Re: Population count in SSE2, again

Terence <spamtrap@[EMAIL PROTECTED]
> wrote in part:
> "James Van Buskirk"  wrote in message
> (code)  [clever XMM bitslice code omitted]
 
> Someone tell me why 21 instructions and a branch, 
> per word, is somehow better/faster than

[byte-wise  LUT pseudocode snipped]

Easy, it just is -- 4.7 bytes/clock .  Your LUT would
be good at 0.5 and have real trouble breaking 1.0 .

Why is equally easy to see -- James "word" is a 128-bit
wide XMM register (presumably zero padded].  That is 16
bytes at a time.  Furthermore the code is unrolled x8,
meaning 128 _bytes_ per 61 instruction inner loop.  
James' timing gives a reasonable 2.2 ipc.

I still have this nagging feeling there's something
faster with multiplying by magic numbers.

-- Robert