ffv1dec: Support frame threading with gop > 1

This is about 20-30% faster than slice threading
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>

libavcodec/ffv1.h

@@ -91,6 +91,7 @@ typedef struct FFV1Context {
     int flags;
     int picture_number;
     ThreadFrame picture, last_picture;
+    struct FFV1Context *fsrc;
 
     AVFrame *cur;
     int plane_count;

libavcodec/ffv1dec.c

@@ -326,6 +326,42 @@ static int decode_slice(AVCodecContext *c, void *arg)
     int width, height, x, y, ret;
     const int ps      = av_pix_fmt_desc_get(c->pix_fmt)->comp[0].step_minus1 + 1;
     AVFrame * const p = f->cur;
+    int i, si;
+
+    for( si=0; fs != f->slice_context[si]; si ++)
+        ;
+
+    if(f->fsrc && !p->key_frame)
+        ff_thread_await_progress(&f->last_picture, si, 0);
+
+    if(f->fsrc && !p->key_frame) {
+        FFV1Context *fssrc = f->fsrc->slice_context[si];
+        FFV1Context *fsdst = f->slice_context[si];
+        av_assert1(fsdst->plane_count == fssrc->plane_count);
+        av_assert1(fsdst == fs);
+
+        if (!p->key_frame)
+            fsdst->slice_damaged |= fssrc->slice_damaged;
+
+        for (i = 0; i < f->plane_count; i++) {
+            PlaneContext *psrc = &fssrc->plane[i];
+            PlaneContext *pdst = &fsdst->plane[i];
+
+            av_free(pdst->state);
+            av_free(pdst->vlc_state);
+            memcpy(pdst, psrc, sizeof(*pdst));
+            pdst->state = NULL;
+            pdst->vlc_state = NULL;
+
+            if (fssrc->ac) {
+                pdst->state = av_malloc(CONTEXT_SIZE * psrc->context_count);
+                memcpy(pdst->state, psrc->state, CONTEXT_SIZE * psrc->context_count);
+            } else {
+                pdst->vlc_state = av_malloc(sizeof(*pdst->vlc_state) * psrc->context_count);
+                memcpy(pdst->vlc_state, psrc->vlc_state, sizeof(*pdst->vlc_state) * psrc->context_count);
+            }
+        }
+    }
 
     if (f->version > 2) {
         if (ffv1_init_slice_state(f, fs) < 0)
@@ -386,6 +422,8 @@ static int decode_slice(AVCodecContext *c, void *arg)
 
     emms_c();
 
+    ff_thread_report_progress(&f->picture, si, 0);
+
     return 0;
 }
 
@@ -724,6 +762,8 @@ static av_cold int decode_init(AVCodecContext *avctx)
     if ((ret = ffv1_init_slice_contexts(f)) < 0)
         return ret;
 
+    avctx->internal->allocate_progress = 1;
+
     return 0;
 }
 
@@ -744,6 +784,7 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac
 
     f->cur = p = f->picture.f;
 
+    f->avctx = avctx;
     ff_init_range_decoder(c, buf, buf_size);
     ff_build_rac_states(c, 0.05 * (1LL << 32), 256 - 8);
 
@@ -770,6 +811,8 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac
         av_log(avctx, AV_LOG_DEBUG, "ver:%d keyframe:%d coder:%d ec:%d slices:%d bps:%d\n",
                f->version, p->key_frame, f->ac, f->ec, f->slice_count, f->avctx->bits_per_raw_sample);
 
+    ff_thread_finish_setup(avctx);
+
     buf_p = buf + buf_size;
     for (i = f->slice_count - 1; i >= 0; i--) {
         FFV1Context *fs = f->slice_context[i];
@@ -822,6 +865,7 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac
         if (fs->slice_damaged && f->last_picture.f->data[0]) {
             const uint8_t *src[4];
             uint8_t *dst[4];
+            ff_thread_await_progress(&f->last_picture, INT_MAX, 0);
             for (j = 0; j < 4; j++) {
                 int sh = (j==1 || j==2) ? f->chroma_h_shift : 0;
                 int sv = (j==1 || j==2) ? f->chroma_v_shift : 0;
@@ -837,6 +881,7 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac
                           fs->slice_height);
         }
     }
+    ff_thread_report_progress(&f->picture, INT_MAX, 0);
 
     f->picture_number++;
 
@@ -854,18 +899,58 @@ static int decode_frame(AVCodecContext *avctx, void *data, int *got_frame, AVPac
 static int init_thread_copy(AVCodecContext *avctx)
 {
     FFV1Context *f = avctx->priv_data;
-    int ret, i;
 
-    for (i = 0; i < f->quant_table_count; i++) {
-        void *p = f->initial_states[i];
-        f->initial_states[i] = av_malloc(f->context_count[i] * sizeof(*f->initial_states[i]));
-        if (!f->initial_states[i])
-            return AVERROR(ENOMEM);
-        memcpy(f->initial_states[i], p, f->context_count[i] * sizeof(*f->initial_states[i]));
+    f->picture.f      = NULL;
+    f->last_picture.f = NULL;
+    f->sample_buffer  = NULL;
+    f->quant_table_count = 0;
+    f->slice_count = 0;
+
+    return 0;
+}
+
+static int update_thread_context(AVCodecContext *dst, const AVCodecContext *src)
+{
+    FFV1Context *fsrc = src->priv_data;
+    FFV1Context *fdst = dst->priv_data;
+    int i, ret;
+
+    if (dst == src)
+        return 0;
+
+    if (!fdst->quant_table_count) {
+        memcpy(fdst, fsrc, sizeof(*fdst));
+
+        for (i = 0; i < fdst->quant_table_count; i++) {
+            fdst->initial_states[i] = av_malloc(fdst->context_count[i] * sizeof(*fdst->initial_states[i]));
+            memcpy(fdst->initial_states[i], fsrc->initial_states[i], fdst->context_count[i] * sizeof(*fdst->initial_states[i]));
+        }
+
+        fdst->picture.f      = av_frame_alloc();
+        fdst->last_picture.f = av_frame_alloc();
+
+        if ((ret = ffv1_init_slice_contexts(fdst)) < 0)
+            return ret;
     }
 
-    if ((ret = ffv1_init_slice_contexts(f)) < 0)
-        return ret;
+    av_assert1(fdst->slice_count == fsrc->slice_count);
+
+    fdst->key_frame_ok = fsrc->key_frame_ok;
+
+    ff_thread_release_buffer(dst, &fdst->picture);
+    if (fsrc->picture.f->data[0]) {
+        if ((ret = ff_thread_ref_frame(&fdst->picture, &fsrc->picture)) < 0)
+            return ret;
+    }
+    for (i = 0; i < fdst->slice_count; i++) {
+        FFV1Context *fsdst = fdst->slice_context[i];
+        FFV1Context *fssrc = fsrc->slice_context[i];
+
+        fsdst->slice_damaged = fssrc->slice_damaged;
+    }
+
+    fdst->fsrc = fsrc;
+
     return 0;
 }
 
@@ -878,7 +963,8 @@ AVCodec ff_ffv1_decoder = {
     .close          = ffv1_close,
     .decode         = decode_frame,
     .init_thread_copy = init_thread_copy,
+    .update_thread_context = update_thread_context,
     .capabilities   = CODEC_CAP_DR1 /*| CODEC_CAP_DRAW_HORIZ_BAND*/ |
-                      CODEC_CAP_SLICE_THREADS,
+                      CODEC_CAP_FRAME_THREADS | CODEC_CAP_SLICE_THREADS,
     .long_name      = NULL_IF_CONFIG_SMALL("FFmpeg video codec #1"),
 };

libavcodec/pthread.c

@@ -652,7 +652,7 @@ int ff_thread_decode_frame(AVCodecContext *avctx,
      */
 
     if (fctx->delaying) {
-        if (fctx->next_decoding >= (avctx->thread_count-1)) fctx->delaying = 0;
+        if (fctx->next_decoding >= (avctx->thread_count-1-(avctx->codec_id == AV_CODEC_ID_FFV1))) fctx->delaying = 0;
 
         *got_picture_ptr=0;
         if (avpkt->size)