math: add fast path for Pow(x, 2)

math: add fast path for Pow(x, 2)

linux/arm. ARMv5

benchmark            old ns/op    new ns/op    delta
BenchmarkPowInt          59195         6505  -89.01%
BenchmarkPowFrac        153932       154104   +0.11%
BenchmarkPow10Pos        13941        14052   +0.80%
BenchmarkPow10Neg        17312        17283   -0.17%

linux/amd64. Core i5

benchmark            old ns/op    new ns/op    delta
BenchmarkPowInt             60            6  -89.65%
BenchmarkPowFrac           127          128   +0.79%
BenchmarkPow10Pos           75           75   -0.13%
BenchmarkPow10Neg           80           79   -0.12%

darwin/386. Core Duo

benchmark            old ns/op    new ns/op    delta
BenchmarkPowInt            191           14  -92.67%
BenchmarkPowFrac           348          348   +0.00%
BenchmarkPow10Pos          336          336   +0.00%
BenchmarkPow10Neg          324          325   +0.31%

Also reduces math/rand test runtime by 6x on linux/arm,Arm5.

[dave_cheney.net, 2012-07-03 00:09:13]

PTAL

[minux1, 2012-07-03 02:31:55]

On Jul 3, 2012 8:09 AM, <dave@cheney.net> wrote:
> Description:
> math: add fast path for Pow(x, 2)
>
> linux/arm. ARMv5
>
> benchmark            old ns/op    new ns/op    delta
> BenchmarkPowInt          59195         6505  -89.01%
> BenchmarkPowFrac        153932       154104   +0.11%
> BenchmarkPow10Pos        13941        14052   +0.80%
> BenchmarkPow10Neg        17312        17283   -0.17%
>
> linux/amd64. Core i5
>
> benchmark            old ns/op    new ns/op    delta
> BenchmarkPowInt             60            6  -89.65%
> BenchmarkPowFrac           127          128   +0.79%
> BenchmarkPow10Pos           75           75   -0.13%
> BenchmarkPow10Neg           80           79   -0.12%
>
> darwin/386. Core Duo
>
> benchmark            old ns/op    new ns/op    delta
> BenchmarkPowInt            191           14  -92.67%
> BenchmarkPowFrac           348          348   +0.00%
> BenchmarkPow10Pos          336          336   +0.00%
> BenchmarkPow10Neg          324          325   +0.31%
>
> Also reduces math/rand test runtime by 6x on linux/arm,Arm5.
the change is great, but the BenchmarkPowInt will no longer reflect the
true quality of the generic pow implementation after that. maybe we need
another generic PowInt benchmark.

[dave_cheney.net, 2012-07-03 02:34:05]

Ironically, the only place in the whole tree that called Pow(x, 2),
was the math/rand test. That is not to say there aren't places that
call Pow with 2 as a variable, but they weren't obvious from a quick
inspection.

Would you prefer to see a Benchmark of the known fast paths, vs
assorted integers?

On Tue, Jul 3, 2012 at 12:31 PM, minux <minux.ma@gmail.com> wrote:
>
> On Jul 3, 2012 8:09 AM, <dave@cheney.net> wrote:
>> Description:
>> math: add fast path for Pow(x, 2)
>>
>> linux/arm. ARMv5
>>
>> benchmark            old ns/op    new ns/op    delta
>> BenchmarkPowInt          59195         6505  -89.01%
>> BenchmarkPowFrac        153932       154104   +0.11%
>> BenchmarkPow10Pos        13941        14052   +0.80%
>> BenchmarkPow10Neg        17312        17283   -0.17%
>>
>> linux/amd64. Core i5
>>
>> benchmark            old ns/op    new ns/op    delta
>> BenchmarkPowInt             60            6  -89.65%
>> BenchmarkPowFrac           127          128   +0.79%
>> BenchmarkPow10Pos           75           75   -0.13%
>> BenchmarkPow10Neg           80           79   -0.12%
>>
>> darwin/386. Core Duo
>>
>> benchmark            old ns/op    new ns/op    delta
>> BenchmarkPowInt            191           14  -92.67%
>> BenchmarkPowFrac           348          348   +0.00%
>> BenchmarkPow10Pos          336          336   +0.00%
>> BenchmarkPow10Neg          324          325   +0.31%
>>
>> Also reduces math/rand test runtime by 6x on linux/arm,Arm5.
> the change is great, but the BenchmarkPowInt will no longer reflect the true
> quality of the generic pow implementation after that. maybe we need another
> generic PowInt benchmark.

[minux1, 2012-07-03 02:43:20]

On Jul 3, 2012 10:34 AM, "Dave Cheney" <dave@cheney.net> wrote:
>
> Ironically, the only place in the whole tree that called Pow(x, 2),
> was the math/rand test. That is not to say there aren't places that
> call Pow with 2 as a variable, but they weren't obvious from a quick
> inspection.
>
> Would you prefer to see a Benchmark of the known fast paths, vs
> assorted integers?
i think the benchmarks should test at least one general case,
renaming the current benchmark to something like BenchPowInt2 is ok with
me, but we need another bench that doesn't trigger the fast path.

[dave_cheney.net, 2012-08-21 01:14:27]

*** Abandoned ***