From 2a11115fd54284772fef10040a4e4b0a39fc6b3a Mon Sep 17 00:00:00 2001 From: Luke Curley Date: Fri, 13 Nov 2026 19:52:04 -0800 Subject: [PATCH] quest: re-check gpu-video 0.5 for the Vulkan encoder Co-Authored-By: Claude Opus 5.5 --- quest/m3/vulkan-encode.md | 13 +++++++++++++ 1 file changed, 13 insertions(+) diff --git a/quest/m3/vulkan-encode.md b/quest/m3/vulkan-encode.md index 27fae9a14ce..3610e43ce4c 100644 --- a/quest/m3/vulkan-encode.md +++ b/quest/m3/vulkan-encode.md @@ -34,6 +34,19 @@ added this quest): lets it import the producer's `OPAQUE_FD` BGRA image as a dedicated allocation with create info identical to the exporter's. It also leaves AV1 possible later. +- Re-checked against gpu-video 0.5.0 on 2026-11-13: still ✅ ash (rejected: + gpu-video 0.5). 0.5 is on wgpu 30 and ash 0.38, and its wgpu-based + `request_device_with_video_support` gets `VK_KHR_external_memory_fd` from + wgpu-hal, so memory import alone would now work. Still missing: + `VK_KHR_external_semaphore_fd` (neither gpu-video nor wgpu enables it, and + `VideoDeviceDescriptor` takes no extra extensions), so the producer's + timeline can't be waited on the GPU, and any live bitrate change, since rate + control is fixed at creation while moq-video follows the send estimate. +- Borrow 0.5's throughput design, which reached about 98% encoder use: let a + few submits stay in flight (it defaults to 3), rotate several input + pictures so conversion overlaps the previous encode, and order consecutive + encodes explicitly. Its completion thread wakes on a semaphore the CPU + signals when it queues work, rather than sleeping and polling. - H.264 and H.265. AV1 is out of scope. - The bar is no CPU round trip; GPU-side copies are allowed.