OpenCL* for CPU
Ask questions and share information on Intel® SDK for OpenCL™ Applications and OpenCL™ implementations for Intel® CPU
Announcements
This forum covers OpenCL* for CPU only. OpenCL* for GPU questions can be asked in the GPU Compute Software forum. Intel® FPGA SDK for OpenCL™ questions can be ask in the FPGA Intel® High Level Design forum.
1663 Discussions

Bug in OpenCL driver on Intel HD Graphics?

Alexander_Karsakov
285 Views

Hello. 

I found a curious bug in Intel OpenCL driver on Intel HD Graphics 4600 (Core i5 4570). 

Driver version: 10.18.10.3412

Intel OpenCL SDK Version: 2013 R3 64-bit

I write following kernel:

    int gidx = get_global_id(0);
    int gidy = get_global_id(1);
    
    __local uchar temp[3];
    uchar r = 255;
    if (gidx == 0 && gidy == 0)
    {
        temp[0] = 255;
        temp[1] = 34;
        temp[2] = 255;
        for (int i=0; i<3; i++)
        {
            printf("temp = %d\n", temp);
            printf("r = %d\n", r);
            r = min(r,temp);
        }
    }

When I run it I get:

temp = 255
r = 255
temp = 34
r = 15
temp = 255
r = 15

Variable 'r' have incorrect value (15 instead 34). What may cause the problem?

 

0 Kudos
9 Replies
Raghupathi_M_Intel
285 Views

I tried reproducing this on a newer driver as well as the driver you mentioned without success. I am using the 2014 SDK though, but I do believe that this wont make any difference.

Is the code you posted the whole kernel or part of a bigger kernel? If its latter, then I suspect the bug is else where in your code.

Raghu

Alexander_Karsakov
285 Views

This code is the whole kernel. This is simplified case from Erode function in OpenCV library. And it correct works on other platforms (Nvidia, AMD, Intel CPU). 

If I define:

#define OP_MIN((A),(B)) (A) < (B) ? (A) : (B)

and replace min() for OP_MIN this example gives correct result (include Intel GPU), but works a little slower. 

I will try to install new version of Intel OpenCL SDK to check it.

Alexander_Karsakov
285 Views

Another strange detail: this error occurs only with min() function. 

Similar example with max() works correct:

    int gidx = get_global_id(0);
    int gidy = get_global_id(1);

    __local uchar temp[3];
    uchar r = 0;
    if (gidx == 0 && gidy == 0)
    {
        temp[0] = 0;
        temp[1] = 34;
        temp[2] = 0;
        for (int i=0; i<3; i++)
        {
            printf("temp = %d\n", temp);
            printf("r = %d\n", r);
            r = max(r,temp);
        }
    }
Alexander_Karsakov
285 Views

Hi Raghu.

I checked this code with Intel OpenCL SDK Version 2014. And this error still occurs for me. 
What do you suggest for solving this issue?

Thanks,

Alexander.

 

Alexander_Karsakov
285 Views

Hi Raghu.

I still wait for your response. If it can help, here is a link to the original code of 'erode' function from OpenCV library - https://github.com/Itseez/opencv/blob/master/modules/imgproc/src/opencl/morph.cl

Thanks, 

Alexander.

Raghupathi_M_Intel
285 Views

Hi Alexander,

Did you update the driver or are you still using 10.18.10.3412?

Thanks,
Raghu

Raghupathi_M_Intel
285 Views

Hi Alexander,

I retested this on another driver and was able to reproduce the behavior you are seeing. I will open a bug and let you know when the fix is in.

Thanks,
Raghu

Raghupathi_M_Intel
285 Views

After further debugging, looks like r gets truncated to 4 bits. This issue is fixed/not reproducible in the internal drivers. I dont have any info on when the next driver will be posted, but I'll keep you updated.

Raghu

Alexander_Karsakov
285 Views

Hi Raghu,

I updated to the latest beta driver (10.18.10.3496).

Thanks for your help!

Alexander.

Reply