Sampling with floating-point distributions seems to be significantly slower about fast-loaded-dice-roller HOT 3 CLOSED

Thanks for the runtime measurements. These numbers are in line with expectation for preprocessing floats, since the floating-point algorithm (which itself is not very optimized) is significantly more expensive than the integer arithmetic algorithm.

The implementation in C is much faster. We should consider using e.g., Cython to create a Python API.

from fast-loaded-dice-roller.

Ok, thanks.
FYI the runtime measurements only include sampling, not preprocessing. I also noted that using the alternative floating-point preprocessing algorithm mentioned in #4 resulted in runtimes much more comparable to the integer algorithm.

python -m timeit -s "from fldr.fldrf import fldr_preprocess_float_c as fldrf_preprocess; from fldr import fldr_sample; l = range(10000); x = fldrf_preprocess([1/len(l) for _ in l])" "fldr_sample(x)"

50000 loops, best of 5: 7.16 usec per loop

Thanks again for the great work.

from fast-loaded-dice-roller.

Thanks again for confirming. I will change the default implementation to use the alternative preprocessing algorithm in light of #4 and #5.

FYI the runtime measurements only include sampling, not preprocessing.

Good point, I missed the fact that your timeit script uses -s for "setup".

from fast-loaded-dice-roller.