Subject: Re: [boost] sorting floats by casting to integer
From: Steven Ross (spreadsort_at_[hidden])
Date: 2009-07-06 17:22:37
Thanks Phil and Vladimir for pointing out the denormalization issue. I have
not seen this issue on PowerPC, which was my main test system for sorting
for years, and I still don't see why the processor treats them differently,
but it clearly does.
The performance impact of denormalized numbers is so huge that it doubles
the runtime of std::sort on a large list of completely random bits being
sorted (200 million), even though only 1 in 256 is denormalized, but
normally people won't use them, so they're not a realistic case.
Given that input, I'm going to scrap my sort-as-integer approach, and stick
with the cast-to-integer approach already in my float_sort implementation,
using no additional memory overhead.
I'll also do my testing with sse optimizations, and make sure to denormalize
my float_sort tests so that performance results are more realistic. I
expect it to drop to a roughly 2X improvement, which is about what the other
algorithms (integer_sort, string_sort) get.
I apologize for being stubborn on this issue.