DPC++ implementation on CPU is faster than C++ implementation on CPU

[student4]

Hi,

Shown below is the code snippet of my program implemented using openmp parallel_for and using SYCL libraries(offload platform: CPU)

Case 1: Parallelization using openmp (standard cpp codes)

#pragma omp parallel for
for (int i = 0; i < loopCount; i++)
{
for (int m = 0; m < Len; m++)
{
OutputParallel[i] += testData[i + m] * Coeff[m];
}
}

case 2: Using DPC++ libraries; targets CPU

// buffer declaration
sycl::buffer testBuffx64(testData);
sycl::bufferBuffx64(Coeff);
sycl::buffer convBuffx64(Outputx64Oneapi);
t4 = std::chrono::high_resolution_clock::now();
try
{
cl::sycl::queue Q(sycl::cpu_selector{}); // device for computation offload

Q.submit([&](sycl::handler& h)
{
// accessor declaration
sycl::accessor testAccess(testBuffx64, h, sycl::read_only);
sycl::accessor Access(Buffx64, h, sycl::read_only);
sycl::accessor cAccess(convBuffx64, h, sycl::read_write);
h.parallel_for(sycl::range{ loopCount }, [=](sycl::id<1> idx) {
for (int m = 0; m < filterLen; m++)
{
cAccess[idx] += testAccess[idx + m] * Access[m];
}
});
});

Q.wait_and_throw();
}

catch (sycl::exception const& e)
{
std::cout << "Caught a SYCL host exception for x64:\n" << e.what() << "\n";
}

The throughput of case 2 is higher than case 1 although both are implemented on CPU.
Is there a reason why the DPC++ implementation is faster?

Thank you

[VarshaS_Intel]

Hi,

Thanks for posting in Intel Communities.

Could you please provide us with the Environment details and Intel OneAPI version you are using?

>>the DPC++ implementation is faster?

Could you please let us know how are you calculating the time? Are you using any specific tool to calculate the time?

Could you please provide us the results(difference in time) you are getting while running the two cases(openmp code and dpcpp code)?

And also, could you please provide us with the complete sample reproducer code and the compiler you are using to run the openmp code(case 1)?

Thanks & Regards,

Varsha

[student4]

Hi,

Intel oneAPI version 2022

IDE: Microsoft Visual Studio 2019.

I have attached the complete project solution.

Thank you.

[student4]

Please let me know if you need any other information

[VarshaS_Intel]

Hi,

We are working on your issue internally. We will get back to you soon.

Thanks & Regards,

Varsha

[Jie_L_Intel]

What's the command line option you used when building the C++/DPC++ code? could you show the "icx or dpcpp --version"?

[Jie_L_Intel]

i do not see DPC++ is faster.

C:\osc05423107>dpcpp -g /EHsc -fiopenmp -fsycl-targets=spir64_x86_64 Filter.cpp

C:\osc05423107>Filter.exe

implementation is correct

Throughput for serial implementation:0.0172098

Throughput for parallel implementation:0.0333564

Throughput for oneAPI x64 implementation:1.9899

Throughput for oneAPI FPGA implementation:inf

success

[student4]

'Throughput for oneAPI x64 implementation:' is the DPC++ implementation which refers to the device code offloaded to x64 platform.

Even in your case you can see that DPC++ implementation is faster than C++ implementation with openmp parallelization(Throughput for parallel implementation).