Welcome to OStack Knowledge Sharing Community for programmer and developer-Open, Learning and Share
Welcome To Ask or Share your Answers For Others

Categories

0 votes
84 views
in Technique[技术] by (71.8m points)

c++ - How should I interpreter these VTune results?

I'm trying to parallelyzing this code using OpenMP. OpenCV (built using IPP for best efficiency) is used as external library.

I'm having problems unbalanced CPU usage in parallel fors, but it seems that there is no load imbalance. As you will see, this could be because of KMP_BLOCKTIME=0, but this could be necessary because of external libraries (IPP, TBB, OpenMP, OpenCV). In the rest of the questions you will find more details and data that you can download.

These are the Google Drive links to my VTune results:

c755823 basic KMP_BLOCKTIME=0 30 runs : basic hotspot with environment variable KMP_BLOCKTIME set to 0 on 30 runs of the same input

c755823 basic 30 runs : same as above, but with default KMP_BLOCKTIME=200

c755823 advanced KMP_BLOCKTIME=0 30 runs : same as first, but advanced hotspot

For those who are interested, I can send you the original code somehow.

On my Intel i7-4700MQ the actual wall-clock time of the application on average on 10 runs is around 0.73 seconds. I compile the code with icpc 2017 update 3 with the following compiler flags:

INTEL_OPT=-O3 -ipo -simd -xCORE-AVX2 -parallel -qopenmp -fargument-noalias -ansi-alias -no-prec-div -fp-model fast=2 -fma -align -finline-functions    
INTEL_PROFILE=-g -qopt-report=5 -Bdynamic -shared-intel -debug inline-debug-info -qopenmp-link dynamic -parallel-source-info=2 -ldl

In addition I set KMP_BLOCKTIME=0 because the default value (200) was generating an huge overhead.

We can divide the code in 3 parallel regions (wrapped in only one #pragma parallel for efficiency) and a previous serial one, which is around 25% of the algorithm (and it can't be parallelized).

I'll try to describe them (or you can skip to the code structure directly):

  1. We create a parallel region in order to avoid the overhead to create a new parallel region. The final result is to populate the rows of a matrix obejct, cv::Mat descriptor. We have 3 shared std::vector objects: (a) blurs which is a chain of blurs (not parallelizable) using GuassianBlur by OpenCV (which uses the IPP implementation of guassian blurs) (b) hessResps (size known, say 32) (c) findAffineShapeArgs (unkown size, but in order of thousands of elements, say 2.3k) (d) cv::Mat descriptors (unkown size, final result). In the serial part, we populate `blurs, which is a read only vector.
  2. In the first parallel region,hessResps is populated using blurs without any synchronization mechanism.
  3. In the second parallel region findLevelKeypoints is populated using hessResps as read only. Since findAffineShapeArgs size is unkown, we need a local vector localfindAffineShapeArgs which will be appended to findAffineShapeArgs in the next step
  4. Since findAffineShapeArgs is shared and its size is unkown, we need a critical section where each localfindAffineShapeArgs is appended to it.
  5. In the third parallel region, each findAffineShapeArgs is used to generate the rows of the final cv::Mat descriptor. Again, since descriptors is shared, we need a local version cv::Mat localDescriptors.
  6. A final critical section push_back each localDescriptors to descriptors. Notice that this is extremely fast since cv::Mat is "kinda" of a smart pointer, so we push_back pointers.

This is the code structure:

cv::Mat descriptors;
std::vector<Mat> blurs(blursSize);
std::vector<Mat> hessResps(32);
std::vector<FindAffineShapeArgs> findAffineShapeArgs;//we don't know its tsize in advance

#pragma omp parallel
{
//compute all the hessianResponses
#pragma omp for collapse(2) schedule(dynamic)
for(int i=0; i<levels; i++)
    for (int j = 1; j <= scaleCycles; j++)
    {
       hessResps[/**/] = hessianResponse(/*...*/);
    }

std::vector<FindAffineShapeArgs> localfindAffineShapeArgs;
#pragma omp for collapse(2) schedule(dynamic) nowait
for(int i=0; i<levels; i++)
    for (int j = 2; j < scaleCycles; j++){
    findLevelKeypoints(localfindAffineShapeArgs, hessResps[/*...*], /*...*/); //populate localfindAffineShapeArgs with push_back
}

#pragma omp critical{
    findAffineShapeArgs.insert(findAffineShapeArgs.end(), localfindAffineShapeArgs.begin(), localfindAffineShapeArgs.end());
}

#pragma omp barrier
#pragma omp for schedule(dynamic) nowait
for(int i=0; i<findAffineShapeArgs.size(); i++){
{
  findAffineShape(findAffineShapeArgs[i]);
}

#pragma omp critical{
  for(size_t i=0; i<localRes.size(); i++)
    descriptors.push_back(localRes[i].descriptor);
}
}

At the end of the question, you can find FindAffineShapeArgs.

I'm using Intel Amplifier to see hotspots and evaluate my application.

The OpenMP Potential Gain analsysis says that the Potential Gain if there would be perfect load balancing would be 5.8%, so we can say that the workload is balanced between different CPUs.

This i the CPU usage histogram for the OpenMP region (remember that this is the result of 10 consecutive runs):

enter image description here

So as you can see, the Average CPU Usage is 7 cores, which is good.

This OpenMP Region Duration Histogram shows that in these 10 runs the parallel region is executed always with the same time (with a spread around 4 milliseconds):

enter image description here

This is the Caller/Calee tab:

enter image description here

For you knowledge:

  • interpolate is called in the last parallel region
  • l9_ownFilter* functions are all called in the last parallel region
  • samplePatch is called in the last parallel region.
  • hessianResponse is called in the second parallel region

Now, my first question is: how should I interpret the data above? As you can see, in many of the functions half of the time the "Effective Time by Utilization` is "ok", which would probably become "Poor" with more cores (for example on a KNL machine, where I'll test the application next).

Finally, this is the Wait and Lock analysis result:

enter image description here

Now, this is the first weird thing: line 276 Join Barrier (which corresponds to the most expensive wait object) is#pragma omp parallel`, so the beginning of the parallel region. So it seems that someone spawned threads before. Am I wrong? In addition, the wait time is longer than the program itself (0.827s vs 1.253s of the Join Barrier that I'm talking about)! But maybe that refers to the waiting of all threads (and not wall-clock time, which is clearly impossible since it's longer than the program itself).

Then, the Explicit Barrier at line 312 is #pragma omp barrier of the code above, and its duration is 0.183s.

Looking at the Caller/Callee tab:

enter image description here

As you can see, most of wait time is poor, so it refers to one thread. But I'm sure that I'm understanding this. My second question is: can we interpret this as "all the threads are waiting just for one thread who is staying behind?".

FindAffineShapeArgs definition:

struct FindAffineShapeArgs
{
    FindAffineShapeArgs(float x, float y, float s, float pixelDistance, float type, float response, const Wrapper &wrapper) :
        x(x), y(y), s(s), pixelDistance(pixelDistance), type(type), response(response), wrapper(std::cref(wrapper)) {}

    float x, y, s;
    float pixelDistance, type, response;
    std::reference_wrapper<Wrapper const> wrapper;
};

The Top 5 Parallel Regions by Potential Gain in the summary view shows only one region (the only one)

enter image description here

Look at the "/OpenMP Region/OpenMP Barrier-to-Barrier" grouping, this is the order of the most expensives loops:

  • The 3th loop:

    pragma omp for schedule(dynamic) nowait

    for(int i=0; i

    is the most expensive one (as I already knew) and here's a screenshot of the expended view:

enter image description here

As you can see, many functions are from OpenCV, which exploits IPP and is (should be) already optimized. Expanding the two other functions (interpolate and samplePatch) shows a [No call stack information]. Same for all the other functions (in other regions too).

The 2nd most expensive region is the second parallel for:

#pragma omp for collapse(2) schedule(dynamic) nowait
for(int i=0; i<levels; i++)
    for (int j = 2; j < scaleCycles; j++){
    findLevelKeypoints(localfindAffineShapeArgs, hessResps[/*...*], /*...*/); //populate localfindAffineShapeArgs with push_back
}

Here's the expanded view:

enter image description here

And finally the 3th most expensive is the first loop:

#pragma omp for collapse(2) schedule(dynamic)
for(int i=0; i<levels; i++)
    for (int j = 1; j <= scaleCycles; j++)
    {
       hessResps[/**/] = hessianResponse(/*...*/);
    }

Here's the expended view:

enter image description here

If you want to know more, please use my attached VTune files or just ask!

See Question&Answers more detail:os

与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome To Ask or Share your Answers For Others

1 Answer

0 votes
by (71.8m points)

Try reading the info in this link, especially the part about "Nested OpenMP" since Intel IPP already uses OpenMP in their implementation. From my experience with Intel IPP and OpenMP, if you are doing some other type of multi-threading and when each of the created threads gets to the OpenMP calls performance was really bad. Also, you can try having #pragma omp parallel for for each of the parallel regions instead of #pragma omp for and getting rid of the outer #pragma omp parallel


与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome to OStack Knowledge Sharing Community for programmer and developer-Open, Learning and Share
Click Here to Ask a Question

...