[FFmpeg-devel] [PATCH 1/5] libavcodec: Implementation of 32 bit fixed point FFT

[Zeljko Lukac]

Below are the comparison tests between three different implementations:

1 - Rockbox
2 - Our implementation with fixed-point multiplication from Rockbox
3 - Our original implementation with high accuracy fixed-point multiplication

Speed test was performed by measuring time required for 20000 iterations of each algorithm on Intel Code 2 Duo @ 2.5 GHz.

size=256
T(1) = 249 ms
T(2) = 235 ms
T(3) = 282 ms

size=1024
T(1) = 1421 ms
T(2) = 1344 ms
T(3) = 1562 ms

size=4096
T(1) = 7515 ms
T(2) = 7188 ms
T(3) = 8360 ms

The accuracy test was performed by comparing each algorithm with the reference double precision floating point implementation with pseudorandom input and maximum supported FFT size (4096).

RMSE(1) = 48.18
RMSE(2) = 43.47
RMSE(3) =  8.70


Best regards,
Zeljko

________________________________________
From: Michael Niedermayer [michaelni at gmx.at]
Sent: Tuesday, June 18, 2013 6:56 PM
To: FFmpeg development discussions and patches
Cc: Mirjana Vulin; Nedeljko Babic; Zeljko Lukac; Bojan Zivkovic
Subject: Re: [FFmpeg-devel] [PATCH 1/5] libavcodec: Implementation of 32 bit fixed point FFT

On Mon, Jun 03, 2013 at 04:11:12PM +0200, Nedeljko Babic wrote:
> From: Nedeljko Babic <nbabic at mips.com>
>
> Iterative implementation of 32 bit fixed point split-radix FFT.
> Max FFT that can be calculated currently is 2^12.

How does this compare to: rockboxs fixed point fft
http://git.rockbox.org/?p=rockbox.git;a=tree;f=apps/codecs/lib;h=4d96a94d432e0d3dbeee88bc5b5573409ef0e662;hb=a0009907de7a0107d49040d8a180f140e2eff299

?

That is, which is faster ?
iam asking because I want the faster one of the 2 :)

[...]
--
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Into a blind darkness they enter who follow after the Ignorance,
they as if into a greater darkness enter who devote themselves
to the Knowledge alone. -- Isha Upanishad