Using a Struct with SIMD Intrinsics

Question

Let's say I have a struct

typdef struct point{
 
  float x;
  float y;
  float z;

} point;

I have an array of these structs and I want to do the following -

std::vector<point> copyArray;

for(auto p : array_of_points){
   
    point newPoint;

    newPoint.x = p.x;
    newPoint.y = p.y;
    newPoint.z = p.z;

    copyArray.push_back(newPoint);
}

Now this particular section in my code can be accelerated using vector operations, if I can operate on multiple structs at once.

I have a two part question

How can this be done using SIMD intrinsic. I am not sure how I would load structs.
Can OpenMP achieve the desired vectorization. I am not that comfortable with assembly and was not able to figure out whether it was indeed being vectorized efficiently or not.

If all you want to do is copy the vector then just do `std::vector copyArray(array_of_points.begin(), array_of_points.end())`, the compiler is likely to optimise this to a SIMD intrinsic for you — Alan Birtles, Oct 23 '21 at 15:27
SIMD generally works better when you have 3 separate arrays, one each for x, y, and z. You *can* do stuff with a geometry vector inside a single SIMD vector, but it's clunky and slower (for stuff other than copying) than processing 4 xyz geometry vectors at a time, even if you pad your SIMD vectors with xyzw with unused w. See https://stackoverflow.com/tags/sse/info, especially [Slides + text: SIMD at Insomniac Games (GDC 2015)](https://deplinenoise.wordpress.com/2015/03/06/slides-simd-at-insomniac-games-gdc-2015/) which specifically covers this SIMD anti-pattern. — Peter Cordes, Oct 23 '21 at 15:31
e.g. [this code optimises](https://godbolt.org/z/7raY43Wcx) down to a `memcpy` which will use intrinsics internally — Alan Birtles, Oct 23 '21 at 15:34
for SIMD you should use [SoA instead of AoS](https://en.wikipedia.org/wiki/AoS_and_SoA). See [Structure of Arrays vs Array of Structures](https://stackoverflow.com/q/17924705/995714), [Improving Vectorization Efficiency using Intel SIMD Data Layout Template](https://www.intel.com/content/dam/www/public/us/en/documents/presentation/improving-vectorization-efficiency.pdf) — phuclv, Oct 23 '21 at 16:05
What is `array_of_points`? If it was also a `std::vector`, just copy-construct `copyArray` from that. If it was a vector of different structs (different type of `x,y,z`, or additional member variables) the question would be different. You should always provide a [mre]! — chtz, Oct 24 '21 at 00:48

score 0 · Answer 1 · answered Oct 23 '21 at 15:43

0

I would expect the loop can not be vectorized as copyArray.push_back(newPoint); is accessing a shared resource.

If you want to speed this up, you might want to look into how to convert arrays into vectors quickly. You could to start your search here.

answered Oct 23 '21 at 15:43

Anthrados

56
1
3

1

`std::vector` isn't thread-safe, so any thread accessing it will assume it's the only thread. (In C++, it would be data-race undefined-behaviour to have multiple threads doing `.push_back()` on the same `std::vector` at the same time, so compilers can and do assume that doesn't happen. That's why compilers are able to optimize normal code well.) – Peter Cordes Oct 23 '21 at 16:07
So TL:DR, even if `std::vector copyArray;` were global, not a local inside the same function, that wouldn't stop auto-vectorization. If the compiler manages to untangle the copy in 3x4 = 12-byte chunks into a simple memcpy, it should vectorize or recognize it as memset. – Peter Cordes Oct 23 '21 at 16:51

Using a Struct with SIMD Intrinsics

1 Answers1