1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
|
TODO before FFTW-$2\pi$:
* DCT/DST codelets? which kinds?
* investigate the addition-chain trig computation
* I can't believe that there isn't a closed form for the omega
array in Rader.
* convolution problem type(s)
* Explore the idea of having n < 0 in tensors, possibly to mean
inverse DFT.
* better estimator: possibly, let "other" cost be coef * n, where
coef is a per-solver constant determined via some big numerical
optimization/fit.
* vector radix, multidimensional codelets
* it may be a good idea to unify all those little loops that do
copying, (X[i], X[n-i]) <- (X[i] + X[n-i], X[i] - X[n-i]),
and multiplication of vectors by twiddle factors.
* Pruned FFTs (basically, a vecloop that skips zeros).
* Try FFTPACK-style back-and-forth (Stockham) FFT. (We tried this a
few years ago and it was slower, but perhaps matters have changed.)
* Generate assembly directly for more processors, or maybe fork gcc. =)
* ensure that threaded solvers generate (block_size % 4 == 0)
to allow SIMD to be used.
* memoize triggen.
* eliminate alignment hacks, which ought to be obsolete by now.
|
