3
$\begingroup$

I am implementing Jerry Tessendorf's ocean waves as described in his paper in c++ and OpenGL.

I implemented two compute shaders, one for the h_tilde_0 and its conjugate, and one for the frequency function for the IFFT.

After performing the real to complex IFFT using GLFFT I obtain the following height map after normalizing the obtained values between 0 and 255:

height map

The first 20 values of the height map before normalizing are:

  • [0] = {float} -24667.7344
  • [1] = {float} 23987.6855
  • [2] = {float} -22951.1094
  • [3] = {float} 22144.1914
  • [4] = {float} -21596.2676
  • [5] = {float} 22228.2578
  • [6] = {float} -23422.1855
  • [7] = {float} 23971.5117
  • [8] = {float} -24944.8633
  • [9] = {float} 26373.8594
  • [10] = {float} -28082.5977
  • [11] = {float} 29593.1387
  • [12] = {float} -30536.7344
  • [13] = {float} 31605.7578
  • [14] = {float} -33134.7813
  • [15] = {float} 34946.293
  • [16] = {float} -36430.6445
  • [17] = {float} 37443.6523
  • [18] = {float} -38190.6758
  • [19] = {float} 39565.5781
  • [20] = {float} -41436.6563

So the values I get after the inverse fourier transform are with inverted signs in an interleaved fashion. And if I were to skip every odd value I get the following height map:

height map

And skipping even values (same IFFT):

enter image description here

It looks like the previous image contains the height map and the opposite sign version of itself. Since it looks like the height field is somehow encoded I don't think that the error is on the generation of the spectrum. I think of two possibilities:

  1. I have wrong coordinates in the frequency domain and somehow produce this result.
  2. I am not understanding how GLFFT expects the input and how the output is given. I have been reading the glsl files in the GLFFT repository and I can't find any mistakes.

I understand that the question is very broad and I am maybe not giving enough information (please tell me if I should upload any other code of result). I have been trying to debug the problem but I am not making any progress.

Thanks in advance for any help!

Appendice:

How I am performing the IFFT:

std::shared_ptr<abstractions::SSBO> update_fft_texture(ssbo_pointer& h_k_t, int N){
    GLFFT::FFTOptions options;
    options.type.fp16 = false;
    options.type.output_fp16 = false;
    options.type.input_fp16 = false;
    GLFFT::GLContext context;

    GLFFT::FFT fft(&context, N, N, GLFFT::ComplexToReal, GLFFT::Inverse, GLFFT::SSBO, GLFFT::SSBO, std::make_shared<GLFFT::ProgramCache>(), options);

    GLuint output_texture, input_texture;

    std::shared_ptr<abstractions::SSBO> out(new abstractions::SSBO(nullptr, 4*N*N, GL_DYNAMIC_COPY));

    input_texture = (*h_k_t).GetRendererId();
    output_texture = (*out).GetRendererId();

    // Adapt raw GL types to types which GLContext uses internally.
    GLFFT::GLBuffer adaptor_output(output_texture);
    GLFFT::GLBuffer adaptor_input(input_texture);

    GLCall(
            {
                // Do the FFT
                GLFFT::CommandBuffer *cmd = context.request_command_buffer();
                fft.process(cmd, &adaptor_output, &adaptor_input);
                context.submit_command_buffer(cmd);
            }
            )


    GLCall(glMemoryBarrier(GL_SHADER_STORAGE_BARRIER_BIT));

    return out;
}

Frequency domain function where red channel represents the real part and green the imaginary (it is not normalized and the values exceed the 0-255 range):

enter image description here

I omitted the h_tidle_0 and frequency shaders' code since it looks to me that the problem is not there.

Two relevant links:

This implementation that uses GLFFT: https://arm-software.github.io/opengl-es-sdk-for-android/ocean_f_f_t.html

This paper that obtains spectrum and frequency textures similar to mines: https://tore.tuhh.de/bitstream/11420/1439/1/GPGPU_FFT_Ocean_Simulation.pdf

$\endgroup$

1 Answer 1

2
$\begingroup$

I found the problem. Maybe this answer helps someone since there isn't much discussion on the topic.

I was doing the following change of coordinates in the frequency domain to fit the values on the 256x256 matrix:

(n, m) = (x-N/2, y-N/2)

where N, n and m appear in the ocean waves' paper and x, y are the coordinates of the matrix.

Then I modified the change of coordinates as is done in this implementation:

(n, m) = (alias(x), alias(y))

where

int alias(int a){
    if(a > N/2) a -= N;
    return N;
}

Then immediately got the correct result (different seed):

enter image description here

$\endgroup$
1
  • 1
    $\begingroup$ Glad you got it. I figured it had to be an indexing issue of some kind, but I'm not familiar with this particular FFT library so couldn't help. $\endgroup$ Commented Jul 17, 2021 at 16:48

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.