coder.loop.parallelize

Parallelize specific for loops in generated code; disable automatic parallelization

Since R2021a

Syntax

coder.loop.parallelize('never')

coder.loop.parallelize("loopID")

loopObj = coder.loop.parallelize(___)

Description

coder.loop.parallelize('never') disables automatic parallelization of the for loop placed immediately after it. This pragma overrides the EnableAutoParallelization configuration setting.

This pragma supports explicit for loops only. For more information on explicit and implicit loops, see Parallelization of Explicit and Implicit for-loops.

coder.loop.parallelize("loopID") parallelizes the for-loop whose index name is loopID.

This prompts the generated code for that loop to execute the iterations in parallel with the threads available for your target. This transforms requires EnableOpenMP to be set to true in your code configuration object.

For more information about loop optimizations, see Optimize Loops in Generated Code.

loopObj = coder.loop.parallelize(___) creates a loop control object with transformSchedule property set to coder.loop.parallelize. Use the apply method to apply the transform to the specified loop.

Examples

Disable Automatic Parallelization of a Specific for loop

This example shows how to disable the automatic parallelization of a specific for loop in your MATLAB^® code by using the coder.loop.parallelize('never') directive.

Define the MATLAB function autoparExample:

function [x, y] = autoparExample(x, y) %#codegen
x = x * 17;
% Pragma to disable automatic parallelization of for-loops
coder.loop.parallelize('never');
for i = 10:numel(x)
    y(i) = sqrt(y(i));
end
end

This function contains two loops:

• The mtimes (matrix multiplication) function contains an implicit for loop that iterates over the elements of the matrix x.

• The explicit for loop that calculates the square root of a subset of the elements of the matrix y. The coder.loop.parallelize('never') directive disables automatic parallelization of this loop.

Create a code generation configuration object for a static library. Set the EnableAutoParallelization property to true. Specify the input types as 1-by-2000 row vectors of doubles and generate code.

cfg = coder.config('lib');
cfg.EnableAutoParallelization = 1;
codegen -config cfg autoparExample -args {rand(1,2000) rand(1,2000)} -report

Code generation successful: View report

Open the code generation report and inspect the generated autoparExample function:

• The generated code contains a parallel loop for the multiplication operation.

• The explicit for loop is converted to a serial loop in the generated code. This loop was preceded by the coder.loop.parallelize('never') directive in your MATLAB code. The code generator does not parallelize it.

void autoparExample(double x[2000], double y[2000])
{
  int b_i;
  int i;
  if (!isInitialized_autoparExample) {
    autoparExample_initialize();
  }
#pragma omp parallel for num_threads(omp_get_max_threads()) private(i)

  for (i = 0; i < 2000; i++) {
    x[i] *= 17.0;
  }
  /*  Pragma to disable automatic parallelization of for-loops */
  for (b_i = 0; b_i < 1991; b_i++) {
    y[b_i + 9] = sqrt(y[b_i + 9]);
  }
}

Selectively Parallelize for Loops in Generated Code

You can choose to parallelize specific for-loops in the generated code by using the coder.loop.parallelize function in your MATLAB code.

Define a MATLAB function forLoopFuncParallel. Within the function, call coder.loop.parallelize with the for-loop index to be parallelized. Also call the function to prevent parallelization of a specific loop.

type forLoopFuncParallel

function [out1,out2] = forLoopFuncParallel
out1 = ones(1,100);
out2 = zeros(1,100);

coder.loop.parallelize("i");
for i = 1:100
    out1(i) = out1(i) * i;
end

coder.loop.parallelize("j","never");
for j = 1:100
    out2(j) = out2(j) + j;
end

Generate code for this function at the command line.

codegen forLoopFuncParallel -config:lib

Warning: Code generation is using a coder.EmbeddedCodeConfig object. Because
Embedded Coder is not installed, this might cause some Embedded Coder features
to fail.


Code generation successful (with warnings): View report

Inspect the generated code. The code generator uses OpenMP to parallelize the loop.

type(fullfile("codegen","lib","forLoopFuncParallel","forLoopFuncParallel.c"))

/*
 * Prerelease License - for engineering feedback and testing purposes
 * only. Not for sale.
 * File: forLoopFuncParallel.c
 *
 * MATLAB Coder version            : 25.1
 * C/C++ source code generated on  : 01-Feb-2025 08:03:47
 */

/* Include Files */
#include "forLoopFuncParallel.h"
#include "forLoopFuncParallel_data.h"
#include "forLoopFuncParallel_initialize.h"
#include "omp.h"
#include <emmintrin.h>

/* Function Definitions */
/*
 * Arguments    : double out1[100]
 *                double out2[100]
 * Return Type  : void
 */
void forLoopFuncParallel(double out1[100], double out2[100])
{
  double dv[2];
  int i;
  int j;
  if (!isInitialized_forLoopFuncParallel) {
    forLoopFuncParallel_initialize();
  }
  for (j = 0; j < 100; j++) {
    out1[j] = 1.0;
    out2[j] = 0.0;
  }
#pragma omp parallel for num_threads(omp_get_max_threads())

  for (i = 0; i < 100; i++) {
    out1[i] *= (double)i + 1.0;
  }
  for (j = 0; j <= 98; j += 2) {
    __m128d r;
    __m128d r1;
    dv[0] = j;
    dv[1] = (double)j + 1.0;
    r = _mm_loadu_pd(&dv[0]);
    r1 = _mm_loadu_pd(&out2[j]);
    _mm_storeu_pd(&out2[j], _mm_add_pd(r1, _mm_add_pd(r, _mm_set1_pd(1.0))));
  }
}

/*
 * File trailer for forLoopFuncParallel.c
 *
 * [EOF]
 */

Input Arguments

loopID — for-loop index name

string scalar | character vector

for loop identifier or index name to parallelize, specified as a character vector a string scalar.

Data Types: char | string

Output Arguments

loopObj — Loop control object

coder.loop.Control object

Loop control object with transformSchedule property set to coder.loop.parallelize.

Extended Capabilities

C/C++ Code Generation
Generate C and C++ code using MATLAB® Coder™.

GPU Code Generation
Generate CUDA® code for NVIDIA® GPUs using GPU Coder™.

Version History

Introduced in R2021a