[Qemu-devel] [PATCH] net: optimize checksum computation

[Qemu-devel] [PATCH] net: optimize checksum computation

[Ladi Prosek]

Very simple loop optimization with a significant performance impact.

Microbenchmark results, modern x86-64:

buffer size | speed up
------------+---------
1500        | 1.7x
64          | 1.5x
8           | 1.15x

Microbenchmark results, POWER7:

buffer size | speed up
------------+---------
1500        | 5x
64          | 3.3x
8           | 1.13x

There is a lot of room for further improvement at the expense of
code complexity - aligned multibyte reads, LE/BE considerations,
architecture-specific optimizations, etc. This patch still keeps
things simple and readable.

Signed-off-by: Ladi Prosek <lprosek@redhat.com>
---
 net/checksum.c | 21 +++++++++++++--------
 1 file changed, 13 insertions(+), 8 deletions(-)

diff --git a/net/checksum.c b/net/checksum.c
index 23323b0..4da72a6 100644
--- a/net/checksum.c
+++ b/net/checksum.c
@@ -22,17 +22,22 @@
 
 uint32_t net_checksum_add_cont(int len, uint8_t *buf, int seq)
 {
-    uint32_t sum = 0;
+    uint32_t sum1 = 0, sum2 = 0;
     int i;
 
-    for (i = seq; i < seq + len; i++) {
-        if (i & 1) {
-            sum += (uint32_t)buf[i - seq];
-        } else {
-            sum += (uint32_t)buf[i - seq] << 8;
-        }
+    for (i = 0; i < len - 1; i += 2) {
+        sum1 += (uint32_t)buf[i];
+        sum2 += (uint32_t)buf[i + 1];
+    }
+    if (i < len) {
+        sum1 += (uint32_t)buf[i];
+    }
+
+    if (seq & 1) {
+        return sum1 + (sum2 << 8);
+    } else {
+        return sum2 + (sum1 << 8);
     }
-    return sum;
 }
 
 uint16_t net_checksum_finish(uint32_t sum)
-- 
2.7.4

Re: [Qemu-devel] [PATCH] net: optimize checksum computation

[Dmitry Fleytman]

> On 6 Jan 2017, at 10:08 AM, Ladi Prosek <lprosek@redhat.com> wrote:
> 
> Very simple loop optimization with a significant performance impact.
> 
> Microbenchmark results, modern x86-64:
> 
> buffer size | speed up
> ------------+---------
> 1500        | 1.7x
> 64          | 1.5x
> 8           | 1.15x
> 
> Microbenchmark results, POWER7:
> 
> buffer size | speed up
> ------------+---------
> 1500        | 5x
> 64          | 3.3x
> 8           | 1.13x
> 
> There is a lot of room for further improvement at the expense of
> code complexity - aligned multibyte reads, LE/BE considerations,
> architecture-specific optimizations, etc. This patch still keeps
> things simple and readable.

Reviewed-by: Dmitry Fleytman <dmitry@daynix.com>

> 
> Signed-off-by: Ladi Prosek <lprosek@redhat.com>
> ---
> net/checksum.c | 21 +++++++++++++--------
> 1 file changed, 13 insertions(+), 8 deletions(-)
> 
> diff --git a/net/checksum.c b/net/checksum.c
> index 23323b0..4da72a6 100644
> --- a/net/checksum.c
> +++ b/net/checksum.c
> @@ -22,17 +22,22 @@
> 
> uint32_t net_checksum_add_cont(int len, uint8_t *buf, int seq)
> {
> -    uint32_t sum = 0;
> +    uint32_t sum1 = 0, sum2 = 0;
>     int i;
> 
> -    for (i = seq; i < seq + len; i++) {
> -        if (i & 1) {
> -            sum += (uint32_t)buf[i - seq];
> -        } else {
> -            sum += (uint32_t)buf[i - seq] << 8;
> -        }
> +    for (i = 0; i < len - 1; i += 2) {
> +        sum1 += (uint32_t)buf[i];
> +        sum2 += (uint32_t)buf[i + 1];
> +    }
> +    if (i < len) {
> +        sum1 += (uint32_t)buf[i];
> +    }
> +
> +    if (seq & 1) {
> +        return sum1 + (sum2 << 8);
> +    } else {
> +        return sum2 + (sum1 << 8);
>     }
> -    return sum;
> }
> 
> uint16_t net_checksum_finish(uint32_t sum)
> -- 
> 2.7.4
> 

Re: [Qemu-devel] [PATCH] net: optimize checksum computation

[Jason Wang]

On 2017年01月08日 17:03, Dmitry Fleytman wrote:
>> On 6 Jan 2017, at 10:08 AM, Ladi Prosek <lprosek@redhat.com> wrote:
>>
>> Very simple loop optimization with a significant performance impact.
>>
>> Microbenchmark results, modern x86-64:
>>
>> buffer size | speed up
>> ------------+---------
>> 1500        | 1.7x
>> 64          | 1.5x
>> 8           | 1.15x
>>
>> Microbenchmark results, POWER7:
>>
>> buffer size | speed up
>> ------------+---------
>> 1500        | 5x
>> 64          | 3.3x
>> 8           | 1.13x
>>
>> There is a lot of room for further improvement at the expense of
>> code complexity - aligned multibyte reads, LE/BE considerations,
>> architecture-specific optimizations, etc. This patch still keeps
>> things simple and readable.
> Reviewed-by: Dmitry Fleytman <dmitry@daynix.com>
>

Applied to -net.

Thanks