OpenCL vectorisation issue in OneAPI driver

[OCLdev]

The OneAPI CPU OpenCL drivers (2021.12.6.0.19_160000) seem to have an issue with vectorization resulting in corrupted data. This simple kernel:

__kernel void test(__global float *f, __global float *r) {
    int i = get_global_id(0);

    r[i] = 0.0;
    if (f[i] == 1.0F) {
        r[i] = 1.0F+pow(1.0F, 1.0F);
    }
}

when run over a buffer of length 16 with f equal to:

1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1

results in the following values in r:

2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 3126, 0, 0, 0, 0, 3126

Note the odd 3126 values. Turning off vectorisation using

CL_CONFIG_CPU_VECTORIZER_MODE=1

Results in the correct values:

2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2

The issue seems to be caused by adding a value to the function (i.e. here 1.0+pow()), for example this gives an expected result:

__kernel void test(__global float *f, __global float *r) {
    int i = get_global_id(0);

    r[i] = 0.0;
    if (f[i] == 1.0F) {
        r[i] = pow(1.0F, 1.0F);
    }
}

This is on a Ubuntu 18.04 VM with an Intel(R) Core(TM) i7-6800K.

Is there anywhere I can raise a bug report/ticket? Or is this the correct forum to report?

[VarshaS_Intel]

Hi,

Thanks for reaching out to us.

Could you please provide the details of the compiler you are using and complete reproducer code with the steps to reproduce the issue?

>>>Is there anywhere I can raise a bug report/ticket? Or is this the correct forum to report?

You can post your queries related to OpenCL in this forum.

Link to the Forum: https://community.intel.com/t5/GPU-Compute-Software/bd-p/gpu-compute-software

Thanks & Regards

Varsha

[OCLdev]

Hi Varsha,

Thanks for getting back. I should start by saying I tried this on an Ubuntu F4s_v2 (Intel Xeon 8272CL) VM on Azure and had no issues. The above came from an Ubuntu VM running on VMWare on Windows (on a Intel Core i7-6800K). So it might be related to just this particular unusual combination. I'll post the steps to run the test in case you need them:

1. Create Ubuntu 18.04 VM.
2. Update VM:
   sudo apt-get update
3. Install g++ and pocl to build:
   sudo apt-get -y install g++ ocl-icd-opencl-dev opencl-headers libpocl2
4. Copy OpenCL_test.cpp and cl2.hpp to VM.
5. Compile test program:
   g++ OpenCL_test.cpp -o test /usr/lib/x86_64-linux-gnu/libOpenCL.so
6. Run test:
   azure:~$./test
   Platform: Portable Computing Language
   OpenCL: Build log:
   1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1
   2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2
7. Install Intel OpenCL:
   wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
   sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
   rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
   sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
   sudo apt update
   sudo apt-get install -y --no-install-recommends intel-oneapi-runtime-opencl
8. Run test, argument is the platform to use: (note it works on the Azure VM):
   ./test 0
   Platform: Intel(R) OpenCL
   OpenCL: Build log: Compilation started
   Compilation done
   Linking started
   Linking done
   Device build started
   Options used by backend compiler: -cl-std=CL1.2 -w
   Device build done
   Kernel <test> was successfully vectorized (16)
   Done.
   1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1
   2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 2

On the VMWare VM this has the error:

./test 0
Platform: Intel(R) OpenCL
OpenCL: Build log: Compilation started
Compilation done
Linking started
Linking done
Device build started
Options used by backend compiler: -cl-std=CL1.2 -w
Device build done
Kernel <test> was successfully vectorized (8)
Done.
1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1
2, 0, 0, 0, 0, 2, 0, 0, 0, 0, 3126, 0, 0, 0, 0, 3126

[Alberto_R_Intel]

OCLdev, Thank you for posting in the Intel® Communities Support.

Just to let you know, the thread was already moved to the proper department, they will further assist you with this matter as soon as possible.

Regards,

Albert R.

Intel Customer Support Technician