mirror of
https://github.com/MobileGL-Dev/MobileGL
synced 2026-11-19 16:01:57 +09:00
[Docs] (docs/Disaggregated, docs/MobileVA, docs/Performance, MG_Impl/Pipe, MG_IntegrationTest/Scenarios, MG_State/GLState, MG_Test/Buffer, MG_Test/Wire, scripts/ci): name the renderer's new files
This commit is contained in:
@@ -72,7 +72,7 @@ that MobileGL owns. Decoded frames then reach GL with **zero CPU copies on any A
|
||||
|---|---|---|
|
||||
| P1 | Device: Lenovo TB321FU (Y700), Adreno 750, Android 15, kernel 6.1. Container: Arch Linux ARM, glibc. libva headers in the cross sysroot are **VA-API 1.24** (`~/sysroots/arch-kde-mgl/usr/include/va`). | **[verified]** |
|
||||
| P2 | Server YUV shared images are AHBs: `Y8Cb8Cr8_420` / `YCbCr_P010`, usage `GPU_SAMPLED_IMAGE \| CPU_WRITE_OFTEN \| CPU_READ_RARELY`. Plane offsets come from `AHardwareBuffer_lockPlanes`. They are exported with modifier `DRM_FORMAT_MOD_INVALID`. A returning fd is identified by `(st_dev, st_ino)`, so a re-import of our own buffer is **zero-copy** (no foreign path). | **[verified]** (`MG_Remote/Server/SharedImageRegistry.cpp`, yuvprobe GBM NV12 case) |
|
||||
| P3 | Espryt converts YUV to RGBA with `GL_EXT_YUV_target` (Adreno 750 has it). Magma converts with `VkSamplerYcbcrConversion` on the AHB external format. Both convert once per reader frame epoch. | **[verified]** (`DirectGLES.cpp ConvertYuvForSampling`, `WireYuvImage.inc`) |
|
||||
| P3 | Espryt converts YUV to RGBA with `GL_EXT_YUV_target` (Adreno 750 has it). Magma converts with `VkSamplerYcbcrConversion` on the AHB external format. Both convert once per reader frame epoch. | **[verified]** (`DirectGLES.cpp ConvertYuvForSampling`, `WireYuvImage.cpp`) |
|
||||
| P4 | The foreign-copy fallback is labelled. Its first use logs "a foreign ... dma-buf is sampled through image N, refreshed by a CPU copy", and `Image::ForeignSource::Copies` counts each copy. That is the proof hook for "our buffers never take the fallback". | **[verified]** (code) |
|
||||
| P5 | The YUV conversion paths call `RefreshForeign` but, from a read of the code, do **not** wait on the image's write fence (`DupWriteFence`) the way the RGBA shared-image path does. | **[unverified]**: code read only; confirm before relying on write fences for YUV |
|
||||
| P6 | The client shared-image C ABI (`MG_Impl/SharedImageApi.h`) makes its round trips on the **process's GL client session**, the same ring its GL calls ride, so "like any GL call they must not race the GL calls of another thread". A VA driver therefore **cannot** use it from a decode thread. It needs its own connection. | **[verified]** (header comment) |
|
||||
|
||||
Reference in New Issue
Block a user