Modular•14mo ago

SIMD produces weird results without print statement at the end

The code below produces weird results:

from algorithm import vectorize
from tensor import Tensor

alias type = DType.int64
alias nelts = simdwidthof[type]()

def main():
    let size = 10
    var a = Tensor[type](size)
    var b = Tensor[type](size)
    var c = Tensor[type](size)

    for i in range(size):
        a[i] = i + 1
        b[i] = i
    
    @parameter
    fn diff[nelts : Int](x : Int):
        #print(a.simd_load[nelts](x))
        c.simd_store[nelts](x, a.simd_load[nelts](x) - b.simd_load[nelts](x))
        #print(nelts, x)
    vectorize[nelts, diff](size)

    for i in range(size):
        print(c[i])

    #print(a.simd_load[nelts](0))

from algorithm import vectorize
from tensor import Tensor

alias type = DType.int64
alias nelts = simdwidthof[type]()

def main():
    let size = 10
    var a = Tensor[type](size)
    var b = Tensor[type](size)
    var c = Tensor[type](size)

    for i in range(size):
        a[i] = i + 1
        b[i] = i
    
    @parameter
    fn diff[nelts : Int](x : Int):
        #print(a.simd_load[nelts](x))
        c.simd_store[nelts](x, a.simd_load[nelts](x) - b.simd_load[nelts](x))
        #print(nelts, x)
    vectorize[nelts, diff](size)

    for i in range(size):
        print(c[i])

    #print(a.simd_load[nelts](0))

For example:

2305843009213693952
4611686018780248321
3
1
1
1
1
1
1
1407374883553281

2305843009213693952
4611686018780248321
3
1
1
1
1
1
1
1407374883553281

If I uncomment last print statement, suddenly results are correct:

Can someone please explain this behaviour? Am I doing something wrong with SIMD?

12 Replies

vmoisOP•14mo ago

I experience the same behaviour on both arm64 and x86

Helehex•14mo ago

I get similar results i think uncommenting #print(nelts, x) should give something like 4,4,2

vmoisOP•14mo ago

To make it work, I uncommeted last print print(a.simd_load[nelts](0)) I have M1, so it gives nelts = 2. Indexes provided by vectorize look fine. 0 2 4 6 8

Helehex•14mo ago

ah yeah i get 4, 4, 4, 4 for the one i mentioned and uncommenting the last line doesnt work for me

vmoisOP•14mo ago

Hmm. Interesting. Try to change size to multiple of 4, like 16?

Helehex•14mo ago

using 16: uncommenting the last print doesnt change it for me, other than the 2 undefined elements at the beginning which are still there

vmoisOP•14mo ago

I see. Surprisingly, my other SIMD implementation for euclidian distance (from Modular blog) is working fine. Maybe I am missing something. When I will have time, I will double check code and maybe look at assembly Thanks for looking

Helehex•14mo ago

the inner nelts looks correct for me actually- 4,4,1,1, and uncommenting does solve the issue in a certain sense my bad but still getting the 2 randos at the beginning: happens with simdload, inside of a function decorated with @parameter, using the tensor after the load solves the issue, so it may have to do with the lifetime to solve, you can do: ` = a` for each tensor (at the end of main in this case)

vmoisOP•14mo ago

What do you mean by simd_load inside the function? Should it still display a correct results after vectorize? But I see where you are pointing. Cause in all examples provided by Modular they returned from function at this point.

ModularBot•14mo ago

Congrats @vmois, you just advanced to level 3!

vmoisOP•14mo ago

_ = b also works. _ = c doesn't 🙂 Oh wait. Just saw your edir So by using _ = a I force compiler to provide a visibility for my tensor after vectorize? Anyway. That is quite interesting. Thank you for the help.

Helehex•14mo ago

no problem

Gaming

Programming

SIMD produces weird results without print statement at the end