Intel MKL (VSL) Correlation 2d error

[imanh]

I'm trying to do a 2D correlation, I get VSL_CC_ERROR_YSHAPE, or VSL_CC_ERROR_DECIMATION for any combination of possible code! I saw an example here: community.intel.com/t5/Intel-oneAPI-Math-Kernel-Library/2D-convolution-using-VSL-Conv/m-p/1007237#M18992 (even the same exact example throws decimation error for me), and here is my code:

"""

MKL_INT xshape[2] = { 128,128 }, yshape[2] = { 3,3 }, zshape[2] = { xshape[0]+yshape[0]-1,xshape[1]+yshape[1]-1 };

status = vslsCorrNewTask(&task, mode, rank, xshape, yshape, zshape);
if (status != VSL_STATUS_OK) {
printf("ERROR: creation of job failed, exit with %d\n", status);
return;
}

status = vslsCorrExec(task, x, NULL, y, NULL, z, NULL);
if (status != VSL_STATUS_OK) {
printf("ERROR: job status bad, exit with %d\n", status);
return;
}

"""
I tried setting vslConvSetStart (set start to [0,0]) -> didnt work
tried setting all the strides to: [1,1] or to [Xshape[1],1] for X \in {x,y,z}
Any combination of the above doesn't work as well.

[MRajesh_intel]

Hi,

The code ran successfully on our end. I used the command: icpc -mkl example.cpp to compile. I am using oneAPI 2021.2 version . Can you please specify your compiler, OS, oneAPI version, so we can reproduce from our end ?

And also here is the sample I tried:

#include <assert.h>
#include <stdio.h>
#include <string.h>
#include "mkl.h"

int main()
{
    VSLConvTaskPtr task;
    
    MKL_INT f_shape[] = { 128, 128};
    MKL_INT g_shape[] = { 3, 3 };
    MKL_INT Rmin[] = { 0, 0 };
    MKL_INT Rmax[] = { f_shape[0] + g_shape[0] - 1, f_shape[1] + g_shape[1] - 1 };
    MKL_INT h_shape[] = { Rmax[0], Rmax[1] };
    MKL_INT h_start[] = { 0, 0 };


    MKL_INT f_stride[] = { f_shape[1], 1 };
    MKL_INT g_stride[] = { g_shape[1], 1 };
    MKL_INT h_stride[] = { h_shape[1], 1 };
    
    double *f = new double[ f_stride[0] * f_shape[0] ];
    double *g = new double[ g_stride[0] * g_shape[0] ];
    double *h = new double[ h_stride[0] * h_shape[0] ];

    for(int i=0; i < f_shape[0]; ++i)
        for(int j=0; j < f_shape[1]; ++j)
            f[ i * f_stride[0] + j ] = 1;

    for(int i=0; i < g_shape[0]; ++i)
        for(int j=0; j < g_shape[1]; ++j)
            g[ i * g_stride[0] + j ] = 1;

    memset( h, 0, sizeof(h[0]) * h_stride[0] * h_shape[0] );

    int status;
    status = vsldConvNewTask( &task, VSL_CONV_MODE_AUTO, 2, f_shape, g_shape, h_shape );
    assert(status == VSL_STATUS_OK);

    status = vslConvSetStart( task, h_start );
    assert(status == VSL_STATUS_OK);

    status = vsldConvExec( task, f, f_stride, g, g_stride, h, h_stride );
    assert(status == VSL_STATUS_OK);

    status = vslConvDeleteTask(&task);
    assert(status == VSL_STATUS_OK);
    
    for (int i=0; i< h_shape[0]; ++i)
    {
        printf("%3i: ",i);
        for (int j=0; j < h_shape[1]; ++j)
        {
            printf("%4.0f ",h[ i * h_stride[0] + j ]);
        }
        printf("\n");
    }
    return 0;
}
 

Regards

Rajesh.

[imanh]

I ran this, and I get all 0s. (the problem is not that it does not run, in this case it does run, but all results are 0)

[MRajesh_intel · Accepted solution]

Hi,

These are the Visual Studio(V16.9.4) settings (attached image below) which ran the sample and the result (attached image below).

Can you also please specify your visual studio version ?

Regards

Rajesh.

[imanh]

VS2019. My settings are the same EXCEPT for:

User interface: DPC++ API (instead of ILP64)

Use oneMKL: Sequential (but this doesn't matter)

: SO I tried and "ILP64" fixed it. But is this okay? I mean, no warning, no compilation error, just results being wrong when you use the wrong interface?

And I found this on ILP64 vs LP64, but I cant find anything about DPC++ API vs (I)LP64.

Thank you.

[MRajesh_intel]

Hi,

Thanks for the confirmation. DPC++ API should be selected if you have DPC++ source code; when working with multiple devices (CPU/GPU) ,using device-specific libraries underneath.

You may try oneMKL DPC++ samples using DPC++ API.

For further reference: https://github.com/oneapi-src/oneMKL

Regards

Rajesh.

[MRajesh_intel]

Hi,

Please let us know if your issues have been resolved. Can we proceed further to close this thread?

Regards

Rajesh.

[imanh]

yes and thanks for your responses. (I did mark your post as solution I thought that closes it.)

[MRajesh_intel]

Hi,

Thanks for the confirmation!

As this issue has been resolved, we will no longer respond to this thread. If you require any additional assistance from Intel, please start a new thread. Any further interaction in this thread will be considered community only.

Have a Good day.

Regards

Rajesh.