Archives Discussions

[Atmapuri]

Hi!

I have a kernel like this:

__kernel void  vzPowx_dc(__global const double *aSrc, const int aSrcIdx,
                              const double bRe, const double bIm,
                      __global double *Dst, const int DstIdx,
                                            const int Length)
{
        dobule Re, Im, XRe, XIm, ExpX;
        size_t i = get_global_id(0);

        XRe = aSrc[aSrcIdx + i];

              Re = log(fabs(XRe));
              Im = atan2(0, XRe);

               XRe = Re * bRe - Im * bIm;
               XIm = Im * bRe + Re * bIm;

               ExpX = exp(XRe);
               Im = sincos(XIm, &Re);

               Re *= ExpX;
               Im *= ExpX;

              Dst[2*(DstIdx + i)] = Re;

              Dst[2*(DstIdx + i)+1] = Im;
}

The code running via AMD Open CL using CPU as Target runs with only 4 numerals of precision (single precision typical) instead of 8.

Thanks!
Atmapuri

[dimkar]

Have you enables khr_fp64 or amd_fp64 extensions?

Try with amd_fp64 on the CPU

#pragma OPENCL EXTENSION cl_amd_fp64: enable

[Atmapuri]

Hi!

Yes, I have this enabled, otherwise the kernel wont even compile. I tried both with Intel and AMD driver and the precision remains equal (very low). I also verified that the results are correct (with the precision mentioned) and that this is in fact what I am computing. When running C++ on CPU I get 7 valid numerals.

Thanks!
Atmapuri

[dimkar]

How do you evaluate your result?

Against the output of a C++ function?

Try might help identify where the problem is exactly:

add cl_amd_printf: enable

printf all the intermediate results of a couple of threads.

printf the equivelant positions in the loop in the reference program