Utility used to pre-compute values of a real-valued function, in order to speedup its evaluation by returning approximate values.

Copy the following files in your project: precomputation.c and precomputation.h. The other files are for demonstration purposes. Below is an example of approximating expf between -10 and 10 with 25000 samples (no SIMD):

Precomputation *precomputation = initPrecomputation(expf, 25000, -10.f, 10.f);
float approx = approximation(precomputation, 1.f); // 2.717305
freePrecomputation(&precomputation);

The following benchmark has been done on the expf function between -10 and 10. Precision tests make use of 100M samples, and time tests use 3.9 GB of generated values aligned on 16-byte boundaries. Note that measured times depend greatly on the ratio between the samples number and the CPU cache size, here of 6 MB (Intel i5-7300HQ 2.50GHz). Values reported per cell are the speedup factor, and the mean relative error on the generated samples:

Compiling is done by running make. Note that if using a custom made makefile, one should use as CFLAGS the option -march=native, for enabling SSE and AVX instructions sets.

For testing the library:

./precomptest.exe