Piik/native/probes/encoder-pool

Encoder Pool Feasibility Probe

Explicit Windows experiment, not a product dependency or CI job. Two real libwebrtc VideoStreamEncoders consume one synthetic 640x360 source through independent stock VideoAdapters. It does not capture the desktop or open ports. Research and evidence own conclusions.

Use Shiguredo m152.7977.0.2 Windows x64 SDK (WebRTC commit 6f37672d358475cd17544121a12494da454d85fb), matching its headers and static library: download. ZIP SHA-256: 3250ee091eb745b9885482d92db352e6bcd5426940021ff90a095747d84d2137. Keep the extracted SDK, official compiler dependencies and executable under ignored build/encoder-pool; preserve the SDK's notices. Do not ship this SDK as part of Piik based on this probe.

With CMake and Visual Studio 2022 x64 tools available:

cmake -S native/probes/encoder-pool -B build/encoder-pool/probe -A x64 `
  -DWEBRTC_SDK="$PWD/build/encoder-pool/webrtc"
cmake --build build/encoder-pool/probe --config Release -j 2
& ./build/encoder-pool/probe/Release/encoder-pool-probe.exe
& ./build/encoder-pool/probe/Release/encoder-pool-probe.exe --pooled
& ./build/encoder-pool/probe/Release/encoder-pool-probe.exe --pooled --group-adjuster

The tested older compiler was MSVC 14.42; linking requires the newer official 14.44 libraries/linker used by the SDK. On such a machine pass MSVC_CRT_DIR to CMake and /p:LinkToolPath=<official tools directory> to MSBuild after --. Do not supply substitute STL implementations or alter global tool settings.

The full trace takes about 51 seconds: healthy, 100 kbps weak B, release to 2 Mbps, one omitted B input, then independent B retirement. --healthy-only omits the weak/release phases. JSONL reports real encoding calls, both WebRTC decoders, raw libvpx hashes without history-dependent postprocessing, dimensions, first transition times, group lifetime and resource observations. integrity checks nonzero decoding, error counts, shared-payload comparisons and retirement; it is not an automatic performance/parity verdict.

--no-adjuster is a diagnostic ablation only. --group-adjuster instead keeps WebRTC's actual adjustment algorithm once per physical codec; it must not run alongside per-VSE correction. --unsafe-skip removes dependency protection as a negative control and should fail integrity with group adjustment and a skipped input. This cache arm is guarded single-source realtime VP8/L1T1 with synchronous codecs; no H264 cache, physical overload or public-network claim follows.

--realtime removes per-input encoder-queue drains. --latency-probe requires that mode and inserts a bounded pixel-proportional 45 ms codec delay, then releases it; it is not a CPU/GPU stress test. --shared-pipeline --realtime --latency-probe is a separate control with one full stock VSE feeding both decoders, not two VSEs borrowing its encoder. Do not combine it with pool or adjuster flags. Source and codec work remain bounded; use a 95-second process deadline for these approximately 70-second traces. The shared-pipeline B resource metrics repeat the common pipeline, while its decoded counters are independent.

Hardware H264 Pipeline

The separate --h264 arm wraps the existing Windows MFT LiveEncoder in a WebRTC encoder adapter. It requires --shared-pipeline --healthy-only --export PATH; one complete stock VSE supplies two output callbacks. It does not test two independent controllers borrowing an H264 encoder. Run it alone, with a 30-second process deadline and no concurrent capture/codec workload:

& ./build/encoder-pool/probe/Release/encoder-pool-probe.exe `
  --h264 --shared-pipeline --healthy-only --realtime `
  --export build/encoder-pool/h264-input.jsonl

This is synthetic I420 upload to NV12/MFT, not WGC capture or zero-copy input. The pinned SDK has no H264 decoder, so this arm records payload equality and encoded dimensions with decoded counters at zero. Validate the exported AUs separately in Chrome. The existing scripts/encoded-group-decode.mjs accepts PIIK_ENCODED_CODEC=avc1.42c033 and PIIK_ENCODED_OUTPUT pointing to two arrays of {Index, Width, Height, PTS, Recovery, Data} frames (nanosecond PTS, base64 Annex-B data). A repeated export in those two arrays proves decoding of that bitstream, not two Pion transports or weak-budget/resource adaptation.

Healthy Relay Chain

Export the first 180 real AUs from sender A for the opt-in Go forwarding check:

& ./build/encoder-pool/probe/Release/encoder-pool-probe.exe `
  --pooled --group-adjuster --healthy-only --export build/encoder-pool/chain-input.jsonl
go test -c -o build/embedded-media/mediaedge.test.exe ./internal/client/mediaedge
$env:PIIK_POOL_CHAIN_FIXTURE = "$PWD/build/encoder-pool/chain-input.jsonl"
$env:PIIK_NATIVE_CAPTURE = "$PWD/build/client-check/piik-client-capture.exe"
$env:PIIK_POOL_CHAIN_OUTPUT = "$PWD/build/encoder-pool/chain-result.json"
& ./build/embedded-media/mediaedge.test.exe `
  '-test.run=^TestRelayChainEncodedFixture$' '-test.v' '-test.timeout=30s'

The real capture executable must already be built, so Native relay capability remains enabled while checking that healthy forwarding does not start it. This replays encoded video through two Native receivers and three leaves, checking recovery-first, contiguous, hash-identical AUs and closure. No frame payload is committed. Optional output contains only the synthetic trace's summary. It is not a live encoder/BWE integration or an internet/NAT test.