keijiro / burstfft Goto Github PK

Thanks for the excellent code keijiro! FWIW, as I was testing your code inside my application I noticed that as fft size continues to be increased the parallel version does eventually become faster than the serial version. For me this was around 16,384 samples (with leak detection off, burst safety checks off, and jobs debugger disabled). I am sharing this in case you want to update your README.

Parallel benchmarks...

BurstFft (length=1024) Millisecond Median:0.10 Min:0.10 Max:0.11 Avg:0.10 Std:0.01 SampleCount: 9 Sum: 0.90
BurstFft (length=8192) Millisecond Median:0.23 Min:0.22 Max:0.24 Avg:0.23 Std:0.00 SampleCount: 9 Sum: 2.04
BurstFft (length=16384) Millisecond Median:0.29 Min:0.28 Max:0.31 Avg:0.29 Std:0.01 SampleCount: 9 Sum: 2.64
BurstFft (length=65536) Millisecond Median:0.53 Min:0.53 Max:0.55 Avg:0.54 Std:0.01 SampleCount: 9 Sum: 4.82

Serial benchmarks...

BurstFft (length=1024) Millisecond Median:0.03 Min:0.03 Max:0.03 Avg:0.03 Std:0.00 SampleCount: 9 Sum: 0.23
BurstFft (length=8192) Millisecond Median:0.14 Min:0.14 Max:0.14 Avg:0.14 Std:0.00 SampleCount: 9 Sum: 1.22
BurstFft (length=16384) Millisecond Median:0.28 Min:0.28 Max:0.28 Avg:0.28 Std:0.00 SampleCount: 9 Sum: 2.49
BurstFft (length=65536) Millisecond Median:1.17 Min:1.17 Max:1.17 Avg:1.17 Std:0.00 SampleCount: 9 Sum: 10.52

Test code...

    [Test, Performance]
    public void BurstFft()
    {
        int[] LENGTHS = new int[] { 1024, 8192, 16384, 65536 };
        for (var i = 0; i < LENGTHS.Length; ++i) {
            var x = new float[LENGTHS[i]];
            var spectrum = new float[LENGTHS[i]];
            var fft = new BurstFft(x.Length);
            var input = new NativeArray<float>(x.Length, Allocator.Persistent);
            Measure.Method(() =>
            {
                AudioListener.GetOutputData(x, 0);
                input.CopyFrom(x);
                fft.Transform(input);
                fft.Spectrum.CopyTo(spectrum);
            }).SampleGroup($"BurstFft (length={LENGTHS[i]})").Run();
            input.Dispose();
            fft.Dispose();
        }
    }

Hello,

I would like to use this project in the creation of a school project and was wondering if you could add a license to this project or if I could get permission to use this code in my project ?

Thanks.

I've had a look at the code trying to use it.

Then I noticed that the jobs should not be able to run in parallel. The following code (simplified):

var handle = new FirstPassJob(...)
for(...)
    handle = new DftPassJob { ... }.Schedule(..., handle);

As you can see, a bunch of jobs are created, which I assume should be run in parallel. However, each of the parallel jobs depends on the job handle of the previous loop iteration, which means all the jobs will be executed in sequence, and not in parallel. Is it correct?