[karthick-mcw]

Summary

AVX-512 register-permute gather for compute_ideal_weights_for_decimation.

The initial-average and refine loops gather ei.weights / ei.weight_error_scale / infilled_weights per texel. The hardware vgatherdps used for this is microcoded and load-latency bound on Zen 5. For texel_count <= 48 (blocks up to 6x6, the size this workload uses) the whole lookup table fits in 3 zmm registers, so the gather is done as an in-register permute (vpermt2ps + vpermps) instead, with no memory access. Larger blocks (e.g. 8x8 = 64 texels) exceed register residency and fall back to the existing AVX2 gather path, unchanged.

The result is bit-identical: it's the same elements, fetched a different way. Adds a self-contained runtime AVX-512F check, since an AVX2 build of this library can still run on a host without AVX-512F — the new path is only taken when both the block is small enough and the host actually supports it.

Files: Source/astcenc_ideal_endpoints_and_weights.cpp

Measured

Asset compression workload, gain in % (positive means faster).

Vendor	CPU	OS	Compiler	Single thread	All threads
AMD	Granite 9600	Windows	clang-cl	6.03	6.78
Intel	Arrow Lake	Windows	clang-cl	-0.14	0.21
Apple	Mac Air	macOS	AppleClang	0	0.2

(Mac Air has no AVX-512, so this change is a no-op there, as expected — the 0%/0.2% is noise. Arrow Lake's AVX-512 was disabled by Intel on this generation, so it also falls back to the AVX2 path.)

Test plan

• Bit-identical output: 6 representative LDR/HDR textures x 6 block sizes (4x4-12x12) x 4 quality presets (144 runs), byte-identical .astc output with and without the patch, on a host without AVX-512F (exercises the unchanged AVX2 fallback path).
• Unit test: astcenc C++ unit test suite (ctest, -DASTCENC_UNITTEST=ON), 223/223 tests pass on the patched build.
• Functional tests: Test/astc_test_functional.py --encoder avx2 gives the identical result with and without the patch (38/93 pass; the remaining 55 error only on a missing ImageMagick dependency in this test environment, not on the code — zero assertion failures either way).
• Timing: single core and all cores on all three machines (table above) — carried over from the original measurement; not independently re-verified in this environment.
• Execution check of the AVX-512 permute path itself on AVX-512F hardware — not available in this environment (dev machine has no AVX-512F); only compile-checked and verified to build cleanly. Worth re-running the bit-identical diff on a Zen 5/AVX-512-capable host before merge.