Comments on: Ridiculous UTF-8 character counting

By: Porges

Porges — Thu, 05 Jun 2008 23:40:28 +0000

Joshua,

Unfortunately, AFAICT, there is no way to do this from the C code. If I edit the assembly manually I lose about 0.001 s

By: Wade

Wade — Thu, 05 Jun 2008 22:09:26 +0000

Shaneal,

It is, and it isnt masturabation. Somewhere, someone will need this code because they need to make their strlens _really_ fast.

By: ast_tree

ast_tree — Thu, 05 Jun 2008 20:08:45 +0000

Are you running an Intel© viral ad? What’s next, multithread/multiprocess/MPI Cluster version?

Maybe if you try the profile mode of icc you can speed up vectorization or lower the cache-miss (if any remains).

By: Joshua Haberman

Joshua Haberman — Thu, 05 Jun 2008 18:30:39 +0000

Inside the loop, you should load with movdqa instead of movdqu, since you know the addresses will be aligned.

By: mark++

mark++ — Thu, 05 Jun 2008 18:05:59 +0000

Core2 and AMD’s Quad Core parts have 128-bit wide SSE execution units. Prior architectures from both companies, including classic Opteron, used double-pumped 64-bit execution units. This will likely make a difference.

By: Matthew Hartman

Matthew Hartman — Thu, 05 Jun 2008 16:59:31 +0000

Keep masturbating guys (sorry, I couldn’t resist). I love reading this kind of constructive banter.

By: Stephen

Stephen — Thu, 05 Jun 2008 16:49:11 +0000

I see your code has way more variance than the others, probably due to the hamming lookup table.

You’d be better off accumulating the counters and doing a horizontal add at the end.

e.g. put a nested loop into the block loop which executes at most 255 times, then accumulate the bytes to the sum and continue.

By: Shaneal Manek

Shaneal Manek — Thu, 05 Jun 2008 15:38:56 +0000

Isn’t this just mental masturbation at some point?

But boy it is fun

I really need to learn x86 assembler – I only know a bit of MIPs. The instruction set seems huge …

By: Vineet Kumar

Vineet Kumar — Thu, 05 Jun 2008 15:37:56 +0000

That oughta be “damn thee”. =)

(Also, the required name and email fields on this comment form are hard to decipher. Icons too small and similar, with not enough meaning.)

By: Colin Percival

Colin Percival — Thu, 05 Jun 2008 15:29:55 +0000

On my Opteron, I get your SSE2 code taking 18.7ms (compared to 14.1ms for my non-sse2 version). Clearly your CPU, whatever it is, has faster SSE2 support than my Opteron.

I’m going to see if I can coax some more performance out of the SSE2 code, though…