Re: Population count in SSE2, again

"James Van Buskirk"  wrote in message
> There was a thread about population count in comp.lang.fortran but
> nobody seemed to like my assembly code over there:

For what it's worth below I give my 32-bit assembler for popcnt.
The format is for lzasm but can easily be changed for nasm.
The two arguments are the pointer and the number of 32-bit words.
The code is from Hackers Delight converted to 32-bit assembler.
On Core2Duo I measured 13 cycles/limb, where limb=4bytes=32bits.
Also I peeked for popcnt in GMP 4.2.2, see:
http://gmplib.org/
They have an IA-64 Itanium popcnt with 1 cycle/limb (64-bit), using the
IA-64 popcnt instruction.
Their table lookup popcnt runs 8 cycle/limb (32-bit).
Clearly your xmm popcnt1 is also very fast, about 1 cycle/limb (32-bit)!
But because it seems more complicated you should also test it for all
possible 128-bit bit patterns.
Regards, Maarten.

PROC    __int_popcnt
ARG     pdata:dword, ndata:dword
push ebp
mov ebp, esp
 push edi
 push esi
 mov edi, [ndata]
 mov esi, [pdata]
 lea esi, [esi-4]
 xor eax, eax
@[EMAIL PROTECTED]
 mov ecx, [esi+4*edi]
 mov edx, ecx
 shr edx, 1
 and ecx, 55555555h
 and edx, 55555555h
 add ecx, edx
 mov edx, ecx
 shr edx, 2
 and ecx, 33333333h
 and edx, 33333333h
 add ecx, edx
 mov edx, ecx
 shr edx, 4
 add ecx, edx
 and ecx, 0F0F0F0Fh
 mov edx, ecx
 shr edx, 8
 add ecx, edx
 mov edx, ecx
 shr edx, 16
 add ecx, edx
 and ecx, 3Fh
 add eax, ecx
 dec edi
 jnz short @[EMAIL PROTECTED]
 pop esi
 pop edi
pop ebp
ret
ENDP    __int_popcnt