The margin was a fixed 768 MiB — 3% of a 24 GiB compute card but 19% of
a 4 GiB card (RTX A400), where free VRAM never clears it, so CanLoad()
refused every engine and UnderPressure() skipped every frame. Now
total/8 clamped to [192, 768] MiB: cards >= 6 GiB keep exactly the old
margin, only small cards change. ANSCENTER_GPU_MARGIN_MB / SetMargin()
still override (0 = auto). Refusal log now reports total VRAM alongside
free, which is what made this diagnosable.
Verified on the A400 test PC: detector + recognizer load and match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Two-pass slot acquisition in runInferenceFromPool: Pass A picks an idle
slot whose deviceIndex matches the input GpuMats' resident device
(zero-copy dispatch); Pass B falls back to plain round-robin, where
runInference migrates the inputs via the pinned-host two-hop copy.
Callers that prepare frames on GPU A no longer pay the ~1 ms/frame PCIe
migration when a same-GPU slot is idle, while spillover to the other GPU
under load still works. Input residency is queried outside the slot
mutex via cudaPointerGetAttributes (cheap driver lookup).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Verified on dual RTX 5080: 355/355 detections, both GPUs in round-robin,
zero CUDA errors. Requires TensorRT >= 10.16.1 in SharedCore (10.15.1 has
a Myelin l2cm crash when creating a context on a second Blackwell device).
1. blobFromGpuMats: thread_local GpuMat caches (tl_blob/tl_floatImg) were
device-blind — a pool thread warming GPU[1] then GPU[0] reused device-1
memory with a device-0 stream, causing an async illegal memory access
that poisoned every CUDA context (sticky 700). Track owning device and
rebuild the cache on device switch.
2. runInference: pool round-robin can dispatch to a slot on a different
GPU than the one the caller prepared inputs on. Detect foreign input
GpuMats via cudaPointerGetAttributes and migrate them through a pinned
host staging buffer (two-hop D2H+H2D; direct cross-device cudaMemcpy2D
is unsupported and GeForce has no P2P).
3. warmUp: now pins its own CUDA device before uploading dummy inputs and
returns success/failure; loadNetwork rejects engines whose warmup fails
so broken slots can never join the pool silently.
4. Blackwell builder workaround: setMaxAuxStreams(0) on SM 12.x plus an
.aux0 engine-cache suffix so pre-workaround caches rebuild.
Also: ANSFR-UnitTest gains --gpu N flag (sets CUDA_VISIBLE_DEVICES) for
single-GPU bisection, and rich diagnostics on input-copy failures.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Cleaned up verbose engine telemetry emitted to stdout/stderr and the
Windows Event Viewer. Removes logEngineEvent/logEvent calls (and their
diagnostic-only locals) across the TensorRT engine load, build, run,
multi-GPU, and pool-manager paths, plus the now-unused logEvent helper
in EnginePoolManager.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>