Results of tbb::parallel_reduce and std::accumulate differ - c++

Question

I am learning Intel's TBB library. When summing all values in a std::vector the result of tbb::parallel_reduce differs from std::accumulate in the case of more than 16.777.220 elements in the vector (errors experienced at 16.777.320 elements). Here is my minimum-working-example:
#include <iostream>
#include <vector>
#include <numeric>
#include <limits>
#include "tbb/tbb.h"
int main(int argc, const char * argv[]) {
int count = std::numeric_limits<int>::max() * 0.0079 - 187800; // - 187900 works
std::vector<float> heights(size);
std::fill(heights.begin(), heights.end(), 1.0f);
float ssum = std::accumulate(heights.begin(), heights.end(), 0);
float psum = tbb::parallel_reduce(tbb::blocked_range<std::vector<float>::iterator>(heights.begin(), heights.end()), 0,
[](tbb::blocked_range<std::vector<float>::iterator> const& range, float init) {
return std::accumulate(range.begin(), range.end(), init);
}, std::plus<float>()
);
std::cout << std::endl << " Heights serial sum: " << ssum << " parallel sum: " << psum;
return 0;
}
which outputs on my OSX 10.10.3 with XCode 6.3.1 and tbb stable 4.3-20141023 (poured from Brew):
Heights serial sum: 1.67772e+07 parallel sum: 1.67773e+07
Why is that? Should I report an error to TBB developers?
Additional testing, applying your answers:
correct value is: 1949700403
cause we add 1.0f to zero 1949700403 times
using (int) init values:
Runtime: 17.407 sec. Heights serial sum: 16777216.000, wrong
Runtime: 8.482 sec. Heights parallel sum: 131127368.000, wrong
using (float) init values:
Runtime: 12.594 sec. Heights serial sum: 16777216.000, wrong
Runtime: 5.044 sec. Heights parallel sum: 303073632.000, wrong
using (double) initial values:
Runtime: 13.671 sec. Heights serial sum: 1949700352.000, wrong
Runtime: 5.343 sec. Heights parallel sum: 263690016.000, wrong
using (double) initial values and tbb::parallel_deterministic_reduce:
Runtime: 13.463 sec. Heights serial sum: 1949700352.000, wrong
Runtime: 99.031 sec. Heights parallel sum: 1949700352.000, wrong >>> almost 10x slower !
Why do all reduce calls produce the wrong sum? Is (double) not sufficient?
Here is my testing code:
#include <iostream>
#include <vector>
#include <numeric>
#include <limits>
#include <sys/time.h>
#include <iomanip>
#include "tbb/tbb.h"
#include <cmath>
class StopWatch {
private:
double elapsedTime;
timeval startTime, endTime;
public:
StopWatch () : elapsedTime(0) {}
void startTimer() {
elapsedTime = 0;
gettimeofday(&startTime, 0);
}
void stopNprintTimer() {
gettimeofday(&endTime, 0);
elapsedTime = (endTime.tv_sec - startTime.tv_sec) * 1000.0; // compute sec to ms
elapsedTime += (endTime.tv_usec - startTime.tv_usec) / 1000.0; // compute us to ms and add
std::cout << " Runtime: " << std::right << std::setw(6) << elapsedTime / 1000 << " sec."; // show in sec
}
};
int main(int argc, const char * argv[]) {
StopWatch watch;
std::cout << std::fixed << std::setprecision(3) << "" << std::endl;
size_t count = std::numeric_limits<int>::max() * 0.9079;
std::vector<float> heights(count);
std::cout << " Vector size: " << count << std::endl;
std::fill(heights.begin(), heights.end(), 1.0f);
watch.startTimer();
float ssum = std::accumulate(heights.begin(), heights.end(), 0.0); // change type of initial value here
watch.stopNprintTimer();
std::cout << " Heights serial sum: " << std::right << std::setw(8) << ssum << std::endl;
watch.startTimer();
float psum = tbb::parallel_reduce(tbb::blocked_range<std::vector<float>::iterator>(heights.begin(), heights.end()), 0.0, // change type of initial value here
[](tbb::blocked_range<std::vector<float>::iterator> const& range, float init) {
return std::accumulate(range.begin(), range.end(), init);
}, std::plus<float>()
);
watch.stopNprintTimer();
std::cout << " Heights parallel sum: " << std::right << std::setw(8) << psum << std::endl;
return 0;
}
Answer to my last question: they all produce wrong results because they are not made for integer addition with large numbers. Switching to int solves that:
[...]
std::vector<int> heights(count);
std::cout << " Vector size: " << count << std::endl;
std::fill(heights.begin(), heights.end(), 1);
watch.startTimer();
int ssum = std::accumulate(heights.begin(), heights.end(), (int)0);
watch.stopNprintTimer();
std::cout << " Heights serial sum: " << std::right << std::setw(8) << ssum << std::endl;
watch.startTimer();
int psum = tbb::parallel_reduce(tbb::blocked_range<std::vector<int>::iterator>(heights.begin(), heights.end()), (int)0,
[](tbb::blocked_range<std::vector<int>::iterator> const& range, int init) {
return std::accumulate(range.begin(), range.end(), init);
}, std::plus<int>()
);
watch.stopNprintTimer();
std::cout << " Heights parallel sum: " << std::right << std::setw(8) << psum << std::endl;
[...]
results in:
Vector size: 1949700403
Runtime: 13.041 sec. Heights serial sum: 1949700403, correct
Runtime: 4.728 sec. Heights parallel sum: 1949700403, correct and almost 4x faster

Answer 1

Your call to std::accumulate is doing integer addition, then transforming the result to float at the end of the calculation. In order to accumulate over floating point numbers, the accumulator should be a float*.
float ssum = std::accumulate(heights.begin(), heights.end(), 0.0f);
^^^^
* Or any other type that can accumulate float correctly.

Answer 2

To other correct answers for the 'why?' part, I'd also add that TBB provides parallel_deterministic_reduce which guarantees reproducible results between two and more runs on the same data (but it still can differ with std::accumulate). See the blog describing the issue and the deterministic algorithm.
Thus regarding 'Should I report an error to TBB developers?' part, the answer is obviously no (unless you find something insufficient on the TBB side).

Answer 3

This may fix this particular problem for you:
Your call to std::accumulate is doing integer addition, then transforming the result to float at the end of the calculation.
BUT floating point addition is NOT an associative operation:
With accumulate: (...((s+a1)+a2)+...)+an
With parralel_reduce: any parenthesis permutation possible.
http://docs.oracle.com/cd/E19957-01/806-3568/ncg_goldberg.html