Enable zero buffer copy method for wayland VDA. #260

shaochangbin · 2014-08-08T05:47:51Z

1080p video playback is quite unsmooth on VTC-1010, with only 15 render fps
and ~70% CPU usage. Root cause is current code spend too much time to
put output vasurface to GL texture for more than 25ms for each
frame, due to buffer copy between driver and user space. As a result,
GPU swap buffer operations are not executed timely and then blocks
render side composting.

This CL enables a zero buffer copy solution that will put vasurface to
CL texture inside GPU process. It creates an EGL image through
EGL_DRM_BUFFER_MESA target, then binds the EGL image to GL texture.

Local test result shows it gets 31 render fps and ~25% CPU usage for 1080p playback.
It also benefits other resolution video playback.

BUG = XWALK-2069

shaochangbin · 2014-08-08T05:48:38Z

ptal, thanks!
@tiagovignatti

Package libva and its driver supports vaLockBuffer & vaUnlockBuffer on Tizen IVI, while the two APIs are not included in Chromium libva ATM. intel/ozone-wayland#260 requires the APIs to get vaimage buffer handle so as to create EGL image. This CL add these API in Chromium libva.

rakuco · 2014-08-08T08:00:26Z

cc @ds-hwang

rakuco · 2014-08-08T08:06:20Z

crosswalk-project/chromium-crosswalk#183 mentions this patch requires custom changes to Chromium's libva, so it won't work on oz-wl by default with a vanilla Chromium?

ds-hwang · 2014-08-08T10:17:24Z

It's cool. However, afaik IVI with oz-wl will use its own vaapi_wrapper.cc impl inside oz-wl project. @tiagovignatti , am I right?

tiagovignatti · 2014-08-11T07:55:46Z

amazing work @shaochangbin. I'll be reviewing this today still. Thanks.

tiagovignatti · 2014-08-11T15:13:31Z

hi back. I've tested the changes here but something is wrong, possibly with my setup. Xwalk crashes here with an output like this:

libva info: VA-API version 0.35.2
libva info: va_getDriverName() returns 0
libva info: Trying to open /home/tiago/install/lib/dri/i965_drv_video.so
libva info: Found init function __vaDriverInit_0_35
libva info: va_openDriver() returns 0
[0811/180845:WARNING:audio_discard_helper.cc(21)] Input timestamps are not monotonically increasing! ts 23220 us diff 0 us
[0811/180845:ERROR:gpu_video_decoder.cc(607)] VDA Error: 4
[0811/180845:FATAL:lock_impl_posix.cc(46)] Check failed: rv == 0 (22 vs. 0). Invalid argument

It might be that the problem is not even in Chromium side. I've tried to isolate it and checked that even libva putsurface doesn't seem to be outputting correctly. It keeps dumping a repeatedly line like this:

surface buffer info: buf_info.handle: 0x0, buf_info.type: -1418922124, buf_info.mem_type: 0x40000000, buf_info.mem_size: 1

I tried it using libva 1.3.1 and on top of that I applied myself the following commits (that I've stolen from review.tizen.org):

dc43b712d903627719e1d2c8073d44518947c386 test: add example code of buffer exports in putsurface
e0756dee475c9c41e04953fbbb76e1c7747ec00d add VA_SURFACE_ATTRIB_MEM_TYPE_KERNEL_DRM_BO in va_drmcommon.h
f71d60097aa085f8fbeacc27168fdb8e427eeefb API: add support for buffer exports.

shaochangbin · 2014-08-12T01:59:11Z

thanks for trying this:) Looks the libva version is quite new, whereas on Tizen IVI its version is 1.2.1.
I guess maybe the patches for libva-driver are not applied here. Could you please try the 'libva' & 'libva-dirver' rpms from http://download.tizen.org/releases/daily/tizen/ivi/ivi/latest/repos/atom/packages/i686/ , they add the support for vaLockBuffer & vaUnlockBuffer.

tiagovignatti · 2014-08-12T05:25:42Z

where are the sources for these packages? as I said previously, I'm not using Tizen to test it.

shaochangbin · 2014-08-12T05:39:58Z

The source code could be found in libva & libva-intel-driver packages from http://download.tizen.org/releases/daily/tizen/ivi/ivi/latest/repos/atom/source/,
Think libva & libva-intel-driver projects from review.tizen.org could also be a reference.
Btw, I'm on crosswalk IRC, we can have chat there:)

tiagovignatti · 2014-08-12T08:13:39Z

media/vaapi_video_decode_accelerator.cc

-  if (!va_wrapper_->PutSurfaceIntoImage(surface, &va_image_)) {
-    DVLOG(1) << "Failed to put va surface to image";
+  if (egl_image_ != EGL_NO_IMAGE_KHR)
+    DestroyEGLImage(egl_display_, egl_image_);


why destroy egl_image_, if it's already null? What's the case that it falls in this conditional?

EGL_NO_IMAGE_KHR equals NULL here. There is GL mapping of eglCreateImageKHR and eglDestroyImageKHR, so release an EGL image first before create a new one.

fine. thanks.

tiagovignatti · 2014-08-12T08:17:12Z

just for the record: I could get the environment up and running after all. All looks great at first -- very nice! Anyhow, I had a offline talk with @shaochangbin and we agreed to talk with libva maintainers first whether we should be using the 'staging' branch of the project, and the Acquire/Release buffer approach instead.

kalyankondapally · 2014-08-22T03:54:05Z

@shaochangbin @tiagovignatti Can one of you please update the Readme with how to set up the environment for taking advantage of HWA decoding. (i.e. Vappi setup and Chromium compiling and run time flags.)

kalyankondapally · 2014-09-02T23:21:39Z

@shaochangbin I realize these changes depend on libva packages in Tizen. Is their a way to determine if the needed support is available or not during run time and if not available use existing code path ?

shaochangbin · 2014-09-03T05:24:43Z

@kalyankondapally thanks a lot for your suggestion, I will add the libva dependency check in the following change.
There is VaapiWrapper::VAAPIVersionLessThan() to check vaapi version, but I think it won't solve this issue due to the vaLockBuffer Apis are customized changes only in Tizen IVI not libva upstream.
IMO the expected behavior would be: If libva supports vaLockBuffer Apis, such as Tizen libva, we can use the proposed VDA method to reach better video playback performance; on the other hand, if libva doesn't support vaLockBuffer Apis, we can still use the previous solution even the performance is bad.
A feasible solution is using linux dlsym() api to check whether the vaLockBuffer apis are supported or not. I will submit a patch later with this solution for your review.

shaochangbin · 2014-09-03T06:29:13Z

@kalyankondapally @tiagovignatti , I've updated patch to add vaLockBuffer check. please take a look, thank you.

kalyankondapally · 2014-09-03T16:09:55Z

@shaochangbin Looks better. Can you please rebase the patch to latest master TOT ?

shaochangbin · 2014-09-04T07:03:05Z

Patch is rebased, ptal, thanks.
@kalyankondapally @tiagovignatti

shaochangbin · 2014-09-04T14:58:39Z

hi, Kalyan & Tiago, any review comments?
Btw, the CL requires the vaLockBuffer function definitions in third_party/libva/va/va.h, will that needs extra changes to support other platforms besides Tizen IVI?

tiagovignatti · 2014-09-04T15:06:52Z

No comments from my side really. I defer now this to @kalyankondapally

kalyankondapally · 2014-09-04T15:47:09Z

@shaochangbin so changes in third_party/libva/va/va.h are needed to get this compiled ?

shaochangbin · 2014-09-05T01:20:05Z

@kalyankondapally yes, i think so. To get this compiled, we need to include the declarations in va.h, like crosswalk-project/chromium-crosswalk#183.

The long term goal would be make those declarations landed in Chromium upstream, after the libva and its driver side change are finished.

kalyankondapally · 2014-09-05T02:45:44Z

@shaochangbin Can you please include those changes as a patch here. See patches/ folder in oz-wl

Reason that add this change: 1080p video playback on Tizen is quite unsmooth on VTC-1010, with only 15 render fps and ~70% CPU usage. Ozone-wl landed a workaround solution that enables a zero buffer copy method that uses libva vaLockBuffer APIs which gains much better performance than previous method. intel/ozone-wayland#260 In order to make the change also benifit Crosswalk/Tizen, we should add the vaLockBuffer API declarations inside Chromium Crosswalk to compile the code successfully. The long term goal of our work would be landing those API declarations inside Chromium upstream, after the API implementations are landed in libva and libva-driver upstream. Therefore, this CL is a temporary solution. It would be reverted after the long term goal is achived.

Reason that add this change: 1080p video playback on Tizen is quite unsmooth on VTC-1010, with only 15 render fps and ~70% CPU usage. Ozone-wl landed a workaround solution that enables a zero buffer copy method that uses libva vaLockBuffer APIs which gains much better performance than previous method. intel/ozone-wayland#260 In order to make the change also benifit Crosswalk/Tizen, we should add the vaLockBuffer API declarations inside Chromium Crosswalk to compile the code successfully. The long term goal of our work would be landing those API declarations inside Chromium upstream, after the API implementations are landed in libva and libva-driver upstream. Therefore, this CL is a temporary solution. It would be reverted after the long term goal is achived. [This is ozone-wayland's 0006-Introduce-vaLockBuffer-APIs-in-libva.patch]

Reason that add this change: 1080p video playback on Tizen is quite unsmooth on VTC-1010, with only 15 render fps and ~70% CPU usage. Ozone-wl landed a workaround solution that enables a zero buffer copy method that uses libva vaLockBuffer APIs which gains much better performance than previous method. intel/ozone-wayland#260 In order to make the change also benifit Crosswalk/Tizen, we should add the vaLockBuffer API declarations inside Chromium Crosswalk to compile the code successfully. The long term goal of our work would be landing those API declarations inside Chromium upstream, after the API implementations are landed in libva and libva-driver upstream. Therefore, this CL is a temporary solution. It would be reverted after the long term goal is achived.

Reason that add this change: 1080p video playback on Tizen is quite unsmooth on VTC-1010, with only 15 render fps and ~70% CPU usage. Ozone-wl landed a workaround solution that enables a zero buffer copy method that uses libva vaLockBuffer APIs which gains much better performance than previous method. intel/ozone-wayland#260 In order to make the change also benifit Crosswalk/Tizen, we should add the vaLockBuffer API declarations inside Chromium Crosswalk to compile the code successfully. The long term goal of our work would be landing those API declarations inside Chromium upstream, after the API implementations are landed in libva and libva-driver upstream. Therefore, this CL is a temporary solution. It would be reverted after the long term goal is achived. [This is ozone-wayland's 0006-Introduce-vaLockBuffer-APIs-in-libva.patch]

shaochangbin mentioned this pull request Aug 8, 2014

Introduce vaLockBuffer & vaUnlockBuffer APIs in libva. crosswalk-project/chromium-crosswalk#183

Merged

tiagovignatti reviewed Aug 12, 2014
View reviewed changes

qjia7 mentioned this pull request Aug 19, 2014

Fixed screen displayed abnormally when playing some size of html5 mp4 vi... #268

Closed

shaochangbin force-pushed the drm-egl-vda branch 2 times, most recently from 45f6768 to e87d301 Compare September 3, 2014 06:26

shaochangbin force-pushed the drm-egl-vda branch from e87d301 to f4fb54f Compare September 4, 2014 06:51

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Enable zero buffer copy method for wayland VDA. #260

Enable zero buffer copy method for wayland VDA. #260

shaochangbin commented Aug 8, 2014

shaochangbin commented Aug 8, 2014

rakuco commented Aug 8, 2014

rakuco commented Aug 8, 2014

ds-hwang commented Aug 8, 2014

tiagovignatti commented Aug 11, 2014

tiagovignatti commented Aug 11, 2014

shaochangbin commented Aug 12, 2014

tiagovignatti commented Aug 12, 2014

shaochangbin commented Aug 12, 2014

tiagovignatti Aug 12, 2014

shaochangbin Aug 12, 2014

tiagovignatti Aug 12, 2014

tiagovignatti commented Aug 12, 2014

kalyankondapally commented Aug 22, 2014

kalyankondapally commented Sep 2, 2014

shaochangbin commented Sep 3, 2014

shaochangbin commented Sep 3, 2014

kalyankondapally commented Sep 3, 2014

shaochangbin commented Sep 4, 2014

shaochangbin commented Sep 4, 2014

tiagovignatti commented Sep 4, 2014

kalyankondapally commented Sep 4, 2014

shaochangbin commented Sep 5, 2014

kalyankondapally commented Sep 5, 2014

Enable zero buffer copy method for wayland VDA. #260

Enable zero buffer copy method for wayland VDA. #260

Conversation

shaochangbin commented Aug 8, 2014

shaochangbin commented Aug 8, 2014

rakuco commented Aug 8, 2014

rakuco commented Aug 8, 2014

ds-hwang commented Aug 8, 2014

tiagovignatti commented Aug 11, 2014

tiagovignatti commented Aug 11, 2014

shaochangbin commented Aug 12, 2014

tiagovignatti commented Aug 12, 2014

shaochangbin commented Aug 12, 2014

tiagovignatti Aug 12, 2014

Choose a reason for hiding this comment

shaochangbin Aug 12, 2014

Choose a reason for hiding this comment

tiagovignatti Aug 12, 2014

Choose a reason for hiding this comment

tiagovignatti commented Aug 12, 2014

kalyankondapally commented Aug 22, 2014

kalyankondapally commented Sep 2, 2014

shaochangbin commented Sep 3, 2014

shaochangbin commented Sep 3, 2014

kalyankondapally commented Sep 3, 2014

shaochangbin commented Sep 4, 2014

shaochangbin commented Sep 4, 2014

tiagovignatti commented Sep 4, 2014

kalyankondapally commented Sep 4, 2014

shaochangbin commented Sep 5, 2014

kalyankondapally commented Sep 5, 2014