On 18/07/16 16:06, Tvrtko Ursulin wrote:
On 18/07/16 14:46, Tvrtko Ursulin wrote:
[snip]
This version generates the smallest code:
static void __memcpy_ntdqa(struct qw2 *dst, const struct qw2 *src, unsigned
long len)
{
unsigned long l4;
kernel_fpu_begin();
l4 = len / 4;
while (l4) {
asm("movntdqa (%0), %%xmm0" :: "r" (src), "m" (src[0]));
asm("movntdqa 16(%0), %%xmm1" :: "r" (src), "m" (src[1]));
asm("movntdqa 32(%0), %%xmm2" :: "r" (src), "m" (src[2]));
asm("movntdqa 48(%0), %%xmm3" :: "r" (src), "m" (src[3]));
asm("movaps %%xmm0, (%1)" : "=m" (dst[0]) : "r" (dst));
asm("movaps %%xmm1, 16(%1)" : "=m" (dst[1]) : "r" (dst));
asm("movaps %%xmm2, 32(%1)" : "=m" (dst[2]) : "r" (dst));
asm("movaps %%xmm3, 48(%1)" : "=m" (dst[3]) : "r" (dst));
src += 4;
dst += 4;
l4--;
}
len %= 4;
while (len) {
asm("movntdqa (%0), %%xmm0" :: "r" (src), "m" (src[0]));
asm("movaps %%xmm0, (%1)" : "=m" (dst[0]) : "r" (dst));
src++;
dst++;
len--;
}
kernel_fpu_end();
}
Although I still haven't figured out a way to convince it to use
the same registers for src and dest between the two loops.
I remembered one famous interview question, along the lines of, "what
is the code below doing". Translated to this example:
static void __memcpy_ntdqa(struct qw2 *dst, const struct qw2 *src, unsigned
long len)
{
unsigned long n;
kernel_fpu_begin();
Bugfix here:
+ len /= 16;
n = (len + 3) / 4;
switch (len % 4) {
case 0: do { asm("movntdqa %1, %%xmm0\n"
"movaps %%xmm0, %0\n" : "=m" (*dst): "m" (*src));
src++; dst++;
case 3: asm("movntdqa %1, %%xmm1\n"
"movaps %%xmm1, %0\n" : "=m" (*dst): "m" (*src));
src++; dst++;
case 2: asm("movntdqa %1, %%xmm2\n"
"movaps %%xmm2, %0\n" : "=m" (*dst): "m" (*src));
src++; dst++;
case 1: asm("movntdqa %1, %%xmm3\n"
"movaps %%xmm3, %0\n" : "=m" (*dst): "m" (*src));
src++; dst++;
} while (--n > 0);
}
kernel_fpu_end();
}
:D
No idea if loads/stores can run async in this case.
It seems to be equally fast as the unrolled loop. And it generates the
smallest code. :)
I would also suggest a "likely (len)" in the original patch since zero
length copies are not expected and that also manages to shrink the code
a bit.
Regards,
Tvrtko
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/intel-gfx
