On Thu, Aug 1, 2019 at 4:04 PM Paul Kocialkowski <paul.kocialkowski@xxxxxxxxxxx> wrote: > > Hi, > > On Thu 01 Aug 19, 13:06, Tomasz Figa wrote: > > Hi Boris, > > > > On Wed, Jun 19, 2019 at 9:15 PM Boris Brezillon > > <boris.brezillon@xxxxxxxxxxxxx> wrote: > > [snip] > > > @@ -533,10 +535,21 @@ hantro_queue_setup(struct vb2_queue *vq, unsigned int *num_buffers, > > > return -EINVAL; > > > } > > > > > > + /* The H264 decoder needs extra size on the output buffer. */ > > > + if (ctx->vpu_src_fmt->fourcc == V4L2_PIX_FMT_H264_SLICE_RAW) > > > + extra_size0 = 128 * DIV_ROUND_UP(pixfmt->width, 16) * > > > + DIV_ROUND_UP(pixfmt->height, 16); > > > + > > > > I wonder if this shouldn't be accounted for already in the sizeimage > > returned by TRY_/S_FMT, so that the application can know the required > > buffer size if it uses some external allocator and DMABUF memory type. > > I know we had it like this in our downstream code, but it wasn't the > > problem because we use minigbm, where we explicitly add the same > > padding in the rockchip backend. Any thoughts? > > Does the extra size have to be allocated along with the buffer? > > On cedrus, we have a need for a similar side-buffer but give it a dedicated CMA > allocation, which should allow dma-buf-imported buffers. Yes, the decoder stores motion vectors (IIRC) after the image data. Best regards, Tomasz
