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 for
s, 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):
- 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.
- In the first parallel region,
hessResps
is populated using blurs
without any synchronization mechanism.
- 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
- Since
findAffineShapeArgs
is shared and its size is unkown, we need a critical
section where each localfindAffineShapeArgs
is appended to it.
- 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
.
- 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):
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):
This is the Caller/Calee tab:
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:
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:
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)
Look at the "/OpenMP Region/OpenMP Barrier-to-Barrier" grouping, this is the order of the most expensives loops:
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:
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:
If you want to know more, please use my attached VTune files or just ask!
See Question&Answers more detail:
os