2 and a -march=native build, 1224 between -O2 and -mavx2 -mfma. This resampler has no bit-exact output across builds to preserve, and this change stays well inside the spread it already has.

Also: resize now reports allocation failure instead of leaving an unfiltered canvas to be normalized. Failing each of the 16 allocation sites in turn returns a clean error with no leak. The added test asserts a uniform image stays uniform across the content area; it passes on the previous implementation too.

Not tested: no Metal, CUDA or ROCm machine and no GGUF here, so make test, the logprob vectors, the vision quality runs and speed-bench are untested. The file is backend independent and contains no GPU code. I can add the differential harness as an in-tree test if you want it.

#1013 stacks a separable two-pass version on top of this one: three times faster again, but it moves tie values by one level under any flags.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VuqnYdebtpz6NpooFpVY1p