shithub: libvpx

Info  •  Files  •  Log  •  Branches

Merge changes from topic 'rm-dec-frame-parallel'

* changes:
  vp9_dx,vpx_codec_alg_priv: rm *worker_id*
  vp9_dx,vpx_codec_alg_priv: rm *cache*
  vp9_dx,vpx_codec_alg_priv: rm frame_parallel_decode

--- a/vp9/vp9_dx_iface.c

+++ b/vp9/vp9_dx_iface.c

@@ -47,9 +47,6 @@

     ctx->priv->init_flags = ctx->init_flags;

     priv->si.sz = sizeof(priv->si);

     priv->flushed = 0;

-    // TODO(jzern): remnants of frame-level parallel decoding should be

-    // removed. cf., https://bugs.chromium.org/p/webm/issues/detail?id=1395

-    priv->frame_parallel_decode = 0;

     if (ctx->config.dec) {

       priv->cfg = *ctx->config.dec;

       ctx->config.dec = &priv->cfg;

@@ -279,25 +276,7 @@

       frame_worker_data->pbi, frame_worker_data->data_size, &data);

   frame_worker_data->data_end = data;

-  if (frame_worker_data->pbi->frame_parallel_decode) {

-    // In frame parallel decoding, a worker thread must successfully decode all

-    // the compressed data.

-    if (frame_worker_data->result != 0 ||

-        frame_worker_data->data + frame_worker_data->data_size - 1 > data) {

-      VPxWorker *const worker = frame_worker_data->pbi->frame_worker_owner;

-      BufferPool *const pool = frame_worker_data->pbi->common.buffer_pool;

-      // Signal all the other threads that are waiting for this frame.

-      vp9_frameworker_lock_stats(worker);

-      frame_worker_data->frame_context_ready = 1;

-      lock_buffer_pool(pool);

-      frame_worker_data->pbi->cur_buf->buf.corrupted = 1;

-      unlock_buffer_pool(pool);

-      frame_worker_data->pbi->need_resync = 1;

-      vp9_frameworker_signal_stats(worker);

-      vp9_frameworker_unlock_stats(worker);

-      return 0;

-    }

-  } else if (frame_worker_data->result != 0) {

+  if (frame_worker_data->result != 0) {

     // Check decode result in serial decode.

     frame_worker_data->pbi->cur_buf->buf.corrupted = 1;

     frame_worker_data->pbi->need_resync = 1;

@@ -310,18 +289,8 @@

   const VPxWorkerInterface *const winterface = vpx_get_worker_interface();

   ctx->last_show_frame = -1;

-  ctx->next_submit_worker_id = 0;

-  ctx->last_submit_worker_id = 0;

-  ctx->next_output_worker_id = 0;

-  ctx->frame_cache_read = 0;

-  ctx->frame_cache_write = 0;

-  ctx->num_cache_frames = 0;

   ctx->need_resync = 1;

-  ctx->num_frame_workers =

-      (ctx->frame_parallel_decode == 1) ? ctx->cfg.threads : 1;

-  if (ctx->num_frame_workers > MAX_DECODE_THREADS)

-    ctx->num_frame_workers = MAX_DECODE_THREADS;

-  ctx->available_threads = ctx->num_frame_workers;

+  ctx->num_frame_workers = 1;

   ctx->flushed = 0;

   ctx->buffer_pool = (BufferPool *)vpx_calloc(1, sizeof(BufferPool));

@@ -375,13 +344,11 @@

 #endif

     // If decoding in serial mode, FrameWorker thread could create tile worker

     // thread or loopfilter thread.

-    frame_worker_data->pbi->max_threads =

-        (ctx->frame_parallel_decode == 0) ? ctx->cfg.threads : 0;

+    frame_worker_data->pbi->max_threads = ctx->cfg.threads;

     frame_worker_data->pbi->inv_tile_order = ctx->invert_tile_order;

-    frame_worker_data->pbi->frame_parallel_decode = ctx->frame_parallel_decode;

-    frame_worker_data->pbi->common.frame_parallel_decode =

-        ctx->frame_parallel_decode;

+    frame_worker_data->pbi->frame_parallel_decode = 0;

+    frame_worker_data->pbi->common.frame_parallel_decode = 0;

     worker->hook = (VPxWorkerHook)frame_worker_hook;

     if (!winterface->reset(worker)) {

       set_error_detail(ctx, "Frame Worker thread creation failed");

@@ -426,7 +393,7 @@

     if (!ctx->si.is_kf && !is_intra_only) return VPX_CODEC_ERROR;

   }

-  if (!ctx->frame_parallel_decode) {

+  {

     VPxWorker *const worker = ctx->frame_workers;

     FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;

     frame_worker_data->data = *data;

@@ -449,80 +416,11 @@

       return update_error_state(ctx, &frame_worker_data->pbi->common.error);

     check_resync(ctx, frame_worker_data->pbi);

-  } else {

-    VPxWorker *const worker = &ctx->frame_workers[ctx->next_submit_worker_id];

-    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;

-    // Copy context from last worker thread to next worker thread.

-    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)

-      vp9_frameworker_copy_context(

-          &ctx->frame_workers[ctx->next_submit_worker_id],

-          &ctx->frame_workers[ctx->last_submit_worker_id]);

-

-    frame_worker_data->pbi->ready_for_new_data = 0;

-    // Copy the compressed data into worker's internal buffer.

-    // TODO(hkuang): Will all the workers allocate the same size

-    // as the size of the first intra frame be better? This will

-    // avoid too many deallocate and allocate.

-    if (frame_worker_data->scratch_buffer_size < data_sz) {

-      vpx_free(frame_worker_data->scratch_buffer);

-      frame_worker_data->scratch_buffer = (uint8_t *)vpx_malloc(data_sz);

-      if (frame_worker_data->scratch_buffer == NULL) {

-        set_error_detail(ctx, "Failed to reallocate scratch buffer");

-        return VPX_CODEC_MEM_ERROR;

-      }

-      frame_worker_data->scratch_buffer_size = data_sz;

-    }

-    frame_worker_data->data_size = data_sz;

-    memcpy(frame_worker_data->scratch_buffer, *data, data_sz);

-

-    frame_worker_data->frame_decoded = 0;

-    frame_worker_data->frame_context_ready = 0;

-    frame_worker_data->received_frame = 1;

-    frame_worker_data->data = frame_worker_data->scratch_buffer;

-    frame_worker_data->user_priv = user_priv;

-

-    if (ctx->next_submit_worker_id != ctx->last_submit_worker_id)

-      ctx->last_submit_worker_id =

-          (ctx->last_submit_worker_id + 1) % ctx->num_frame_workers;

-

-    ctx->next_submit_worker_id =

-        (ctx->next_submit_worker_id + 1) % ctx->num_frame_workers;

-    --ctx->available_threads;

-    worker->had_error = 0;

-    winterface->launch(worker);

   }

   return VPX_CODEC_OK;

 }

-static void wait_worker_and_cache_frame(vpx_codec_alg_priv_t *ctx) {

-  YV12_BUFFER_CONFIG sd;

-  vp9_ppflags_t flags = { 0, 0, 0 };

-  const VPxWorkerInterface *const winterface = vpx_get_worker_interface();

-  VPxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];

-  FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;

-  ctx->next_output_worker_id =

-      (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;

-  // TODO(hkuang): Add worker error handling here.

-  winterface->sync(worker);

-  frame_worker_data->received_frame = 0;

-  ++ctx->available_threads;

-

-  check_resync(ctx, frame_worker_data->pbi);

-

-  if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {

-    VP9_COMMON *const cm = &frame_worker_data->pbi->common;

-    RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;

-    ctx->frame_cache[ctx->frame_cache_write].fb_idx = cm->new_fb_idx;

-    yuvconfig2image(&ctx->frame_cache[ctx->frame_cache_write].img, &sd,

-                    frame_worker_data->user_priv);

-    ctx->frame_cache[ctx->frame_cache_write].img.fb_priv =

-        frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;

-    ctx->frame_cache_write = (ctx->frame_cache_write + 1) % FRAME_CACHE_SIZE;

-    ++ctx->num_cache_frames;

-  }

-}

-

 static vpx_codec_err_t decoder_decode(vpx_codec_alg_priv_t *ctx,

                                       const uint8_t *data, unsigned int data_sz,

                                       void *user_priv, long deadline) {

@@ -553,91 +451,37 @@

   if (ctx->svc_decoding && ctx->svc_spatial_layer < frame_count - 1)

     frame_count = ctx->svc_spatial_layer + 1;

-  if (ctx->frame_parallel_decode) {

-    // Decode in frame parallel mode. When decoding in this mode, the frame

-    // passed to the decoder must be either a normal frame or a superframe with

-    // superframe index so the decoder could get each frame's start position

-    // in the superframe.

-    if (frame_count > 0) {

-      int i;

+  // Decode in serial mode.

+  if (frame_count > 0) {

+    int i;

-      for (i = 0; i < frame_count; ++i) {

-        const uint8_t *data_start_copy = data_start;

-        const uint32_t frame_size = frame_sizes[i];

-        if (data_start < data ||

-            frame_size > (uint32_t)(data_end - data_start)) {

-          set_error_detail(ctx, "Invalid frame size in index");

-          return VPX_CODEC_CORRUPT_FRAME;

-        }

-

-        if (ctx->available_threads == 0) {

-          // No more threads for decoding. Wait until the next output worker

-          // finishes decoding. Then copy the decoded frame into cache.

-          if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {

-            wait_worker_and_cache_frame(ctx);

-          } else {

-            // TODO(hkuang): Add unit test to test this path.

-            set_error_detail(ctx, "Frame output cache is full.");

-            return VPX_CODEC_ERROR;

-          }

-        }

-

-        res =

-            decode_one(ctx, &data_start_copy, frame_size, user_priv, deadline);

-        if (res != VPX_CODEC_OK) return res;

-        data_start += frame_size;

+    for (i = 0; i < frame_count; ++i) {

+      const uint8_t *data_start_copy = data_start;

+      const uint32_t frame_size = frame_sizes[i];

+      vpx_codec_err_t res;

+      if (data_start < data || frame_size > (uint32_t)(data_end - data_start)) {

+        set_error_detail(ctx, "Invalid frame size in index");

+        return VPX_CODEC_CORRUPT_FRAME;

       }

-    } else {

-      if (ctx->available_threads == 0) {

-        // No more threads for decoding. Wait until the next output worker

-        // finishes decoding. Then copy the decoded frame into cache.

-        if (ctx->num_cache_frames < FRAME_CACHE_SIZE) {

-          wait_worker_and_cache_frame(ctx);

-        } else {

-          // TODO(hkuang): Add unit test to test this path.

-          set_error_detail(ctx, "Frame output cache is full.");

-          return VPX_CODEC_ERROR;

-        }

-      }

-      res = decode_one(ctx, &data, data_sz, user_priv, deadline);

+      res = decode_one(ctx, &data_start_copy, frame_size, user_priv, deadline);

       if (res != VPX_CODEC_OK) return res;

+

+      data_start += frame_size;

     }

   } else {

-    // Decode in serial mode.

-    if (frame_count > 0) {

-      int i;

+    while (data_start < data_end) {

+      const uint32_t frame_size = (uint32_t)(data_end - data_start);

+      const vpx_codec_err_t res =

+          decode_one(ctx, &data_start, frame_size, user_priv, deadline);

+      if (res != VPX_CODEC_OK) return res;

-      for (i = 0; i < frame_count; ++i) {

-        const uint8_t *data_start_copy = data_start;

-        const uint32_t frame_size = frame_sizes[i];

-        vpx_codec_err_t res;

-        if (data_start < data ||

-            frame_size > (uint32_t)(data_end - data_start)) {

-          set_error_detail(ctx, "Invalid frame size in index");

-          return VPX_CODEC_CORRUPT_FRAME;

-        }

-

-        res =

-            decode_one(ctx, &data_start_copy, frame_size, user_priv, deadline);

-        if (res != VPX_CODEC_OK) return res;

-

-        data_start += frame_size;

-      }

-    } else {

+      // Account for suboptimal termination by the encoder.

       while (data_start < data_end) {

-        const uint32_t frame_size = (uint32_t)(data_end - data_start);

-        const vpx_codec_err_t res =

-            decode_one(ctx, &data_start, frame_size, user_priv, deadline);

-        if (res != VPX_CODEC_OK) return res;

-

-        // Account for suboptimal termination by the encoder.

-        while (data_start < data_end) {

-          const uint8_t marker =

-              read_marker(ctx->decrypt_cb, ctx->decrypt_state, data_start);

-          if (marker) break;

-          ++data_start;

-        }

+        const uint8_t marker =

+            read_marker(ctx->decrypt_cb, ctx->decrypt_state, data_start);

+        if (marker) break;

+        ++data_start;

       }

     }

   }

@@ -645,80 +489,42 @@

   return res;

 }

-static void release_last_output_frame(vpx_codec_alg_priv_t *ctx) {

-  RefCntBuffer *const frame_bufs = ctx->buffer_pool->frame_bufs;

-  // Decrease reference count of last output frame in frame parallel mode.

-  if (ctx->frame_parallel_decode && ctx->last_show_frame >= 0) {

-    BufferPool *const pool = ctx->buffer_pool;

-    lock_buffer_pool(pool);

-    decrease_ref_count(ctx->last_show_frame, frame_bufs, pool);

-    unlock_buffer_pool(pool);

-  }

-}

-

 static vpx_image_t *decoder_get_frame(vpx_codec_alg_priv_t *ctx,

                                       vpx_codec_iter_t *iter) {

   vpx_image_t *img = NULL;

-  // Only return frame when all the cpu are busy or

-  // application fluhsed the decoder in frame parallel decode.

-  if (ctx->frame_parallel_decode && ctx->available_threads > 0 &&

-      !ctx->flushed) {

-    return NULL;

-  }

-

-  // Output the frames in the cache first.

-  if (ctx->num_cache_frames > 0) {

-    release_last_output_frame(ctx);

-    ctx->last_show_frame = ctx->frame_cache[ctx->frame_cache_read].fb_idx;

-    if (ctx->need_resync) return NULL;

-    img = &ctx->frame_cache[ctx->frame_cache_read].img;

-    ctx->frame_cache_read = (ctx->frame_cache_read + 1) % FRAME_CACHE_SIZE;

-    --ctx->num_cache_frames;

-    return img;

-  }

-

   // iter acts as a flip flop, so an image is only returned on the first

   // call to get_frame.

   if (*iter == NULL && ctx->frame_workers != NULL) {

-    do {

-      YV12_BUFFER_CONFIG sd;

-      vp9_ppflags_t flags = { 0, 0, 0 };

-      const VPxWorkerInterface *const winterface = vpx_get_worker_interface();

-      VPxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];

-      FrameWorkerData *const frame_worker_data =

-          (FrameWorkerData *)worker->data1;

-      ctx->next_output_worker_id =

-          (ctx->next_output_worker_id + 1) % ctx->num_frame_workers;

-      if (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC)

-        set_ppflags(ctx, &flags);

-      // Wait for the frame from worker thread.

-      if (winterface->sync(worker)) {

-        // Check if worker has received any frames.

-        if (frame_worker_data->received_frame == 1) {

-          ++ctx->available_threads;

-          frame_worker_data->received_frame = 0;

-          check_resync(ctx, frame_worker_data->pbi);

-        }

-        if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {

-          VP9_COMMON *const cm = &frame_worker_data->pbi->common;

-          RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;

-          release_last_output_frame(ctx);

-          ctx->last_show_frame = frame_worker_data->pbi->common.new_fb_idx;

-          if (ctx->need_resync) return NULL;

-          yuvconfig2image(&ctx->img, &sd, frame_worker_data->user_priv);

-          ctx->img.fb_priv = frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;

-          img = &ctx->img;

-          return img;

-        }

-      } else {

-        // Decoding failed. Release the worker thread.

+    YV12_BUFFER_CONFIG sd;

+    vp9_ppflags_t flags = { 0, 0, 0 };

+    const VPxWorkerInterface *const winterface = vpx_get_worker_interface();

+    VPxWorker *const worker = &ctx->frame_workers[0];

+    FrameWorkerData *const frame_worker_data = (FrameWorkerData *)worker->data1;

+    if (ctx->base.init_flags & VPX_CODEC_USE_POSTPROC) set_ppflags(ctx, &flags);

+    // Wait for the frame from worker thread.

+    if (winterface->sync(worker)) {

+      // Check if worker has received any frames.

+      if (frame_worker_data->received_frame == 1) {

         frame_worker_data->received_frame = 0;

-        ++ctx->available_threads;

-        ctx->need_resync = 1;

-        if (ctx->flushed != 1) return NULL;

+        check_resync(ctx, frame_worker_data->pbi);

       }

-    } while (ctx->next_output_worker_id != ctx->next_submit_worker_id);

+      if (vp9_get_raw_frame(frame_worker_data->pbi, &sd, &flags) == 0) {

+        VP9_COMMON *const cm = &frame_worker_data->pbi->common;

+        RefCntBuffer *const frame_bufs = cm->buffer_pool->frame_bufs;

+        ctx->last_show_frame = frame_worker_data->pbi->common.new_fb_idx;

+        if (ctx->need_resync) return NULL;

+        yuvconfig2image(&ctx->img, &sd, frame_worker_data->user_priv);

+        ctx->img.fb_priv = frame_bufs[cm->new_fb_idx].raw_frame_buffer.priv;

+        img = &ctx->img;

+        return img;

+      }

+    } else {

+      // Decoding failed. Release the worker thread.

+      frame_worker_data->received_frame = 0;

+      ctx->need_resync = 1;

+      if (ctx->flushed != 1) return NULL;

+    }

   }

   return NULL;

 }

@@ -744,12 +550,6 @@

                                           va_list args) {

   vpx_ref_frame_t *const data = va_arg(args, vpx_ref_frame_t *);

-  // Only support this function in serial decode.

-  if (ctx->frame_parallel_decode) {

-    set_error_detail(ctx, "Not supported in frame parallel decode");

-    return VPX_CODEC_INCAPABLE;

-  }

-

   if (data) {

     vpx_ref_frame_t *const frame = (vpx_ref_frame_t *)data;

     YV12_BUFFER_CONFIG sd;

@@ -768,12 +568,6 @@

                                            va_list args) {

   vpx_ref_frame_t *data = va_arg(args, vpx_ref_frame_t *);

-  // Only support this function in serial decode.

-  if (ctx->frame_parallel_decode) {

-    set_error_detail(ctx, "Not supported in frame parallel decode");

-    return VPX_CODEC_INCAPABLE;

-  }

-

   if (data) {

     vpx_ref_frame_t *frame = (vpx_ref_frame_t *)data;

     YV12_BUFFER_CONFIG sd;

@@ -791,12 +585,6 @@

                                           va_list args) {

   vp9_ref_frame_t *data = va_arg(args, vp9_ref_frame_t *);

-  // Only support this function in serial decode.

-  if (ctx->frame_parallel_decode) {

-    set_error_detail(ctx, "Not supported in frame parallel decode");

-    return VPX_CODEC_INCAPABLE;

-  }

-

   if (data) {

     YV12_BUFFER_CONFIG *fb;

     VPxWorker *const worker = ctx->frame_workers;

@@ -842,12 +630,6 @@

                                                  va_list args) {

   int *const update_info = va_arg(args, int *);

-  // Only support this function in serial decode.

-  if (ctx->frame_parallel_decode) {

-    set_error_detail(ctx, "Not supported in frame parallel decode");

-    return VPX_CODEC_INCAPABLE;

-  }

-

   if (update_info) {

     if (ctx->frame_workers) {

       VPxWorker *const worker = ctx->frame_workers;

@@ -891,12 +673,6 @@

                                            va_list args) {

   int *const frame_size = va_arg(args, int *);

-  // Only support this function in serial decode.

-  if (ctx->frame_parallel_decode) {

-    set_error_detail(ctx, "Not supported in frame parallel decode");

-    return VPX_CODEC_INCAPABLE;

-  }

-

   if (frame_size) {

     if (ctx->frame_workers) {

       VPxWorker *const worker = ctx->frame_workers;

@@ -918,12 +694,6 @@

                                             va_list args) {

   int *const render_size = va_arg(args, int *);

-  // Only support this function in serial decode.

-  if (ctx->frame_parallel_decode) {

-    set_error_detail(ctx, "Not supported in frame parallel decode");

-    return VPX_CODEC_INCAPABLE;

-  }

-

   if (render_size) {

     if (ctx->frame_workers) {

       VPxWorker *const worker = ctx->frame_workers;

@@ -944,7 +714,7 @@

 static vpx_codec_err_t ctrl_get_bit_depth(vpx_codec_alg_priv_t *ctx,

                                           va_list args) {

   unsigned int *const bit_depth = va_arg(args, unsigned int *);

-  VPxWorker *const worker = &ctx->frame_workers[ctx->next_output_worker_id];

+  VPxWorker *const worker = &ctx->frame_workers[0];

   if (bit_depth) {

     if (worker) {

--- a/vp9/vp9_dx_iface.h

+++ b/vp9/vp9_dx_iface.h

@@ -15,15 +15,6 @@

 typedef vpx_codec_stream_info_t vp9_stream_info_t;

-// This limit is due to framebuffer numbers.

-// TODO(hkuang): Remove this limit after implementing ondemand framebuffers.

-#define FRAME_CACHE_SIZE 6  // Cache maximum 6 decoded frames.

-

-typedef struct cache_frame {

-  int fb_idx;

-  vpx_image_t img;

-} cache_frame;

-

 struct vpx_codec_alg_priv {

   vpx_codec_priv_t base;

   vpx_codec_dec_cfg_t cfg;

@@ -41,17 +32,8 @@

   int skip_loop_filter;

   // Frame parallel related.

-  int frame_parallel_decode;  // frame-based threading.

   VPxWorker *frame_workers;

   int num_frame_workers;

-  int next_submit_worker_id;

-  int last_submit_worker_id;

-  int next_output_worker_id;

-  int available_threads;

-  cache_frame frame_cache[FRAME_CACHE_SIZE];

-  int frame_cache_write;

-  int frame_cache_read;

-  int num_cache_frames;

   int need_resync;  // wait for key/intra-only frame

   // BufferPool that holds all reference frames. Shared by all the FrameWorkers.

   BufferPool *buffer_pool;