[FFmpeg-devel] [PATCH 1/2] avcodec: add av1 hardware acceleration decoder

Wang, Fei W fei.w.wang at intel.com
Mon Aug 24 09:57:47 EEST 2020



> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of James
> Almer
> Sent: Friday, August 21, 2020 10:43 PM
> To: ffmpeg-devel at ffmpeg.org
> Subject: Re: [FFmpeg-devel] [PATCH 1/2] avcodec: add av1 hardware
> acceleration decoder
> 
> On 8/21/2020 2:29 AM, Fei Wang wrote:
> > This av1 decoder is now only used for av1 hardware acceleration
> > decoder. Consider it can be extend to a local decoder like hevc or vp9
> > in the future, so define its name as "av1" and put it into external
> > libraries codec list.
> >
> > Signed-off-by: Fei Wang <fei.w.wang at intel.com>
> > ---
> >  Changelog              |   1 +
> >  configure              |   1 +
> >  libavcodec/Makefile    |   1 +
> >  libavcodec/allcodecs.c |   1 +
> >  libavcodec/av1dec.c    | 746
> +++++++++++++++++++++++++++++++++++++++++
> >  libavcodec/av1dec.h    |  89 +++++
> >  libavcodec/version.h   |   2 +-
> >  7 files changed, 840 insertions(+), 1 deletion(-)  create mode 100644
> > libavcodec/av1dec.c  create mode 100644 libavcodec/av1dec.h
> >
> > diff --git a/Changelog b/Changelog
> > index 1efc768387..3ff88cc12f 100644
> > --- a/Changelog
> > +++ b/Changelog
> > @@ -14,6 +14,7 @@ version <next>:
> >  - ADPCM Argonaut Games encoder
> >  - Argonaut Games ASF muxer
> >  - AV1 Low overhead bitstream format demuxer
> > +- AV1 decoder (Hardware acceleration used only)
> >
> >
> >  version 4.3:
> > diff --git a/configure b/configure
> > index 6faff9bc7b..ef71e47c4e 100755
> > --- a/configure
> > +++ b/configure
> > @@ -2685,6 +2685,7 @@ atrac3al_decoder_select="mdct"
> >  atrac3p_decoder_select="mdct sinewin"
> >  atrac3pal_decoder_select="mdct sinewin"
> >  atrac9_decoder_select="mdct"
> > +av1_decoder_select="cbs_av1"
> >  avrn_decoder_select="exif jpegtables"
> >  bink_decoder_select="blockdsp hpeldsp"
> >  binkaudio_dct_decoder_select="mdct rdft dct sinewin wma_freqs"
> > diff --git a/libavcodec/Makefile b/libavcodec/Makefile index
> > 3431ba2dca..2371039bd2 100644
> > --- a/libavcodec/Makefile
> > +++ b/libavcodec/Makefile
> > @@ -764,6 +764,7 @@ OBJS-$(CONFIG_ZLIB_DECODER)            += lcldec.o
> >  OBJS-$(CONFIG_ZLIB_ENCODER)            += lclenc.o
> >  OBJS-$(CONFIG_ZMBV_DECODER)            += zmbv.o
> >  OBJS-$(CONFIG_ZMBV_ENCODER)            += zmbvenc.o
> > +OBJS-$(CONFIG_AV1_DECODER)             += av1dec.o
> >
> >  # (AD)PCM decoders/encoders
> >  OBJS-$(CONFIG_PCM_ALAW_DECODER)           += pcm.o
> > diff --git a/libavcodec/allcodecs.c b/libavcodec/allcodecs.c index
> > 4bd830e5d0..00799d3431 100644
> > --- a/libavcodec/allcodecs.c
> > +++ b/libavcodec/allcodecs.c
> > @@ -761,6 +761,7 @@ extern AVCodec ff_idf_decoder;
> >   * above is available */
> >  extern AVCodec ff_aac_mf_encoder;
> >  extern AVCodec ff_ac3_mf_encoder;
> > +extern AVCodec ff_av1_decoder;
> >  extern AVCodec ff_h263_v4l2m2m_encoder;  extern AVCodec
> > ff_libaom_av1_decoder;  extern AVCodec ff_libopenh264_encoder; diff
> > --git a/libavcodec/av1dec.c b/libavcodec/av1dec.c new file mode 100644
> > index 0000000000..ab88ed987a
> > --- /dev/null
> > +++ b/libavcodec/av1dec.c
> > @@ -0,0 +1,746 @@
> > +/*
> > + * AV1video decoder
> > + *
> > + * This file is part of FFmpeg.
> > + *
> > + * FFmpeg is free software; you can redistribute it and/or
> > + * modify it under the terms of the GNU Lesser General Public
> > + * License as published by the Free Software Foundation; either
> > + * version 2.1 of the License, or (at your option) any later version.
> > + *
> > + * FFmpeg is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> > + * Lesser General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU Lesser General Public
> > + * License along with FFmpeg; if not, write to the Free Software
> > + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
> > +02110-1301 USA  */
> > +
> > +#include "avcodec.h"
> > +#include "get_bits.h"
> > +#include "hwconfig.h"
> > +#include "internal.h"
> > +#include "profiles.h"
> > +#include "thread.h"
> > +#include "videodsp.h"
> > +#include "libavutil/avassert.h"
> > +#include "libavutil/pixdesc.h"
> > +
> > +#include "av1dec.h"
> > +#include "libavformat/av1.h"
> > +
> > +static void setup_past_independence(AV1Frame *f) {
> > +    f->loop_filter_delta_enabled = 1;
> > +
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_INTRA] = 1;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_LAST] = 0;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_LAST2] = 0;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_LAST3] = 0;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_GOLDEN] = -1;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_BWDREF] = 0;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_ALTREF2] = -1;
> > +    f->loop_filter_ref_deltas[AV1_REF_FRAME_ALTREF] = -1;
> > +
> > +    f->loop_filter_mode_deltas[0] = 0;
> > +    f->loop_filter_mode_deltas[1] = 0; }
> > +
> > +static void load_previous_and_update(AV1DecContext *s) {
> > +    int i;
> > +
> > +    memcpy(s->cur_frame.loop_filter_ref_deltas,
> > +           s->ref[s-
> >raw_frame_header.primary_ref_frame].loop_filter_ref_deltas,
> > +           AV1_TOTAL_REFS_PER_FRAME * sizeof(int8_t));
> > +    memcpy(s->cur_frame.loop_filter_mode_deltas,
> > +           s->ref[s-
> >raw_frame_header.primary_ref_frame].loop_filter_mode_deltas,
> > +           2 * sizeof(int8_t));
> > +
> > +    if (s->raw_frame_header.loop_filter_delta_update) {
> > +        for (i = 0; i < AV1_TOTAL_REFS_PER_FRAME; i++) {
> > +            if (s->raw_frame_header.update_ref_delta[i])
> > +                s->cur_frame.loop_filter_ref_deltas[i] =
> > +                    s->raw_frame_header.loop_filter_ref_deltas[i];
> > +        }
> > +
> > +        for (i = 0; i < 2; i++) {
> > +            if (s->raw_frame_header.update_mode_delta[i])
> > +                s->cur_frame.loop_filter_mode_deltas[i] =
> > +                    s->raw_frame_header.loop_filter_mode_deltas[i];
> > +        }
> > +    }
> > +
> > +    s->cur_frame.loop_filter_delta_enabled =
> > +        s->raw_frame_header.loop_filter_delta_enabled;
> > +}
> > +
> > +static uint32_t inverse_recenter(int r, uint32_t v) {
> > +    if (v > 2 * r)
> > +        return v;
> > +    else if (v & 1)
> > +        return r - ((v + 1) >> 1);
> > +    else
> > +        return r + (v >> 1);
> > +}
> > +
> > +static uint32_t decode_unsigned_subexp_with_ref(uint32_t sub_exp,
> > +                                                int mx, int r) {
> > +    if ((r << 1) <= mx) {
> > +        return inverse_recenter(r, sub_exp);
> > +    } else {
> > +        return mx - 1 - inverse_recenter(mx - 1 - r, sub_exp);
> > +    }
> > +}
> > +
> > +static int32_t decode_signed_subexp_with_ref(uint32_t sub_exp, int low,
> > +                                             int high, int r) {
> > +    int32_t x = decode_unsigned_subexp_with_ref(sub_exp, high - low, r - low);
> > +    return x + low;
> > +}
> > +
> > +static void read_global_param(AV1DecContext *s, int type, int ref,
> > +int idx) {
> > +    uint32_t abs_bits, prec_bits, round, prec_diff, sub, mx, r;
> > +    abs_bits = AV1_GM_ABS_ALPHA_BITS;
> > +    prec_bits = AV1_GM_ALPHA_PREC_BITS;
> > +
> > +    if (idx < 2) {
> > +        if (type == AV1_WARP_MODEL_TRANSLATION) {
> > +            abs_bits = AV1_GM_ABS_TRANS_ONLY_BITS -
> > +                !s->raw_frame_header.allow_high_precision_mv;
> > +            prec_bits = AV1_GM_TRANS_ONLY_PREC_BITS -
> > +                !s->raw_frame_header.allow_high_precision_mv;
> > +        } else {
> > +            abs_bits = AV1_GM_ABS_TRANS_BITS;
> > +            prec_bits = AV1_GM_TRANS_PREC_BITS;
> > +        }
> > +    }
> > +    round = (idx % 3) == 2 ? (1 << AV1_WARPEDMODEL_PREC_BITS) : 0;
> > +    prec_diff = AV1_WARPEDMODEL_PREC_BITS - prec_bits;
> > +    sub = (idx % 3) == 2 ? (1 << prec_bits) : 0;
> > +    mx = 1 << abs_bits;
> > +    r = (s->ref[s->raw_frame_header.primary_ref_frame].gm_params[ref][idx]
> > +         >> prec_diff) - sub;
> > +
> > +    s->cur_frame.gm_params[ref][idx] =
> > +        (decode_signed_subexp_with_ref(s-
> >raw_frame_header.gm_params[ref][idx],
> > +                                       -mx, mx + 1, r) << prec_diff)
> > ++ round; }
> > +
> > +/**
> > +* update gm type/params, since cbs already implemented part of this
> > +funcation,
> > +* so we don't need to full implement spec.
> 
> FYI, the gm params code in cbs_av1 is currently buggy for some streams, and it
> needs fixing.
Thanks for your information. I will keep it in my mind. For clips(more than 200+) 
that I tested, it looks good and I didn't meet any bug yet from cbs_av1. 

Thanks
Fei
> 
> > +*/
> > +static void global_motion_params(AV1DecContext *s) {
> > +    const AV1RawFrameHeader *header = &s->raw_frame_header;
> > +    int type;
> > +    int i, j;
> > +    for (i = AV1_REF_FRAME_LAST; i <= AV1_REF_FRAME_ALTREF; i++) {
> > +        s->cur_frame.gm_type[i] = AV1_WARP_MODEL_IDENTITY;
> > +        for (j = 0; j <= 5; j++)
> > +            s->cur_frame.gm_params[i][j] = (j % 3 == 2) ?
> > +                                           1 << AV1_WARPEDMODEL_PREC_BITS : 0;
> > +    }
> > +    if (header->frame_type == AV1_FRAME_KEY ||
> > +        header->frame_type == AV1_FRAME_INTRA_ONLY)
> > +        return;
> > +
> > +    for (i = AV1_REF_FRAME_LAST; i <= AV1_REF_FRAME_ALTREF; i++) {
> > +        if (header->is_global[i]) {
> > +            if (header->is_rot_zoom[i]) {
> > +                type = AV1_WARP_MODEL_ROTZOOM;
> > +            } else {
> > +                type = header->is_translation[i] ?
> AV1_WARP_MODEL_TRANSLATION
> > +                                                 : AV1_WARP_MODEL_AFFINE;
> > +            }
> > +        } else {
> > +            type = AV1_WARP_MODEL_IDENTITY;
> > +        }
> > +        s->cur_frame.gm_type[i] = type;
> > +
> > +        if (type >= AV1_WARP_MODEL_ROTZOOM) {
> > +            read_global_param(s, type, i, 2);
> > +            read_global_param(s, type, i, 3);
> > +            if (type == AV1_WARP_MODEL_AFFINE) {
> > +                read_global_param(s, type, i, 4);
> > +                read_global_param(s, type, i, 5);
> > +            } else {
> > +                s->cur_frame.gm_params[i][4] = - s->cur_frame.gm_params[i][3];
> > +                s->cur_frame.gm_params[i][5] = - s->cur_frame.gm_params[i][2];
> > +            }
> > +        }
> > +        if (type >= AV1_WARP_MODEL_TRANSLATION) {
> > +            read_global_param(s, type, i, 0);
> > +            read_global_param(s, type, i, 1);
> > +        }
> > +    }
> > +}
> > +
> > +static int get_tiles_info(AVCodecContext *avctx,
> > +                          AV1RawTileGroup *tile_group,
> > +                          uint32_t tile_group_offset) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    GetBitContext gb;
> > +    uint16_t tile_num, tile_row, tile_col;
> > +    uint16_t mi_cols, mi_rows, sb_cols, sb_rows, tile_width_sb, tile_height_sb;
> > +    uint32_t size = 0, size_bytes = 0, offset = 0;
> > +    int ret = 0;
> > +
> > +    if ((ret = init_get_bits8(&gb,
> > +                              tile_group->tile_data.data,
> > +                              tile_group->tile_data.data_size)) < 0) {
> > +        av_log(avctx, AV_LOG_ERROR, "Fail to initialize bitstream reader.\n");
> > +        return ret;
> > +    }
> > +
> > +    s->tg_start = tile_group->tg_start;
> > +    s->tg_end = tile_group->tg_end;
> > +
> > +    if (s->raw_frame_header.uniform_tile_spacing_flag) {
> > +        mi_cols = 2 * ((s->raw_seq.max_frame_width_minus_1 + 8) >> 3);
> > +        mi_rows = 2 * ((s->raw_seq.max_frame_height_minus_1 + 8) >> 3);
> > +        sb_cols = s->raw_seq.use_128x128_superblock ? ((mi_cols + 31) >> 5)
> > +                                                    : ((mi_cols + 15) >> 4);
> > +        sb_rows = s->raw_seq.use_128x128_superblock ? ((mi_rows + 31) >> 5)
> > +                                                    : ((mi_rows + 15) >> 4);
> > +        tile_width_sb = (sb_cols + (1 << s->raw_frame_header.tile_cols_log2)
> > +                         -1) >> s->raw_frame_header.tile_cols_log2;
> > +        tile_height_sb = (sb_rows + (1 << s->raw_frame_header.tile_rows_log2)
> > +                         -1) >> s->raw_frame_header.tile_rows_log2;
> > +    }
> > +
> > +    for (tile_num = tile_group->tg_start; tile_num <= tile_group->tg_end;
> tile_num++) {
> > +        tile_row = tile_num / s->raw_frame_header.tile_cols;
> > +        tile_col = tile_num % s->raw_frame_header.tile_cols;
> > +
> > +        if (tile_num == tile_group->tg_end) {
> > +            s->tile_group_info[tile_num].tile_group_offset = tile_group_offset;
> > +            s->tile_group_info[tile_num].tile_size = get_bits_left(&gb) / 8;
> > +            s->tile_group_info[tile_num].tile_offset = offset;
> > +            s->tile_group_info[tile_num].tile_row = tile_row;
> > +            s->tile_group_info[tile_num].tile_column = tile_col;
> > +            s->tile_group_info[tile_num].tile_size_bytes = size_bytes;
> > +            if (s->raw_frame_header.uniform_tile_spacing_flag) {
> > +                s->tile_group_info[tile_num].tile_width_sbs =
> > +                    (tile_col + 1) == s->raw_frame_header.tile_cols ?
> > +                    sb_cols - tile_col * tile_width_sb :
> > +                    tile_width_sb;
> > +                s->tile_group_info[tile_num].tile_height_sbs =
> > +                    (tile_row + 1) == s->raw_frame_header.tile_rows ?
> > +                    sb_rows - tile_row * tile_height_sb :
> > +                    tile_height_sb;
> > +            } else {
> > +                s->tile_group_info[tile_num].tile_width_sbs =
> > +                    s->raw_frame_header.width_in_sbs_minus_1[tile_col] + 1;
> > +                s->tile_group_info[tile_num].tile_height_sbs =
> > +                    s->raw_frame_header.height_in_sbs_minus_1[tile_row] + 1;
> > +            }
> > +            return 0;
> > +        }
> > +        size_bytes = s->raw_frame_header.tile_size_bytes_minus1 + 1;
> > +        size = get_bits_le(&gb, size_bytes * 8) + 1;
> > +        skip_bits(&gb, size * 8);
> > +
> > +        offset += size_bytes;
> > +
> > +        s->tile_group_info[tile_num].tile_group_offset = tile_group_offset;
> > +        s->tile_group_info[tile_num].tile_size = size;
> > +        s->tile_group_info[tile_num].tile_offset = offset;
> > +        s->tile_group_info[tile_num].tile_row = tile_row;
> > +        s->tile_group_info[tile_num].tile_column = tile_col;
> > +        s->tile_group_info[tile_num].tile_size_bytes = size_bytes;
> > +        if (s->raw_frame_header.uniform_tile_spacing_flag) {
> > +            s->tile_group_info[tile_num].tile_width_sbs =
> > +                (tile_col + 1) == s->raw_frame_header.tile_cols ?
> > +                sb_cols - tile_col * tile_width_sb :
> > +                tile_width_sb;
> > +            s->tile_group_info[tile_num].tile_height_sbs =
> > +                (tile_row + 1) == s->raw_frame_header.tile_rows ?
> > +                sb_rows - tile_row * tile_height_sb :
> > +                tile_height_sb;
> > +        } else {
> > +            s->tile_group_info[tile_num].tile_width_sbs =
> > +                s->raw_frame_header.width_in_sbs_minus_1[tile_col] + 1;
> > +            s->tile_group_info[tile_num].tile_height_sbs =
> > +                s->raw_frame_header.height_in_sbs_minus_1[tile_row] + 1;
> > +        }
> > +        offset += size;
> > +    }
> > +
> > +    return 0;
> > +
> > +}
> > +
> > +static int get_pixel_format(AVCodecContext *avctx) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    const AV1RawSequenceHeader *seq = &s->raw_seq;
> > +    uint8_t bit_depth;
> > +    int ret = 0;
> > +    enum AVPixelFormat pix_fmt = AV_PIX_FMT_NONE; #define
> HWACCEL_MAX
> > +(0)
> > +    enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmtp = pix_fmts;
> > +
> > +    if (seq->seq_profile == 2 && seq->color_config.high_bitdepth)
> > +        bit_depth = seq->color_config.twelve_bit ? 12 : 10;
> > +    else if (seq->seq_profile <= 2)
> > +        bit_depth = seq->color_config.high_bitdepth ? 10 : 8;
> > +    else {
> > +        av_log(avctx, AV_LOG_ERROR,
> > +               "Unknow av1 profile %d.\n", seq->seq_profile);
> > +        return -1;
> > +    }
> > +
> > +    if (!seq->color_config.mono_chrome) {
> > +        // 4:4:4 x:0 y:0, 4:2:2 x:1 y:0, 4:2:0 x:1 y:1
> > +        if (seq->color_config.subsampling_x == 0 &&
> > +            seq->color_config.subsampling_y == 0) {
> > +            if (bit_depth == 8)
> > +                pix_fmt = AV_PIX_FMT_YUV444P;
> > +            else if (bit_depth == 10)
> > +                pix_fmt = AV_PIX_FMT_YUV444P10;
> > +            else if (bit_depth == 12)
> > +                pix_fmt = AV_PIX_FMT_YUV444P12;
> > +            else
> > +                av_log(avctx, AV_LOG_WARNING, "Unknow av1 pixel format.\n");
> > +        } else if (seq->color_config.subsampling_x == 1 &&
> > +                   seq->color_config.subsampling_y == 0) {
> > +            if (bit_depth == 8)
> > +                pix_fmt = AV_PIX_FMT_YUV422P;
> > +            else if (bit_depth == 10)
> > +                pix_fmt = AV_PIX_FMT_YUV422P10;
> > +            else if (bit_depth == 12)
> > +                pix_fmt = AV_PIX_FMT_YUV422P12;
> > +            else
> > +                av_log(avctx, AV_LOG_WARNING, "Unknow av1 pixel format.\n");
> > +        } else if (seq->color_config.subsampling_x == 1 &&
> > +                   seq->color_config.subsampling_y == 1) {
> > +            if (bit_depth == 8)
> > +                pix_fmt = AV_PIX_FMT_YUV420P;
> > +            else if (bit_depth == 10)
> > +                pix_fmt = AV_PIX_FMT_YUV420P10;
> > +            else if (bit_depth == 12)
> > +                pix_fmt = AV_PIX_FMT_YUV420P12;
> > +            else
> > +                av_log(avctx, AV_LOG_WARNING, "Unknow av1 pixel format.\n");
> > +        }
> > +    } else {
> > +        if (seq->color_config.subsampling_x == 1 &&
> > +            seq->color_config.subsampling_y == 1)
> > +            pix_fmt = AV_PIX_FMT_YUV440P;
> > +        else
> > +            av_log(avctx, AV_LOG_WARNING, "Unknow av1 pixel format.\n");
> > +    }
> > +
> > +    av_log(avctx, AV_LOG_DEBUG, "Av1 decode get format: %s.\n",
> > +           av_get_pix_fmt_name(pix_fmt));
> > +
> > +    if (pix_fmt == AV_PIX_FMT_NONE)
> > +        return -1;
> > +    s->pix_fmt = pix_fmt;
> > +
> > +    *fmtp++ = s->pix_fmt;
> > +    *fmtp = AV_PIX_FMT_NONE;
> > +    ret = ff_thread_get_format(avctx, pix_fmts);
> > +    if (ret < 0)
> > +        return ret;
> > +
> > +    avctx->pix_fmt = ret;
> 
> s->pix_fmt and avctx->pix_fmt may not be in sync after this.
> 
> Do you really need the former at all?

s->pix_fmt can be regard as a reference fmt which av1 decoder want to
decode current stream to. But the HW may not be decode the stream
to the fmt same with reference fmt. So ff_thread_get_format will query
HW and compare with the reference fmt and pick up a best one that HW
Can provided. avctx->pix_fmt will be the actually the HW will decode to. So
they can be different. 

> 
> > +
> > +    return 0;
> > +}
> > +
> > +static void av1_frame_unref(AVCodecContext *avctx, AV1Frame *f) {
> > +    ff_thread_release_buffer(avctx, &f->tf);
> > +    av_buffer_unref(&f->hwaccel_priv_buf);
> > +    f->hwaccel_picture_private = NULL; }
> > +
> > +static int av1_frame_ref(AVCodecContext *avctx, AV1Frame *dst,
> > +AV1Frame *src) {
> > +    int ret = 0;
> > +
> > +    ret = ff_thread_ref_frame(&dst->tf, &src->tf);
> > +    if (ret < 0)
> > +        return ret;
> > +
> > +    if (src->hwaccel_picture_private) {
> > +        dst->hwaccel_priv_buf = av_buffer_ref(src->hwaccel_priv_buf);
> > +        if (!dst->hwaccel_priv_buf)
> > +            goto fail;
> > +        dst->hwaccel_picture_private = dst->hwaccel_priv_buf->data;
> > +    }
> > +
> > +    dst->loop_filter_delta_enabled = src->loop_filter_delta_enabled;
> > +    memcpy(dst->loop_filter_ref_deltas,
> > +           src->loop_filter_ref_deltas,
> > +           AV1_TOTAL_REFS_PER_FRAME * sizeof(int8_t));
> > +    memcpy(dst->loop_filter_mode_deltas,
> > +           src->loop_filter_mode_deltas,
> > +           2 * sizeof(int8_t));
> > +    memcpy(dst->gm_type,
> > +           src->gm_type,
> > +           AV1_TOTAL_REFS_PER_FRAME * sizeof(uint8_t));
> > +    memcpy(dst->gm_params,
> > +           src->gm_params,
> > +           AV1_TOTAL_REFS_PER_FRAME * 6 * sizeof(int8_t));
> > +
> > +    return 0;
> > +
> > +fail:
> > +    av1_frame_unref(avctx, dst);
> > +    return AVERROR(ENOMEM);
> > +}
> > +
> > +static av_cold int av1_decode_free(AVCodecContext *avctx) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    int i;
> > +
> > +    for (i = 0; i < FF_ARRAY_ELEMS(s->ref); i++) {
> > +        if (s->ref[i].tf.f->buf[0])
> > +            av1_frame_unref(avctx, &s->ref[i]);
> > +        av_frame_free(&s->ref[i].tf.f);
> > +    }
> > +    if (s->cur_frame.tf.f->buf[0])
> > +        av1_frame_unref(avctx, &s->cur_frame);
> > +    av_frame_free(&s->cur_frame.tf.f);
> > +
> > +    ff_cbs_fragment_free(&s->current_obu);
> > +    ff_cbs_close(&s->cbc);
> > +
> > +    return 0;
> > +}
> > +
> > +static av_cold int av1_decode_init(AVCodecContext *avctx) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    int i;
> > +    int ret = 0;
> > +
> > +    for (i = 0; i < FF_ARRAY_ELEMS(s->ref); i++) {
> > +        s->ref[i].tf.f = av_frame_alloc();
> > +        if (!s->ref[i].tf.f) {
> > +            av1_decode_free(avctx);
> > +            av_log(avctx, AV_LOG_ERROR,
> > +                   "Fail to allocate frame buffer %d.\n", i);
> > +            return AVERROR(ENOMEM);
> > +        }
> > +    }
> > +
> > +        s->cur_frame.tf.f = av_frame_alloc();
> > +        if (!s->cur_frame.tf.f) {
> > +            av1_decode_free(avctx);
> > +            av_log(avctx, AV_LOG_ERROR, "Fail to allocate frame buffer.\n");
> > +            return AVERROR(ENOMEM);
> > +        }
> 
> Wrong indentation.


Thanks, I will fix this in my next patch version.

> 
> > +
> > +    s->avctx = avctx;
> > +    s->pix_fmt = AV_PIX_FMT_NONE;
> > +
> > +    ret = ff_cbs_init(&s->cbc, AV_CODEC_ID_AV1, avctx);
> > +    if (ret < 0)
> > +        return ret;
> > +    return 0;
> > +}
> > +
> > +static int av1_frame_alloc(AVCodecContext *avctx, AV1Frame *f) {
> > +    int ret;
> > +    if ((ret = ff_thread_get_buffer(avctx, &f->tf, AV_GET_BUFFER_FLAG_REF))
> < 0)
> > +        return ret;
> > +
> > +    if (avctx->hwaccel) {
> > +        const AVHWAccel *hwaccel = avctx->hwaccel;
> > +        if (hwaccel->frame_priv_data_size) {
> > +            f->hwaccel_priv_buf =
> > +                av_buffer_allocz(hwaccel->frame_priv_data_size);
> > +            if (!f->hwaccel_priv_buf)
> > +                goto fail;
> > +            f->hwaccel_picture_private = f->hwaccel_priv_buf->data;
> > +        }
> > +    }
> > +    return 0;
> > +
> > +fail:
> > +    av1_frame_unref(avctx, f);
> > +    return AVERROR(ENOMEM);
> > +}
> > +
> > +static int set_output_frame(AVFrame *srcframe, AVFrame *frame,
> > +AVPacket *pkt) {
> > +    int ret = 0;
> > +    if ((ret = av_frame_ref(frame, srcframe)) < 0) {
> > +        return ret;
> > +    }
> > +
> > +    frame->pts = pkt->pts;
> > +    frame->pkt_dts = pkt->dts;
> 
> Did you add this because you saw bogus timestamps filled by
> ff_thread_get_buffer? If so, could you try applying
> https://patchwork.ffmpeg.org/project/ffmpeg/list/?series=2057 and removing
> this chunk, to see if it fixes it?
Will add FF_CODEC_CAP_SETS_PKT_DTS to caps_internal as you mentioned in
another mail thread. Thanks.

> 
> > +
> > +    return 0;
> > +}
> > +
> > +static int update_reference_list (AVCodecContext *avctx) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    AV1RawFrameHeader *header= &s->raw_frame_header;
> > +    int i;
> > +    int ret = 0;
> > +
> > +    for (i = 0; i < 8; i++) {
> > +        if (header->frame_type == AV1_FRAME_KEY ||
> > +            (header->refresh_frame_flags & (1 << i))) {
> > +            if (s->ref[i].tf.f->buf[0])
> > +                av1_frame_unref(avctx, &s->ref[i]);
> > +            if ((ret = av1_frame_ref(avctx, &s->ref[i], &s->cur_frame)) < 0) {
> > +                av_log(avctx, AV_LOG_DEBUG, "Ref frame error:%d.\n", ret);
> > +                return ret;
> > +            }
> > +        }
> > +    }
> > +    return 0;
> > +}
> > +
> > +static int get_current_frame(AVCodecContext *avctx) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    int ret = 0;
> > +
> > +    if (s->cur_frame.tf.f->buf[0])
> > +        av1_frame_unref(avctx, &s->cur_frame);
> > +
> > +    ret = av1_frame_alloc(avctx, &s->cur_frame);
> > +    if (ret < 0) {
> > +        av_log(avctx, AV_LOG_ERROR, "Get current frame fail.\n");
> > +        return ret;
> > +    }
> > +
> > +    if (s->raw_frame_header.primary_ref_frame == AV1_PRIMARY_REF_NONE)
> > +        setup_past_independence(&s->cur_frame);
> > +    else
> > +        load_previous_and_update(s);
> > +
> > +    global_motion_params(s);
> > +
> > +    return ret;
> > +}
> > +
> > +static int decode_current_frame(AVCodecContext *avctx) {
> > +    int ret = 0;
> > +
> > +    if (avctx->hwaccel) {
> > +        ret = avctx->hwaccel->start_frame(avctx, NULL, 0);
> 
> You should pass this a buffer. Some hwaccel implementation may need it.

Thanks, I will modify this in my next patch version.

> 
> > +        if (ret < 0) {
> > +            av_log(avctx, AV_LOG_ERROR, "HW accel strat frame fail.\n");
> > +            return ret;
> > +        }
> > +
> > +        ret = avctx->hwaccel->end_frame(avctx);
> 
> Shouldn't hwaccel->decode_slice() be called before this, and not after?

Actually decode_slice() will be called first, then start_frame() and end_frame.
For vaapi driver, it doesn't care the order of sequence info(filled in start_frame)
and slice info(filled in decode_slice). Consider other HW codecs, I will modify
it to general order like start_frame, decode_slice, then end_frame. Thanks.

> 
> > +        if (ret < 0) {
> > +            av_log(avctx, AV_LOG_ERROR, "HW accel end frame fail.\n");
> > +            return ret;
> > +        }
> > +    } else {
> > +        av_log(avctx, AV_LOG_ERROR, "Your platform doesn't suppport
> hardware "
> > +               "acceleration AV1 decode.\n");
> > +        return -1;
> > +    }
> > +
> > +    ret = update_reference_list(avctx);
> > +    if (ret < 0) {
> > +        av_log(avctx, AV_LOG_ERROR, "Fail to update reference list.\n");
> > +        return ret;
> > +    }
> > +    return 0;
> > +}
> > +
> > +static int av1_decode_frame(AVCodecContext *avctx, void *frame,
> > +                            int *got_frame, AVPacket *pkt) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    AV1RawSequenceHeader* raw_seq = NULL;
> > +    AV1RawFrameHeader *raw_frame_header = NULL;
> > +    AV1RawTileGroup *raw_tile_group = NULL;
> > +    uint32_t tile_group_offset = 0;
> 
> Should be ptrdiff_t.

Thanks, I will fix this in my next patch version.

> 
> > +    int i;
> 
> Just do for (int i...) below.

OK, I will modify this. Thanks.

> 
> > +    int ret = 0;
> > +
> > +    ret = ff_cbs_read_packet(s->cbc, &s->current_obu, pkt);
> > +    if (ret < 0) {
> > +        av_log(avctx, AV_LOG_ERROR, "Fail to read packet.\n");
> > +        goto end;
> > +    }
> > +    av_log(avctx, AV_LOG_DEBUG, "Total obu for this frame:%d.\n",
> > +           s->current_obu.nb_units);
> > +
> > +    for (i = 0; i < s->current_obu.nb_units; i++) {
> > +        CodedBitstreamUnit *unit = &s->current_obu.units[i];
> > +        AV1RawOBU *obu = unit->content;
> > +        av_log(avctx, AV_LOG_DEBUG, "Obu idx:%d, obu type:%d.\n", i,
> > + unit->type);
> > +
> > +        switch (unit->type) {
> > +        case AV1_OBU_SEQUENCE_HEADER:
> > +            raw_seq = &obu->obu.sequence_header;
> > +            memcpy(&s->raw_seq, raw_seq,
> > + sizeof(AV1RawSequenceHeader));
> 
> You can use
> (CodedBitstreamAV1Context*)(s->cbc->priv_data)->sequence_header
> for this instead. And if you'd rather not, you can always create a new reference
> to unit->content_ref and store it in AV1DecContext.

Create a new reference will be better and store them in AV1DecContext. VAAPI
need both sequence header and frame header in hwaccel->start_frame. 

> 
> Also, consider adding a avctx->hwaccel->decode_params() call here as well, in
> case some future hwaccel decides/requires to implement it.

OK, will add it in my next patch version. Thanks.

> 
> > +            if (s->pix_fmt == AV_PIX_FMT_NONE) {
> > +                ret = get_pixel_format(avctx);
> > +                if (ret < 0) {
> > +                    av_log(avctx, AV_LOG_ERROR,
> > +                           "Fail to get format from sequence header.\n");
> > +                    goto end;
> > +                }
> > +            }
> > +            break;
> > +        case AV1_OBU_FRAME_HEADER:
> > +        case AV1_OBU_REDUNDANT_FRAME_HEADER:
> > +            raw_frame_header = &obu->obu.frame_header;
> > +            memcpy (&s->raw_frame_header,
> > +                    raw_frame_header,
> > +                    sizeof(AV1RawFrameHeader));
> 
> Same as above, you can create a new reference to unit->content_ref instead.
> memcpy should be avoided as much as possible, especially in this case which is
> run on every frame.

Will modify this, Thanks.

> 
> > +            s->tile_num =
> > +                raw_frame_header->tile_cols * raw_frame_header->tile_rows;
> > +            if (raw_frame_header->show_existing_frame) {
> > +                if (set_output_frame(s->ref[raw_frame_header-
> >frame_to_show_map_idx].tf.f,
> > +                                     frame, pkt) < 0) {
> > +                    av_log(avctx, AV_LOG_DEBUG, "Set output frame error:%d.\n",
> > +                           ret);
> > +                    goto end;
> > +                }
> > +                *got_frame = 1;
> > +                goto end;
> > +            } else {
> > +                ret = get_current_frame(avctx);
> > +                if (ret < 0) {
> > +                    av_log(avctx, AV_LOG_DEBUG, "Get current frame error:%d.\n",
> > +                           ret);
> > +                    goto end;
> > +                }
> > +            }
> > +            break;
> > +        case AV1_OBU_FRAME:
> > +            raw_frame_header = &obu->obu.frame.header;
> > +            raw_tile_group = &obu->obu.frame.tile_group;
> > +            memcpy(&s->raw_frame_header,
> > +                   raw_frame_header,
> > +                   sizeof(AV1RawFrameHeader));
> 
> Ditto.

Will modify this, Thanks.

> 
> > +            s->tile_num =
> > +                raw_frame_header->tile_cols * raw_frame_header->tile_rows;
> > +            tile_group_offset = raw_tile_group->tile_data.data - pkt->data;
> > +            get_tiles_info(avctx, raw_tile_group, tile_group_offset);
> > +
> > +            ret = get_current_frame(avctx);
> > +            if (ret < 0) {
> > +                av_log(avctx, AV_LOG_DEBUG, "Get current frame error:%d.\n",
> > +                       ret);
> > +                goto end;
> > +            }
> > +            if (avctx->hwaccel) {
> > +                ret = avctx->hwaccel->decode_slice(avctx, pkt->data,
> > + pkt->size);
> 
> Can't you pass raw_tile_group->tile_data.data or something less broad than the
> entire packet?

Will try to modify this in my next patch version.

> 
> > +                if (ret < 0) {
> > +                    av_log(avctx, AV_LOG_ERROR,
> > +                           "HW accel decode slice fail.\n");
> > +                    return ret;
> > +                }
> > +            }
> > +            ret = decode_current_frame(avctx);
> > +            if (ret < 0) {
> > +                av_log(avctx, AV_LOG_DEBUG, "Decode current frame error:%d.\n",
> > +                       ret);
> > +                goto end;
> > +            }
> > +            if (raw_frame_header->show_frame) {
> > +                if (set_output_frame(s->cur_frame.tf.f, frame, pkt) < 0) {
> > +                    av_log(avctx, AV_LOG_DEBUG, "Set output frame error:%d.\n",
> > +                           ret);
> > +                    goto end;
> > +                }
> > +
> > +                *got_frame = 1;
> > +            }
> > +            break;
> > +        case AV1_OBU_TILE_GROUP:
> 
> You can merge this and the frame/frame_header cases above, letting it fall
> through at the point it would start reading the AV1_OBU_TILE_GROUP bits in
> the former case.

Will try to modify this. Thanks.

> 
> > +            raw_tile_group = &obu->obu.tile_group;
> > +            tile_group_offset = raw_tile_group->tile_data.data - pkt->data;
> > +            get_tiles_info(avctx, raw_tile_group, tile_group_offset);
> > +
> > +            if (avctx->hwaccel) {
> > +                ret = avctx->hwaccel->decode_slice(avctx, pkt->data, pkt->size);
> > +                if (ret < 0) {
> > +                    av_log(avctx, AV_LOG_ERROR,
> > +                           "HW accel decode slice fail.\n");
> > +                    return ret;
> > +                }
> > +            }
> > +            if (s->tile_num != raw_tile_group->tg_end + 1)
> > +                break;
> > +            ret = decode_current_frame(avctx);
> > +            if (ret < 0) {
> > +                av_log(avctx, AV_LOG_DEBUG, "Decode current frame error:%d.\n",
> > +                       ret);
> > +                goto end;
> > +            }
> > +
> > +            if (raw_frame_header->show_frame) {
> > +                if (set_output_frame(s->cur_frame.tf.f, frame, pkt) < 0) {
> > +                    av_log(avctx, AV_LOG_DEBUG, "Set output frame error:%d.\n",
> > +                           ret);
> > +                    goto end;
> > +                }
> > +
> > +                *got_frame = 1;
> > +            }
> > +            break;
> > +        case AV1_OBU_TILE_LIST:
> > +        case AV1_OBU_TEMPORAL_DELIMITER:
> > +        case AV1_OBU_PADDING:
> > +        case AV1_OBU_METADATA:
> > +            break;
> > +        default:
> > +            av_log(avctx, AV_LOG_DEBUG,
> > +                   "Un-implemented obu type: %d (%zd bits).\n",
> 
> Unknown, rather than unimplemented.

Will try to modify this in my next patch version. Thanks.

> 
> > +                   unit->type, unit->data_size);
> > +        }
> > +    }
> > +
> > +end:
> > +    ff_cbs_fragment_reset(&s->current_obu);
> > +    return ret;
> > +}
> > +
> > +static void av1_decode_flush(AVCodecContext *avctx) {
> > +    AV1DecContext *s = avctx->priv_data;
> > +    int i;
> > +
> > +    for (i = 0; i < FF_ARRAY_ELEMS(s->ref); i++)
> > +        av1_frame_unref(avctx, &s->ref[i]);
> > +
> > +    av1_frame_unref(avctx, &s->cur_frame); }
> > +
> > +AVCodec ff_av1_decoder = {
> > +    .name                  = "av1",
> > +    .long_name             = NULL_IF_CONFIG_SMALL("Hardware Acceleration
> AV1"),
> 
> "Alliance for Open Media AV1" or something like that. The fact the decoder is
> currently hwaccel backed only is irrelevant to its name.

Actually I want to make it obviously to let user know this av1 decoder is just for
hwaccel by using cmd like "ffmpeg --help decoder=av1". If you think we can ignore
this, I will change it. Thanks.

> 
> > +    .type                  = AVMEDIA_TYPE_VIDEO,
> > +    .id                    = AV_CODEC_ID_AV1,
> > +    .priv_data_size        = sizeof(AV1DecContext),
> > +    .init                  = av1_decode_init,
> > +    .close                 = av1_decode_free,
> > +    .decode                = av1_decode_frame,
> > +    .capabilities          = AV_CODEC_CAP_DR1,
> 
> Probably should add caps_internal like FF_CODEC_CAP_INIT_THREADSAFE and
> FF_CODEC_CAP_INIT_CLEANUP.

Will add it in my next patch version. Thanks.

> 
> > +    .flush                 = av1_decode_flush,
> > +    .profiles              = NULL_IF_CONFIG_SMALL(ff_av1_profiles),
> > +    .hw_configs            = (const AVCodecHWConfigInternal * []) {
> > +        NULL
> > +    },
> > +};
> > diff --git a/libavcodec/av1dec.h b/libavcodec/av1dec.h new file mode
> > 100644 index 0000000000..5cb0e896b7
> > --- /dev/null
> > +++ b/libavcodec/av1dec.h
> > @@ -0,0 +1,89 @@
> > +/*
> > + * AV1 video decoder
> > + * *
> > + * This file is part of FFmpeg.
> > + *
> > + * FFmpeg is free software; you can redistribute it and/or
> > + * modify it under the terms of the GNU Lesser General Public
> > + * License as published by the Free Software Foundation; either
> > + * version 2.1 of the License, or (at your option) any later version.
> > + *
> > + * FFmpeg is distributed in the hope that it will be useful,
> > + * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
> > + * Lesser General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU Lesser General Public
> > + * License along with FFmpeg; if not, write to the Free Software
> > + * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
> > +02110-1301 USA  */
> > +
> > +#ifndef AVCODEC_AV1DEC_H
> > +#define AVCODEC_AV1DEC_H
> > +
> > +#include <stdatomic.h>
> > +#include "libavutil/buffer.h"
> > +#include "libavutil/md5.h"
> > +#include "avcodec.h"
> > +#include "bswapdsp.h"
> > +#include "get_bits.h"
> > +#include "av1_parse.h"
> > +#include "libavformat/avformat.h"
> > +#include "libavformat/av1.h"
> > +#include "thread.h"
> > +#include "cbs.h"
> > +#include "cbs_av1.h"
> > +
> > +typedef enum AV1FrameRestorationType {
> > +    AV1_RESTORE_NONE,
> > +    AV1_RESTORE_WIENER,
> > +    AV1_RESTORE_SGRPROJ,
> > +    AV1_RESTORE_SWITCHABLE,
> > +} AV1FrameRestorationType;
> 
> Should be added to av1.h instead.
> 
> > +
> > +typedef struct AV1Frame {
> > +    ThreadFrame tf;
> > +
> > +    AVBufferRef *hwaccel_priv_buf;
> > +    void *hwaccel_picture_private;
> > +
> > +    uint8_t loop_filter_delta_enabled;
> > +    int8_t  loop_filter_ref_deltas[AV1_TOTAL_REFS_PER_FRAME];
> > +    int8_t  loop_filter_mode_deltas[2];
> > +    uint8_t gm_type[AV1_TOTAL_REFS_PER_FRAME];
> > +    int32_t gm_params[AV1_TOTAL_REFS_PER_FRAME][6];
> > +} AV1Frame;
> > +
> > +typedef struct TileGroupInfo {
> > +    uint32_t tile_group_offset;
> > +    uint32_t tile_offset;
> > +    uint32_t tile_size;
> > +    uint16_t tile_row;
> > +    uint16_t tile_column;
> > +    uint16_t tile_width_sbs;
> > +    uint16_t tile_height_sbs;
> > +    uint32_t tile_size_bytes;
> > +} TileGroupInfo;
> > +
> > +typedef struct AV1DecContext {
> > +    const AVClass *class;
> > +    AVCodecContext *avctx;
> > +
> > +    enum AVPixelFormat pix_fmt;
> > +    CodedBitstreamContext *cbc;
> > +    CodedBitstreamFragment current_obu;
> > +
> > +    AV1RawSequenceHeader raw_seq;
> > +    AV1RawFrameHeader raw_frame_header;
> > +    AV1RawTileGroup raw_tile_group;
> > +    TileGroupInfo tile_group_info[128];
> > +    uint16_t tile_num;
> > +    uint16_t tg_start;
> > +    uint16_t tg_end;
> > +
> > +    AV1Frame ref[8];
> > +    AV1Frame cur_frame;
> > +
> > +} AV1DecContext;
> > +
> > +#endif /* AVCODEC_AV1DEC_H */
> > diff --git a/libavcodec/version.h b/libavcodec/version.h index
> > a3f9f828ee..5bdfdce363 100644
> > --- a/libavcodec/version.h
> > +++ b/libavcodec/version.h
> > @@ -28,7 +28,7 @@
> >  #include "libavutil/version.h"
> >
> >  #define LIBAVCODEC_VERSION_MAJOR  58
> > -#define LIBAVCODEC_VERSION_MINOR 100
> > +#define LIBAVCODEC_VERSION_MINOR 101
> >  #define LIBAVCODEC_VERSION_MICRO 100
> >
> >  #define LIBAVCODEC_VERSION_INT
> > AV_VERSION_INT(LIBAVCODEC_VERSION_MAJOR, \
> >
> 
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> 
> To unsubscribe, visit link above, or email ffmpeg-devel-request at ffmpeg.org
> with subject "unsubscribe".


More information about the ffmpeg-devel mailing list