Unexpected DftiComputeForward failures using larger inputs

Cohen__Idan — Tue, 15 Oct 2019 13:47:29 GMT

Hey,

I'm using MKL 2019u3 to compute 3D FFT.

This is the function:

void CheckMklFFT(const std::string &err_prefix, MKL_LONG status) {
        if (status != 0) {
            std::string error_message(DftiErrorMessage(status));
            std::cerr << err_prefix + error_message << std::endl;
            throw std::runtime_error(err_prefix + error_message);
        }
 }

std::shared_ptr<Ipp64fc> FFT3D(float *image, RPP::DataDimensions dims) {

        std::shared_ptr<Ipp64fc> image_fft = std::shared_ptr<Ipp64fc>(ippsMalloc_64fc((dims.x / 2 + 1) * dims.y * dims.z), [](Ipp64fc *x) { ippFree(x); });
        ippsZero_64fc(image_fft.get(), (dims.x / 2 + 1) * dims.x * dims.z);

        long lengths[] = {dims.z, dims.y, dims.x};
        long strides_in[] = {0, dims.x * dims.y, dims.x, 1};
        long strides_out[] = {0, (dims.x / 2 + 1) * dims.y, dims.x / 2 + 1, 1};

        std::string err_prefix = "Error: Failed with FFT due to ";

        // create descriptor
        DFTI_DESCRIPTOR_HANDLE Desc_Handle;
        CheckMklFFT(err_prefix, DftiCreateDescriptor(&Desc_Handle, DFTI_SINGLE, DFTI_REAL, 3, lengths));

        CheckMklFFT(err_prefix, DftiSetValue(Desc_Handle, DFTI_PLACEMENT, DFTI_NOT_INPLACE));
        CheckMklFFT(err_prefix, DftiSetValue(Desc_Handle, DFTI_PACKED_FORMAT, DFTI_CCE_FORMAT));
        CheckMklFFT(err_prefix, DftiSetValue(Desc_Handle, DFTI_CONJUGATE_EVEN_STORAGE, DFTI_COMPLEX_COMPLEX));
        CheckMklFFT(err_prefix, DftiSetValue(Desc_Handle, DFTI_INPUT_STRIDES, strides_in));
        CheckMklFFT(err_prefix, DftiSetValue(Desc_Handle, DFTI_OUTPUT_STRIDES, strides_out));
        CheckMklFFT(err_prefix, DftiCommitDescriptor(Desc_Handle));

        // direct FFT
        CheckMklFFT(err_prefix, DftiComputeForward(Desc_Handle, image, image_fft.get()));

        // Free descriptor
        CheckMklFFT(err_prefix, DftiFreeDescriptor(&Desc_Handle));

        return image_fft;
    }

My problem occurs when I use a very large array (for example, an image with the dimensions of 814x814x814 when each value is float, the size of the array is about 2³¹, when moving to Fourier domain (to complex numbers) the size of the array grows to ~2³²).
I compile on Linux system with Intel (R) Xeon (R) CPU E5-4627 v3 @ 2.60GHz processor and add the following flags:

  -DMKL_ILP64 -lmkl_intel_ilp64 -lmkl_core -lmkl_intel_thread

However, as soon as I get to DftiComputeForward, an error value returned with the following error message:
Intel MKL DFTI ERROR: Inconsistent configuration parameters.
I've red at the forum that someone was having a similar problem and on MKL 11 updated 5 it fixed, link to the post:

Unexpected DftiCommitDescriptor failures/interactions using larger inputs with MKL 11.1

Anyone know how to deal with the problem?

Thank you :)

I've found the problem.

Cohen__Idan — Thu, 17 Oct 2019 07:29:26 GMT

I've found the problem.

When I allocate memory I use ippsMalloc_64fc, but this allocation allocate memory for blocks of 32-bit length, and the blocks that I use are bigger, so the memory should be allocated by ippsMalloc_64fc_L.

thanks for the update. that's

Gennady_F_Intel — Thu, 17 Oct 2019 08:11:39 GMT

thanks for the update. that's true, for work with 64-bit object sizes, IPP suggest to use platform-aware API

topic Unexpected DftiComputeForward failures using larger inputs in Intel® oneAPI Math Kernel Library

Unexpected DftiComputeForward failures using larger inputs

I've found the problem.

thanks for the update. that's