SSE & AVX Vectorization

Marchete

93.5K views

01 What is SSE and AVX?

02 Prerequisites

03 Autovectorization

04 SSE and AVX Usage

05 First AVX Code: SQRT calculation

06 SSE/AVX C++ Frameworks

07 Masking and Conditional Load

08 Controlling the Data Flow

09 Final Words

Previous: SSE and AVX Usage Next: SSE/AVX C++ Frameworks

First AVX Code: SQRT calculation

Now that we have reviewed all the requirements, the autovectorization, and AVX intrinsics, we can create our first manually vectorized program. In this exercise, you need to vectorize a sqrt calculation of float numbers. We will explicitly use the __m256 datatype to store our floats, reducing the overhead in data loading.

Vectorized SQRT

#pragma GCC optimize("O3","unroll-loops","omit-frame-pointer","inline") //Optimization flags
#pragma GCC option("arch=native","tune=native","no-zeroupper") //Enable AVX
#pragma GCC target("avx")  //Enable AVX
#include <x86intrin.h> //SSE Extensions
#include <bits/stdc++.h> //All main STD libraries
const int N = 64000000; //Number of tests
const int V = N/8;      //Vectorized size
//linear function, 
float linear[N];
__attribute__((optimize("no-tree-vectorize"))) //Force disable auto-vectorization
inline void normal_sqrt()
{
    for (int i = 0; i < N; ++i)
        linear[i] = sqrtf(linear[i]);
}
//Exercise 1: Create a vectorized version of the "linear" function.
//Please note the following:
// "vectorized" array is size V=N/8, because each __m256 variable holds 8 floats.
// sqrtf(const float& f) vectorized function is: _mm256_sqrt_ps(const __m256& v)
__m256 __attribute__((aligned(32))) vectorized[V]; //Vectorized array
inline void avx_sqrt()
{
  //****** Add AVX code here*******
}
 
 
using namespace std;
using namespace std::chrono;
 
high_resolution_clock::time_point now = high_resolution_clock::now();
#define TIME duration_cast<duration<double>>(high_resolution_clock::now() - now).count()
int main()
{
//Data initialization   
    for (int i = 0; i < N; ++i) { linear[i] = ((float)i)+ 0.1335f; }
    for (int i = 0; i < V; ++i) {
        for (int v=0;v<8;++v)
         {  vectorized[i][v] = ((float)(i*8+v))+ 0.1335f; }
    }
 
 //Normal sqrt benchmarking. 20*64 Million Sqrts
    now = high_resolution_clock::now();
    for (int i = 0; i < 20; ++i)
    normal_sqrt();
    double linear_time = TIME;
    cerr << "Normal sqrtf: "<< linear_time << endl;
 

You will probably see a 600% performance improvement or more. That is, once you have the data loaded, AVX will perform up to 7 times faster than normal sqrtf. The theoretical limit is 800%, but it's rarely achieved. You can expect between a 300% and 600% average increase.