Author: Gerd Isenberg
Date: 11:05:33 07/21/04
Go up one level in this thread
On July 21, 2004 at 13:54:12, Gerd Isenberg wrote:
>On July 20, 2004 at 12:31:59, Anthony Cozzie wrote:
>
>>What do you think of the following C code:
>>
>>int bb_dot_product(bitboard a, unsigned char *weights)
>>{
>> bitboard t, t1, *_weights = weights;
>> static bitboard table[256] = {correct translations, e.g. 0xFF -> 0xffffffff}
>>
>> //we count on the compiler to unroll this loop.
>> for(i = 0; i < 8; i++, a ) {
>> t = table[(a >> i*8) & 0xFF] & weights[i];
>> t1 = t;
>> t << 8;
>> sum += (t & 0x00FF00FF00FF00FF) + (t1 & 0x00FF00FF00FF00FF);
>> }
>>
>> sum = (sum & 0x0000FFFF0000FFFF) + ((sum >> 16) & 0x0000FFFF0000FFFF);
>> sum = ((sum >> 32) + sum & 0x00000000FFFFFFFF);
>> return sum;
>>}
>>
>>It has several advantages: Can use full 0-255 for each weight, the table does
>>not have to be rotated, and there is no penalty for moving between the integer
>>and MMX pipes.
>>
>>OTOH, this solution is also much less cache friendly, requiring maybe 2x the
>>number of instructions and also needed 2KB of data cache.
>>
>>anthony
>
>probably some minor improvements.
>Save the inner shift by building two intermediate results.
>Whether the byte access it worth instead of shift/and?
>On x86-32 it was.
>
> unsigned char *bptr = & (unsigned char) a;
> sum0 = 0, sum1 = 0;
> for(i = 0; i < 8; i++, ptr++ ) {
> t = table[*bptr] & weights[i];
> sum0 += t & 0x00FF00FF00FF00FF
> sum1 += t & 0xFF00FF00FF00FF00;
Ok, that doesn't work correctly because the highest byte has no overflow bits.
Therefore the >> 8 is necessary inside the loop!
> }
> sum = sum0 + (sum1>>8);
> sum = (sum & 0x0000FFFF0000FFFF) + ((sum >> 16) & 0x0000FFFF0000FFFF);
> sum = ((sum >> 32) + sum & 0x00000000FFFFFFFF);
> return sum;
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.