[Mesa-dev] [PATCH 1/2] st/va: ensure linear memory for dmabuf

[Christian König]

On 29.11.2015 18:26, Emil Velikov wrote:
> On 27 November 2015 at 08:57, Julien Isorce <j.isorce at samsung.com> wrote:
>> In order to do zero-copy between two different devices
>> the memory should not be tiled.
>>
>> This is currently no way to set pipe_resource template's flag
>> from pipe_video_buffer template. So disabled_tiling is added.
>>
>> Choosed "disable" prefix so that CALLOC keeps tiling enabled
>> by default.
>>
>> Tested with GStreamer on a laptop that has 2 GPUs:
>> 1- gstvaapidecode:
>>     HW decoding and dmabuf export with nouveau driver on Nvidia GPU.
>> 2- glimagesink:
>>     EGLImage imports dmabuf on Intel GPU.
>>
>> Note that tiling is working if 1 and 2 are done on the same GPU.
>> So it is up to the application to set or not the flag:
>> VA_SURFACE_EXTBUF_DESC_ENABLE_TILING
>>
>> Signed-off-by: Julien Isorce <j.isorce at samsung.com>
>> ---
>>   src/gallium/auxiliary/vl/vl_video_buffer.c | 3 +++
>>   src/gallium/include/pipe/p_video_codec.h   | 1 +
>>   src/gallium/state_trackers/va/surface.c    | 5 +++++
>>   3 files changed, 9 insertions(+)
>>
>> diff --git a/src/gallium/auxiliary/vl/vl_video_buffer.c b/src/gallium/auxiliary/vl/vl_video_buffer.c
>> index 6cd2557..62f4aa9 100644
>> --- a/src/gallium/auxiliary/vl/vl_video_buffer.c
>> +++ b/src/gallium/auxiliary/vl/vl_video_buffer.c
>> @@ -253,6 +253,9 @@ vl_video_buffer_template(struct pipe_resource *templ,
>>      templ->bind = PIPE_BIND_SAMPLER_VIEW | PIPE_BIND_RENDER_TARGET;
>>      templ->usage = usage;
>>
>> +   if (tmpl->disable_tiling)
>> +      templ->bind |= PIPE_BIND_LINEAR;
>> +
>>      if (plane > 0) {
>>         if (tmpl->chroma_format == PIPE_VIDEO_CHROMA_FORMAT_420) {
>>            templ->width0 /= 2;
>> diff --git a/src/gallium/include/pipe/p_video_codec.h b/src/gallium/include/pipe/p_video_codec.h
>> index 196d00b..dbfffd9 100644
>> --- a/src/gallium/include/pipe/p_video_codec.h
>> +++ b/src/gallium/include/pipe/p_video_codec.h
>> @@ -125,6 +125,7 @@ struct pipe_video_buffer
>>      enum pipe_video_chroma_format chroma_format;
>>      unsigned width;
>>      unsigned height;
>> +   bool disable_tiling;
>>      bool interlaced;
>>
>>      /**
>> diff --git a/src/gallium/state_trackers/va/surface.c b/src/gallium/state_trackers/va/surface.c
>> index c052c8f..f7043ad 100644
>> --- a/src/gallium/state_trackers/va/surface.c
>> +++ b/src/gallium/state_trackers/va/surface.c
>> @@ -616,6 +616,11 @@ vlVaCreateSurfaces2(VADriverContextP ctx, unsigned int format,
>>
>>         switch (memory_type) {
>>         case VA_SURFACE_ATTRIB_MEM_TYPE_VA:
>> +         /* The application will clear the TILING flag when the surface is
>> +          * intended to be exported as dmabuf. */
>> +         templat.disable_tiling = memory_attibute &&
>> +            !(memory_attibute->flags & VA_SURFACE_EXTBUF_DESC_ENABLE_TILING);
> The condition seems to be flipped, no ? Currently it's doing
> "disable_tiling = ENABLE_TILING_BIT_SET"
>
> Other than that, the idea is ok imho, although I'd appreciate
> Christian and others' feedback.
>
> A few things worth mentioning while looking around for this:
>   - suface_from_external_memory should (must) also know about disable_tiling.
>   - missing R in function name ^^ suRface_ ...
>   - sometimes radeon (see r600_video_buffer_create) completely
> overwrites the existing flags, as opposed to just set the linear bit.
> Bug, intentional, worth adding a comment ?
>   - in many cases nouveau won't create a linear surface as it's not
> using the above vl helper (hint, earlier suggesting about
> reworking/cleaning things up a bit, hint)

Yeah, agree deduplicating that code would probably be a good idea.

I would also prefer to have all the bind flags in the pipe_video_buffer 
as well.

Regards,
Christian.

>
> Thanks
> Emil
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev