code review 28460043: crypto/sha256: block implementation in amd64 assembly

crypto/sha256: block implementation in amd64 assembly

Benchmark on Intel(R) Xeon(R) CPU X5650  @ 2.67GHz

benchmark              old ns/op    new ns/op    delta
BenchmarkHash8Bytes         1259          677  -46.23%
BenchmarkHash1K            14387         6749  -53.09%
BenchmarkHash8K           106006        50107  -52.73%

benchmark               old MB/s     new MB/s  speedup
BenchmarkHash8Bytes         6.35        11.81    1.86x
BenchmarkHash1K            71.17       151.72    2.13x
BenchmarkHash8K            77.28       163.49    2.12x

[jsing, 2013-12-02 14:25:06]

Hello agl@golang.org (cc: golang-dev@googlegroups.com),

I'd like you to review this change to
https://go.googlecode.com/hg/

[agl1, 2013-12-02 17:56:25]

LGTM with nits.

Just to confirm - the addition of |s| to |digest| is just to avoid stack
allocation in the block function, right?

https://codereview.appspot.com/28460043/diff/60001/src/pkg/crypto/sha256/sha2...
File src/pkg/crypto/sha256/sha256.go (right):

https://codereview.appspot.com/28460043/diff/60001/src/pkg/crypto/sha256/sha2...
src/pkg/crypto/sha256/sha256.go:49: type schedule struct {
The use of a new type for this seems excessive.

[jsing, 2013-12-10 12:34:23]

On 2013/12/02 17:56:25, agl1 wrote:
> LGTM with nits.
> 
> Just to confirm - the addition of |s| to |digest| is just to avoid stack
> allocation in the block function, right?

Correct. I've reworked it so that we use stack-based allocation (which means we
have to enable stack splitting) - this resolves your other dislike and seems to
give a small performance gain.

>
https://codereview.appspot.com/28460043/diff/60001/src/pkg/crypto/sha256/sha2...
> File src/pkg/crypto/sha256/sha256.go (right):
> 
>
https://codereview.appspot.com/28460043/diff/60001/src/pkg/crypto/sha256/sha2...
> src/pkg/crypto/sha256/sha256.go:49: type schedule struct {
> The use of a new type for this seems excessive.

PTAL

[agl1, 2013-12-11 16:41:45]

*** Submitted as https://code.google.com/p/go/source/detail?r=5de0f8743adf ***

crypto/sha256: block implementation in amd64 assembly

Benchmark on Intel(R) Xeon(R) CPU X5650  @ 2.67GHz

benchmark              old ns/op    new ns/op    delta
BenchmarkHash8Bytes         1259          677  -46.23%
BenchmarkHash1K            14387         6749  -53.09%
BenchmarkHash8K           106006        50107  -52.73%

benchmark               old MB/s     new MB/s  speedup
BenchmarkHash8Bytes         6.35        11.81    1.86x
BenchmarkHash1K            71.17       151.72    2.13x
BenchmarkHash8K            77.28       163.49    2.12x

R=agl
CC=golang-dev
https://codereview.appspot.com/28460043

Committer: Adam Langley <agl@golang.org>

[donovanhide_gmail.com, 2013-12-21 17:07:26]

Hi,

are they any plans for giving the same treatment to SHA512? The proof of
work algorithm for Ripple uses an unusual hash that involves taking the
first 32 bytes of a SHA512 hash. Would be great if we could get similar
speedups :-)

https://ripple.com/wiki/Proof_of_Work

All the best,
Donovan.


On 11 December 2013 16:41, <agl@golang.org> wrote:

> *** Submitted as
> https://code.google.com/p/go/source/detail?r=5de0f8743adf ***
>
> crypto/sha256: block implementation in amd64 assembly
>
> Benchmark on Intel(R) Xeon(R) CPU X5650  @ 2.67GHz
>
> benchmark              old ns/op    new ns/op    delta
> BenchmarkHash8Bytes         1259          677  -46.23%
> BenchmarkHash1K            14387         6749  -53.09%
> BenchmarkHash8K           106006        50107  -52.73%
>
> benchmark               old MB/s     new MB/s  speedup
> BenchmarkHash8Bytes         6.35        11.81    1.86x
> BenchmarkHash1K            71.17       151.72    2.13x
> BenchmarkHash8K            77.28       163.49    2.12x
>
> R=agl
> CC=golang-dev
> https://codereview.appspot.com/28460043
>
> Committer: Adam Langley <agl@golang.org>
>
>
>
> https://codereview.appspot.com/28460043/
>
> --
>
> ---You received this message because you are subscribed to the Google
> Groups "golang-dev" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to golang-dev+unsubscribe@googlegroups.com.
> For more options, visit https://groups.google.com/groups/opt_out.
>

[jsing, 2013-12-22 06:46:41]

I had planned on looking at it, but it was low on my list of priorities...
however, since you asked :)

https://codereview.appspot.com/37150044


On 22 December 2013 04:07, Donovan Hide <donovanhide@gmail.com> wrote:

> Hi,
>
> are they any plans for giving the same treatment to SHA512? The proof of
> work algorithm for Ripple uses an unusual hash that involves taking the
> first 32 bytes of a SHA512 hash. Would be great if we could get similar
> speedups :-)
>
> https://ripple.com/wiki/Proof_of_Work
>
> All the best,
> Donovan.
>
>
> On 11 December 2013 16:41, <agl@golang.org> wrote:
>
>> *** Submitted as
>> https://code.google.com/p/go/source/detail?r=5de0f8743adf ***
>>
>> crypto/sha256: block implementation in amd64 assembly
>>
>> Benchmark on Intel(R) Xeon(R) CPU X5650  @ 2.67GHz
>>
>> benchmark              old ns/op    new ns/op    delta
>> BenchmarkHash8Bytes         1259          677  -46.23%
>> BenchmarkHash1K            14387         6749  -53.09%
>> BenchmarkHash8K           106006        50107  -52.73%
>>
>> benchmark               old MB/s     new MB/s  speedup
>> BenchmarkHash8Bytes         6.35        11.81    1.86x
>> BenchmarkHash1K            71.17       151.72    2.13x
>> BenchmarkHash8K            77.28       163.49    2.12x
>>
>> R=agl
>> CC=golang-dev
>> https://codereview.appspot.com/28460043
>>
>> Committer: Adam Langley <agl@golang.org>
>>
>>
>>
>> https://codereview.appspot.com/28460043/
>>
>> --
>>
>> ---You received this message because you are subscribed to the Google
>> Groups "golang-dev" group.
>> To unsubscribe from this group and stop receiving emails from it, send an
>> email to golang-dev+unsubscribe@googlegroups.com.
>> For more options, visit https://groups.google.com/groups/opt_out.
>>
>
>

[donovanhide_gmail.com, 2013-12-22 11:54:53]

Wow, that looks extremely good to me :-)

Thank you very, very much! I'm contacting the Guinness Books to ask them to
record the fastest delivery of an assembly version of a crypto hasher!

Cheers,
Donovan (busy proving work!)


On 22 December 2013 06:46, Joel Sing <jsing@google.com> wrote:

> I had planned on looking at it, but it was low on my list of priorities...
> however, since you asked :)
>
> https://codereview.appspot.com/37150044
>
>
> On 22 December 2013 04:07, Donovan Hide <donovanhide@gmail.com> wrote:
>
>> Hi,
>>
>> are they any plans for giving the same treatment to SHA512? The proof of
>> work algorithm for Ripple uses an unusual hash that involves taking the
>> first 32 bytes of a SHA512 hash. Would be great if we could get similar
>> speedups :-)
>>
>> https://ripple.com/wiki/Proof_of_Work
>>
>> All the best,
>> Donovan.
>>
>>
>> On 11 December 2013 16:41, <agl@golang.org> wrote:
>>
>>>  *** Submitted as
>>> https://code.google.com/p/go/source/detail?r=5de0f8743adf ***
>>>
>>> crypto/sha256: block implementation in amd64 assembly
>>>
>>> Benchmark on Intel(R) Xeon(R) CPU X5650  @ 2.67GHz
>>>
>>> benchmark              old ns/op    new ns/op    delta
>>> BenchmarkHash8Bytes         1259          677  -46.23%
>>> BenchmarkHash1K            14387         6749  -53.09%
>>> BenchmarkHash8K           106006        50107  -52.73%
>>>
>>> benchmark               old MB/s     new MB/s  speedup
>>> BenchmarkHash8Bytes         6.35        11.81    1.86x
>>> BenchmarkHash1K            71.17       151.72    2.13x
>>> BenchmarkHash8K            77.28       163.49    2.12x
>>>
>>> R=agl
>>> CC=golang-dev
>>> https://codereview.appspot.com/28460043
>>>
>>> Committer: Adam Langley <agl@golang.org>
>>>
>>>
>>>
>>> https://codereview.appspot.com/28460043/
>>>
>>> --
>>>
>>> ---You received this message because you are subscribed to the Google
>>> Groups "golang-dev" group.
>>> To unsubscribe from this group and stop receiving emails from it, send
>>> an email to golang-dev+unsubscribe@googlegroups.com.
>>> For more options, visit https://groups.google.com/groups/opt_out.
>>>
>>
>>
>