Chris@19: TODO before FFTW-$2\pi$: Chris@19: Chris@19: * Wisdom: make it clear that it is specific to the exact fftw version Chris@19: and configuration. Report error codes when reading wisdom. Maybe Chris@19: have multiple system wisdom files, one per version? Chris@19: Chris@19: * DCT/DST codelets? which kinds? Chris@19: Chris@19: * investigate the addition-chain trig computation Chris@19: Chris@19: * I can't believe that there isn't a closed form for the omega Chris@19: array in Rader. Chris@19: Chris@19: * convolution problem type(s) Chris@19: Chris@19: * Explore the idea of having n < 0 in tensors, possibly to mean Chris@19: inverse DFT. Chris@19: Chris@19: * better estimator: possibly, let "other" cost be coef * n, where Chris@19: coef is a per-solver constant determined via some big numerical Chris@19: optimization/fit. Chris@19: Chris@19: * vector radix, multidimensional codelets Chris@19: Chris@19: * it may be a good idea to unify all those little loops that do Chris@19: copying, (X[i], X[n-i]) <- (X[i] + X[n-i], X[i] - X[n-i]), Chris@19: and multiplication of vectors by twiddle factors. Chris@19: Chris@19: * Pruned FFTs (basically, a vecloop that skips zeros). Chris@19: Chris@19: * Try FFTPACK-style back-and-forth (Stockham) FFT. (We tried this a Chris@19: few years ago and it was slower, but perhaps matters have changed.) Chris@19: Chris@19: * Generate assembly directly for more processors, or maybe fork gcc. =) Chris@19: Chris@19: * ensure that threaded solvers generate (block_size % 4 == 0) Chris@19: to allow SIMD to be used. Chris@19: Chris@19: * memoize triggen.