Re: Multithreaded FFTW3 in octave?

[Andreas Weber]

On 17.12.2012 01:05, Andreas Weber wrote:
> Hi Mike, I've attached a "first shot" patch which
I made some simple 1D FFT R2C (real to complex) benchmarks (code follows).
Here is the plot: http://i.imgur.com/omFUi.png
Here you can see the increase of execution time for small FFTs, probably 
due to splitting for multithreading.
Should we therefore disable multithreading per default or is this 
negligible?

Has anyone done a complete FFT benchmark for octave which I can use for 
further tests?

By saying "complete" I mean "real to complex", "complex to complex", 1D, 
2D, 3D, double and single precision,
with sizes powers of 2, powers of 10, sizes of 2^a * 3^b * 5^c * 7^d * 
11^e * 13^f, where e+f is either 0 or
and all combinations.

Here the result of attached benchmark on an AMD Phenom(tm) II X6 1055T 
Processor @ 2.8GHz

Number of FFTW threads = 1
time for 262144x1 FFT 0.024335 s
time for 524288x1 FFT 0.050655 s
time for 1048576x1 FFT 0.116450 s
time for 2097152x1 FFT 0.206864 s
time for 4194304x1 FFT 0.461818 s
time for 8388608x1 FFT 0.874915 s
time for 16777216x1 FFT 1.948641 s
time for 33554432x1 FFT 3.694970 s
time for 100000x1 FFT 0.020179 s
time for 1000000x1 FFT 0.070333 s
time for 10000000x1 FFT 0.875357 s
time for 8421875x1 FFT 0.643188 s
time for 11790625x1 FFT 1.021335 s
time for 13934375x1 FFT 1.186604 s
time for 42109375x1 FFT 3.866208 s

Number of FFTW threads = 6
time for 262144x1 FFT 0.027172 s
time for 524288x1 FFT 0.044611 s
time for 1048576x1 FFT 0.083005 s
time for 2097152x1 FFT 0.152968 s
time for 4194304x1 FFT 0.264076 s
time for 8388608x1 FFT 0.523213 s
time for 16777216x1 FFT 1.020191 s
time for 33554432x1 FFT 2.161175 s
time for 100000x1 FFT 0.041788 s
time for 1000000x1 FFT 0.076238 s
time for 10000000x1 FFT 0.560943 s
time for 8421875x1 FFT 0.263840 s
time for 11790625x1 FFT 0.383473 s
time for 13934375x1 FFT 0.462033 s
time for 42109375x1 FFT 1.154992 s

Regards, Andy

fftw_bench.m
Description: Text Data