Re: [PATCH v1] x86/lib: Optimize 8x loop and memory clobbers in csum_partial.c

From: Noah Goldstein
Date: Fri Nov 26 2021 - 15:07:48 EST


On Fri, Nov 26, 2021 at 1:21 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Fri, Nov 26, 2021 at 11:14 AM Noah Goldstein <goldstein.w.n@xxxxxxxxx> wrote:
> >
> > On Fri, Nov 26, 2021 at 12:50 PM Noah Goldstein <goldstein.w.n@xxxxxxxxx> wrote:
> > >
> > > On Fri, Nov 26, 2021 at 12:27 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
> > > >
> > > > On Fri, Nov 26, 2021 at 10:17 AM Noah Goldstein <goldstein.w.n@xxxxxxxxx> wrote:
> > > > >
> > > >
> > > > >
> > > > > Makes sense. Although if you inline I think you definitely will want a more
> > > > > conservative clobber than just "memory". Also I think with 40 you also will
> > > > > get some value from two counters.
> > > > >
> > > > > Did you see the number/question I posted about two accumulators for 32
> > > > > byte case?
> > > > > Its a judgement call about latency vs throughput that I don't really have an
> > > > > answer for.
> > > > >
> > > >
> > > > The thing I do not know is if using more units would slow down the
> > > > hyper thread ?
> >
> > Did some quick tests with the latency/throughput benchmarks running
> > in parallel on two hyperthreads on the same processors. The 32 byte case
> > latency advantage goes with 2 accum and there is still a slight regression
> > in throughput. The larger cases that hit the loop still still have improvements
> > both in tput and latency with 2 accum.
>
> Great. I also played with rorx instruction, because it removes one
> "adcl $0,..." step

Bright :) but it will need a BMI support check.

>
> __wsum ipv6_csum_partial(const void *buff, int len, __wsum sum)
> {
> u64 temp64;
> u64 tmp;
> u32 res32;
>
> if (unlikely(len != 40))
> return csum_partial(buff, len, sum);
>
> temp64 = (__force u64)sum;
> asm("addq 0*8(%[src]),%[temp64]\n\t"
> "adcq 1*8(%[src]),%[temp64]\n\t"
> "adcq 2*8(%[src]),%[temp64]\n\t"
> "adcq 3*8(%[src]),%[temp64]\n\t"
> "adcq 4*8(%[src]),%[temp64]\n\t"
> "mov %k[temp64],%[res32]\n\t"
> "rorx $32,%[temp64],%[temp64]\n\t"
> "adcl %k[temp64],%[res32]\n\t"
> "adcl $0,%[res32]"
> : [temp64] "+r" (temp64), [res32] "=r" (res32)
> : [src] "r" (buff)
> : "memory");
> return (__force __wsum)res32;
> }

I actually get better performance in hyperthread benchmarks with 2 accum:

Used:

u64 res;
temp64 = (__force uint64_t)sum;
asm("movq 0*8(%[src]),%[res]\n\t"
"addq 1*8(%[src]),%[res]\n\t"
"adcq 2*8(%[src]),%[res]\n\t"
"adcq $0, %[res]\n"
"addq 3*8(%[src]),%[temp64]\n\t"
"adcq 4*8(%[src]),%[temp64]\n\t"
"adcq %[res], %[temp64]\n\t"
"mov %k[temp64],%k[res]\n\t"
"rorx $32,%[temp64],%[temp64]\n\t"
"adcl %k[temp64],%k[res]\n\t"
"adcl $0,%k[res]"
: [temp64] "+r"(temp64), [res] "=&r"(res)
: [src] "r"(buff)
: "memory");
return (__force __wsum)res;

w/ hyperthread:
size, 2acc lat, 1acc lat, 2acc tput, 1acc tput
40, 6.511, 7.863, 6.177, 6.157

w/o hyperthread:
size, 2acc lat, 1acc lat, 2acc tput, 1acc tput
40, 5.577, 6.764, 3.150, 3.210