[v5,1/4] asm-generic: Improve csum_fold

Message ID 20230914-optimize_checksum-v5-1-c95b82a2757e@rivosinc.com
State New
Headers
Series riscv: Add fine-tuned checksum functions |

Commit Message

Charlie Jenkins Sept. 15, 2023, 3:49 a.m. UTC
  This csum_fold implementation introduced into arch/arc by Vineet Gupta
is better than the default implementation on at least arc, x86, arm, and
riscv. Using GCC trunk and compiling non-inlined version, this
implementation has 41.6667%, 25%, 16.6667% fewer instructions on
riscv64, x86-64, and arm64 respectively with -O3 optimization.

Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
---
 include/asm-generic/checksum.h | 5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
  

Comments

David Laight Sept. 15, 2023, 7:29 a.m. UTC | #1
From: Charlie Jenkins
> Sent: 15 September 2023 04:50
> 
> This csum_fold implementation introduced into arch/arc by Vineet Gupta
> is better than the default implementation on at least arc, x86, arm, and
> riscv. Using GCC trunk and compiling non-inlined version, this
> implementation has 41.6667%, 25%, 16.6667% fewer instructions on
> riscv64, x86-64, and arm64 respectively with -O3 optimization.

Nit-picking the commit message...
Some of those architectures have their own asm implementation.
The arm one is better than the C code below, the x86 ones aren't.

I think that only sparc32 (carry flag but no rotate) and
arm/arm64 (barrel shifter on every instruction) have versions
that are better than the one here.

Since I suggested it to Charlie:

Reviewed-by: David Laight <david.laight@aculab.com>

> 
> Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
> ---
>  include/asm-generic/checksum.h | 5 +----
>  1 file changed, 1 insertion(+), 4 deletions(-)
> 
> diff --git a/include/asm-generic/checksum.h b/include/asm-generic/checksum.h
> index 43e18db89c14..adab9ac4312c 100644
> --- a/include/asm-generic/checksum.h
> +++ b/include/asm-generic/checksum.h
> @@ -30,10 +30,7 @@ extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
>   */
>  static inline __sum16 csum_fold(__wsum csum)
>  {
> -	u32 sum = (__force u32)csum;

You'll need to re-instate that line to stop sparse complaining.

> -	sum = (sum & 0xffff) + (sum >> 16);
> -	sum = (sum & 0xffff) + (sum >> 16);
> -	return (__force __sum16)~sum;
> +	return (__force __sum16)((~csum - ror32(csum, 16)) >> 16);
>  }
>  #endif
> 
> 
> --
> 2.42.0

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)
  
Charlie Jenkins Sept. 15, 2023, 3:45 p.m. UTC | #2
On Fri, Sep 15, 2023 at 07:29:23AM +0000, David Laight wrote:
> From: Charlie Jenkins
> > Sent: 15 September 2023 04:50
> > 
> > This csum_fold implementation introduced into arch/arc by Vineet Gupta
> > is better than the default implementation on at least arc, x86, arm, and
> > riscv. Using GCC trunk and compiling non-inlined version, this
> > implementation has 41.6667%, 25%, 16.6667% fewer instructions on
> > riscv64, x86-64, and arm64 respectively with -O3 optimization.
> 
> Nit-picking the commit message...
> Some of those architectures have their own asm implementation.
> The arm one is better than the C code below, the x86 ones aren't.
I can clean up the commit message to be more accurate.
> 
> I think that only sparc32 (carry flag but no rotate) and
> arm/arm64 (barrel shifter on every instruction) have versions
> that are better than the one here.
> 
> Since I suggested it to Charlie:
> 
> Reviewed-by: David Laight <david.laight@aculab.com>
> 
> > 
> > Signed-off-by: Charlie Jenkins <charlie@rivosinc.com>
> > ---
> >  include/asm-generic/checksum.h | 5 +----
> >  1 file changed, 1 insertion(+), 4 deletions(-)
> > 
> > diff --git a/include/asm-generic/checksum.h b/include/asm-generic/checksum.h
> > index 43e18db89c14..adab9ac4312c 100644
> > --- a/include/asm-generic/checksum.h
> > +++ b/include/asm-generic/checksum.h
> > @@ -30,10 +30,7 @@ extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
> >   */
> >  static inline __sum16 csum_fold(__wsum csum)
> >  {
> > -	u32 sum = (__force u32)csum;
> 
> You'll need to re-instate that line to stop sparse complaining.

I will add that back, thanks.

- Charlie
> 
> > -	sum = (sum & 0xffff) + (sum >> 16);
> > -	sum = (sum & 0xffff) + (sum >> 16);
> > -	return (__force __sum16)~sum;
> > +	return (__force __sum16)((~csum - ror32(csum, 16)) >> 16);
> >  }
> >  #endif
> > 
> > 
> > --
> > 2.42.0
> 
> -
> Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
> Registration No: 1397386 (Wales)
  

Patch

diff --git a/include/asm-generic/checksum.h b/include/asm-generic/checksum.h
index 43e18db89c14..adab9ac4312c 100644
--- a/include/asm-generic/checksum.h
+++ b/include/asm-generic/checksum.h
@@ -30,10 +30,7 @@  extern __sum16 ip_fast_csum(const void *iph, unsigned int ihl);
  */
 static inline __sum16 csum_fold(__wsum csum)
 {
-	u32 sum = (__force u32)csum;
-	sum = (sum & 0xffff) + (sum >> 16);
-	sum = (sum & 0xffff) + (sum >> 16);
-	return (__force __sum16)~sum;
+	return (__force __sum16)((~csum - ror32(csum, 16)) >> 16);
 }
 #endif