Re: [Qemu-devel] [PATCH v1 12/14] hostfloat: support float32/64 square root

[Emilio G. Cota]

On Thu, Mar 22, 2018 at 01:29:23 +0000, Alex Bennée wrote:
> Emilio G. Cota <address@hidden> writes:
> 
> > Performance results for fp-bench run under aarch64-linux-user
> > on an Intel(R) Core(TM) i7-4790K CPU @ 4.00GHz host:
> >
> > - before:
> > sqrt-single: 13.23 MFlops
> > sqrt-double: 13.24 MFlops
> >
> > - after:
> > sqrt-single: 15.02 MFlops
> > sqrt-double: 15.07 MFlops
> >
> > Note that sqrt in soft-ft is relatively fast, which means
> > that fp-bench is not very sensitive to changes to sqrt's
> > emulation speed.
> 
> Weird, I thought we had slowed it down quite a bit in the re-factor as
> we eschewed the estimate step for an easier to read but slower iterative
> process. That's why I chose sqrt for my hostfp hack experiment.

Yes, my first statement ("soft-ft is relatively fast") is
wrong. Sorry about that, I thought I had deleted it but it
slipped through.

What I should have said (but decided against to keep the commit log
short) is that fp-bench doesn't do a good job in being sensitive
to the performance of the sqrt instruction, so even if got it
to take 0 time we'd still get a small speedup.

Just realised that this happens because ~50% of the inputs are
negative, which will go through some very slow paths. This ends
up showing in perf like this:

# Overhead  Command   Shared Object      Symbol
# ........  ........  .................  ...........................
#
    61.74%  fp-bench  fp-bench           [.] main
    22.58%  fp-bench  libm-2.23.so       [.] __kernel_standard
     6.22%  fp-bench  libm-2.23.so       [.] __kernel_standard_f
     5.21%  fp-bench  libm-2.23.so       [.] __sqrtf
     2.17%  fp-bench  fp-bench           [.] _init
     1.91%  fp-bench  [kernel.kallsyms]  [k] __call_rcu.constprop.70
     0.18%  fp-bench  [kernel.kallsyms]  [k] cpumask_any_but
     0.01%  perf      [kernel.kallsyms]  [k] native_iret
     0.00%  perf      [kernel.kallsyms]  [k] native_write_msr_safe

__sqrtf (which does 'sqrtss %xmm0,%xmm0') only takes 5% of the time!

I just fixed fp-bench to discard negative inputs. This looks
much better: (Note that this is fp-test-x86_64 instead of -aarch64,
which explains why the "before" throughput is different than
the one reported above)

[...]
+fma: (patch 11, i.e. sqrt still in soft-fp)
sqrt-single: 27.11 MFlops
sqrt-double: 27.17 MFlops
+sqrt: (12)
sqrt-single: 66.67 MFlops
sqrt-double: 66.79 MFlops
+cmp: (13)
sqrt-single: 126.46 MFlops
sqrt-double: 126.06 MFlops
+f32f64: (patch 14)
sqrt-single: 122.75 MFlops
sqrt-double: 126.57 MFlops

We get a >2x speedup, which is consistent with the fact
that now perf shows that sqrt takes ~60% of execution time.

Compare does matter here as well because libm is checking
sqrt's result against NaN.

I'll include this fix to fp-bench in v2.

Thanks,

                E.