Reduction of Gibbs Phenomenon on FFT compression #124
Conversation
|
@rukai asking for your help on this review due to the way vectors/arrays are being padded, sliced and cloned. Any suggestions to improve the performance of this code is welcome! Thanks, Carlos |
| /// Given a number N, checks what is the next number that is in the form of (2^N * 3^M) | ||
| pub fn next_size(mut n: usize) -> usize { | ||
| n += 1; | ||
| while !is_decomposable(n) { |
There was a problem hiding this comment.
all this looping looks slow, are you sure there isnt a way to get the same result with an expression instead?
There was a problem hiding this comment.
Prime numbers are a pain! And I don't see how I can improve this. Let me see if I find a better solution.
This is only run once per frame on compression and decompression.
Given that we force frames into a subset of sizes (and power of 2^N) I don't expect this to loop for long. On larger numbers it might become problematic.
Alternative is providing a table of next number for the expected frame sizes and calculate for everything else? (We have some pre-defined frame sizes hardcoded anyway).
There was a problem hiding this comment.
A table does sound like a good idea.
Could be left for follow up work.
Everyone makes mistakes, the tests did their job and that is why I caught them! :) |
Current FFT implementation disregards the Gibbs phenomenon at the boundaries of the frame. This creates artifacts that need correction at those transitions.
This PR fixes the Gibbs phenomenon at the frame edge by expanding the frames during the FFT calculation (and inverse calculation) with either the first or last value of the frame.
This creates an attenuation effect without impacting the compression ratio. The tradeoff is more memory consumption (up to 5%) on the use of the FFT compressor.
This image illustrates a frame transition with the current code base and the one purposed by this fix.
