Hi all,

There are two first [1] benchmarks for upcoming ndslice.algorithm [2].
Recent LDC alpha based on LLVM 3.8 and recent Mir v0.16.0-alpha3 are required. @fasmath syntax may be changed a little bit and will be simplified anyway.

Dot Product:

       ndReduce vectorized = 3 ms, 314 μs
                  ndReduce = 14 ms, 767 μs
numeric.dotProduct, arrays = 7 ms, 260 μs
numeric.dotProduct, slices = 14 ms, 782 μs
              zip & reduce = 74 ms, 280 μs

Euclidean Distance:

                ndReduce vectorized = 3 ms, 668 μs
                           ndReduce = 14 ms, 595 μs
  numeric.euclideanDistance, arrays = 14 ms, 463 μs
  numeric.euclideanDistance, slices = 14 ms, 465 μs
                       zip & reduce = 73 ms, 678 μs

[1] https://github.com/libmir/mir/tree/master/benchmarks/ndslice
[2] https://github.com/dlang/phobos/pull/4652

Best regards,
Ilya

On Wednesday, 3 August 2016 at 20:53:59 UTC, Ilya Yaroshenko wrote:

Update:

> Dot Product:
>               zip & reduce = 74 ms, 280 μs

              zip & reduce = 44 ms, 57 μs

>
> Euclidean Distance:
>                        zip & reduce = 73 ms, 678 μs

                       zip & reduce = 44 ms, 646 μs

The tests above are for double precision floating point numbers. The results for single precision are below.

Dot Product (single precision):

       ndReduce vectorized = 2 ms, 200 μs
                  ndReduce = 14 ms, 543 μs
numeric.dotProduct, arrays = 7 ms, 208 μs
numeric.dotProduct, slices = 14 ms, 414 μs
              zip & reduce = 43 ms, 657 μs

Euclidean Distance (single precisoin):

                ndReduce vectorized = 2 ms, 226 μs
                           ndReduce = 14 ms, 661 μs
  numeric.euclideanDistance, arrays = 14 ms, 597 μs
  numeric.euclideanDistance, slices = 14 ms, 581 μs
                       zip & reduce = 46 ms, 759 μs

On Wednesday, 3 August 2016 at 20:53:59 UTC, Ilya Yaroshenko wrote:
> Hi all,
>
> There are two first [1] benchmarks for upcoming ndslice.algorithm [2].
> Recent LDC alpha based on LLVM 3.8 and recent Mir v0.16.0-alpha3 are required. @fasmath syntax may be changed a little bit and will be simplified anyway.
>
> [...]

Ilya: The result are awesome!!
Let's make some noise:

https://www.reddit.com/r/programming/comments/4w16i5/ndslicealgorithm_speed_up_your_matrix/

On Wednesday, 3 August 2016 at 20:53:59 UTC, Ilya Yaroshenko wrote:
> Hi all,
>
> There are two first [1] benchmarks for upcoming ndslice.algorithm [2].
> Recent LDC alpha based on LLVM 3.8 and recent Mir v0.16.0-alpha3 are required. @fasmath syntax may be changed a little bit and will be simplified anyway.
>
> Dot Product:
>
>        ndReduce vectorized = 3 ms, 314 μs
>                   ndReduce = 14 ms, 767 μs

**That's** the difference with or without fastmath??

(awesome work of course!)

-Johan

On Wednesday, 3 August 2016 at 22:22:19 UTC, Johan Engelen wrote:
> On Wednesday, 3 August 2016 at 20:53:59 UTC, Ilya Yaroshenko
>> Dot Product:
>>
>>        ndReduce vectorized = 3 ms, 314 μs
>>                   ndReduce = 14 ms, 767 μs
>
> **That's** the difference with or without fastmath??

The first one is with @fastmath and addition execution branch for iteration in case of stride equal to 1.

On Wednesday, 3 August 2016 at 20:53:59 UTC, Ilya Yaroshenko wrote:
> Hi all,
>
> There are two first [1] benchmarks for upcoming ndslice.algorithm [2].
> Recent LDC alpha based on LLVM 3.8 and recent Mir v0.16.0-alpha3 are required. @fasmath syntax may be changed a little bit and will be simplified anyway.


Keep up the good work!

On Wednesday, 3 August 2016 at 20:53:59 UTC, Ilya Yaroshenko wrote:
> Hi all,
>
> There are two first [1] benchmarks for upcoming ndslice.algorithm [2].
> Recent LDC alpha based on LLVM 3.8 and recent Mir v0.16.0-alpha3 are required. @fasmath syntax may be changed a little bit and will be simplified anyway.
>
> [...]

The PR and Mir v0.16.0-alpha7 have half and triangular selections. They are very helpful to work with matrixes.