第26讲 DequeueBuffer Latency - Android Camera性能分析

本讲是Android Camera性能分析专题的第26讲，我们介绍DequeueBuffer Latency，包括如下内容：

1. DequeueBuffer Latency是什么
2. DequeueBuffer Latency配置
3. DequeeuBuffer Latency实战

DequeueBuffer Latency是什么

DequeueBuffer Latency 用于衡量/记录从BufferQueue里面获取Buffer的耗时。

影响：

• 如果该耗时太长，会拉低CaptureRequest的帧率，具体体现在送给HAL的CaptureRequest速度变慢。

每一个Camera3OutputStream都会统计DequeueBuffer Latency。

• 这里的Output/Input Stream是相对HAL而言，HAL输出的流是Output Stream，输入给HAL的流是Input Stream

DequeueBuffer：

• 通过GraphicBufferProducer从BufferQueue里面获取一张未使用的Buffer

DequeueBuffer Latency配置

DequeueBuffer Latency直方图配置

• 分成10份（默认值），间隔为5ms
• static const int32_t kDequeueLatencyBinSize = 5; // in ms

添加Sample的时机

• 在每个CaptureRequest的prepareHalRequests中会调用getBuffer，进而触发到对BufferQueue的DequeueBuffer操作。
• Camera3OutputStream会对每次dequeueBuffer操作耗时做记录，然后作为一个Sample添加到CameraLatencyHistogram。

DequeeuBuffer Latency实战

• 代码讲解

  nsecs_t dequeueStart = systemTime(SYSTEM_TIME_MONOTONIC);
  
  size_t batchSize = mBatchSize.load();
  if (batchSize == 1) {
    sp anw = consumer;
    res = anw->dequeueBuffer(anw.get(), anb, fenceFd);
  } else {
    std::unique_lock batchLock(mBatchLock);
    res = OK;
    if (mBatchedBuffers.size() == 0) {
        if (remainingBuffers == 0) {
            ALOGE("%s: cannot get buffer while all buffers are handed out", __FUNCTION__);
            return INVALID_OPERATION;
        }
        if (batchSize > remainingBuffers) {
            batchSize = remainingBuffers;
        }
        batchLock.unlock();
        // Refill batched buffers
        std::vector batchedBuffers;
        batchedBuffers.resize(batchSize);
        res = consumer->dequeueBuffers(&batchedBuffers);
        batchLock.lock();
        if (res != OK) {
            ALOGE("%s: batch dequeueBuffers call failed! %s (%d)",
                    __FUNCTION__, strerror(-res), res);
        } else {
            mBatchedBuffers = std::move(batchedBuffers);
        }
    }
  
    if (res == OK) {
        // Dispatch batch buffers
        *anb = mBatchedBuffers.back().buffer;
        *fenceFd = mBatchedBuffers.back().fenceFd;
        mBatchedBuffers.pop_back();
    }
  }
  
  nsecs_t dequeueEnd = systemTime(SYSTEM_TIME_MONOTONIC);
  mDequeueBufferLatency.add(dequeueStart, dequeueEnd);

• 实战演示，请参阅视频

本文链接：http://so.lmcjl.com/news/16551/