IntelLabs/ParallelAccelerator.jl

The ParallelAccelerator package, part of the High Performance Scripting project at Intel Labs

f

@f_kazemian_twitter

I worte

using ParallelAccelerator

function sum(x::Int64)
    sum=0
    for i=1:5
        sum=sum+x
    end
    return sum
end

@acc function sum2(x::Int64)
    sum=0
    for i=1:5
        sum=sum+x
    end
    return sum
end

and then:

@time sum(2)
0.003541 seconds (1.02 k allocations: 57.919 KiB)

@time sum2(2)
16.701655 seconds (15.16 M allocations: 751.828 MiB, 2.44% gc time)

why?

Christopher Rackauckas

@ChrisRackauckas

you ran them twice?

f

@f_kazemian_twitter

I thought the second one is going to be faster!

I run them by jupyter and clear all and run all

and I got this result and I tried it again and again I got the same result

why??

Do I have to set number of processor or something?

Christopher Rackauckas

@ChrisRackauckas

you timed each one twice? Or used BenchmarkTools.jl?

f

@f_kazemian_twitter

for the second one I got more memory allocation and more time

I'm sorry Ican't understand

at firs I add this package:

Pkg.add("ParallelAccelerator")

and then I wrote that code I just sent it here

and then I got that result that I sent

Christopher Rackauckas

@ChrisRackauckas

running time how many times?

f

@f_kazemian_twitter

one time

Christopher Rackauckas

@ChrisRackauckas

we've gone over this

https://docs.julialang.org/en/latest/manual/performance-tips/#Measure-performance-with-[@time](@ref)-and-pay-attention-to-memory-allocation-1

time it twice.

f

@f_kazemian_twitter

If I run it for twice the first one is also better than the second one

It's not fair we can run it twice but the first one is important

you mean parallel accelerator has good spead up for twice run?

Christopher Rackauckas

@ChrisRackauckas

it always works by compiling the first time

it is fair to time it without timing compilation which is unrelated to the actual performance.

f

@f_kazemian_twitter

Actually I want to know why using @acc doesn't improve the running time?

Christopher Rackauckas

@ChrisRackauckas

because in this case it isn't doing anything

it speeds up vectorized functions

this function isn't vectorized at all

f

@f_kazemian_twitter

can we use @acc and @parallel together?

Christopher Rackauckas

@ChrisRackauckas

of course, but I wouldn't put multiple processes on the same machine though

ParallelAccelerator should be pretty efficient for shared memory.

f

@f_kazemian_twitter

I think it's not correct

Christopher Rackauckas

@ChrisRackauckas

it is...?

f

@f_kazemian_twitter

I use the main example in that link you just sent to me

yes the last sentences of you is correct

another thing is not correct

I mean

I use this:

f(x)=x.+x.*x

@acc f2(x)=x.+x.*x

@time f([1,2,3,4,5])

is 0.020449

but

@time f2([1,2,3,4,5]) is 14.322378