Handle nv12 pix fmt.

[topilski]

Hi, according code https://github.com/linux-sunxi/libvdpau-sunxi/blob/master/decoder.c#L149 used VDP_YCBCR_FORMAT_YV12 pixformat for cedar version grather or equal 0x1680, but when executed https://github.com/linux-sunxi/libvdpau-sunxi/blob/master/surface_video.c#L160 not handled case when vs->source_format == VDP_YCBCR_FORMAT_NV12 only handled for old cedar library https://github.com/linux-sunxi/libvdpau-sunxi/blob/master/surface_video.c#L208 , why nv12 for new cedar library are not work out? Also wiil be cool have conversion copy_yv12_to_nv12 and veca versa.

static INLINE void
u_copy_nv12_to_yv12(void *const *destination_data,
                    uint32_t const *destination_pitches,
                    int src_plane, int src_field,
                    int src_stride, int num_fields,
                    uint8_t const *src,
                    int width, int height)
{
   int x, y;
   unsigned u_stride = destination_pitches[2] * num_fields;
   unsigned v_stride = destination_pitches[1] * num_fields;
   uint8_t *u_dst = (uint8_t *)destination_data[2] + destination_pitches[2] * src_field;
   uint8_t *v_dst = (uint8_t *)destination_data[1] + destination_pitches[1] * src_field;

   /* TODO: SIMD */
   for (y = 0; y < height; y++) {
      for (x = 0; x < width; x++) {
         u_dst[x] = src[2*x];
         v_dst[x] = src[2*x+1];
      }
      u_dst += u_stride;
      v_dst += v_stride;
      src += src_stride;
   }
}

static INLINE void
u_copy_yv12_to_nv12(void *const *destination_data,
                    uint32_t const *destination_pitches,
                    int src_plane, int src_field,
                    int src_stride, int num_fields,
                    uint8_t const *src,
                    int width, int height)
{
   int x, y;
   unsigned offset = 2 - src_plane;
   unsigned stride = destination_pitches[1] * num_fields;
   uint8_t *dst = (uint8_t *)destination_data[1] + destination_pitches[1] * src_field;

   /* TODO: SIMD */
   for (y = 0; y < height; y++) {
      for (x = 0; x < 2 * width; x += 2) {
         dst[x+offset] = src[x>>1];
      }
      dst += stride;
      src += src_stride;
   }
}

[rellla]

Because nobody implemented the get_bits_ycbcr case, if you want to get NV12 from the decoder's YV12 surface.

vs->source_format == VDP_YCBCR_FORMAT_YV12 && destination_ycbcr_format == VDP_YCBCR_FORMAT_NV12

is simply not implemented. We have more than just YV12 and NV12 that should be implemented according to the API: http://http.download.nvidia.com/XFree86/vdpau/doxygen/html/group__misc__types.html#gab7550cf65e6d46f4fd7a1e322372e207

Feel free to do a pull request ;)

[topilski]

Hi @rellla i with pleasure will be implement this functionality but i don't understand how convert your pointers src = cedrus_mem_get_pointer(vs->yuv->data); dst = destination_data[0]; into common yv12 to nv12 conversion algorithm.

[rellla]

I have not read up about this algorithm, but cedrus_mem_get_pointer(vs->yuv->data) will return a pointer to the beginning of the yuv bitstream, which you can use with CPU. destination_data[0] points to the beginning of the first plane (depending on the output format)
http://http.download.nvidia.com/XFree86/vdpau/doxygen/html/group___vdp_video_surface.html#gae3550dc62e4e4c390c1874b37f96e7c0
For the planes, check https://wiki.videolan.org/YUV/

This is YV12 (for 1 pixel)
YYYYYYYY UU VV (3 planes)
NV12 (for 1 pixel)
YYYYYYYY UVUV

which means, when converting NV12 to YV12 you just need to split up cedrus_mem_get_pointer(vs->yuv->data)+ vs->luma_size into separate U (destination_data[1]) and V (destination_data[2]) planes.

This is, what basically tiled_deinterleave_to_planar does - but without doing it with the tiled input ....
Where did the above code snippets come from?

[topilski]

https://github.com/olvaffe/mesa/blob/master/src/gallium/auxiliary/util/u_video.h#L103

[rellla]

just asked because of the license ...