[Mesa-dev] [PATCH v3 3/4] nouveau: fix chunk decoding by updating number of slices

[Ilia Mirkin]

On Mon, Sep 7, 2015 at 8:15 AM, Julien Isorce <j.isorce at samsung.com> wrote:
> https://bugs.freedesktop.org/show_bug.cgi?id=89969
>
> Signed-off-by: Julien Isorce <j.isorce at samsung.com>
> ---
>  src/gallium/drivers/nouveau/nouveau_vp3_video.h     | 3 +++
>  src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c | 8 ++++++++
>  2 files changed, 11 insertions(+)
>
> diff --git a/src/gallium/drivers/nouveau/nouveau_vp3_video.h b/src/gallium/drivers/nouveau/nouveau_vp3_video.h
> index 43b9556..87ae012 100644
> --- a/src/gallium/drivers/nouveau/nouveau_vp3_video.h
> +++ b/src/gallium/drivers/nouveau/nouveau_vp3_video.h
> @@ -121,6 +121,9 @@ struct nouveau_vp3_decoder {
>
>     /* Total data appended so far after last begin frame. */
>     unsigned bsp_size;
> +
> +   /* Number of slices so far since last begin frame */
> +   unsigned int nb_slices;
>  };
>
>  struct comm {
> diff --git a/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c b/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
> index c1d70ea..7776e4e 100644
> --- a/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
> +++ b/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
> @@ -128,6 +128,8 @@ nouveau_vp3_fill_picparm_mpeg12_bsp(struct nouveau_vp3_decoder *dec,
>     for (i = 0; i < 4; ++i)
>        pic_bsp->f_code[i/2][i%2] = desc->f_code[i/2][i%2] + 1; // FU
>
> +   desc->num_slices = dec->nb_slices;
> +
>     return (desc->num_slices << 4) | (dec->base.profile != PIPE_VIDEO_PROFILE_MPEG1);
>  }
>
> @@ -196,6 +198,8 @@ nouveau_vp3_fill_picparm_h264_bsp(struct nouveau_vp3_decoder *dec,
>     struct h264_picparm_bsp stub_h = {}, *h = &stub_h;
>     uint32_t caps = (d->slice_count << 4)&0xfff0;
>
> +   d->slice_count = dec->nb_slices;

But caps above is already set based on d->slice_count. With this diff,
vaapi works *much* better:

diff --git a/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
b/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
index 7776e4e..7aa70e2 100644
--- a/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
+++ b/src/gallium/drivers/nouveau/nouveau_vp3_video_bsp.c
@@ -196,11 +196,12 @@ nouveau_vp3_fill_picparm_h264_bsp(struct
nouveau_vp3_decoder *dec,
                                   char *map)
 {
    struct h264_picparm_bsp stub_h = {}, *h = &stub_h;
-   uint32_t caps = (d->slice_count << 4)&0xfff0;
+   uint32_t caps;

-   d->slice_count = dec->nb_slices;
+   d->slice_count += dec->nb_slices;

    assert(!(d->slice_count & ~0xfff));
+   caps = (d->slice_count << 4)&0xfff0;
    if (d->slice_count & 0x1000)
       caps |= 1 << 20;


I'm not sure about the += -- can you investigate what the proper way
of handling this is? Also note that you appear to have forgotten to do
this for vc1.

Very interesting by the way -- the vaapi decode does not have the same
problem as the vdpau decode on that simpsons trailer. (About 28s in,
when the sleigh is flying over the ravine.)

> +
>     assert(!(d->slice_count & ~0xfff));
>     if (d->slice_count & 0x1000)
>        caps |= 1 << 20;
> @@ -239,6 +243,7 @@ nouveau_vp3_bsp_begin(struct nouveau_vp3_decoder *dec)
>
>     dec->bsp_ptr = bsp_bo->map;
>     dec->bsp_size = NOUVEAU_VP3_BSP_RESERVED_SIZE;
> +   dec->nb_slices = 0;
>
>     dec->bsp_ptr += 0x100;
>
> @@ -266,6 +271,8 @@ nouveau_vp3_bsp_next(struct nouveau_vp3_decoder *dec, unsigned num_buffers,
>     struct strparm_bsp *str_bsp = NULL;
>     int i = 0;
>
> +   ++dec->nb_slices;
> +
>     bsp_origin += 0x100;
>     str_bsp = (struct strparm_bsp *)bsp_origin;
>
> @@ -331,6 +338,7 @@ nouveau_vp3_bsp_end(struct nouveau_vp3_decoder *dec, union pipe_desc desc)
>
>     dec->bsp_ptr = NULL;
>     dec->bsp_size = 0;
> +   dec->nb_slices = 0;
>
>     return caps;
>  }
> --
> 1.9.1
>
> _______________________________________________
> mesa-dev mailing list
> mesa-dev at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/mesa-dev