Rewrite native FFT benchmarks

[bibikar]

@oleksandr-pavlyk

This PR includes a complete rewrite of native FFT benchmarks using MKL DFTI:

• Unify all benchmarks into one file
• Add support for single-precision dtypes and 4-, 5-, 6-, 7-dimensional FFTs
• Switch to command-line options (getopt_long) instead of env
• Add help message

Copying superfluous harmonics for conjugate-even outputs produced by the FFT for real inputs as is done in mkl_fft is not yet implemented, but using the --rfft argument allows skipping that step for now. I don't think our native benchmarks have ever supported FFTs of real inputs with full outputs.

The benchmark still works on Windows, but only short options are supported there.

usage: ./fft_bench [args] size
Benchmark FFT using Intel(R) MKL DFTI.

FFT problem arguments:
  -t, --threads=THREADS    use THREADS threads for FFT execution
                           (default: use MKL's default)
  -d, --dtype=DTYPE        use DTYPE as the FFT domain. For a list of
                           understood dtypes, use '-d help'.
                           (default: complex128)
  -r, --rfft               do not copy superfluous harmonics when FFT
                           output is even-conjugate, i.e. for real inputs
  -P, --in-place           allow overwriting the input buffer with the
                           FFT outputs
  -c, --cached             use the same DFTI descriptor for the same
                           outer loop, i.e. "cache" the descriptor

Timing arguments:
  -i, --inner-loops=IL     time the benchmark IL times for each printed
                           measurement. Copies are not included in the
                           measurements. (default: 16)
  -o, --outer-loops=OL     print OL measurements. (default: 5)

Output arguments:
  -p, --prefix=PREFIX      output PREFIX as the first value in outputs
                           (default: 'Native-C')
  -H, --no-header          do not output CSV header. This can be useful
                           if running multiple benchmarks back-to-back.
  -h, --help               print this message and exit

The size argument specifies the input matrix size as a tuple of positive
decimal integers, delimited by any non-digit. For example, both
(101, 203, 305) and 101x203x305 denote the same 3D FFT.

[oleksandr-pavlyk]

What is the point of having -DDEBUG alongside -DNDEBUG ?

[bibikar]

I had written some debug messages in an earlier revision enabled with -DDEBUG, but they've been since removed, so this flag is now useless. I went ahead and removed it - thanks!

[oleksandr-pavlyk]

Help seems to suggest that option -i and -o are valid, yet:

(t_scipy-1.4.0) [10:14:08 vmlin native]$ ./fft_bench -t 1 -d float32 -P -c --inner-loops=10 --outer-loops=7 7  # works as advertised
prefix,function,threads,dtype,size,place,cached,time
Native-C,fft,1,float32,7,in-place,cached,0.0092834
Native-C,fft,1,float32,7,in-place,cached,1.9109e-05
Native-C,fft,1,float32,7,in-place,cached,1.8982e-05
Native-C,fft,1,float32,7,in-place,cached,1.9297e-05
Native-C,fft,1,float32,7,in-place,cached,1.8928e-05
Native-C,fft,1,float32,7,in-place,cached,1.8946e-05
Native-C,fft,1,float32,7,in-place,cached,1.8844e-05

(t_scipy-1.4.0) [10:14:15 vmlin native]$ ./fft_bench -t 1 -d float32 -P -c --i=10 --outer-loops=7 7
./fft_bench: option '--i=10' is ambiguous; possibilities: '--inner-loops' '--in-place'
(t_scipy-1.4.0) [10:14:21 vmlin native]$ ./fft_bench -t 1 -d float32 -P -c -i=10 --outer-loops=7 7
./fft_bench: invalid option -- 'i'  # I was expecting this to actually work

[bibikar]

Short options -ioH were missing from the getopt_long call. I added them, but instead of -i=10, it is required to specify -i10 or -i 10.

[oleksandr-pavlyk]

LGTM. Thanks @bibikar

[oleksandr-pavlyk]

👍

[bibikar]

Thanks!