OpenMP and (Rcpp)Eigen - parallel-processing

I am wondering how to write code that at times makes use of OpenMP parallelization built into the Eigen library while at other times uses Parallelization that I specify. Hopefully, the below code snippet should provide background into my problem.
I am asking this question at the design stage of my library (sorry I don't have a working / broken code example).
#ifdef _OPENMP
#include <omp.h>
#endif
#include <RcppEigen.h>
void fxn(..., int ncores=-1){
if (ncores > 0) omp_set_num_threads(ncores);
/*
* Code with matrix products
* where I would like to use Eigen's
* OpenMP parallelization
*/
#pragma omp parallel for
for (int i=0; i < iter; i++){
/*
* Code I would like to parallelize "myself"
* even though it involves matrix products
*/
}
}
What is best practice for controlling the balance between Eigen's own parallelization with OpenMP and my own.
UPDATE:
I wrote a simple example and tested ggael's suggestion. In short, I am skeptical that it solves the problem I was posing (or I am doing something else wrong - apologies if its the latter). Notice that with explicit parallelization of the for loop there is no change in run-time (not even a slow
#ifdef _OPENMP
#include <omp.h>
#endif
#include <RcppEigen.h>
using namespace Rcpp;
// [[Rcpp::plugins(openmp)]]
// [[Rcpp::export]]
Eigen::MatrixXd testing(Eigen::MatrixXd A, Eigen::MatrixXd B, int n_threads=1){
Eigen::setNbThreads(n_threads);
Eigen::MatrixXd C = A*B;
Eigen::setNbThreads(1);
for (int i=0; i < A.cols(); i++){
A.col(i).array() = A.col(i).array()*B.col(i).array();
}
return A;
}
// [[Rcpp::export]]
Eigen::MatrixXd testing_omp(Eigen::MatrixXd A, Eigen::MatrixXd B, int n_threads=1){
Eigen::setNbThreads(n_threads);
Eigen::MatrixXd C = A*B;
Eigen::setNbThreads(1);
#pragma omp parallel for num_threads(n_threads)
for (int i=0; i < A.cols(); i++){
A.col(i).array() = A.col(i).array()*B.col(i).array();
}
return A;
}
/*** R
A <- matrix(rnorm(1000*1000), 1000, 1000)
B <- matrix(rnorm(1000*1000), 1000, 1000)
microbenchmark::microbenchmark(testing(A,B, n_threads=1),
testing_omp(A,B, n_threads=1),
testing(A,B, n_threads=8),
testing_omp(A,B, n_threads=8),
times=10)
*/
Unit: milliseconds
expr min lq mean median uq max neval cld
testing(A, B, n_threads = 1) 169.74272 183.94500 212.83868 218.15756 236.97049 264.52183 10 b
testing_omp(A, B, n_threads = 1) 166.53132 178.48162 210.54195 227.65258 234.16727 238.03961 10 b
testing(A, B, n_threads = 8) 56.03258 61.16001 65.15763 62.67563 67.37089 83.43565 10 a
testing_omp(A, B, n_threads = 8) 54.18672 57.78558 73.70466 65.36586 67.24229 167.90310 10 a

The easiest is probably to disable/enable Eigen's multi-threading at runtime:
Eigen::setNbThreads(1); // single thread mode
#pragma omp parallel for
for (int i=0; i < iter; i++){
// Code I would like to parallelize "myself"
// even though it involves matrix products
}
Eigen::setNbThreads(0); // restore default

Related

Difference between mutual exclusion like atomic and reduction in OpenMP

I'm am following video lectures of Tim Mattson on OpenMP and there was one exercise to find errors in provided code that count area of the Mandelbrot. So here is the solution that was provided:
#define NPOINTS 1000
#define MAXITER 1000
void testpoint(struct d_complex);
struct d_complex{
double r;
double i;
};
struct d_complex c;
int numoutside = 0;
int main(){
int i,j;
double area, error, eps = 1.0e-5;
#pragma omp parallel for default(shared) private(c,j) firstprivate(eps)
for(i = 0; i<NPOINTS; i++){
for(j=0; j < NPOINTS; j++){
c.r = -2.0+2.5*(double)(i)/(double)(NPOINTS)+eps;
c.i = 1.125*(double)(j)/(double)(NPOINTS)+eps;
testpoint(c);
}
}
area=2.0*2.5*1.125*(double)(NPOINTS*NPOINTS-numoutside)/(double)(NPOINTS*NPOINTS);
error=area/(double)NPOINTS;
printf("Area of Mandlebrot set = %12.8f +/- %12.8f\n",area,error);
printf("Correct answer should be around 1.510659\n");
}
void testpoint(struct d_complex c){
// Does the iteration z=z*z+c, until |z| > 2 when point is known to be outside set
// If loop count reaches MAXITER, point is considered to be inside the set
struct d_complex z;
int iter;
double temp;
z=c;
for (iter=0; iter<MAXITER; iter++){
temp = (z.r*z.r)-(z.i*z.i)+c.r;
z.i = z.r*z.i*2+c.i;
z.r = temp;
if ((z.r*z.r+z.i*z.i)>4.0) {
#pragma omp atomic
numoutside++;
break;
}
}
}
The question I have is, could we use reduction in #pragma omp parallel of variable numoutside like:
#pragma omp parallel for default(shared) private(c,j) firstprivate(eps) reduction(+:numoutside)
without atomic construct in testpoint function?
I tested the function without atomic, and the result was different from the one I got in the first place. Why does that happen? And while I understand the concept of mutual exclusion and use of it because of race conditioning, isn't reduction just another form of solving that problem with private variables?
Thank You in advance.

Matrix multiplication via std::vector is 10 times slower than numpy

Although it is known that using nested std::vector to represent matrices is a bad idea, let's use it for now since it is flexible and many existing functions can handle std::vector.
I thought, in small cases, the speed difference can be ignored. But it turned out that vector<vector<double>> is 10+ times slower than numpy.dot().
Let A and B be matrices whose size is sizexsize. Assuming square matrices is just for simplicity. (We don't intend to limit discussion to the square matrices case.) We initialize each matrix in a deterministic way, and finally calculate C = A * B.
We define "calculation time" as the time elapsed just to calculate C = A * B. In other words, various overheads are not included.
Python3 code
import numpy as np
import time
import sys
if (len(sys.argv) != 2):
print("Pass `size` as an argument.", file = sys.stderr);
sys.exit(1);
size = int(sys.argv[1]);
A = np.ndarray((size, size));
B = np.ndarray((size, size));
for i in range(size):
for j in range(size):
A[i][j] = i * 3.14 + j
B[i][j] = i * 3.14 - j
start = time.time()
C = np.dot(A, B);
print("{:.3e}".format(time.time() - start), file = sys.stderr);
C++ code
using namespace std;
#include <iostream>
#include <vector>
#include <chrono>
int main(int argc, char **argv) {
if (argc != 2) {
cerr << "Pass `size` as an argument.\n";
return 1;
}
const unsigned size = atoi(argv[1]);
vector<vector<double>> A(size, vector<double>(size));
vector<vector<double>> B(size, vector<double>(size));
for (int i = 0; i < size; ++i) {
for (int j = 0; j < size; ++j) {
A[i][j] = i * 3.14 + j;
B[i][j] = i * 3.14 - j;
}
}
auto start = chrono::system_clock::now();
vector<vector<double>> C(size, vector<double>(size, /* initial_value = */ 0));
for (int i = 0; i < size; ++i) {
for (int j = 0; j < size; ++j) {
for (int k = 0; k < size; ++k) {
C[i][j] += A[i][k] * B[k][j];
}
}
}
cerr << scientific;
cerr.precision(3);
cerr << chrono::duration<double>(chrono::system_clock::now() - start).count() << "\n";
}
C++ code (multithreaded)
We also wrote a multithreaded version of C++ code since numpy.dot() is automatically calculated in parallel.
You can get all the codes from GitHub.
Result
C++ version is 10+ times slower than Python 3 (with numpy) version.
matrix_size: 200x200
--------------- Time in seconds ---------------
C++ (not multithreaded): 8.45e-03
C++ (1 thread): 8.66e-03
C++ (2 threads): 4.68e-03
C++ (3 threads): 3.14e-03
C++ (4 threads): 2.43e-03
Python 3: 4.07e-04
-----------------------------------------------
matrix_size: 400x400
--------------- Time in seconds ---------------
C++ (not multithreaded): 7.011e-02
C++ (1 thread): 6.985e-02
C++ (2 threads): 3.647e-02
C++ (3 threads): 2.462e-02
C++ (4 threads): 1.915e-02
Python 3: 1.466e-03
-----------------------------------------------
Question
Is there any way to make the C++ implementation faster?
Optimizations I Tried
swap calculation order -> at most 3.5 times faster (not than numpy code but than C++ code)
optimization 1 plus partial unroll -> at most 4.5 times faster, but this can be done only when size is known in advance No. As pointed out in this comment, size is not needed to be known. We can just limit the max value of loop variables of unrolled loops and process remaining elements with normal loops. See my implementation for example.
optimization 2, plus minimizing the call of C[i][j] by introducing a simple variable sum -> at most 5.2 times faster. The implementation is here. This result implies std::vector::operator[] is un-ignorably slow.
optimization 3, plus g++ -march=native flag -> at most 6.2 times faster (By the way, we use -O3 of course.)
Optimization 3, plus reducing the call of operator [] by introducing a pointer to an element of A since A's elements are sequentially accessed in the unrolled loop. -> At most 6.2 times faster, and a little little bit faster than Optimization 4. The code is shown below.
g++ -funroll-loops flag to unroll for loops -> no change
g++ #pragma GCC unroll n -> no change
g++ -flto flag to turn on link time optimizations -> no change
Block Algorithm -> no change
transpose B to avoid cache miss -> no change
long linear std::vector instead of nested std::vector<std::vector>, swap calculation order, block algorithm, and partial unroll -> at most 2.2 times faster
Optimization 1, plus PGO(profile-guided optimization) -> 4.7 times faster
Optimization 3, plus PGO -> same as Optimization 3
Optimization 3, plus g++ specific __builtin_prefetch() -> same as Optimization 3
Current Status
(originally) 13.06 times slower -> (currently) 2.10 times slower
Again, you can get all the codes on GitHub. But let us cite some codes, all of which are functions called from the multithreaded version of C++ code.
Original Code (GitHub)
void f(const vector<vector<double>> &A, const vector<vector<double>> &B, vector<vector<double>> &C, unsigned row_start, unsigned row_end) {
const unsigned j_max = B[0].size();
const unsigned k_max = B.size();
for (int i = row_start; i < row_end; ++i) {
for (int j = 0; j < j_max; ++j) {
for (int k = 0; k < k_max; ++k) {
C[i][j] += A[i][k] * B[k][j];
}
}
}
}
Current Best Code (GitHub)
This is the implementation of the Optimization 5 above.
void f(const vector<vector<double>> &A, const vector<vector<double>> &B, vector<vector<double>> &C, unsigned row_start, unsigned row_end) {
static const unsigned num_unroll = 5;
const unsigned j_max = B[0].size();
const unsigned k_max_for_unrolled_loop = B.size() / num_unroll * num_unroll;
const unsigned k_max = B.size();
for (int i = row_start; i < row_end; ++i) {
for (int k = 0; k < k_max_for_unrolled_loop; k += num_unroll) {
for (int j = 0; j < j_max; ++j) {
const double *p = A[i].data() + k;
double sum;
sum = *p++ * B[k][j];
sum += *p++ * B[k+1][j];
sum += *p++ * B[k+2][j];
sum += *p++ * B[k+3][j];
sum += *p++ * B[k+4][j];
C[i][j] += sum;
}
}
for (int k = k_max_for_unrolled_loop; k < k_max; ++k) {
const double a = A[i][k];
for (int j = 0; j < j_max; ++j) {
C[i][j] += a * B[k][j];
}
}
}
}
We've tried many optimizations since we first posted this question. We spent whole two days struggling with this problem, and finally reached the point where we have no more idea how to optimize the current best code. We doubt more complex algorithms like Strassen's will do it better since cases we handle are not large and each operation on std::vector is so expensive that, as we've seen, just reducing the call of [] improved the performance well.
We (want to) believe we can make it better, though.
Matrix multiplication is relativly easy to optimize. However if you want to get to decent cpu utilization it becomes tricky because you need deep knowledge of the hardware you are using. The steps to implement a fast matmul kernel are the following:
Use SIMDInstructions
Use Register Blocking and fetch multiple data at once
Optimize for your chache lines (mainly L2 and L3)
Parallelize your code to use multiple threads
Under this linke is a very good ressource, that explains all the nasty details:
https://gist.github.com/nadavrot/5b35d44e8ba3dd718e595e40184d03f0
If you want more indepth advise leave a comment.

Matrix Multiplication OpenMP Counter-Intuitive Results

I am currently porting some code over to OpenMP at my place of work. One of the tasks I am doing is figuring out how to speed up matrix multiplication for one of our applications.
The matrices are stored in row-major format, so A[i*cols +j] gives the A_i_j element of the matrix A.
The code looks like this (uncommenting the pragma parallelises the code):
#include <omp.h>
#include <iostream>
#include <iomanip>
#include <stdio.h>
#define NUM_THREADS 8
#define size 500
#define num_iter 10
int main (int argc, char *argv[])
{
// omp_set_num_threads(NUM_THREADS);
int *A = new int [size*size];
int *B = new int [size*size];
int *C = new int [size*size];
for (int i=0; i<size; i++)
{
for (int j=0; j<size; j++)
{
A[i*size+j] = j*1;
B[i*size+j] = i*j+2;
C[i*size+j] = 0;
}
}
double total_time = 0;
double start = 0;
for (int t=0; t<num_iter; t++)
{
start = omp_get_wtime();
int i, k;
// #pragma omp parallel for num_threads(10) private(i, k) collapse(2) schedule(dynamic)
for (int j=0; j<size; j++)
{
for (i=0; i<size; i++)
{
for (k=0; k<size; k++)
{
C[i*size+j] += A[i*size+k] * B[k*size+j];
}
}
}
total_time += omp_get_wtime() - start;
}
std::setprecision(5);
std::cout << total_time/num_iter << std::endl;
delete[] A;
delete[] B;
delete[] C;
return 0;
}
What is confusing me is the following: why is dynamic scheduling faster than static scheduling for this task? Timing the runs and taking an average shows that static scheduling is slower, which to me is a bit counterintuitive since each thread is doing the same amount of work.
Also, am I correctly speeding up my matrix multiplication code?
Parallel matrix multiplication is non-trivial (have you even considered cache-blocking?). Your best bet is likely to be to use a BLAS Library for this, rather than writing it yourself. (Remember, "The best code is the code I do not have to write").
Wikipedia: Basic Linear Algebra Subprograms points to many implementations, a lot of which (including Intel Math Kernel Library) have free licenses.

parallelizing in openMP

I have the following code that I want to paralleize using OpenMP
for(m=0; m<r_c; m++)
{
for(n=0; n<c_c; n++)
{
double value = 0.0;
for(j=0; j<r_b; j++)
for(k=0; k<c_b; k++)
{
double a;
if((m-j)<0 || (n-k)<0 || (m-j)>r_a || (n-k)>c_a)
a = 0.0;
else
a = h_a[((m-j)*c_a) + (n-k)];
//printf("%lf\t", a);
value += h_b[(j*c_b) + k] * a;
}
h_c[m*c_c + n] = value;
//printf("%lf\t", h_c[m*c_c + n]);
}
//cout<<"row "<<m<<" completed"<<endl;
}
In this I want every thread to perform "for j" and "for k" simultaneouly.
I am trying to do using pragma omp parallel for before the "for m" loop but not getting the correct result.
How can I do this in an optimized manner. thanks in advance.
Depending exactly from which loop you want to parallelize, you have three options:
#pragma omp parallel
{
#pragma omp for // Option #1
for(m=0; m<r_c; m++)
{
for(n=0; n<c_c; n++)
{
double value = 0.0;
#pragma omp for // Option #2
for(j=0; j<r_b; j++)
for(k=0; k<c_b; k++)
{
double a;
if((m-j)<0 || (n-k)<0 || (m-j)>r_a || (n-k)>c_a)
a = 0.0;
else
a = h_a[((m-j)*c_a) + (n-k)];
//printf("%lf\t", a);
value += h_b[(j*c_b) + k] * a;
}
h_c[m*c_c + n] = value;
//printf("%lf\t", h_c[m*c_c + n]);
}
//cout<<"row "<<m<<" completed"<<endl;
}
}
//////////////////////////////////////////////////////////////////////////
// Option #3
for(m=0; m<r_c; m++)
{
for(n=0; n<c_c; n++)
{
#pragma omp parallel
{
double value = 0.0;
#pragma omp for
for(j=0; j<r_b; j++)
for(k=0; k<c_b; k++)
{
double a;
if((m-j)<0 || (n-k)<0 || (m-j)>r_a || (n-k)>c_a)
a = 0.0;
else
a = h_a[((m-j)*c_a) + (n-k)];
//printf("%lf\t", a);
value += h_b[(j*c_b) + k] * a;
}
h_c[m*c_c + n] = value;
//printf("%lf\t", h_c[m*c_c + n]);
}
}
//cout<<"row "<<m<<" completed"<<endl;
}
Test and profile each. You might find that option #1 is fastest if there isn't a lot of work for each thread, or you may find that with optimizations on, there is no difference (or even a slowdown) when enabling OMP.
Edit
I've adopted the MCVE supplied in the comments as follows:
#include <iostream>
#include <chrono>
#include <omp.h>
#include <algorithm>
#include <vector>
#define W_OMP
int main(int argc, char *argv[])
{
std::vector<double> h_a(9);
std::generate(h_a.begin(), h_a.end(), std::rand);
int r_b = 500;
int c_b = r_b;
std::vector<double> h_b(r_b * c_b);
std::generate(h_b.begin(), h_b.end(), std::rand);
int r_c = 500;
int c_c = r_c;
int r_a = 3, c_a = 3;
std::vector<double> h_c(r_c * c_c);
auto start = std::chrono::system_clock::now();
#ifdef W_OMP
#pragma omp parallel
{
#endif
int m,n,j,k;
#ifdef W_OMP
#pragma omp for
#endif
for(m=0; m<r_c; m++)
{
for(n=0; n<c_c; n++)
{
double value = 0.0,a;
for(j=0; j<r_b; j++)
{
for(k=0; k<c_b; k++)
{
if((m-j)<0 || (n-k)<0 || (m-j)>r_a || (n-k)>c_a)
a = 0.0;
else a = h_a[((m-j)*c_a) + (n-k)];
value += h_b[(j*c_b) + k] * a;
}
}
h_c[m*c_c + n] = value;
}
}
#ifdef W_OMP
}
#endif
auto end = std::chrono::system_clock::now();
auto elapsed = std::chrono::duration_cast<std::chrono::milliseconds>(end - start);
std::cout << elapsed.count() << "ms"
#ifdef W_OMP
"\t with OMP"
#else
"\t without OMP"
#endif
"\n";
return 0;
}
As a reference, I'm using VS2012 (OMP 2.0, grrr). I'm not sure when collapse was introduced, but apparently after 2.0. Optimizations were /O2 and compiled in Release x64.
Benchmarks
Using the original sizes of the loops (7,7,5,5) and therefore arrays, the results were 0ms without OMP and 1ms with. Verdict: optimizations were better, and the added overhead wasn't worth it. Also, the measurements are not reliable (too short).
Using the slightly larger sizes of the loops (100, 100, 100, 100) and therefore arrays, the results were about equal at about 108ms. Verdict: still not worth the naive effort, tweaking OMP parameters might tip the scale. Definitely not the x4 speedup I would hope for.
Using an even larger sizes of the loops (500, 500, 500, 500) and therefore arrays, OMP started to pull ahead. Without OMP 74.3ms, with 15s. Verdict: Worth it. Weird. I got a x5 speedup with four threads and four cores on an i5. I'm not going to try and figure out how that happened.
Summary
As has been stated in countless answers here on SO, it's not always a good idea to parallelize every for loop you come across. Things that can screw up your desired xN speedup:
Not enough work per thread to justify the overhead of creating the additional threads
The work itself is memory bound. This means that the CPU can be running at 1petaHz and you still won't see a speedup.
Memory access patterns. I'm not going to go there. Feel free to edit in the relevant info if you want it.
OMP parameters. The best choice of parameters will often be a result of this entire list (not including this item, to avoid recursion issues).
SIMD operations. Depending on what and how you're doing, the compiler may vectorize your operations. I have no idea if OMP will usurp the SIMD operations, but it is possible. Check your assembly (foreign language to me) to confirm.

Why does this OpenMP code work on dynamic scheduling but not on static?

I'm learning OpenMP by building a simple program to calculate pi using the following algorithm:
pi = 4/1 - 4/3 + 4/5 - 4/7 + 4/9...
The problem is that it does not work correctly when I change the scheduling to static. It works perfectly when the thread count is one. It also runs correctly under dynamic scheduling despite the result differing slightly every time it's run. Any idea what could be the problem?
#include <stdio.h>
#include <stdlib.h>
#include <omp.h>
#define N 100
#define CSIZE 1
#define nthread 2
int pi()
{
int i, chunk;
float pi = 0, x = 1;
chunk = CSIZE;
omp_set_num_threads(nthread);
#pragma omp parallel shared(i, x, chunk)
{
if (omp_get_num_threads() == 0)
{
printf("Number of threads = %d\n", omp_get_num_threads());
}
printf("Thread %d starting...\n", omp_get_thread_num());
#pragma omp for schedule(dynamic, chunk)
for (i = 1; i <= N; i++)
{
if (i % 2 == 0)
pi = pi - 4/x;
else
pi = pi + 4/x;
x = x + 2;
printf("Pi is currently %f at iteration %d with x = %0.0f on thread %d\n",
pi, i, x, omp_get_thread_num());
}
}
return EXIT_SUCCESS;
}
Using printf in the loop when I test your code makes dynamic do all the work on the first thread and none on the second (making the program effectively serial). If you remove the printf statement then you will find that the value of pi is random. This is because you have race conditions in x and pi.
Instead of using x you can divide by 2*i+1 (for i starting at zero). Also instead of using a branch to get the sign you can use sign = -2*(i%2)+1. To get pi you need to do a reduction using #pragma omp for schedule(static) reduction(+:pi).
#include <stdio.h>
#define N 10000
int main() {
float pi;
int i;
pi = 0;
#pragma omp parallel for schedule(static) reduction(+:pi)
for (i = 0; i < N; i++) {
pi += (-2.0f*(i&1)+1)/(2*i+1);
}
pi*=4.0f;
printf("%f\n", pi);
}

Resources