Things you have to decide before starting:

* Where the library will be installed, normally /usr/local/djbfft. To
change this directory, edit conf-home now.

* What compiler to use, normally gcc -O1 -fomit-frame-pointer. The
performance of djbfft depends on your choice of compiler. To change the
compiler, edit conf-cc now.


How to install:

 1. Compile the library and test programs:
       % make

 2. Install the library, include files, and man pages:
       # make setup


How to test:

 3. Check the cumulative errors for a spanning set of transforms:
       % ./accuracy
    Typical totals range up to 10^{-8} for single-precision transforms
    and 10^{-25} for double-precision transforms. If you see any numbers
    larger than 10^{-5}, please let me know.

 4. See how fast the transforms are:
       % ./speed
    The results are for a complex DFT, then for the scaled inverse, for
    each size; first for single precision, then for double precision.
    speed does 10 timings at each size and displays the last 5.

 5. Check the cumulative errors for some pseudorandom real convolutions:
       % ./accuconv
    Typical results range up to 10^{-12} for single-precision
    convolutions and 10^{-29} for double-precision convolutions. If you
    see any numbers larger than 10^{-9}, please let me know.

 6. See how fast the convolutions are:
       % ./speedconv
    The results are for a complex cyclic convolution against a
    precomputed filter, then a real cyclic convolution twice as large,
    for each size.


That's it! To report success:
       % ( echo 'First M. Last'; cat `cat SYSDEPS`; ./speed; ./speedconv ) \
         | mail djb-qst@cr.yp.to
Replace First M. Last with your name.
