[pypy-dev] Automatic SIMD vectorization

[Richard Plangger]

Hi,

glad you liked the post! See the answers below...

On 10/20/2015 04:20 PM, Tuom Larsen wrote:
> Hello all!
> 
> I just read with great interest the blog post "Automatic SIMD
> vectorization support in PyPy".
> 
> Please, I have a few questions:
> 
> - Does regular Python code benefit from the vectorization? I mean, the
> article on one hand says "it is not specifically targeted for the
> NumPy library" but on the other it says "Any interpreter (written in
> RPython)".

Speaking about 'regular' Python code, there is potential !BUT! only if
enough time is spent in numeric code. What I meant in the article was:
E.g. you have a vector construct in your language (like the one the R
language has) you could use the optimization to vectorize operations on
the variables that represent.

Take a look at my test virtual machine implementing a small subset of R.
https://bitbucket.org/plan_rich/vecopt-test-vm

> 
> - I would like to write a vector class as much suitable for PyPy as
> possible, what approach should I take in order to implement it? For
> example, what would suit PyPy JIT the best:
> 
>     class Vector3d:
>         def __init__(a, x, y, z):
>             (a.x, a.y, a.z) = x, y, z
>         def __add__(a, b):
>             return Vector3d(a.x + b.x, a.y+b.y, a.z+b.z)
> 
>     def add1(a, b):
>         (ax, ay, az) = a
>         (bx, by, bz) =  b
>         return [ax + bx, ay + by, az + bz]
> 
>    def add2(a, b):
>         (ax, ay, az) = a
>         (bx, by, bz) =  b
>         return (ax + bx, ay + by, az + bz)
> 
>     def add3((ax, ay, az), (bx, by, bz)):
>         return (ax + bx, ay + by, az + bz)
> 
>     def add3:
>         ???
> 

I have made some tests with this already. You would need to use the
array module. Python lists would also work, but they leave behind some
instructions that are not well optimized. It is described in this post:

http://pypyvecopt.blogspot.co.at/2015/08/gsoc-vec-little-brother-of-numpy-array.html

The missing piece is then the --jit vec_all=1 parameter, that you must
specify on the command line.

Be aware:

1) vec_all=1 parameter might lead to a crash, I have tested it, but it
occurred to me that it is not really ready for production. I'm still
working on this, thus it is disabled by default.

2) that with a very low number of vector elements (e.g 3) the overhead
to iterate the first iteration is quite significant. So I do not think
that there is much you can get just executing 1 vector add in parallel
on x86.

> - Is NumPyPy going to be included with regular PyPy download/install?
> 

NumPyPy is included in a normal PyPy release version

> Thanks a lot in advance!
> _______________________________________________
> pypy-dev mailing list
> pypy-dev at python.org
> https://mail.python.org/mailman/listinfo/pypy-dev
> 

Cheers,
Richard