FD_Plate_Class

A finite difference Kirchoff Thin-Plate Scheme. Based on the MATLAB code from my master's thesis

FDTD CLasses

Included are two classes, one for the stiff string and the other for the Kirchoff Plate.

FDPlate

Class wrapped method of the inline code; The number of grid points is currently limited by the class member

const int maxXgrid = 300;

Minimal Code

Minimal code to get a sound from the plate:

#include <iostream>
#include <string>
#include "../../FDTDClasses/FDPlate.cpp"
#include "../../FDTDClasses/FDString.cpp"
#include "/../../cpp-cli-audio-tools/src/CliAudioTools.h"

int main (int argc, const char *argv[])
{
    const char *fileName = "/absolute/path/to/file.wav";
    const double sampleRate = 44.1e3;
    const double duration = 4;       // duration (seconds)

    FDPlate plate(sampleRate);
    plate.setInitialCondition();

    const double Nf = duration * sampleRate; // duration (samples)
    double *out = new double[Nf];

    for(int n = 0; n < Nf; ++n)
    {
        plate.updateScheme();
        out[n] = plate.getOutput();
    }

    WavCodec wc;
    wc.writeWavMS(out, fileName, Nf, sampleRate);
    return EXIT_SUCCESS;
}

ToDo

• Methods to couple FDTD systems together.
• Cleaner method to set system limits on total grid points
• Move to std::unique_ptr
• complete addForce (float force)
• complete Doxygen documentation

OpenCL

Out of morbid curiosity, and to force a little code refactoring, included is an OpenCL implementation of the plate. Currently Using the Apple OpenCL API which is now deprecated. There have been the odd machine freeze-ups so it shouldn't be treated as stable.

ToDo

Output One of the main bottlenecks in the ^kernelBlock is the gcl_memcpy at the end. This was just copying the methodology from Apples own OpenCL Hello World Example. To get output efficiently, the next step is to allocate memory on the GPU and read audio data to it with a final read out at the end.

Work Group Size

Work Group size appears happiest around 32 rather than the max size of 256. Defining workgroup size with

gcl_get_kernel_block_workgroup_info(fd_plate_kernel,
                                            CL_KERNEL_WORK_GROUP_SIZE,
                                            sizeof(wgs),
                                            &wgs, NULL);

always seems to report back 256. Stripping out the polling for workgroup size and having a minimal ^kernelBlock appears to make things run a little quicker. i.e.

void (^kernelBlock)() =
    ^{        
        cl_ndrange range =
        {
            1,              // Dimensions
            {0, 0, 0},      // Offset in each dimension.
            {clItems,0, 0}, // global range process.
            {32, 0, 0}      // Workgroup Size
        };

        fd_plate_kernel(&range,
                        (cl_float*)cl_u,
                        (cl_float*)cl_u1,
                        (cl_float*)cl_u2,
                        (cl_float*)cl_B,
                        (cl_float*)cl_C,
                        (cl_int*)cl_Ny,
                        (cl_int*)cl_Nx,
                        (cl_int*)cl_ss);

        void* dummy = cl_u2; cl_u2 = cl_u1; cl_u1 = cl_u; cl_u = dummy; // swap pointers
    };

Cli Audio

Included are a few tools for gerenrating audio from the command line. For the finer details check out the cpp-cli-audio-tools repo.

Notes

Derivations of schemes can be found at the accompanying GitHub pages site

All code tested with Xcode Version 8.2.1 (8C1002)
4.2.1 Compatible Apple LLVM 8.0.0 (clang-800.0.42.1)

Remember to add /usr/local/include to Header Search Paths