[Mesa-dev] [PATCH] radeon/vcn: move message buffer to vram for now

[Christian König]

Am 25.07.2017 um 14:13 schrieb Leo Liu:
>
>
> On 07/25/2017 05:19 AM, Christian König wrote:
>> Am 25.07.2017 um 08:27 schrieb Michel Dänzer:
>>> On 25/07/17 02:54 AM, Leo Liu wrote:
>>>> To workaround an unknown bug.
>>>>
>>>> Signed-off-by: Leo Liu <leo.liu at amd.com>
>>>> ---
>>>>   src/gallium/drivers/radeon/radeon_vcn_dec.c | 3 ++-
>>>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/src/gallium/drivers/radeon/radeon_vcn_dec.c 
>>>> b/src/gallium/drivers/radeon/radeon_vcn_dec.c
>>>> index bd93b849db..a60b969a27 100644
>>>> --- a/src/gallium/drivers/radeon/radeon_vcn_dec.c
>>>> +++ b/src/gallium/drivers/radeon/radeon_vcn_dec.c
>>>> @@ -1237,8 +1237,9 @@ struct pipe_video_codec 
>>>> *radeon_create_decoder(struct pipe_context *context,
>>>>           unsigned msg_fb_it_size = FB_BUFFER_OFFSET + FB_BUFFER_SIZE;
>>>>           if (have_it(dec))
>>>>               msg_fb_it_size += IT_SCALING_TABLE_SIZE;
>>>> +        /* use vram to improve performance, workaround an unknown 
>>>> bug */
>>>>           if (!rvid_create_buffer(dec->screen, 
>>>> &dec->msg_fb_it_buffers[i],
>>>> -                    msg_fb_it_size, PIPE_USAGE_STAGING)) {
>>>> +                    msg_fb_it_size, PIPE_USAGE_DEFAULT)) {
>>>>               RVID_ERR("Can't allocated message buffers.\n");
>>>>               goto error;
>>>>           }
>>>>
>>> Does PIPE_USAGE_STREAM help as well? That would be system memory but
>>> with write-combined CPU access, whereas PIPE_USAGE_STAGING is 
>>> cacheable.
>>
>> No, unfortunately not. It's a hardware bug which is independent of 
>> the caching mode.
>>
>> Leo can you limit this workaround to APUs? Shouldn't matter for dGPUs.
>
> This change is on vcn_dec, and that is for Raven only now.

Ah, of course. In this case feel free to go ahead.

>
>>
>> Additional to that I would rather like to have that in the kernel, so 
>> that we can easily remove it when the correct workaround from the 
>> hardware guys for this issue lands.
>
> Do you mean disabling Gart Cacheable completely in kernel, and using 
> VRAM globally instead? if that's the case, not sure if it would have 
> side effect on other part of the driver.

The hardware guys suggested to clear the system bit in the page tables 
and handle it like stolen VRAM instead, but of hand I couldn't get this 
to work.

> On the current issue, Msg buffer and Bitstream buffer are both in 
> Gart, but only Msg buffer affect the performance, since it interacts 
> more often with GPU.

Good point. In this case the patch is Acked-by: Christian König 
<christian.koenig at amd.com>.

Regards,
Christian.

>
>
> Leo
>
>>
>> Christian.
>>
>