4 ms·

The Shootout's C implementations do occasionally turn into a festival of SSE intrinsics and the associated unrolling. Some benchmarks, anyway. Still, not using

The Shootout's C implementations do occasionally turn into a festival of SSE intrinsics and the associated unrolling. Some benchmarks, anyway. Still, not using SSE when it's available is dumb and perhaps all the other languages need to start playing, too. A little tricker in Python, of course....

kingkilr 16y ago
Well, that's what the compiler is for! PyPy uses SSE under the hood for various float handing things.
- onan_barbarian 16y ago
  Unfortunately, SSE - especially in integer-land, so I suppose we're talking SSE2 and beyond - often requires more specialist care and feeding than any automatic method (compiler or run-time) can provide. Some cases aren't hard to pick up (e.g. bulk operations on big arrays) but others require trickery of the kind that compilers usually don't (or couldn't) have. This isn't made easier by the notoriously non-orthogonal nature of the SSE integer operations and the rather limited number of ways that you can get in and out of SSE-land (to, say, affect a conditional or get something into a GPR).
- igouy 16y ago
  Also "J2SE platform version 1.4.2 now uses SSE and SSE2 instruction sets for floating point computations".
Luyt 16y ago
I once wrote a function which uses MMX to convert a (possibly very large) string to lowercase. That was pretty much faster than the library's strlwr() function. http://codepad.org/BeDqS1Ws http://codepad.org/BeDqS1Ws
- onan_barbarian 16y ago
  Great stuff. I think there are some potential tricks here to reduce the number of comparisons - maybe do a parallel subtract by k to pull down 'A' to -128 (smallest possible byte) then do your comparison against (ord('Z')-k). Or maybe push up 'Z' to +127..? That way you can get a single comparison and can replace a pcmpgtb and pand with a single subtract. Then switch it to SSE2, unroll and you're good to go. Alternately, http://www.azillionmonkeys.com/qed/asmexample.html http://www.azillionmonkeys.com/qed/asmexample.html in section 11 ("Converting Uppercase") contains a brainsmashing version of this entirely in SWAR ("SIMD Within a Register"), which could be adapted with a certain amount of pain (largely due to the absence of a double-quadword bitshift in SSE2, which is retardlepated).

kingkilr 16y ago

Well, that's what the compiler is for! PyPy uses SSE under the hood for various float handing things.

onan_barbarian 16y ago
Unfortunately, SSE - especially in integer-land, so I suppose we're talking SSE2 and beyond - often requires more specialist care and feeding than any automatic method (compiler or run-time) can provide. Some cases aren't hard to pick up (e.g. bulk operations on big arrays) but others require trickery of the kind that compilers usually don't (or couldn't) have. This isn't made easier by the notoriously non-orthogonal nature of the SSE integer operations and the rather limited number of ways that you can get in and out of SSE-land (to, say, affect a conditional or get something into a GPR).
igouy 16y ago
Also "J2SE platform version 1.4.2 now uses SSE and SSE2 instruction sets for floating point computations".

Luyt 16y ago

I once wrote a function which uses MMX to convert a (possibly very large) string to lowercase. That was pretty much faster than the library's strlwr() function. http://codepad.org/BeDqS1Ws http://codepad.org/BeDqS1Ws

onan_barbarian 16y ago
Great stuff. I think there are some potential tricks here to reduce the number of comparisons - maybe do a parallel subtract by k to pull down 'A' to -128 (smallest possible byte) then do your comparison against (ord('Z')-k). Or maybe push up 'Z' to +127..? That way you can get a single comparison and can replace a pcmpgtb and pand with a single subtract. Then switch it to SSE2, unroll and you're good to go. Alternately, http://www.azillionmonkeys.com/qed/asmexample.html http://www.azillionmonkeys.com/qed/asmexample.html in section 11 ("Converting Uppercase") contains a brainsmashing version of this entirely in SWAR ("SIMD Within a Register"), which could be adapted with a certain amount of pain (largely due to the absence of a double-quadword bitshift in SSE2, which is retardlepated).