В итоге я наковырял такую функцию, она хоть и значительно быстрее простого цикла, но все равно на ПК медленнее где-то на 30-40%, а на нетбуке на 50-70% чем memcpy из VS2010
| Код |
static void TestMemcpy_sse(void* pDest, void* pSrc, int numOfBytes)
{
if(numOfBytes < CACHE_LINE_SIZE)
{
_asm
{
mov ecx, numOfBytes
mov edi, pDest
mov esi, pSrc
rep movsb
}
return;
}
int destOffset = (int)pDest % 16;
if(destOffset > 0)
{
_asm
{
mov ecx, destOffset
mov edi, numOfBytes
sub edi, ecx
mov numOfBytes, edi
mov edi, pDest
mov esi, pSrc
rep movsb
add edi, destOffset
add esi, destOffset
mov pDest, edi
mov pSrc, esi
}
}
if((int)pSrc % 16 == 0)
{
_asm
{
mov esi, pSrc
prefetchnta [esi]
mov ecx, numOfBytes
mov edi, pDest
cmp ecx, 128
jb A64
A128:
prefetchnta XMMWORD PTR[esi + 16 * 4]
movdqa xmm0, XMMWORD PTR[esi]
movdqa xmm1, XMMWORD PTR[esi + 16 * 1]
movdqa xmm2, XMMWORD PTR[esi + 16 * 2]
movdqa xmm3, XMMWORD PTR[esi + 16 * 3]
movdqa XMMWORD PTR[edi], xmm0
movdqa XMMWORD PTR[edi + 16 * 1], xmm1
movdqa XMMWORD PTR[edi + 16 * 2], xmm2
movdqa XMMWORD PTR[edi + 16 * 3], xmm3
prefetchnta XMMWORD PTR[esi + 16 * 8]
movdqa xmm4, XMMWORD PTR[esi + 16 * 4]
movdqa xmm5, XMMWORD PTR[esi + 16 * 5]
movdqa xmm6, XMMWORD PTR[esi + 16 * 6]
movdqa xmm7, XMMWORD PTR[esi + 16 * 7]
movdqa XMMWORD PTR[edi + 16 * 4], xmm4
movdqa XMMWORD PTR[edi + 16 * 5], xmm5
movdqa XMMWORD PTR[edi + 16 * 6], xmm6
movdqa XMMWORD PTR[edi + 16 * 7], xmm7
add esi, 128
prefetchnta [esi]
add edi, 128
sub ecx, 128
cmp ecx, 128
jae A128
A64:
cmp ecx, 64
jb A32
prefetchnta XMMWORD PTR[esi]
sub ecx, 64
movdqa xmm0, XMMWORD PTR[esi]
movdqa xmm1, XMMWORD PTR[esi + 16 * 1]
movdqa xmm2, XMMWORD PTR[esi + 16 * 2]
movdqa xmm3, XMMWORD PTR[esi + 16 * 3]
movdqa XMMWORD PTR[edi], xmm0
movdqa XMMWORD PTR[edi + 16 * 1], xmm1
movdqa XMMWORD PTR[edi + 16 * 2], xmm2
movdqa XMMWORD PTR[edi + 16 * 3], xmm3
add esi, 64
add edi, 64
A32:
prefetchnta XMMWORD PTR[esi]
cmp ecx, 32
jb A16
sub ecx, 32
movdqa xmm4, XMMWORD PTR[esi]
movdqa xmm5, XMMWORD PTR[esi + 16 * 1]
movdqa XMMWORD PTR[edi], xmm4
movdqa XMMWORD PTR[edi + 16 * 1], xmm5
add esi, 32
add edi, 32
A16:
cmp ecx, 16
jb A
sub ecx, 16
movdqa xmm6, XMMWORD PTR[esi]
movdqa XMMWORD PTR[edi], xmm6
A:
sfence
}
}
else
{
_asm
{
mov ecx, numOfBytes
mov esi, pSrc
prefetchnta [esi]
mov edi, pDest
cmp ecx, 128
jb U64
U128:
prefetchnta XMMWORD PTR[esi + 16 * 4]
movdqu xmm0, XMMWORD PTR[esi]
movdqu xmm1, XMMWORD PTR[esi + 16 * 1]
movdqu xmm2, XMMWORD PTR[esi + 16 * 2]
movdqu xmm3, XMMWORD PTR[esi + 16 * 3]
movdqu XMMWORD PTR[edi], xmm0
movdqu XMMWORD PTR[edi + 16 * 1], xmm1
movdqu XMMWORD PTR[edi + 16 * 2], xmm2
movdqu XMMWORD PTR[edi + 16 * 3], xmm3
prefetchnta XMMWORD PTR[esi + 16 * 8]
movdqu xmm4, XMMWORD PTR[esi + 16 * 4]
movdqu xmm5, XMMWORD PTR[esi + 16 * 5]
movdqu xmm6, XMMWORD PTR[esi + 16 * 6]
movdqu xmm7, XMMWORD PTR[esi + 16 * 7]
movdqu XMMWORD PTR[edi + 16 * 4], xmm4
movdqu XMMWORD PTR[edi + 16 * 5], xmm5
movdqu XMMWORD PTR[edi + 16 * 6], xmm6
movdqu XMMWORD PTR[edi + 16 * 7], xmm7
add esi, 128
prefetchnta [esi]
add edi, 128
sub ecx, 128
cmp ecx, 128
jae U128
U64:
cmp ecx, 64
jb U32
prefetchnta XMMWORD PTR[esi]
sub ecx, 64
movdqu xmm0, XMMWORD PTR[esi]
movdqu xmm1, XMMWORD PTR[esi + 16 * 1]
movdqu xmm2, XMMWORD PTR[esi + 16 * 2]
movdqu xmm3, XMMWORD PTR[esi + 16 * 3]
movdqu XMMWORD PTR[edi], xmm0
movdqu XMMWORD PTR[edi + 16 * 1], xmm1
movdqu XMMWORD PTR[edi + 16 * 2], xmm2
movdqu XMMWORD PTR[edi + 16 * 3], xmm3
add esi, 64
add edi, 64
U32:
prefetchnta XMMWORD PTR[esi]
cmp ecx, 32
jb U16
sub ecx, 32
movdqu xmm4, XMMWORD PTR[esi]
movdqu xmm5, XMMWORD PTR[esi + 16 * 1]
movdqu XMMWORD PTR[edi], xmm4
movdqu XMMWORD PTR[edi + 16 * 1], xmm5
add esi, 32
add edi, 32
U16:
cmp ecx, 16
jb U
sub ecx, 16
movdqu xmm6, XMMWORD PTR[esi]
movdqu XMMWORD PTR[edi], xmm6
U:
sfence
}
}
int residue = numOfBytes % 16;
if(residue > 0)
{
_asm
{
mov ecx, residue
mov ebx, numOfBytes
mov edi, pDest
add edi, ebx
sub edi, ecx
mov esi, pSrc
add esi, ebx
sub esi, ecx
rep movsb
}
}
return;
}
|
|
)
