[Addin]

Summary

Optimizes the Windows D3D11 frame-conversion hot path and adds the instrumentation used to verify it.

• Reuse cached video-processor input/output views and staging textures instead of recreating them per frame.
• Share one D3D11 converter/device across pool workers on Windows, with a staging ring sized for four workers.
• Poll Map with D3D11_MAP_FLAG_DO_NOT_WAIT and release the shared context between retries so other workers can submit GPU work.
• Keep one pair of staging textures for standalone converters and grow lazily up to the four-pair cap.
• Add benchmark controls (BENCH_INPUT, BENCH_WORKERS, BENCH_SINGLE_ONLY, BENCH_POOL_ONLY, BENCH_D3D11_ONLY) and report the selected adapter.
• Request D3D11_CREATE_DEVICE_VIDEO_SUPPORT on capture devices.
• Instrument the Media Foundation encoder loop and the Windows M4S readback/encode stages, and report readback misses against the frames that actually attempted a readback.

Performance comparison

GTX 950M / i7-6700HQ, YUYV422 -> NV12, release build. Before is 8a71a02c; after is e8333db6, the last converter commit before the readback Flush correctness fix. Medians of three four-worker runs and two single-thread runs.

Workload	Before	After	Improvement	Speedup
Single-thread, 720p	2.23 ms	2.13 ms	4.2%	1.04x
Single-thread, 1080p	4.29 ms	3.88 ms	9.6%	1.11x
4-worker shared converter, 720p	2.30 ms	1.11 ms	51.8%	2.07x
4-worker shared converter, 1080p	4.48 ms	2.05 ms	54.3%	2.19x
4-worker from_config pool, 720p	1.07 ms	1.09 ms	~2% (noise)	0.98x
4-worker from_config pool, 1080p	1.99 ms	2.04 ms	~2% (noise)	0.98x

from_config is the production camera pool. At maximum submitted throughput it is within run-to-run noise of the old four-device configuration while using one D3D11 device instead of four; at the 30 fps camera rate the pool does not contend, so the single-thread latency and contention rows are the relevant gains. Without the readback polling changes, sharing the converter would have been about 2x slower.

Additional after-only measurements:

• Real camera 640x480 @ 30 fps: conversion 1.65 ms avg / 2.08 ms P99, encode 0.84 ms, 29.6 fps, zero drops
• Screen readback > 5 ms: 1/251 frames; encoder > 5 ms: 2.4%; no drops
• DO_NOT_WAIT polling at maximum throughput uses 148% of one core (18.5% of the 8-thread machine); at the 30 fps production rate the workers idle between frames, so no backoff was added

The MF hardware encoder still falls back on this GPU: NVIDIA MFT returns 0x8000FFFF and Intel QSV returns E_INVALIDARG. The FFmpeg fallback remains.

Greptile review fixes

• frame-converter: flush the immediate context after queuing the output copy and before the DO_NOT_WAIT polling loop, so standalone conversions cannot spin on work that was never submitted.
• capture_pipeline: build the WARP fallback without D3D11_CREATE_DEVICE_VIDEO_SUPPORT.
• win_fragmented_m4s: after draining, return the Media Foundation encoder error so a failed encoder cannot be finalized as a successful recording.
• scap-ffmpeg (base perf(recording): cut Windows readback copies and encode from textures #2940): create the mapped readback AVBufferRef with AV_BUFFER_FLAG_READONLY.

Dependency

Stacked on #2940 (readback/MF groundwork). Until that merges, this diff includes its readback/MF commits. If #2940 lands with a merge commit, GitHub will reduce this PR to the converter/capture commits automatically; if it is squash-merged, this branch will need a rebase onto the new main.

Validation

• cargo fmt --all -- --check
• cargo clippy -p cap-frame-converter -p cap-enc-mediafoundation -p cap-recording --all-targets -- -D warnings
• cargo test -p cap-recording --lib: 686 passed, 0 failed, 4 ignored
• cap-frame-converter and cap-enc-mediafoundation tests pass
• Release d3d11-benchmark rerun after the lazy-staging change; numbers above are unchanged from the eager-allocation baseline

Confidence Score: 5/5

The PR appears safe to merge; all previous findings are addressed.

Summary

Optimizes Windows frame conversion by sharing a D3D11 converter across workers and reusing views and staging textures.

• Windows converter pools share D3D11 resources across workers.
• D3D11 readback reuses mapped staging memory for FFmpeg frames.
• Windows screen recordings can encode captured textures with Media Foundation.
• Windows recording logs show time spent in capture and encoding stages.
• The older Windows segmented muxers are removed.

_{Reviews (3) · Last reviewed commit: "fix(recording): report MF encoder failur..." · Reviewed by Greptile}

[vercel]

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

A member of the Team first needs to authorize it.

[Addin]

@greptileai review

[Addin]

@greptileai review