On 21/04/15 9:54 PM, Ronald S. Bultje wrote: > --- > libavcodec/vp9.c | 316 > ++++++++++++++++++++++--------------------- > libavcodec/vp9_mc_template.c | 171 +++++++++++++++++++++++ > libavcodec/vp9_parser.c | 5 +- > libavcodec/vp9dsp.c | 205 ++++++++++++++++++++++++++-- > libavcodec/vp9dsp.h | 9 ++ > 5 files changed, 535 insertions(+), 171 deletions(-) > create mode 100644 libavcodec/vp9_mc_template.c > > diff --git a/libavcodec/vp9.c b/libavcodec/vp9.c > index 89257fa..ee73325 100644 > --- a/libavcodec/vp9.c > +++ b/libavcodec/vp9.c > @@ -242,7 +242,7 @@ typedef struct VP9Context { > // whole-frame cache > uint8_t *intra_pred_data[3]; > struct VP9Filter *lflvl; > - DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[71*80]; > + DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135*144]; > > // block reconstruction intermediates > int block_alloc_using_2pass; > @@ -251,6 +251,8 @@ typedef struct VP9Context { > struct { int x, y; } min_mv, max_mv; > DECLARE_ALIGNED(32, uint8_t, tmp_y)[64*64]; > DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][32*32]; > + uint16_t mvscale[3][2]; > + uint8_t mvstep[3][2]; > } VP9Context; > > static const uint8_t bwh_tab[2][N_BS_SIZES][2] = { > @@ -577,6 +579,26 @@ static int decode_frame_header(AVCodecContext *ctx, > s->varcompref[1] = 2; > } > } > + > + for (i = 0; i < 3; i++) { > + AVFrame *ref = s->refs[s->refidx[i]].f; > + int refw = ref->width, refh = ref->height; > + > + if (refw == w && refh == h) { > + s->mvscale[i][0] = s->mvscale[i][1] = 0; > + } else { > + if (w * 2 < refw || h * 2 < refh || w > 16 * refw || h > > 16 * refh) { > + av_log(ctx, AV_LOG_ERROR, > + "Invalid ref frame dimensions %dx%d for frame > size %dx%d\n", > + refw, refh, w, h); > + return AVERROR_INVALIDDATA; > + } > + s->mvscale[i][0] = (refw << 14) / w; > + s->mvscale[i][1] = (refh << 14) / h; > + s->mvstep[i][0] = 16 * s->mvscale[i][0] >> 14; > + s->mvstep[i][1] = 16 * s->mvscale[i][1] >> 14; > + } > + } > } > } > s->refreshctx = s->errorres ? 0 : get_bits1(&s->gb); > @@ -2524,12 +2546,118 @@ static void intra_recon(AVCodecContext *ctx, > ptrdiff_t y_off, ptrdiff_t uv_off) > } > } > > -static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2], > - uint8_t *dst, ptrdiff_t dst_stride, > - const uint8_t *ref, ptrdiff_t > ref_stride, > - ThreadFrame *ref_frame, > - ptrdiff_t y, ptrdiff_t x, const > VP56mv *mv, > - int bw, int bh, int w, int h) > +static av_always_inline void mc_luma_scaled(VP9Context *s, > vp9_scaled_mc_func smc, > + uint8_t *dst, ptrdiff_t > dst_stride, > + const uint8_t *ref, ptrdiff_t > ref_stride, > + ThreadFrame *ref_frame, > + ptrdiff_t y, ptrdiff_t x, const > VP56mv *mv, > + int bw, int bh, int w, int h, > + const uint16_t *scale, const > uint8_t *step) > +{ > +#define scale_mv(n, dim) (((int64_t)n * scale[dim]) >> 14) > + // BUG libvpx seems to scale the two components separately. This > introduces > + // rounding errors but we have to reproduce them to be exactly compatible > + // with the output from libvpx...
Did you report it to libvpx? It's probably a better idea to fix the bug in the reference library than purposely reproduce the buggy behavior in our implementation to remain compatible. _______________________________________________ ffmpeg-devel mailing list ffmpeg-devel@ffmpeg.org http://ffmpeg.org/mailman/listinfo/ffmpeg-devel