[dpdk-dev] hash: optimize the softrss computation

Message ID 1503403355-4917-1-git-send-email-zhouyates@gmail.com (mailing list archive)
State Accepted, archived
Headers

Checks

Context Check Description
ci/checkpatch warning coding style issues
ci/Intel-compilation success Compilation OK

Commit Message

Matt Aug. 22, 2017, 12:02 p.m. UTC
  Use rte_bsf32 and fast bit unset operation to optimize the softrss computation.
The following measurements shows improvement over the default
softrss computation function.

tuple lens old(cycles) new(cycles)
    3        1225         337
    9        3743         992

Signed-off-by: Yangchao Zhou <zhouyates@gmail.com>
---
 lib/librte_hash/rte_thash.h | 22 ++++++++++------------
 1 file changed, 10 insertions(+), 12 deletions(-)
  

Comments

Vladimir Medvedkin Aug. 22, 2017, 3:57 p.m. UTC | #1
Hi,

2017-08-22 15:02 GMT+03:00 Yangchao Zhou <zhouyates@gmail.com>:

> Use rte_bsf32 and fast bit unset operation to optimize the softrss
> computation.
> The following measurements shows improvement over the default
> softrss computation function.
>
> tuple lens old(cycles) new(cycles)
>     3        1225         337
>     9        3743         992
>
> Signed-off-by: Yangchao Zhou <zhouyates@gmail.com>
> ---
>  lib/librte_hash/rte_thash.h | 22 ++++++++++------------
>  1 file changed, 10 insertions(+), 12 deletions(-)
>
> diff --git a/lib/librte_hash/rte_thash.h b/lib/librte_hash/rte_thash.h
> index 2fffd61..4fa5e07 100644
> --- a/lib/librte_hash/rte_thash.h
> +++ b/lib/librte_hash/rte_thash.h
> @@ -207,15 +207,14 @@ static inline uint32_t
>  rte_softrss(uint32_t *input_tuple, uint32_t input_len,
>                 const uint8_t *rss_key)
>  {
> -       uint32_t i, j, ret = 0;
> +       uint32_t i, j, map, ret = 0;
>
>         for (j = 0; j < input_len; j++) {
> -               for (i = 0; i < 32; i++) {
> -                       if (input_tuple[j] & (1 << (31 - i))) {
> -                               ret ^= rte_cpu_to_be_32(((const uint32_t
> *)rss_key)[j]) << i |
> +               for (map = input_tuple[j]; map; map &= (map - 1)) {
> +                       i = rte_bsf32(map);
> +                       ret ^= rte_cpu_to_be_32(((const uint32_t
> *)rss_key)[j]) << (31 - i) |
>                                         (uint32_t)((uint64_t)(rte_cpu_to_be_32(((const
> uint32_t *)rss_key)[j + 1])) >>
> -                                       (32 - i));
> -                       }
> +                                       (i + 1));
>                 }
>         }
>         return ret;
> @@ -238,14 +237,13 @@ static inline uint32_t
>  rte_softrss_be(uint32_t *input_tuple, uint32_t input_len,
>                 const uint8_t *rss_key)
>  {
> -       uint32_t i, j, ret = 0;
> +       uint32_t i, j, map, ret = 0;
>
>         for (j = 0; j < input_len; j++) {
> -               for (i = 0; i < 32; i++) {
> -                       if (input_tuple[j] & (1 << (31 - i))) {
> -                               ret ^= ((const uint32_t *)rss_key)[j] << i
> |
> -                                       (uint32_t)((uint64_t)(((const
> uint32_t *)rss_key)[j + 1]) >> (32 - i));
> -                       }
> +               for (map = input_tuple[j]; map; map &= (map - 1)) {
> +                       i = rte_bsf32(map);
> +                       ret ^= ((const uint32_t *)rss_key)[j] << (31 - i) |
> +                               (uint32_t)((uint64_t)(((const uint32_t
> *)rss_key)[j + 1]) >> (i + 1));
>                 }
>         }
>         return ret;
> --
> 2.7.4
>
>
Looks good for me. Thanks!

Reviewed-by:  Medvedkin Vladimir <medvedkinv@gmail.com>
  
De Lara Guarch, Pablo Sept. 22, 2017, 3:33 p.m. UTC | #2
Hi,

> -----Original Message-----

> From: dev [mailto:dev-bounces@dpdk.org] On Behalf Of Vladimir

> Medvedkin

> Sent: Tuesday, August 22, 2017 4:57 PM

> To: Yangchao Zhou <zhouyates@gmail.com>

> Cc: dev@dpdk.org

> Subject: Re: [dpdk-dev] [PATCH] hash: optimize the softrss computation

> 

> Hi,

> 

> 2017-08-22 15:02 GMT+03:00 Yangchao Zhou <zhouyates@gmail.com>:

> 

> > Use rte_bsf32 and fast bit unset operation to optimize the softrss

> > computation.

> > The following measurements shows improvement over the default softrss

> > computation function.

> >

> > tuple lens old(cycles) new(cycles)

> >     3        1225         337

> >     9        3743         992

> >

> > Signed-off-by: Yangchao Zhou <zhouyates@gmail.com>


Commit message is a bit long. Could you wrap it to have less than 75 characters per line?
Although, the patch itself has lines with more than 90 characters, which should be avoided:

http://dpdk.org/ml/archives/test-report/2017-August/027283.html

Last thing, the "Reviewed-by" has an extra space, before the name.

Thanks,
Pablo

> >

> Looks good for me. Thanks!

> 

> Reviewed-by:  Medvedkin Vladimir <medvedkinv@gmail.com>

> 

> --

> Regards,

> Vladimir
  
Thomas Monjalon Oct. 6, 2017, 10:31 p.m. UTC | #3
22/09/2017 17:33, De Lara Guarch, Pablo:
> From: Vladimir Medvedkin
> > 2017-08-22 15:02 GMT+03:00 Yangchao Zhou <zhouyates@gmail.com>:
> > 
> > > Use rte_bsf32 and fast bit unset operation to optimize the softrss
> > > computation.
> > > The following measurements shows improvement over the default softrss
> > > computation function.
> > >
> > > tuple lens old(cycles) new(cycles)
> > >     3        1225         337
> > >     9        3743         992
> > >
> > > Signed-off-by: Yangchao Zhou <zhouyates@gmail.com>
> 
> Commit message is a bit long. Could you wrap it to have less than 75 characters per line?
> Although, the patch itself has lines with more than 90 characters, which should be avoided:
> 
> http://dpdk.org/ml/archives/test-report/2017-August/027283.html
> 
> Last thing, the "Reviewed-by" has an extra space, before the name.

Applied, thanks
  

Patch

diff --git a/lib/librte_hash/rte_thash.h b/lib/librte_hash/rte_thash.h
index 2fffd61..4fa5e07 100644
--- a/lib/librte_hash/rte_thash.h
+++ b/lib/librte_hash/rte_thash.h
@@ -207,15 +207,14 @@  static inline uint32_t
 rte_softrss(uint32_t *input_tuple, uint32_t input_len,
 		const uint8_t *rss_key)
 {
-	uint32_t i, j, ret = 0;
+	uint32_t i, j, map, ret = 0;
 
 	for (j = 0; j < input_len; j++) {
-		for (i = 0; i < 32; i++) {
-			if (input_tuple[j] & (1 << (31 - i))) {
-				ret ^= rte_cpu_to_be_32(((const uint32_t *)rss_key)[j]) << i |
+		for (map = input_tuple[j]; map;	map &= (map - 1)) {
+			i = rte_bsf32(map);
+			ret ^= rte_cpu_to_be_32(((const uint32_t *)rss_key)[j]) << (31 - i) |
 					(uint32_t)((uint64_t)(rte_cpu_to_be_32(((const uint32_t *)rss_key)[j + 1])) >>
-					(32 - i));
-			}
+					(i + 1));
 		}
 	}
 	return ret;
@@ -238,14 +237,13 @@  static inline uint32_t
 rte_softrss_be(uint32_t *input_tuple, uint32_t input_len,
 		const uint8_t *rss_key)
 {
-	uint32_t i, j, ret = 0;
+	uint32_t i, j, map, ret = 0;
 
 	for (j = 0; j < input_len; j++) {
-		for (i = 0; i < 32; i++) {
-			if (input_tuple[j] & (1 << (31 - i))) {
-				ret ^= ((const uint32_t *)rss_key)[j] << i |
-					(uint32_t)((uint64_t)(((const uint32_t *)rss_key)[j + 1]) >> (32 - i));
-			}
+		for (map = input_tuple[j]; map;	map &= (map - 1)) {
+			i = rte_bsf32(map);
+			ret ^= ((const uint32_t *)rss_key)[j] << (31 - i) |
+				(uint32_t)((uint64_t)(((const uint32_t *)rss_key)[j + 1]) >> (i + 1));
 		}
 	}
 	return ret;