当前位置：网站首页>The principle and use of AOSP CameraLatencyHistogram

The principle and use of AOSP CameraLatencyHistogram

2022-08-03 22:45:00 【InfoQ】

AOSPCode reference reference：

www.aospxref.com/

可以搜索到android各个版本的代码,比较方便,不必下载AOSP代码.

CameraLatencyHistogram是什么

CameraLatencyHistogram是AOSP CameraA utility class for modules,The main function is to perform histogram statistics,For some delayed operations,比如request执行的时间,dequeue buffer的时间等,Do histogram statistics,Last for performance related issuesdebug.

Core source code location：

/frameworks/av/services/camera/libcameraservice/utils/LatencyHistogram.cpp

源码解析

头文件定义

其中,

mBinSizeMs The record is the standard time of the current record table,Such as preset oncerequestThe standard duration is 40ms

mBinCount Records the number of levels of the histogram,通常默认是10个等级

mBins Records the number of levels in the histogram

mTotalCount Records the total number of statistics for the histogram

类初始化

Initialize the above variables,注意,

这里对vector进行的初始化,It's actually a settingvector的size大小

,Instead of setting the first element tobinCount.

增加一个数据/清空所有数据

add一个数据时：Calculate the time difference first,注意,这里是ns,然后转换为ms,Then divide by standard time（A computer's integer division takes only the integer part,For example, standard time is40ms,Time consuming this time30ms,Then the division is obtained0,耗时70ms,Then the division is obtained1）,这样就能保证0<=t<mBinSizeMs的时长index为0,And so on9

mbinSizeMs<=t<10

mBinSizeMs的index为9;

Then for less than0和大于等于10do the normalization operation;

Then add one to the corresponding storage location,总数加一.

dump/logand internal processing functionsformatHistogramText

dump的逻辑：

先dumpThe line header of the data,标识dumpamount of recorded data;

The histogram data is then constructed as a string;

Write the histogram data to the corresponding fd文件中.

logThe operation of the function ANDdump基本一致,Only this time is calledAndroid log 打印出来.

The function of this function is to represent the statistical result as a string：

Print the elapsed time period first,如33ms/66ms/99ms,The maximum value is printed as inf (max ms)

Then print the percentage of each segment data to the total data volume
100.0*mBins[i]/mTotalCount

Print one last%号.

Where to use the tool

mDequeueBufferLatency：dequeue buffer的时延,5ms

///frameworks/av/services/camera/libcameraservice/device3/Camera3OutputStream.h
//dequebuffer默认耗时5ms,进行记录 
static const int32_t kDequeueLatencyBinSize = 5; // in ms C
ameraLatencyHistogram mDequeueBufferLatency; 

status_t Camera3OutputStream::getBuffersLocked(std::vector<OutstandingBuffer>* outBuffers) { 
 ... 
 nsecs_t dequeueStart = systemTime(SYSTEM_TIME_MONOTONIC); 
 res = consumer->dequeueBuffers(&buffers); 
 nsecs_t dequeueEnd = systemTime(SYSTEM_TIME_MONOTONIC);
 mDequeueBufferLatency.add(dequeueStart, dequeueEnd); 
 ... 
}

mBufferLimitLatency：buffer signal时延,33ms

///frameworks/av/services/camera/libcameraservice/device3/Camera3Stream.h
// Latency histogram of the wait time for handout buffer count to drop below 
// max_buffers. 
static const int32_t kBufferLimitLatencyBinSize = 33; //in ms 
CameraLatencyHistogram mBufferLimitLatency; 
status_t Camera3Stream::getBuffer(camera_stream_buffer *buffer, nsecs_t waitBufferTimeout, const std::vector<size_t>& surface_ids) { 
 ... 
 nsecs_t waitStart = systemTime(SYSTEM_TIME_MONOTONIC); 
 if (waitBufferTimeout < kWaitForBufferDuration) { 
 waitBufferTimeout = kWaitForBufferDuration; 
 } 
 res = mOutputBufferReturnedSignal.waitRelative(mLock, waitBufferTimeout); 
 nsecs_t waitEnd = systemTime(SYSTEM_TIME_MONOTONIC);
 mBufferLimitLatency.add(waitStart, waitEnd); 
 ... 
}

mRequestLatency：requestexecution delay,40ms

///frameworks/av/services/camera/libcameraservice/device3/Camera3Device.h
//记录request的时延,Default one framerequest时长40ms 
static const int32_t kRequestLatencyBinSize = 40; // in ms 
CameraLatencyHistogram mRequestLatency; 
bool Camera3Device::RequestThread::threadLoop() { 
 ... 
 bool submitRequestSuccess = false; 
 nsecs_t tRequestStart = systemTime(SYSTEM_TIME_MONOTONIC); 
 submitRequestSuccess = sendRequestsBatch(); 
 nsecs_t tRequestEnd = systemTime(SYSTEM_TIME_MONOTONIC);
 mRequestLatency.add(tRequestStart, tRequestEnd); 
 ... 
}