Non-deterministic results with unaligned reductions

Submitted by Jose Rubio

Assigned to Nobody

Link to original bugzilla bug (#1728)
Version: 3.3 (current stable)

Description

OS: Debian 4.9.30-2+deb9u5 (2017-09-19) x86_64 GNU/Linux

I observed subtle differences when summing a sparse matrix across different runs.

This test reproduces the issue (fails around 50% of the time it's run)

#include <Eigen/Sparse>
#include <gtest/gtest.h>
#include <stdlib.h>
#include <time.h>

TEST(Sparse, Reduction)
{
srand(time(NULL));
int nrows = 11300;
int ncols = 600;
int num_non_zeros = 100;

std::vector<Eigen::Triplet<float, int> > triplets;  
for (int i = 0; i < num_non_zeros; i++) {  
    int row = rand() % nrows;  
    int col = rand() % ncols;  
    float value = static_cast&lt;float&gt;(rand()) / static_cast&lt;float&gt;(RAND_MAX);  
    triplets.push_back(Eigen::Triplet<float, int>(row, col, value));  
}  

Eigen::SparseMatrix<float, 0, int> mat = Eigen::SparseMatrix<float, 0, int>(nrows, ncols);  
mat.reserve(num_non_zeros);  
mat.setFromTriplets(triplets.begin(), triplets.end());  

int num_trials = 10000;  
for (int tr = 0; tr < num_trials; tr++) {  
    Eigen::SparseMatrix<float, 0, int> mat2 = Eigen::SparseMatrix<float, 0, int>(nrows, ncols);  
    mat2.reserve(num_non_zeros);  
    mat2.setFromTriplets(triplets.begin(), triplets.end());  
    EXPECT_TRUE(mat.sum() == mat2.sum());  
}  

}

Blocking

#1608