...
- Jerry Zhao - Panasonic
- Jan-Simon Moeller - Linux Foundation
- Scott Murray - Konsulko
- Mikhail Golubev - Open Synergy
- Vasyl Vavrychuk - OpenSynergy
- Andriy Tryshnivsky - OpenSynergy
- Laurent Cremmer - Carmeq
- Mark Silberberg - Volkswagen
Victor Duan - Linaro- Alex Bennée (Stratos Tech Lead) - Linaro
- Peter Griffin (Multimedia Tech Lead) - Linaro
- Tadao Tanikawa - Panasonic
- Binghua - Qualcomm
- Marius Vlad - Collabora
- Tadao Tanikawa - Panasonic
- Kenji Hosokawa - ADIT
Harunobu Kurokawa - Renesas
Masahiro Hasegawa - Renesas
- Tomeu Vizoso
Venkata Ramalinga Prasad Tadepalli
- Parag Borkar - OpenSynergy
Agenda:
- 1st VirtIO Workshop: VirtIO GPU-3d Performance - How to achieve Zero-Copy with AGL
- Open Synergy Slides (How to achieve zero-copy on Renesas SoC)
VirtIO_GPU_3D_Workshop.pdf- cache coherence between CPU/GPU
- Discussion Points:
- Does VirtIO GPU supports zero-copy in QEMU, crosvm, etc?
- crosvm uses host allocations, is it requirement for something?
- progress in crosvm is on-going and realted to Vulkan.
- crosvm use the virtio-pci (to expose host memory to guest - which also avoid memory copy)
- Any available standard solution/platform to solve this zero-copy for the AGL community
- Tomeu: Collabora is having project in ChromeBook (which is more focusing on gaming) but expose host memory to guest
- Alex: not using virtio-pci may be the limitation (kiran explained why it is not because limitation →with 64 bit PCI addressing the windows are more than large enough)
- Conclusion: there's no good solution available at the moment for zero-copy mechanism for automotive use case
- Note: sometimes, copying smaller buffers ( < 4k) is faster then using zero copying
- Does virglrender supports zero-copy of OpenGL buffers (vertex buffer, element buffer, etc.)?
- there are some mechanisms in virglrender. no copy happened. (although some limitation may exist)
- EXT_image_dma_buf_import fails with EGL_BAD_ACCESS
- Tomeu: Reason is that specific platform stride requirements are not propagated to the guest. New IOCTL exist to propagate stride info to the guest, but not mainlined yet.
https://gitlab.freedesktop.org/olv/drm-misc-next/-/commit/6b2013a48528fc9ea1e82c6b68c85521f6d03041
- why stride information is needed in guest? This info will be pitched to MESA3d.
- GPU/display supports only 32-bit addresses. Issue: ensure device guest (mesa) allocates memory form lower addresses
- Bounce buffers:
- Stratos prj is interested in a mechanism to limit guest's allocations to a particular mem window (potentially, breaking virtio driver mem model)
- Kernel dma pools are being looked at. This will result in memcpy from userspace to kernel dma capable pool (page cache as well?)
- Bounce buffers:
- Tomeu: Reason is that specific platform stride requirements are not propagated to the guest. New IOCTL exist to propagate stride info to the guest, but not mainlined yet.
- Virglrenderer tests fails on R-CAR
- Does VirtIO GPU supports zero-copy in QEMU, crosvm, etc?
...
- Jerry Zhao - Panasonic
- Jan-Simon Moeller - Linux Foundation
- Scott Murray - Konsulko
- Mikhail Golubev - Open Synergy
- Vasyl Vavrychuk - OpenSynergy
- Parag Borkar - OpenSynergy
Andriy Tryshnivsky - OpenSynergyLaurent Cremmer - CarmeqMark Silberberg - VolkswagenVictor Duan - Linaro- Francis Ozog - Linaro
- Masami Hiramatsu - Linaro
- Alex Bennée (Stratos Tech Lead) - Linaro
- Peter Griffin (Multimedia Tech Lead) - Linaro
- Tadao Tanikawa - Panasonic
- Binghua - Qualcomm
Marius Vlad - Collabora- Tadao Tanikawa - Panasonic
- Kenji Hosokawa - ADIT
...