[PATCH 1/2] MIPS: lib: csum_partial: more instruction paral

[PATCH 1/2] MIPS: lib: csum_partial: more instruction paral

[chenj]

This will bring at most 50% performance gain on loongson3a.
See
http://dev.lemote.com/files/upload/software/csum-opti/csum-opti-benchmark.html

The benchmark is done in userspace through
http://dev.lemote.com/files/upload/software/csum-opti/csum-test.tar.gz
---
 arch/mips/lib/csum_partial.S | 38 +++++++++++++++++++-------------------
 1 file changed, 19 insertions(+), 19 deletions(-)

diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
index 9901237..6cea101 100644
--- a/arch/mips/lib/csum_partial.S
+++ b/arch/mips/lib/csum_partial.S
@@ -76,10 +76,10 @@
 	LOAD	_t1, (offset + UNIT(1))(src);			\
 	LOAD	_t2, (offset + UNIT(2))(src);			\
 	LOAD	_t3, (offset + UNIT(3))(src);			\
+	ADDC(_t0, _t1);						\
+	ADDC(_t2, _t3);						\
 	ADDC(sum, _t0);						\
-	ADDC(sum, _t1);						\
-	ADDC(sum, _t2);						\
-	ADDC(sum, _t3)
+	ADDC(sum, _t2)
 
 #ifdef USE_DOUBLE
 #define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3)	\
@@ -501,21 +501,21 @@ LEAF(csum_partial)
 	SUB	len, len, 8*NBYTES
 	ADD	src, src, 8*NBYTES
 	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
-	ADDC(sum, t0)
+	ADDC(t0, t1)
 	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
-	ADDC(sum, t1)
+	ADDC(sum, t0)
 	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
-	ADDC(sum, t2)
+	ADDC(t2, t3)
 	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
-	ADDC(sum, t3)
+	ADDC(sum, t2)
 	STORE(t4, UNIT(4)(dst),	.Ls_exc\@)
-	ADDC(sum, t4)
+	ADDC(t4, t5)
 	STORE(t5, UNIT(5)(dst),	.Ls_exc\@)
-	ADDC(sum, t5)
+	ADDC(sum, t4)
 	STORE(t6, UNIT(6)(dst),	.Ls_exc\@)
-	ADDC(sum, t6)
+	ADDC(t6, t7)
 	STORE(t7, UNIT(7)(dst),	.Ls_exc\@)
-	ADDC(sum, t7)
+	ADDC(sum, t6)
 	.set	reorder				/* DADDI_WAR */
 	ADD	dst, dst, 8*NBYTES
 	bgez	len, 1b
@@ -541,13 +541,13 @@ LEAF(csum_partial)
 	SUB	len, len, 4*NBYTES
 	ADD	src, src, 4*NBYTES
 	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
-	ADDC(sum, t0)
+	ADDC(t0, t1)
 	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
-	ADDC(sum, t1)
+	ADDC(sum, t0)
 	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
-	ADDC(sum, t2)
+	ADDC(t2, t3)
 	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
-	ADDC(sum, t3)
+	ADDC(sum, t2)
 	.set	reorder				/* DADDI_WAR */
 	ADD	dst, dst, 4*NBYTES
 	beqz	len, .Ldone\@
@@ -646,13 +646,13 @@ LEAF(csum_partial)
 	nop				# improves slotting
 #endif
 	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
-	ADDC(sum, t0)
+	ADDC(t0, t1)
 	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
-	ADDC(sum, t1)
+	ADDC(sum, t0)
 	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
-	ADDC(sum, t2)
+	ADDC(t2, t3)
 	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
-	ADDC(sum, t3)
+	ADDC(sum, t2)
 	.set	reorder				/* DADDI_WAR */
 	ADD	dst, dst, 4*NBYTES
 	bne	len, rem, 1b
-- 
1.9.0

[PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[chenj]

wsbh & movn are available on loongson3 CPU.
---
 arch/mips/lib/csum_partial.S | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)

diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
index 6cea101..ed88647 100644
--- a/arch/mips/lib/csum_partial.S
+++ b/arch/mips/lib/csum_partial.S
@@ -277,9 +277,12 @@ LEAF(csum_partial)
 #endif
 
 	/* odd buffer alignment? */
-#ifdef CONFIG_CPU_MIPSR2
+#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
+	.set	push
+	.set	arch=mips32r2
 	wsbh	v1, sum
 	movn	sum, v1, t7
+	.set	pop
 #else
 	beqz	t7, 1f			/* odd buffer alignment? */
 	 lui	v1, 0x00ff
@@ -726,9 +729,12 @@ LEAF(csum_partial)
 	addu	sum, v1
 #endif
 
-#ifdef CONFIG_CPU_MIPSR2
+#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
+	.set	push
+	.set	arch=mips32r2
 	wsbh	v1, sum
 	movn	sum, v1, odd
+	.set	pop
 #else
 	beqz	odd, 1f			/* odd buffer alignment? */
 	 lui	v1, 0x00ff
-- 
1.9.0

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Paul Burton]

[-- Attachment #1: Type: text/plain, Size: 1217 bytes --]

On Thu, May 15, 2014 at 03:09:03PM +0800, chenj wrote:
> wsbh & movn are available on loongson3 CPU.
> ---
>  arch/mips/lib/csum_partial.S | 10 ++++++++--
>  1 file changed, 8 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
> index 6cea101..ed88647 100644
> --- a/arch/mips/lib/csum_partial.S
> +++ b/arch/mips/lib/csum_partial.S
> @@ -277,9 +277,12 @@ LEAF(csum_partial)
>  #endif
>  
>  	/* odd buffer alignment? */
> -#ifdef CONFIG_CPU_MIPSR2
> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)

Is there some reason CPU_LOONGSON3 can't select CPU_MIPSR2?

Thanks,
    Paul

> +	.set	push
> +	.set	arch=mips32r2
>  	wsbh	v1, sum
>  	movn	sum, v1, t7
> +	.set	pop
>  #else
>  	beqz	t7, 1f			/* odd buffer alignment? */
>  	 lui	v1, 0x00ff
> @@ -726,9 +729,12 @@ LEAF(csum_partial)
>  	addu	sum, v1
>  #endif
>  
> -#ifdef CONFIG_CPU_MIPSR2
> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
> +	.set	push
> +	.set	arch=mips32r2
>  	wsbh	v1, sum
>  	movn	sum, v1, odd
> +	.set	pop
>  #else
>  	beqz	odd, 1f			/* odd buffer alignment? */
>  	 lui	v1, 0x00ff
> -- 
> 1.9.0
> 
> 

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Paul Burton]

[-- Attachment #1: Type: text/plain, Size: 1217 bytes --]

On Thu, May 15, 2014 at 03:09:03PM +0800, chenj wrote:
> wsbh & movn are available on loongson3 CPU.
> ---
>  arch/mips/lib/csum_partial.S | 10 ++++++++--
>  1 file changed, 8 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
> index 6cea101..ed88647 100644
> --- a/arch/mips/lib/csum_partial.S
> +++ b/arch/mips/lib/csum_partial.S
> @@ -277,9 +277,12 @@ LEAF(csum_partial)
>  #endif
>  
>  	/* odd buffer alignment? */
> -#ifdef CONFIG_CPU_MIPSR2
> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)

Is there some reason CPU_LOONGSON3 can't select CPU_MIPSR2?

Thanks,
    Paul

> +	.set	push
> +	.set	arch=mips32r2
>  	wsbh	v1, sum
>  	movn	sum, v1, t7
> +	.set	pop
>  #else
>  	beqz	t7, 1f			/* odd buffer alignment? */
>  	 lui	v1, 0x00ff
> @@ -726,9 +729,12 @@ LEAF(csum_partial)
>  	addu	sum, v1
>  #endif
>  
> -#ifdef CONFIG_CPU_MIPSR2
> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
> +	.set	push
> +	.set	arch=mips32r2
>  	wsbh	v1, sum
>  	movn	sum, v1, odd
> +	.set	pop
>  #else
>  	beqz	odd, 1f			/* odd buffer alignment? */
>  	 lui	v1, 0x00ff
> -- 
> 1.9.0
> 
> 

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Chen Jie]

2014-05-15 19:40 GMT+08:00 Paul Burton <paul.burton@imgtec.com>:
> On Thu, May 15, 2014 at 03:09:03PM +0800, chenj wrote:
>> wsbh & movn are available on loongson3 CPU.
>> ---
>>  arch/mips/lib/csum_partial.S | 10 ++++++++--
>>  1 file changed, 8 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
>> index 6cea101..ed88647 100644
>> --- a/arch/mips/lib/csum_partial.S
>> +++ b/arch/mips/lib/csum_partial.S
>> @@ -277,9 +277,12 @@ LEAF(csum_partial)
>>  #endif
>>
>>       /* odd buffer alignment? */
>> -#ifdef CONFIG_CPU_MIPSR2
>> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
>
> Is there some reason CPU_LOONGSON3 can't select CPU_MIPSR2?

Loongson 3a is not fully mips32r2 compatible, but Loongson 3b is.

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Paul Burton]

[-- Attachment #1: Type: text/plain, Size: 870 bytes --]

On Fri, May 16, 2014 at 09:29:39PM +0800, Chen Jie wrote:
> >> -#ifdef CONFIG_CPU_MIPSR2
> >> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
> >
> > Is there some reason CPU_LOONGSON3 can't select CPU_MIPSR2?
> 
> Loongson 3a is not fully mips32r2 compatible, but Loongson 3b is.

Interesting, I was led to believe that the 3A was MIPS64r2 compliant
(and thus by nature MIPS32r2 compliant) with the one difference with
other implementations being how it handles Status.FR=0 (appearing as
16x64b registers rather than the 32x32b registers of other CPUs). Are
there any other ways in which Loongson 3A is not MIPS64r2/MIPS32r2
compliant?

If not then wouldn't it make more sense to select CPU_MIPSR2 & then
special case any FPU differences? That would get you all the R2 bits of
the kernel with minimal #ifdef-ery.

Thanks,
    Paul

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Ralf Baechle]

On Thu, May 15, 2014 at 03:09:03PM +0800, chenj wrote:

> wsbh & movn are available on loongson3 CPU.
> ---
>  arch/mips/lib/csum_partial.S | 10 ++++++++--

Does Loongson 3 also have both the DSBH and DSHD instructions?

  Ralf

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Chen Jie]

2014-06-03 19:03 GMT+08:00 Ralf Baechle <ralf@linux-mips.org>:
> On Thu, May 15, 2014 at 03:09:03PM +0800, chenj wrote:
>
>> wsbh & movn are available on loongson3 CPU.
>> ---
>>  arch/mips/lib/csum_partial.S | 10 ++++++++--
>
> Does Loongson 3 also have both the DSBH and DSHD instructions?
Both of these two instructions are available on Loongson 3a.

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Ralf Baechle]

On Thu, May 15, 2014 at 03:09:03PM +0800, chenj wrote:

> wsbh & movn are available on loongson3 CPU.

I think there are a few more case that need to be fixed than just
this file to make best use of WSBH and similar on Loongson 3A.  How
about below patch?

As I don't have Loongson 3 hardware I am not able to runtime test this.

Thanks,

  Ralf

 arch/mips/include/asm/cpu-features.h                           | 10 ++++++++++
 .../include/asm/mach-cavium-octeon/cpu-feature-overrides.h     |  1 +
 arch/mips/include/asm/mach-loongson/cpu-feature-overrides.h    |  2 ++
 arch/mips/include/uapi/asm/swab.h                              |  5 +++--
 arch/mips/lib/csum_partial.S                                   | 10 ++++++++--
 arch/mips/net/bpf_jit.c                                        |  2 +-
 6 files changed, 25 insertions(+), 5 deletions(-)

diff --git a/arch/mips/include/asm/cpu-features.h b/arch/mips/include/asm/cpu-features.h
index c7d8c99..d927bda 100644
--- a/arch/mips/include/asm/cpu-features.h
+++ b/arch/mips/include/asm/cpu-features.h
@@ -222,6 +222,16 @@
 #define cpu_has_clo_clz	cpu_has_mips_r
 #endif
 
+/*
+ * MIPS32 R2, MIPS64 R2, Loongson 3A and Octeon have WSBH.
+ * MIPS64 R2, Loongson 3A and Octeon have WSBH, DSBH and DSHD.
+ * This indicates the availability of WSBH and in case of 64 bit CPUs also
+ * DSBH and DSHD.
+ */
+#ifndef cpu_has_wsbh
+#define cpu_has_wsbh		cpu_has_mips_r2
+#endif
+
 #ifndef cpu_has_dsp
 #define cpu_has_dsp		(cpu_data[0].ases & MIPS_ASE_DSP)
 #endif
diff --git a/arch/mips/include/asm/mach-cavium-octeon/cpu-feature-overrides.h b/arch/mips/include/asm/mach-cavium-octeon/cpu-feature-overrides.h
index cf80228..fa1f3cf 100644
--- a/arch/mips/include/asm/mach-cavium-octeon/cpu-feature-overrides.h
+++ b/arch/mips/include/asm/mach-cavium-octeon/cpu-feature-overrides.h
@@ -57,6 +57,7 @@
 #define cpu_has_vint		0
 #define cpu_has_veic		0
 #define cpu_hwrena_impl_bits	0xc0000000
+#define cpu_has_wsbh            1
 
 #define cpu_has_rixi		(cpu_data[0].cputype != CPU_CAVIUM_OCTEON)
 
diff --git a/arch/mips/include/asm/mach-loongson/cpu-feature-overrides.h b/arch/mips/include/asm/mach-loongson/cpu-feature-overrides.h
index c0f3ef4..7d28f95 100644
--- a/arch/mips/include/asm/mach-loongson/cpu-feature-overrides.h
+++ b/arch/mips/include/asm/mach-loongson/cpu-feature-overrides.h
@@ -59,4 +59,6 @@
 #define cpu_has_watch		1
 #define cpu_has_local_ebase	0
 
+#define cpu_has_wsbh		IS_ENABLED(CONFIG_CPU_LOONGSON3)
+
 #endif /* __ASM_MACH_LOONGSON_CPU_FEATURE_OVERRIDES_H */
diff --git a/arch/mips/include/uapi/asm/swab.h b/arch/mips/include/uapi/asm/swab.h
index ac9a8f9..b2ab2cf 100644
--- a/arch/mips/include/uapi/asm/swab.h
+++ b/arch/mips/include/uapi/asm/swab.h
@@ -13,7 +13,8 @@
 
 #define __SWAB_64_THRU_32__
 
-#if defined(__mips_isa_rev) && (__mips_isa_rev >= 2)
+#if (defined(__mips_isa_rev) && (__mips_isa_rev >= 2)) ||		\
+    defined(_MIPS_ARCH_LOONGSON3A)
 
 static inline __attribute_const__ __u16 __arch_swab16(__u16 x)
 {
@@ -55,5 +56,5 @@ static inline __attribute_const__ __u64 __arch_swab64(__u64 x)
 }
 #define __arch_swab64 __arch_swab64
 #endif /* __mips64 */
-#endif /* MIPS R2 or newer  */
+#endif /* MIPS R2 or newer or Loongson 3A  */
 #endif /* _ASM_SWAB_H */
diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
index 9901237..4c721e2 100644
--- a/arch/mips/lib/csum_partial.S
+++ b/arch/mips/lib/csum_partial.S
@@ -277,9 +277,12 @@ LEAF(csum_partial)
 #endif
 
 	/* odd buffer alignment? */
-#ifdef CONFIG_CPU_MIPSR2
+#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
+	.set	push
+	.set	arch=mips32r2
 	wsbh	v1, sum
 	movn	sum, v1, t7
+	.set	pop
 #else
 	beqz	t7, 1f			/* odd buffer alignment? */
 	 lui	v1, 0x00ff
@@ -726,9 +729,12 @@ LEAF(csum_partial)
 	addu	sum, v1
 #endif
 
-#ifdef CONFIG_CPU_MIPSR2
+#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
+	.set	push
+	.set	arch=mips32r2
 	wsbh	v1, sum
 	movn	sum, v1, odd
+	.set	pop
 #else
 	beqz	odd, 1f			/* odd buffer alignment? */
 	 lui	v1, 0x00ff
diff --git a/arch/mips/net/bpf_jit.c b/arch/mips/net/bpf_jit.c
index a67b975..b2a560b 100644
--- a/arch/mips/net/bpf_jit.c
+++ b/arch/mips/net/bpf_jit.c
@@ -1240,7 +1240,7 @@ jmp_cmp:
 			emit_half_load(r_A, r_skb, off, ctx);
 #ifdef CONFIG_CPU_LITTLE_ENDIAN
 			/* This needs little endian fixup */
-			if (cpu_has_mips_r2) {
+			if (cpu_has_wsbh) {
 				/* R2 and later have the wsbh instruction */
 				emit_wsbh(r_A, r_A, ctx);
 			} else {

Re: [PATCH 2/2] MIPS: lib: csum_partial: use wsbh/movn on ls3

[Chen Jie]

I've tested the patch:
1. arch/mips/net/bpf_jit.c is missing in master branch of linux-mips,
so it was ignored.
2. Other things are fine.

2014-06-04 2:44 GMT+08:00 Ralf Baechle <ralf@linux-mips.org>:
> diff --git a/arch/mips/include/uapi/asm/swab.h b/arch/mips/include/uapi/asm/swab.h
> index ac9a8f9..b2ab2cf 100644
> --- a/arch/mips/include/uapi/asm/swab.h
> +++ b/arch/mips/include/uapi/asm/swab.h
> @@ -13,7 +13,8 @@
>
>  #define __SWAB_64_THRU_32__
>
> -#if defined(__mips_isa_rev) && (__mips_isa_rev >= 2)
> +#if (defined(__mips_isa_rev) && (__mips_isa_rev >= 2)) ||              \
> +    defined(_MIPS_ARCH_LOONGSON3A)
>
>  static inline __attribute_const__ __u16 __arch_swab16(__u16 x)
>  {
> @@ -55,5 +56,5 @@ static inline __attribute_const__ __u64 __arch_swab64(__u64 x)
Should we add ".set arch=mips64r2" for using of wsbh/dsbh?
I can compile 3.15-rc8 with the patch applied, but when I tried the
patch on 3.10, it reported "unrecognized op dsbh..."

> diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
> index 9901237..4c721e2 100644
> --- a/arch/mips/lib/csum_partial.S
> +++ b/arch/mips/lib/csum_partial.S
> @@ -277,9 +277,12 @@ LEAF(csum_partial)
>  #endif
>
>         /* odd buffer alignment? */
> -#ifdef CONFIG_CPU_MIPSR2
> +#if defined(CONFIG_CPU_MIPSR2) || defined(CONFIG_CPU_LOONGSON3)
Should above line be replaced with the following line?
"#if cpu_has_wsbh"

Re: [PATCH 1/2] MIPS: lib: csum_partial: more instruction paral

[Markos Chandras]

On 05/15/2014 08:09 AM, chenj wrote:
> This will bring at most 50% performance gain on loongson3a.
> See
> http://dev.lemote.com/files/upload/software/csum-opti/csum-opti-benchmark.html
> 
> The benchmark is done in userspace through
> http://dev.lemote.com/files/upload/software/csum-opti/csum-test.tar.gz
> ---
>  arch/mips/lib/csum_partial.S | 38 +++++++++++++++++++-------------------
>  1 file changed, 19 insertions(+), 19 deletions(-)

Hi chenj,

My opinion is that the commit message is not ideal. If the http links
ever go away, the commit message will be mostly useless for someone
trying to understand the reason for this patch. I would suggest to
explain the reason for the optimizations in the commit message and put
the http links as a comment to this patch which will still be visible in
the mailing list archives. Or you can keep them in the commit message
but I think the reason for the patch should be explained as well.

-- 
markos

Re: [PATCH 1/2] MIPS: lib: csum_partial: more instruction paral

[Markos Chandras]

On 05/15/2014 08:09 AM, chenj wrote:
> This will bring at most 50% performance gain on loongson3a.
> See
> http://dev.lemote.com/files/upload/software/csum-opti/csum-opti-benchmark.html
> 
> The benchmark is done in userspace through
> http://dev.lemote.com/files/upload/software/csum-opti/csum-test.tar.gz
> ---
>  arch/mips/lib/csum_partial.S | 38 +++++++++++++++++++-------------------
>  1 file changed, 19 insertions(+), 19 deletions(-)

Hi chenj,

My opinion is that the commit message is not ideal. If the http links
ever go away, the commit message will be mostly useless for someone
trying to understand the reason for this patch. I would suggest to
explain the reason for the optimizations in the commit message and put
the http links as a comment to this patch which will still be visible in
the mailing list archives. Or you can keep them in the commit message
but I think the reason for the patch should be explained as well.

-- 
markos

Re: [PATCH 1/2] MIPS: lib: csum_partial: more instruction paral

[Ralf Baechle]

On Thu, May 15, 2014 at 09:20:39AM +0100, Markos Chandras wrote:

> On 05/15/2014 08:09 AM, chenj wrote:
> > This will bring at most 50% performance gain on loongson3a.
> > See
> > http://dev.lemote.com/files/upload/software/csum-opti/csum-opti-benchmark.html
> > 
> > The benchmark is done in userspace through
> > http://dev.lemote.com/files/upload/software/csum-opti/csum-test.tar.gz
> > ---
> >  arch/mips/lib/csum_partial.S | 38 +++++++++++++++++++-------------------
> >  1 file changed, 19 insertions(+), 19 deletions(-)
> 
> Hi chenj,
> 
> My opinion is that the commit message is not ideal. If the http links
> ever go away, the commit message will be mostly useless for someone
> trying to understand the reason for this patch. I would suggest to
> explain the reason for the optimizations in the commit message and put
> the http links as a comment to this patch which will still be visible in
> the mailing list archives. Or you can keep them in the commit message
> but I think the reason for the patch should be explained as well.

I can certainly offer a FTP space for such things.  Or alternatively,
post the archive's content as a shar archive (GNU sharutils) which will
end up in the mailing list archives itself and which provides the best
guarantee for a stable URL.

That said, tinkering with the checksum code shouldn't be done lightly.
The results on some CPUs are quite surprising and the current
impleemntation which is working well on R4000 or Sibyte class cores
has worked reasonably well for most more modern cores and some of the
attempted tweaks have benefited a few but hurt many cores.

But time progresses and we optimize the kernel for modern hardware so
I'm open to reevaluate things, including runtime generation.

  Ralf

[PATCH, v2] MIPS: lib: csum_partial: more instruction paral

[chenj]

Computing sum introduces true data dependency, e.g.
	ADDC(sum, t0)
	ADDC(sum, t1)
	ADDC(sum, t2)
	ADDC(sum, t3)
Here, each ADDC(sum, ...) references the sum value updated by previous ADDC.

In this patch, above sequence is adjusted as following:
	ADDC(t0, t1)
	ADDC(t2, t3)
	ADDC(sum, t0)
	ADDC(sum, t2)
The first two ADDC operations are independent, hence can be executed
simultaneously if possible.

This patch improves instruction level parallelism, and brings at most 50%
csum performance gain on Loongson 3a processor[1].

---
1. The result can be found at
http://dev.lemote.com/files/upload/software/csum-opti/csum-opti-benchmark.html
And is generated by a userspace test program:
http://dev.lemote.com/files/upload/software/csum-opti/csum-test.tar.gz

[v2: amend commit message]

 arch/mips/lib/csum_partial.S | 38 +++++++++++++++++++-------------------
 1 file changed, 19 insertions(+), 19 deletions(-)

diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
index 9901237..6cea101 100644
--- a/arch/mips/lib/csum_partial.S
+++ b/arch/mips/lib/csum_partial.S
@@ -76,10 +76,10 @@
 	LOAD	_t1, (offset + UNIT(1))(src);			\
 	LOAD	_t2, (offset + UNIT(2))(src);			\
 	LOAD	_t3, (offset + UNIT(3))(src);			\
+	ADDC(_t0, _t1);						\
+	ADDC(_t2, _t3);						\
 	ADDC(sum, _t0);						\
-	ADDC(sum, _t1);						\
-	ADDC(sum, _t2);						\
-	ADDC(sum, _t3)
+	ADDC(sum, _t2)
 
 #ifdef USE_DOUBLE
 #define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3)	\
@@ -501,21 +501,21 @@ LEAF(csum_partial)
 	SUB	len, len, 8*NBYTES
 	ADD	src, src, 8*NBYTES
 	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
-	ADDC(sum, t0)
+	ADDC(t0, t1)
 	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
-	ADDC(sum, t1)
+	ADDC(sum, t0)
 	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
-	ADDC(sum, t2)
+	ADDC(t2, t3)
 	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
-	ADDC(sum, t3)
+	ADDC(sum, t2)
 	STORE(t4, UNIT(4)(dst),	.Ls_exc\@)
-	ADDC(sum, t4)
+	ADDC(t4, t5)
 	STORE(t5, UNIT(5)(dst),	.Ls_exc\@)
-	ADDC(sum, t5)
+	ADDC(sum, t4)
 	STORE(t6, UNIT(6)(dst),	.Ls_exc\@)
-	ADDC(sum, t6)
+	ADDC(t6, t7)
 	STORE(t7, UNIT(7)(dst),	.Ls_exc\@)
-	ADDC(sum, t7)
+	ADDC(sum, t6)
 	.set	reorder				/* DADDI_WAR */
 	ADD	dst, dst, 8*NBYTES
 	bgez	len, 1b
@@ -541,13 +541,13 @@ LEAF(csum_partial)
 	SUB	len, len, 4*NBYTES
 	ADD	src, src, 4*NBYTES
 	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
-	ADDC(sum, t0)
+	ADDC(t0, t1)
 	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
-	ADDC(sum, t1)
+	ADDC(sum, t0)
 	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
-	ADDC(sum, t2)
+	ADDC(t2, t3)
 	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
-	ADDC(sum, t3)
+	ADDC(sum, t2)
 	.set	reorder				/* DADDI_WAR */
 	ADD	dst, dst, 4*NBYTES
 	beqz	len, .Ldone\@
@@ -646,13 +646,13 @@ LEAF(csum_partial)
 	nop				# improves slotting
 #endif
 	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
-	ADDC(sum, t0)
+	ADDC(t0, t1)
 	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
-	ADDC(sum, t1)
+	ADDC(sum, t0)
 	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
-	ADDC(sum, t2)
+	ADDC(t2, t3)
 	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
-	ADDC(sum, t3)
+	ADDC(sum, t2)
 	.set	reorder				/* DADDI_WAR */
 	ADD	dst, dst, 4*NBYTES
 	bne	len, rem, 1b
-- 
1.9.0

Re: [PATCH, v2] MIPS: lib: csum_partial: more instruction paral

[James Hogan]

[-- Attachment #1: Type: text/plain, Size: 3679 bytes --]

On Monday 19 May 2014 11:14:07 chenj wrote:
> Computing sum introduces true data dependency, e.g.
> 	ADDC(sum, t0)
> 	ADDC(sum, t1)
> 	ADDC(sum, t2)
> 	ADDC(sum, t3)
> Here, each ADDC(sum, ...) references the sum value updated by previous ADDC.
> 
> In this patch, above sequence is adjusted as following:
> 	ADDC(t0, t1)
> 	ADDC(t2, t3)
> 	ADDC(sum, t0)
> 	ADDC(sum, t2)
> The first two ADDC operations are independent, hence can be executed
> simultaneously if possible.

The actual patch appears to change it to this:
ADDC(t0, t1)
ADDC(sum, t0)
ADDC(t2, t3)
ADDC(sum, t2)

which is slightly different (presumably due to the interleaved stores in some 
of the cases).

> This patch improves instruction level parallelism, and brings at most 50%
> csum performance gain on Loongson 3a processor[1].

Nice results.

The stuff below the --- will get dropped when the patch is applied though, 
after which the "[1]" won't refer to anything.

Cheers
James

> 
> ---
> 1. The result can be found at
> http://dev.lemote.com/files/upload/software/csum-opti/csum-opti-benchmark.ht
> ml And is generated by a userspace test program:
> http://dev.lemote.com/files/upload/software/csum-opti/csum-test.tar.gz
> 
> [v2: amend commit message]
> 
>  arch/mips/lib/csum_partial.S | 38 +++++++++++++++++++-------------------
>  1 file changed, 19 insertions(+), 19 deletions(-)
> 
> diff --git a/arch/mips/lib/csum_partial.S b/arch/mips/lib/csum_partial.S
> index 9901237..6cea101 100644
> --- a/arch/mips/lib/csum_partial.S
> +++ b/arch/mips/lib/csum_partial.S
> @@ -76,10 +76,10 @@
>  	LOAD	_t1, (offset + UNIT(1))(src);			\
>  	LOAD	_t2, (offset + UNIT(2))(src);			\
>  	LOAD	_t3, (offset + UNIT(3))(src);			\
> +	ADDC(_t0, _t1);						\
> +	ADDC(_t2, _t3);						\
>  	ADDC(sum, _t0);						\
> -	ADDC(sum, _t1);						\
> -	ADDC(sum, _t2);						\
> -	ADDC(sum, _t3)
> +	ADDC(sum, _t2)
> 
>  #ifdef USE_DOUBLE
>  #define CSUM_BIGCHUNK(src, offset, sum, _t0, _t1, _t2, _t3)	\
> @@ -501,21 +501,21 @@ LEAF(csum_partial)
>  	SUB	len, len, 8*NBYTES
>  	ADD	src, src, 8*NBYTES
>  	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
> -	ADDC(sum, t0)
> +	ADDC(t0, t1)
>  	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
> -	ADDC(sum, t1)
> +	ADDC(sum, t0)
>  	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
> -	ADDC(sum, t2)
> +	ADDC(t2, t3)
>  	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
> -	ADDC(sum, t3)
> +	ADDC(sum, t2)
>  	STORE(t4, UNIT(4)(dst),	.Ls_exc\@)
> -	ADDC(sum, t4)
> +	ADDC(t4, t5)
>  	STORE(t5, UNIT(5)(dst),	.Ls_exc\@)
> -	ADDC(sum, t5)
> +	ADDC(sum, t4)
>  	STORE(t6, UNIT(6)(dst),	.Ls_exc\@)
> -	ADDC(sum, t6)
> +	ADDC(t6, t7)
>  	STORE(t7, UNIT(7)(dst),	.Ls_exc\@)
> -	ADDC(sum, t7)
> +	ADDC(sum, t6)
>  	.set	reorder				/* DADDI_WAR */
>  	ADD	dst, dst, 8*NBYTES
>  	bgez	len, 1b
> @@ -541,13 +541,13 @@ LEAF(csum_partial)
>  	SUB	len, len, 4*NBYTES
>  	ADD	src, src, 4*NBYTES
>  	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
> -	ADDC(sum, t0)
> +	ADDC(t0, t1)
>  	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
> -	ADDC(sum, t1)
> +	ADDC(sum, t0)
>  	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
> -	ADDC(sum, t2)
> +	ADDC(t2, t3)
>  	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
> -	ADDC(sum, t3)
> +	ADDC(sum, t2)
>  	.set	reorder				/* DADDI_WAR */
>  	ADD	dst, dst, 4*NBYTES
>  	beqz	len, .Ldone\@
> @@ -646,13 +646,13 @@ LEAF(csum_partial)
>  	nop				# improves slotting
>  #endif
>  	STORE(t0, UNIT(0)(dst),	.Ls_exc\@)
> -	ADDC(sum, t0)
> +	ADDC(t0, t1)
>  	STORE(t1, UNIT(1)(dst),	.Ls_exc\@)
> -	ADDC(sum, t1)
> +	ADDC(sum, t0)
>  	STORE(t2, UNIT(2)(dst),	.Ls_exc\@)
> -	ADDC(sum, t2)
> +	ADDC(t2, t3)
>  	STORE(t3, UNIT(3)(dst),	.Ls_exc\@)
> -	ADDC(sum, t3)
> +	ADDC(sum, t2)
>  	.set	reorder				/* DADDI_WAR */
>  	ADD	dst, dst, 4*NBYTES
>  	bne	len, rem, 1b

[-- Attachment #2: This is a digitally signed message part. --]
[-- Type: application/pgp-signature, Size: 836 bytes --]

Re: [PATCH, v2] MIPS: lib: csum_partial: more instruction paral

[Chen Jie]

2014-05-19 14:59 GMT+08:00 James Hogan <james.hogan@imgtec.com>:
> On Monday 19 May 2014 11:14:07 chenj wrote:
>> Computing sum introduces true data dependency, e.g.
>>       ADDC(sum, t0)
>>       ADDC(sum, t1)
>>       ADDC(sum, t2)
>>       ADDC(sum, t3)
>> Here, each ADDC(sum, ...) references the sum value updated by previous ADDC.
>>
>> In this patch, above sequence is adjusted as following:
>>       ADDC(t0, t1)
>>       ADDC(t2, t3)
>>       ADDC(sum, t0)
>>       ADDC(sum, t2)
>> The first two ADDC operations are independent, hence can be executed
>> simultaneously if possible.
>
> The actual patch appears to change it to this:
> ADDC(t0, t1)
> ADDC(sum, t0)
> ADDC(t2, t3)
> ADDC(sum, t2)
>
> which is slightly different (presumably due to the interleaved stores in some
> of the cases).
>
>> This patch improves instruction level parallelism, and brings at most 50%
>> csum performance gain on Loongson 3a processor[1].
>
> Nice results.
>
> The stuff below the --- will get dropped when the patch is applied though,
> after which the "[1]" won't refer to anything.
>
Thanks for your suggestion, I'll amend the commit message further later.

Basically, the patch reduces the case of one ADDC depending on the
result of the previous ADDC.

BTW, I'm not sure whether the sum value of the new implementation is
equivalent to the original one, but in my test(make run_test of the
csum_test.tar.gz, and a comparing patch in kernel) it is.

Re: [PATCH, v2] MIPS: lib: csum_partial: more instruction paral

[Chen Jie]

2014-05-19 23:32 GMT+08:00 Chen Jie <chenj@lemote.com>:
> 2014-05-19 14:59 GMT+08:00 James Hogan <james.hogan@imgtec.com>:
>> On Monday 19 May 2014 11:14:07 chenj wrote:
>>> Computing sum introduces true data dependency, e.g.
>>>       ADDC(sum, t0)
>>>       ADDC(sum, t1)
>>>       ADDC(sum, t2)
>>>       ADDC(sum, t3)
>>> Here, each ADDC(sum, ...) references the sum value updated by previous ADDC.
>>>
>>> In this patch, above sequence is adjusted as following:
>>>       ADDC(t0, t1)
>>>       ADDC(t2, t3)
>>>       ADDC(sum, t0)
>>>       ADDC(sum, t2)
>>> The first two ADDC operations are independent, hence can be executed
>>> simultaneously if possible.
>>
>> The actual patch appears to change it to this:
>> ADDC(t0, t1)
>> ADDC(sum, t0)
>> ADDC(t2, t3)
>> ADDC(sum, t2)
>>
>> which is slightly different (presumably due to the interleaved stores in some
>> of the cases).
>>
>>> This patch improves instruction level parallelism, and brings at most 50%
>>> csum performance gain on Loongson 3a processor[1].
>>
>> Nice results.
>>
>> The stuff below the --- will get dropped when the patch is applied though,
>> after which the "[1]" won't refer to anything.
>>
> Thanks for your suggestion, I'll amend the commit message further later.
>
> Basically, the patch reduces the case of one ADDC depending on the
> result of the previous ADDC.
>
> BTW, I'm not sure whether the sum value of the new implementation is
> equivalent to the original one, but in my test(make run_test of the
> csum_test.tar.gz, and a comparing patch in kernel) it is.

It is equivalent to the original one.
More explanation about the math behind "x ADDC y ADDC z == x ADDC (y ADDC z)":

Let C = the value of '(uint64_t) -1 + 1' in 64bit case, then
0 <= x <= C-1
0 <= y <= C-1
0 <= z <= C-1

Here 'x ADDC y' is defined as
x + y >= C ? x + y - C + 1 : x + y


Case 1: x + y >= C && x + y - C + 1 + z < C (i.e. x ADDC y ADDC z = x
+ y - C +1 + z)
if y + z >= C:
=> y ADDC z = y + z - C + 1
C > x + y + z - C + 1
=> The result is x + y + z - C + 1

if y + z < C:
=> y ADDC z = y + z
x + y >= C
=> x + (y + z) >= C + z >= C
=> The result is x + y + z - C + 1

Case 2: x + y < C && x + y + z >= C (i.e. x ADDC y ADDC z = x + y + z - C + 1)
if y + z >= C:
=> y ADDC z = y + z - C + 1
C > x + y
=> C + z - C + 1 > x + y + z - C + 1
=> z + 1 > x + y + z - C + 1
=> C >= z + 1 > x + y + z - C + 1
=> The result is x + y + z - C + 1

if y + z < C:
=> y ADDC z = y + z
x + y + z >= C
=> The result is x + y + z - C + 1

Case 3: x + y >= C && x + y - C + 1 + z >= C (i.e. x ADDC y ADDC z = x
+ y + z - 2C + 2)
x + y - C + 1 + z >= C
=> y + z >= 2C - 1 - x
C >= 1 + x
=> 2C - 1 - x >= C
=> y + z >= C
=> y ADDC z = y + z - C + 1

x + y - C + 1 + z >= C
=> The result is  x + y + z - 2C + 2


Case 4: x + y < C && x + y + z < C (i.e. x ADDC y ADDC z = x + y + z)
x + y + z < C
=> y + z < C - x <= C
=> y + z < C
=> y ADDC z = y + z

x + y + z < C
=> The result is x + y + z