TODO before FFTW-$2\pi$:

* figure out how to autodetect NEON at runtime

* figure out the arm cycle counter business

* Wisdom: make it clear that it is specific to the exact fftw version
  and configuration.  Report error codes when reading wisdom.  Maybe
  have multiple system wisdom files, one per version?

* 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.