How to use std::vector<nsimd::pack> like std::vector<__m128> <p dir="a

Hi ustcsq, I am not sure how you get the segfault --> <a href="ht

Vector<nsimd::pack<T>>,about agenium-scale/nsimd

Comments (4)

gquintin commented on June 12, 2024

Hi ustcsq,

I am not sure how you get the segfault --> https://godbolt.org/z/z1se7oGe3
To help you I will need a fully functionnal snippet of code along with its compilation command-line.

My advice

Do not write code like this. The code you write won't even compile for certain SIMD extensions such as Arm SVE. You have to separate concerns : data structure and computations. You should use a std::vector<float> and in you computation code something like this:

template <typename T>
int computations(std::vector<T> &tmp) {
    typedef nsimd::pack<T> vec; // shortcut
    int s = nsimd::len<vec>(); // shortcut
    int n = (int)tmp.size();  // shortcut
    T *ptr = tmp.data(); // always work with raw pointers
    for (int i = 0; i + s <= tmp.size(); i += s) {
        vec v = nsimd::loadu<vec>(ptr + i);
        // ... do some work with v
        nsimd::storeu(ptr + i, v);
    }
}

from nsimd.

ustcsq commented on June 12, 2024

https://godbolt.org/z/z9fMd3o8n
I got the segfault under avx2 gcc version <=10.3
thks

from nsimd.

gquintin commented on June 12, 2024

It seems to be a bug of those versions of GCC. As you can see when disassembling the resulting binary and doing a run with GDB you can see the problem. A lot of people think that doing a std::vector<nsimd::pack<float>> or a std::vector<__m256> will be more optimized as there will be no need for loads/stores. But this is of course wrong. When writing tmp[0] a load (or store) instruction is generated by the compiler. And this is where is the problem. You cannot assume that data is properly aligned and obviously it is not but the compiler wrongly generated a movaps as you can see below

  friend std::ostream &operator<<(std::ostream &os, pack const &a0) {
    12b0:       4c 8d 54 24 08          lea    0x8(%rsp),%r10
    12b5:       48 83 e4 e0             and    $0xffffffffffffffe0,%rsp
        __ostream_insert(__out, __s,
    12b9:       ba 02 00 00 00          mov    $0x2,%edx
    12be:       41 ff 72 f8             pushq  -0x8(%r10)
    12c2:       55                      push   %rbp
    12c3:       48 89 e5                mov    %rsp,%rbp
    12c6:       41 56                   push   %r14
    12c8:       41 55                   push   %r13
    12ca:       41 54                   push   %r12
    12cc:       49 89 fc                mov    %rdi,%r12
    12cf:       41 52                   push   %r10
    12d1:       53                      push   %rbx
    12d2:       48 81 ec 28 01 00 00    sub    $0x128,%rsp
    T buf[max_len_t<T>::value];
    storeu(buf, a0.car, T(), SimdExt());
    12d9:       c5 fc 28 06             vmovaps (%rsi),%ymm0
    12dd:       48 8d 35 20 0d 00 00    lea    0xd20(%rip),%rsi        # 2004 <_IO_stdin_used+0x4>
}

extern __inline void __attribute__((__gnu_inline__, __always_inline__, __artificial__))
_mm256_storeu_ps (float *__P, __m256 __A)
{
  *(__m256_u *)__P = __A;
    12e4:       c5 fc 29 85 d0 fe ff    vmovaps %ymm0,-0x130(%rbp)
    12eb:       ff 
    12ec:       c5 fc 29 85 b0 fe ff    vmovaps %ymm0,-0x150(%rbp)
    12f3:       ff

To bypass this GCC bug you can do the following:

#include <nsimd/nsimd-all.hpp>
#include <iostream>

int main() {
    std::vector<nsimd::pack<float>, nsimd::allocator<float>> tmp(8);
    std::cout << "tmp : " << tmp.size() << ", " << tmp.capacity() << std::endl;
    std::cout << tmp[0] << std::endl;
    return 0;
}

The nsimd::allocator will force the proper alignment of data and the wrongly generated movaps by GCC will be given aligned pointers and it will work.

But again, please, do not write code like this.

from nsimd.

ustcsq commented on June 12, 2024

thks

from nsimd.

Vector<nsimd::pack<T>> about nsimd HOT 4 CLOSED

Comments (4)

My advice

Related Issues (20)

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent