[kyzer1023]

What

Add an AVX2 path for software RGBA→NV12 in crates/rendering (cpu_yuv / ProjectRenderer::convert_rgba_to_nv12). Dispatches when AVX2 is present and stride/length checks pass. 8-pixel vector kernel; width % 8 and leftover odd row stay scalar. Chroma still averages each complete 2×2 with (sum + 2) / 4 before the matrix (BT.709). Single-threaded (no rayon on this path).

Why

Export / software-adapter encode converts RGBA→NV12 on CPU. Inverse of #3281 (YUV→RGBA). Exact match vs existing scalar (max abs diff 0).

Measurement

Hardware: Intel Xeon family 6 model 207 (KVM, 4 vCPU), AVX2, rustc 1.88.0. Release profile (opt-level=s, codegen-units=1, lto=thin), generic x86-64 (no -C target-cpu=native). Warmup 8, samples 31, median. Stride = renderer 256-byte row padding.

Size	Stride	Scalar	AVX2	Speedup
1920×1080	7680	3.803 ms	1.242 ms	3.06×
3024×1964	12288	10.829 ms	3.582 ms	3.02×
3840×2160	15360	15.064 ms	4.926 ms	3.06×

Max abs diff vs scalar: 0 (incl. odd widths / padded strides).

Test plan

• cargo test -p cap-rendering --lib cpu_yuv:: (21 passed)
• Timed benches above under release flags
• Manual export / software-adapter encode spot check

No overlap with #3281 (YUV→RGBA), #3227 (YUV plane pack), or Windows MF encode PRs.

Confidence Score: 5/5

The PR appears safe to merge, with the previous build issue fixed and no new actionable findings.

Summary

Adds an AVX2 path for software RGBA→NV12 conversion and moves the renderer’s scalar conversion into cpu_yuv.

• Software-rendered frames convert to NV12 faster on AVX2 CPUs.

_{Reviews (2) · Last reviewed commit: "fix(rendering): gate RGBA to NV12 bounds..." · Reviewed by Greptile}

[vercel]

@kyzer1023 is attempting to deploy a commit to the Cap Software Inc Team on Vercel.

A member of the Team first needs to authorize it.

[kyzer1023]

@greptileai review

[richiemcilroy]

Thanks @kyzer1023! Before we can merge, this needs a 5/5 from Greptile on its latest commit and green CI. It's at 4/5, and there are three things to sort out:

• It now conflicts with main in crates/rendering/src/cpu_yuv.rs after perf(rendering): AVX2 the software YUV to RGBA path #3281 landed, so it needs a rebase.
• rgba_nv12_bounds_ok is only called from the x86 block, so on Apple Silicon it's dead code. That fails macOS Clippy (-D warnings). Gate it with the same #[cfg] as its caller.
• Engine verify and macos / test 1/2 failed on this head. Please check whether that's still the case after the rebase.

Once it's rebased and green, request a re-review.

[kyzer1023]

@greptileai review

[kyzer1023]

@richiemcilroy Ready for another look: rebased onto main after #3281, rgba_nv12_bounds_ok is now cfg-gated so aarch64 Clippy is clean, and the Engine verify/macOS failures are gone. Greptile 5/5 and CI green on af48fba63 (Vercel auth checks aside).