libgav1/src/threading_strategy.cc

// Copyright 2019 The libgav1 Authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

#include "src/threading_strategy.h"

#include <algorithm>
#include <cassert>
#include <memory>

#include "src/frame_scratch_buffer.h"
#include "src/utils/constants.h"
#include "src/utils/logging.h"
#include "src/utils/vector.h"

namespace libgav1 {
namespace {

#if !defined(LIBGAV1_FRAME_PARALLEL_THRESHOLD_MULTIPLIER)
constexpr int kFrameParallelThresholdMultiplier = 4;
#else
constexpr int kFrameParallelThresholdMultiplier =
    LIBGAV1_FRAME_PARALLEL_THRESHOLD_MULTIPLIER;
#endif

// Computes the number of frame threads to be used based on the following
// heuristic:
//   * If |thread_count| == 1, return 0.
//   * If |thread_count| <= |tile_count| * 4, return 0.
//   * Otherwise, return the largest value of i which satisfies the following
//     condition: i + i * tile_columns <= thread_count. This ensures that there
//     are at least |tile_columns| worker threads for each frame thread.
//   * This function will never return 1 or a value > |thread_count|.
//
//  This heuristic is based empirical performance data. The in-frame threading
//  model (combination of tile multithreading, superblock row multithreading and
//  post filter multithreading) performs better than the frame parallel model
//  until we reach the threshold of |thread_count| > |tile_count| *
//  kFrameParallelThresholdMultiplier.
//
//  It is a function of |tile_count| since tile threading and superblock row
//  multithreading will scale only as a factor of |tile_count|. The threshold 4
//  is arrived at based on empirical data. The general idea is that superblock
//  row multithreading plateaus at 4 * |tile_count| because in most practical
//  cases there aren't more than that many superblock rows and columns available
//  to work on in parallel.
int ComputeFrameThreadCount(int thread_count, int tile_count,
                            int tile_columns) {
  assert(thread_count > 0);
  if (thread_count == 1) return 0;
  return (thread_count <= tile_count * kFrameParallelThresholdMultiplier)
             ? 0
             : std::max(2, thread_count / (1 + tile_columns));
}

}  // namespace

bool ThreadingStrategy::Reset(const ObuFrameHeader& frame_header,
                              int thread_count) {
  assert(thread_count > 0);
  frame_parallel_ = false;

  if (thread_count == 1) {
    thread_pool_.reset(nullptr);
    tile_thread_count_ = 0;
    max_tile_index_for_row_threads_ = 0;
    return true;
  }

  // We do work in the current thread, so it is sufficient to create
  // |thread_count|-1 threads in the threadpool.
  thread_count = std::min(thread_count, static_cast<int>(kMaxThreads)) - 1;

  if (thread_pool_ == nullptr || thread_pool_->num_threads() != thread_count) {
    thread_pool_ = ThreadPool::Create("libgav1", thread_count);
    if (thread_pool_ == nullptr) {
      LIBGAV1_DLOG(ERROR, "Failed to create a thread pool with %d threads.",
                   thread_count);
      tile_thread_count_ = 0;
      max_tile_index_for_row_threads_ = 0;
      return false;
    }
  }

  // Prefer tile threads first (but only if there is more than one tile).
  const int tile_count = frame_header.tile_info.tile_count;
  if (tile_count > 1) {
    // We want 1 + tile_thread_count_ <= tile_count because the current thread
    // is also used to decode tiles. This is equivalent to
    // tile_thread_count_ <= tile_count - 1.
    tile_thread_count_ = std::min(thread_count, tile_count - 1);
    thread_count -= tile_thread_count_;
    if (thread_count == 0) {
      max_tile_index_for_row_threads_ = 0;
      return true;
    }
  } else {
    tile_thread_count_ = 0;
  }

#if defined(__ANDROID__)
  // Assign the remaining threads for each Tile. The heuristic used here is that
  // we will assign two threads for each Tile. So for example, if |thread_count|
  // is 2, for a stream with 2 tiles the first tile would get both the threads
  // and the second tile would have row multi-threading turned off. This
  // heuristic is based on the fact that row multi-threading is fast enough only
  // when there are at least two threads to do the decoding (since one thread
  // always does the parsing).
  //
  // This heuristic might stop working when SIMD optimizations make the decoding
  // much faster and the parsing thread is only as fast as the decoding threads.
  // So we will have to revisit this later to make sure that this is still
  // optimal.
  //
  // Note that while this heuristic significantly improves performance on high
  // end devices (like the Pixel 3), there are some performance regressions in
  // some lower end devices (in some cases) and that needs to be revisited as we
  // bring in more optimizations. Overall, the gains because of this heuristic
  // seems to be much larger than the regressions.
  for (int i = 0; i < tile_count; ++i) {
    max_tile_index_for_row_threads_ = i + 1;
    thread_count -= 2;
    if (thread_count <= 0) break;
  }
#else  // !defined(__ANDROID__)
  // Assign the remaining threads to each Tile.
  for (int i = 0; i < tile_count; ++i) {
    const int count = thread_count / tile_count +
                      static_cast<int>(i < thread_count % tile_count);
    if (count == 0) {
      // Once we see a 0 value, all subsequent values will be 0 since it is
      // supposed to be assigned in a round-robin fashion.
      break;
    }
    max_tile_index_for_row_threads_ = i + 1;
  }
#endif  // defined(__ANDROID__)
  return true;
}

bool ThreadingStrategy::Reset(int thread_count) {
  assert(thread_count > 0);
  frame_parallel_ = true;

  // In frame parallel mode, we simply access the underlying |thread_pool_|
  // directly. So ensure all the other threadpool getter functions return
  // nullptr. Also, superblock row multithreading is always disabled in frame
  // parallel mode.
  tile_thread_count_ = 0;
  max_tile_index_for_row_threads_ = 0;

  if (thread_pool_ == nullptr || thread_pool_->num_threads() != thread_count) {
    thread_pool_ = ThreadPool::Create("libgav1-fp", thread_count);
    if (thread_pool_ == nullptr) {
      LIBGAV1_DLOG(ERROR, "Failed to create a thread pool with %d threads.",
                   thread_count);
      return false;
    }
  }
  return true;
}

bool InitializeThreadPoolsForFrameParallel(
    int thread_count, int tile_count, int tile_columns,
    std::unique_ptr<ThreadPool>* const frame_thread_pool,
    FrameScratchBufferPool* const frame_scratch_buffer_pool) {
  assert(*frame_thread_pool == nullptr);
  thread_count = std::min(thread_count, static_cast<int>(kMaxThreads));
  const int frame_threads =
      ComputeFrameThreadCount(thread_count, tile_count, tile_columns);
  if (frame_threads == 0) return true;
  *frame_thread_pool = ThreadPool::Create(frame_threads);
  if (*frame_thread_pool == nullptr) {
    LIBGAV1_DLOG(ERROR, "Failed to create frame thread pool with %d threads.",
                 frame_threads);
    return false;
  }
  int remaining_threads = thread_count - frame_threads;
  if (remaining_threads == 0) return true;
  int threads_per_frame = remaining_threads / frame_threads;
  const int extra_threads = remaining_threads % frame_threads;
  Vector<std::unique_ptr<FrameScratchBuffer>> frame_scratch_buffers;
  if (!frame_scratch_buffers.reserve(frame_threads)) return false;
  // Create the tile thread pools.
  for (int i = 0; i < frame_threads && remaining_threads > 0; ++i) {
    std::unique_ptr<FrameScratchBuffer> frame_scratch_buffer =
        frame_scratch_buffer_pool->Get();
    if (frame_scratch_buffer == nullptr) {
      return false;
    }
    // If the number of tile threads cannot be divided equally amongst all the
    // frame threads, assign one extra thread to the first |extra_threads| frame
    // threads.
    const int current_frame_thread_count =
        threads_per_frame + static_cast<int>(i < extra_threads);
    if (!frame_scratch_buffer->threading_strategy.Reset(
            current_frame_thread_count)) {
      return false;
    }
    remaining_threads -= current_frame_thread_count;
    frame_scratch_buffers.push_back_unchecked(std::move(frame_scratch_buffer));
  }
  // We release the frame scratch buffers in reverse order so that the extra
  // threads are allocated to buffers in the top of the stack.
  for (int i = static_cast<int>(frame_scratch_buffers.size()) - 1; i >= 0;
       --i) {
    frame_scratch_buffer_pool->Release(std::move(frame_scratch_buffers[i]));
  }
  return true;
}

}  // namespace libgav1