Re: [FFmpeg-devel] [PATCH 11/11] avcodec/vvcdec: move frame tab memset from the main thread to worker threads

[Nuo Mi <nuomi2021@gmail.com>]

On Sun, Jul 28, 2024 at 11:19 AM Nuo Mi <nuomi2021@gmail.com> wrote:

> memset tables in the main thread can become a bottleneck for the decoder.
> For example, if it takes 1% of the processing time for one core, the
> maximum achievable FPS will be 100.
> Move the memeset to worker threads will fix the issue.
>
will apply next week if there are no objections

> ---
>  libavcodec/vvc/dec.c    |  13 ++++-
>  libavcodec/vvc/thread.c | 122 ++++++++++++++++++++++++----------------
>  libavcodec/vvc/thread.h |   1 +
>  3 files changed, 85 insertions(+), 51 deletions(-)
>
> diff --git a/libavcodec/vvc/dec.c b/libavcodec/vvc/dec.c
> index 575bcfa33d..d34713296d 100644
> --- a/libavcodec/vvc/dec.c
> +++ b/libavcodec/vvc/dec.c
> @@ -82,7 +82,13 @@ static int tl_create(TabList *l)
>              if (!*t->tab)
>                  return AVERROR(ENOMEM);
>          }
> -    } else if (l->zero) {
> +    }
> +    return 0;
> +}
> +
> +static int tl_zero(TabList *l)
> +{
> +    if (l->zero) {
>          for (int i = 0; i < l->nb_tabs; i++) {
>              Tab *t = l->tabs + i;
>              memset(*t->tab, 0, t->size);
> @@ -404,6 +410,11 @@ static int pic_arrays_init(VVCContext *s,
> VVCFrameContext *fc)
>      return 0;
>  }
>
> +int ff_vvc_per_frame_init(VVCFrameContext *fc)
> +{
> +    return frame_context_for_each_tl(fc, tl_zero);
> +}
> +
>  static int min_positive(const int idx, const int diff, const int min_diff)
>  {
>      return diff > 0 && (idx < 0 || diff < min_diff);
> diff --git a/libavcodec/vvc/thread.c b/libavcodec/vvc/thread.c
> index 28065d726f..74f8e4e9d0 100644
> --- a/libavcodec/vvc/thread.c
> +++ b/libavcodec/vvc/thread.c
> @@ -40,6 +40,7 @@ typedef struct ProgressListener {
>  } ProgressListener;
>
>  typedef enum VVCTaskStage {
> +    VVC_TASK_STAGE_INIT,                    // for CTU(0, 0) only
>      VVC_TASK_STAGE_PARSE,
>      VVC_TASK_STAGE_INTER,
>      VVC_TASK_STAGE_RECON,
> @@ -175,10 +176,14 @@ static int task_has_target_score(VVCTask *t, const
> VVCTaskStage stage, const uin
>      uint8_t target = 0;
>      VVCFrameContext *fc = t->fc;
>
> +    if (stage == VVC_TASK_STAGE_INIT)
> +        return 1;
> +
>      if (stage == VVC_TASK_STAGE_PARSE) {
> -        const H266RawSPS *rsps = fc->ps.sps->r;
> -        const int wpp = rsps->sps_entropy_coding_sync_enabled_flag &&
> !is_first_row(fc, t->rx, t->ry);
> -        target = 2 + wpp - 1;                           //left parse +
> colocation + wpp - no previous stage
> +        const H266RawSPS *rsps   = fc->ps.sps->r;
> +        const int wpp            =
> rsps->sps_entropy_coding_sync_enabled_flag && !is_first_row(fc, t->rx,
> t->ry);
> +        const int no_prev_stage  = t->rs > 0;
> +        target = 2 + wpp - no_prev_stage;
>  //left parse + colocation + wpp - no_prev_stage
>      } else if (stage == VVC_TASK_STAGE_INTER) {
>          target = atomic_load(&t->target_inter_score);
>      } else {
> @@ -399,6 +404,55 @@ static int task_priority_higher(const AVTask *_a,
> const AVTask *_b)
>      return a->ry < b->ry;
>  }
>
> +static void check_colocation(VVCContext *s, VVCTask *t)
> +{
> +    const VVCFrameContext *fc = t->fc;
> +
> +    if (fc->ps.ph.r->ph_temporal_mvp_enabled_flag ||
> fc->ps.sps->r->sps_sbtmvp_enabled_flag) {
> +        VVCFrame *col       = fc->ref->collocated_ref;
> +        const int first_col = t->rx == fc->ps.pps->ctb_to_col_bd[t->rx];
> +        if (col && first_col) {
> +            //we depend on bottom and right boundary, do not - 1 for y
> +            const int y = (t->ry << fc->ps.sps->ctb_log2_size_y);
> +            add_progress_listener(col, &t->col_listener, t, s,
> VVC_PROGRESS_MV, y);
> +            return;
> +        }
> +    }
> +    frame_thread_add_score(s, fc->ft, t->rx, t->ry, VVC_TASK_STAGE_PARSE);
> +}
> +
> +static void submit_entry_point(VVCContext *s, VVCFrameThread *ft,
> SliceContext *sc, EntryPoint *ep)
> +{
> +    const int rs = sc->sh.ctb_addr_in_curr_slice[ep->ctu_start];
> +    VVCTask *t   = ft->tasks + rs;
> +
> +    frame_thread_add_score(s, ft, t->rx, t->ry, VVC_TASK_STAGE_PARSE);
> +}
> +
> +static int run_init(VVCContext *s, VVCLocalContext *lc, VVCTask *t)
> +{
> +    VVCFrameContext *fc = lc->fc;
> +    VVCFrameThread *ft  = fc->ft;
> +    const int ret       = ff_vvc_per_frame_init(fc);
> +
> +    if (ret < 0)
> +        return ret;
> +
> +    for (int i = 0; i < fc->nb_slices; i++) {
> +        SliceContext *sc = fc->slices[i];
> +        for (int j = 0; j < sc->nb_eps; j++) {
> +            EntryPoint *ep = sc->eps + j;
> +            for (int k = ep->ctu_start; k < ep->ctu_end; k++) {
> +                const int rs = sc->sh.ctb_addr_in_curr_slice[k];
> +                VVCTask *t   = ft->tasks + rs;
> +                check_colocation(s, t);
> +            }
> +            submit_entry_point(s, ft, sc, ep);
> +        }
> +    }
> +    return 0;
> +}
> +
>  static void report_frame_progress(VVCFrameContext *fc,
>     const int ry, const VVCProgress idx)
>  {
> @@ -547,6 +601,7 @@ static int run_alf(VVCContext *s, VVCLocalContext *lc,
> VVCTask *t)
>  #define VVC_THREAD_DEBUG
>  #ifdef VVC_THREAD_DEBUG
>  const static char* task_name[] = {
> +    "INIT",
>      "P",
>      "I",
>      "R",
> @@ -567,6 +622,7 @@ static void task_run_stage(VVCTask *t, VVCContext *s,
> VVCLocalContext *lc)
>      VVCFrameThread *ft       = fc->ft;
>      const VVCTaskStage stage = t->stage;
>      static const run_func run[] = {
> +        run_init,
>          run_parse,
>          run_inter,
>          run_recon,
> @@ -726,7 +782,7 @@ int ff_vvc_frame_thread_init(VVCFrameContext *fc)
>
>      for (int rs = 0; rs < ft->ctu_count; rs++) {
>          VVCTask *t = ft->tasks + rs;
> -        task_init(t, VVC_TASK_STAGE_PARSE, fc, rs % ft->ctu_width, rs /
> ft->ctu_width);
> +        task_init(t, rs ? VVC_TASK_STAGE_PARSE : VVC_TASK_STAGE_INIT, fc,
> rs % ft->ctu_width, rs / ft->ctu_width);
>      }
>
>      memset(&ft->row_progress[0], 0, sizeof(ft->row_progress));
> @@ -745,59 +801,25 @@ fail:
>      return AVERROR(ENOMEM);
>  }
>
> -static void check_colocation(VVCContext *s, VVCTask *t)
> -{
> -    const VVCFrameContext *fc = t->fc;
> -
> -    if (fc->ps.ph.r->ph_temporal_mvp_enabled_flag ||
> fc->ps.sps->r->sps_sbtmvp_enabled_flag) {
> -        VVCFrame *col       = fc->ref->collocated_ref;
> -        const int first_col = t->rx == fc->ps.pps->ctb_to_col_bd[t->rx];
> -        if (col && first_col) {
> -            //we depend on bottom and right boundary, do not - 1 for y
> -            const int y = (t->ry << fc->ps.sps->ctb_log2_size_y);
> -            add_progress_listener(col, &t->col_listener, t, s,
> VVC_PROGRESS_MV, y);
> -            return;
> -        }
> -    }
> -    frame_thread_add_score(s, fc->ft, t->rx, t->ry, VVC_TASK_STAGE_PARSE);
> -}
> -
> -static void submit_entry_point(VVCContext *s, VVCFrameThread *ft,
> SliceContext *sc, EntryPoint *ep)
> -{
> -    const int rs = sc->sh.ctb_addr_in_curr_slice[ep->ctu_start];
> -    VVCTask *t   = ft->tasks + rs;
> -
> -    frame_thread_add_score(s, ft, t->rx, t->ry, VVC_TASK_STAGE_PARSE);
> -}
> -
>  int ff_vvc_frame_submit(VVCContext *s, VVCFrameContext *fc)
>  {
>      VVCFrameThread *ft = fc->ft;
>
> -    // We'll handle this in two passes:
> -    // Pass 0 to initialize tasks with parser, this will help detect bit
> stream error
> -    // Pass 1 to shedule location check and submit the entry point
> -    for (int pass = 0; pass < 2; pass++) {
> -        for (int i = 0; i < fc->nb_slices; i++) {
> -            SliceContext *sc = fc->slices[i];
> -            for (int j = 0; j < sc->nb_eps; j++) {
> -                EntryPoint *ep = sc->eps + j;
> -                for (int k = ep->ctu_start; k < ep->ctu_end; k++) {
> -                    const int rs = sc->sh.ctb_addr_in_curr_slice[k];
> -                    VVCTask *t   = ft->tasks + rs;
> -                    if (pass) {
> -                        check_colocation(s, t);
> -                    } else {
> -                        const int ret = task_init_parse(t, sc, ep, k);
> -                        if (ret < 0)
> -                            return ret;
> -                    }
> -                }
> -                if (pass)
> -                    submit_entry_point(s, ft, sc, ep);
> +    for (int i = 0; i < fc->nb_slices; i++) {
> +        SliceContext *sc = fc->slices[i];
> +        for (int j = 0; j < sc->nb_eps; j++) {
> +            EntryPoint *ep = sc->eps + j;
> +            for (int k = ep->ctu_start; k < ep->ctu_end; k++) {
> +                const int rs = sc->sh.ctb_addr_in_curr_slice[k];
> +                VVCTask *t   = ft->tasks + rs;
> +                const int ret = task_init_parse(t, sc, ep, k);
> +                if (ret < 0)
> +                    return ret;
>              }
>          }
>      }
> +    frame_thread_add_score(s, ft, 0, 0, VVC_TASK_STAGE_INIT);
> +
>      return 0;
>  }
>
> diff --git a/libavcodec/vvc/thread.h b/libavcodec/vvc/thread.h
> index 8ac59b2ecf..7b15dbee59 100644
> --- a/libavcodec/vvc/thread.h
> +++ b/libavcodec/vvc/thread.h
> @@ -32,5 +32,6 @@ int ff_vvc_frame_thread_init(VVCFrameContext *fc);
>  void ff_vvc_frame_thread_free(VVCFrameContext *fc);
>  int ff_vvc_frame_submit(VVCContext *s, VVCFrameContext *fc);
>  int ff_vvc_frame_wait(VVCContext *s, VVCFrameContext *fc);
> +int ff_vvc_per_frame_init(VVCFrameContext *fc);
>
>  #endif // AVCODEC_VVC_THREAD_H
> --
> 2.34.1
>
>
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel

To unsubscribe, visit link above, or email
ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".