-
Notifications
You must be signed in to change notification settings - Fork 1
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Add minigbm driver that provides generic gralloc functionality based on mesa/gbm library #2
base: minigbm-staging-sep-2022
Are you sure you want to change the base?
Add minigbm driver that provides generic gralloc functionality based on mesa/gbm library #2
Conversation
Allow backends with custom DRM probing logic or backends that does not rely on DRM (dma-heap, ION). Signed-off-by: Roman Stratiienko <r.stratiienko@gmail.com> Change-Id: I7bcaf10205ca051eb109d6e220b8a2af38267442
Non-DRM drivers shouldn't rely on handles and DRM API. Add hook to allow drivers create custom implementation. Signed-off-by: Roman Stratiienko <r.stratiienko@gmail.com> Change-Id: I458dae38f80697184070019606b125992a9aa01d
... to avoid compile-time error on C++: error: ISO C++11 does not allow conversion from string literal to 'char *' [-Werror,-Wwritable-strings] Signed-off-by: Roman Stratiienko <r.stratiienko@gmail.com> Change-Id: I176eb657f72e92d6b5c7c3b25c78c56f776c20ab
Some drivers may copy/convert the buffer during mapping and in some cases stride of copied image can be different from original. Android uses pixel_stride for CPU access and need map_time stride instead of original stride in this cases. Signed-off-by: Roman Stratiienko <r.stratiienko@gmail.com>
This is adaptation of the Rob Herring's gbm_gralloc HAL [1] to work as minigbm backend, optimised to work in modern Android conditions. gbm_gralloc has a huge potential and can provide more-or-less optimal allocation by using mesa3d allocation APIs. It should work out of the box for the all hardware mesa3d is supported. Limitations: This backend doesn't care of any other SOC-specific graphical components such as camera, hardware video codecs, etc. Differences between gbm_gralloc: 1. Designed to distinguish between Allocator and Mapper-sphal users. For Allocator gbm driver is initialized using standalone KMS card node (but only if 'lima', 'panfrost' or 'v3d' GPU was detected). For Mapper requests driver is initialized only using GPU render node. (which require less permissions for regular apps but suffitient for mapping). 2. GBM driver is initialized and bo is imported into gbm_mesa only if Mapper imports bo with software usage flags. (no time wasted in case SW access isn't required) 3. gbm_gralloc supported only HAL_PIXEL_FORMAT_YV12 video format. This driver aims to support more formats (using linear modifier only). * NOTE: This may also require: 1. Adding secomp rule into the mediaswcodec.policy file "sched_getaffinity: 1" 2. Adding records to selinux vendor/file_contexts file "/vendor/lib{64}?/libgbm_mesa_wrapper.so u:object_r:same_process_hal_file:s0" "/vendor/bin/hw/android\.hardware\.graphics\.allocator@4\.0-service\.minigbm_gbm_mesa u:object_r:hal_graphics_allocator_default_exec:s0" 3. Kernel v5.3+ for using fstat(dma-buf)->inode as unique buffer id [1]: https://github.com/robherring/gbm_gralloc v2: - Fixed incorrect size calculations for NV12 buffers - Add etnaviv, freedreno and vc4 to gpus list which require kmsro entry - Rebased v3: Squashed local fix commits: - RPI4: Add alignment for CSI camera - gbm_mesa: Handle DRM_FORMAT_R8 format to handle BLOBS - gbm_mesa: Always use DRM_FORMAT_R8 for buffers that unsupported by GBM - gbm_mesa: Add RGB888 format support for CPU access - gbm_mesa: Remove incorrect negation - gbm_mesa: Fallback COMPOSER buffers into VRAM - gbm_mesa: Add combinations to support external camera v4: - Use dlopen/dlsym to access gbm_wrapper - Add fallback allocation for unsupported by mesa3d formats v5: - Obtain map-time stride and report it to Android as pixel-stride. Map-time strides are different after gbm_create and gbm_import. Use map stride after gbm_import. This fixes artifacts on Intel and Nouveau. - GBM wrapper has been converted from cpp to c - Code refactor and cleanup - Licence headers added Signed-off-by: Roman Stratiienko <r.stratiienko@gmail.com> Change-Id: If0f3dac42bea74f97a87e5a682380c33f4ff6837
WIP at this moment. Working on RPI4. (v3d+vc4). Signed-off-by: Roman Stratiienko <r.stratiienko@gmail.com> Change-Id: I517f095543b031e35f0d48bc03f6010a67012865
v5:
|
Review / feedback are welcome. |
v5 verified working on waydroid - intel |
Tested this on Pi 4, Android 13 & 5.15 kernel. Using minigbm_gbm_mesa everything is working just the same as with the previous minigbm version I've used. OpenGL, Vulkan, libcamera, AOSP external camera HAL and v4l2_codec2 all work to the same extent as before. Using minigbm_dmabuf seems still overall quite unstable. There's random crashes on occasion that cause a soft reboot.
Great work as always! Just a small note that files in dmabuf_driver are missing copyright headers. |
Would it be possible to handle gbm_mesa and dmabuf like the other backends in
This could be handy on Android-x86, which targets multiple hardware and tries to stay generic, but also for testing purpose for a specific platform. |
It's planned for next iterations in some or another forms. |
I tested video playback (using codec2) in QEMU/virgl running my Android 13 build, using the 3 flavors: minigbm_intel, minigbm_gbm_mesa and minigbm_dmabuf. minigbm_intelThis produces recognizable, but incorrect, result: minigbm_gbm_mesaAlso there is these logs constantly repeated in QEMU standard error:
And also these logcat outputs at the start of the playback (not constantly repeated):
minigbm_dmabufNote: this is the first time I actually try minigbm_dmabuf, so I might be missing something.
|
That actually looks pretty similar to what I have on Pi 4 using minigbm_gbm_mesa (GloDroid/glodroid_manifest#160 (comment)). There's also some other artifacts present in the playback that are likely because hw codecs on Pi 4 produce buffers that use their own custom modifier. So, this makes me now think there's two separate issues with this.
Kernel needs to be compiled with following options enabled.
Default ueventd.rc already has permission for these (https://cs.android.com/android/platform/superproject/+/master:system/core/rootdir/ueventd.rc;l=46-48). In addition I also have Could be due to some entirely different issue of course but that's at least easy to check. |
Thanks. I already have these enabled, because codec2 depends on it (and ION drivers were removed from kernel 5.15, as you already know :)). |
dmabuf driver is very dumb right now and in future would require external constraints configuration file |
Same for hwcodecs and camera. There's no way to solve this issue with gbm/mesa, at least it would require platform-specific constraints configuration file. |
It's a bit off-topic, but since it's been mentioned, does somebody know whether v4l2_codec2 can be used with Intel GPU? |
No, it can't ADD: But artifacts you attached can be fixed with proper allocation constraints adjustment. |
Mainline CL: https://chromium-review.googlesource.com/c/chromiumos/platform/minigbm/+/2982542