powerpc: Tweak copy selection parameter in __copy_tofrom_user_power7()

Submitted by Andrew Jeffery on May 12, 2017, 3:58 a.m.

Details

Message ID 20170512035810.15070-1-andrew@aj.id.au
State Accepted
Commit a3f952df3c6d205fe3b1e0e4848d3c3db4a28ef0
Headers show

Commit Message

Andrew Jeffery May 12, 2017, 3:58 a.m.
Experiments with the netperf benchmark indicated that the size selecting
VMX-based copies in __copy_tofrom_user_power7() was suboptimal on POWER8.
Measurements showed that parity was in the neighbourhood of 3328 bytes,
rather than greater than 4096. The change gives a 1.5-2.0% improvement in
performance for 4096-byte buffers, reducing the relative time spent in
__copy_tofrom_user_power7() from approximately 7% to approximately 5% in
the TCP_RR benchmark.

Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
---
 arch/powerpc/lib/copyuser_power7.S | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

Comments

Anton Blanchard May 18, 2017, 5:08 p.m.
Hi Andrew,

> Experiments with the netperf benchmark indicated that the size
> selecting VMX-based copies in __copy_tofrom_user_power7() was
> suboptimal on POWER8. Measurements showed that parity was in the
> neighbourhood of 3328 bytes, rather than greater than 4096. The
> change gives a 1.5-2.0% improvement in performance for 4096-byte
> buffers, reducing the relative time spent in
> __copy_tofrom_user_power7() from approximately 7% to approximately 5%
> in the TCP_RR benchmark.

Nice work! All our context switch optimisations we've made over
the last year has likely moved the break even point for this.

Acked-by: Anton Blanchard <anton@samba.org>

Anton

> Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
> ---
>  arch/powerpc/lib/copyuser_power7.S | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/powerpc/lib/copyuser_power7.S
> b/arch/powerpc/lib/copyuser_power7.S index a24b4039352c..706b7cc19846
> 100644 --- a/arch/powerpc/lib/copyuser_power7.S
> +++ b/arch/powerpc/lib/copyuser_power7.S
> @@ -82,14 +82,14 @@
>  _GLOBAL(__copy_tofrom_user_power7)
>  #ifdef CONFIG_ALTIVEC
>  	cmpldi	r5,16
> -	cmpldi	cr1,r5,4096
> +	cmpldi	cr1,r5,3328
>  
>  	std	r3,-STACKFRAMESIZE+STK_REG(R31)(r1)
>  	std	r4,-STACKFRAMESIZE+STK_REG(R30)(r1)
>  	std	r5,-STACKFRAMESIZE+STK_REG(R29)(r1)
>  
>  	blt	.Lshort_copy
> -	bgt	cr1,.Lvmx_copy
> +	bge	cr1,.Lvmx_copy
>  #else
>  	cmpldi	r5,16
>
Michael Ellerman May 30, 2017, 9:11 a.m.
On Fri, 2017-05-12 at 03:58:10 UTC, Andrew Jeffery wrote:
> Experiments with the netperf benchmark indicated that the size selecting
> VMX-based copies in __copy_tofrom_user_power7() was suboptimal on POWER8.
> Measurements showed that parity was in the neighbourhood of 3328 bytes,
> rather than greater than 4096. The change gives a 1.5-2.0% improvement in
> performance for 4096-byte buffers, reducing the relative time spent in
> __copy_tofrom_user_power7() from approximately 7% to approximately 5% in
> the TCP_RR benchmark.
> 
> Signed-off-by: Andrew Jeffery <andrew@aj.id.au>
> Acked-by: Anton Blanchard <anton@samba.org>

Applied to powerpc next, thanks.

https://git.kernel.org/powerpc/c/a3f952df3c6d205fe3b1e0e4848d3c

cheers

Patch hide | download patch | download mbox

diff --git a/arch/powerpc/lib/copyuser_power7.S b/arch/powerpc/lib/copyuser_power7.S
index a24b4039352c..706b7cc19846 100644
--- a/arch/powerpc/lib/copyuser_power7.S
+++ b/arch/powerpc/lib/copyuser_power7.S
@@ -82,14 +82,14 @@ 
 _GLOBAL(__copy_tofrom_user_power7)
 #ifdef CONFIG_ALTIVEC
 	cmpldi	r5,16
-	cmpldi	cr1,r5,4096
+	cmpldi	cr1,r5,3328
 
 	std	r3,-STACKFRAMESIZE+STK_REG(R31)(r1)
 	std	r4,-STACKFRAMESIZE+STK_REG(R30)(r1)
 	std	r5,-STACKFRAMESIZE+STK_REG(R29)(r1)
 
 	blt	.Lshort_copy
-	bgt	cr1,.Lvmx_copy
+	bge	cr1,.Lvmx_copy
 #else
 	cmpldi	r5,16