Documentation

Enable SIMD in Generated Code

To enable SIMD replacement, in the MATLAB Coder app, after selecting the required Device vendor and Device type, on the Custom Code tab, set the Code replacement library parameter. This table shows the code replacement libraries for the supported Device vendor and Device type.

Alternatively, in a coder.EmbeddedCodeConfig configuration object, set the parameter CodeReplacementLibrary parameter to a library such as 'Intel AVX (Windows)'.

cfg = coder.config('lib');
cfg.CodeReplacementLibrary = 'Intel AVX (Windows)';

Generate SIMD Code for Loops and Arrays

Consider the MATLAB function dynamic that has a sum of products expression and variables with data type single. The loop bound is not set at compile time. The table displays the generated code when no code replacement library is selected and the loop executes one iteration at a time. When you choose Intel SSE(Windows) as code replacement library, the generated code contains SIMD instructions. SIMD instructions can process the loop in increments of four through the computational function, such as _mm_add_ps and mm_mul_ps.The data management instructions _mm_storeu_ps and _mm_loadu_ps store and load data from the SIMD registers. If the data type is double, the loop executes in increments of two. Variables that cannot be vectorized are processed in the loop by using the iterator scalarLB.

function C = dynamic(A, B)
   assert(all(size(A) <= [100 100]));
   assert(all(size(B) <= [100 100]));
   assert(isa(A, 'single'));
   assert(isa(B, 'single'));

   C = zeros(size(A), 'like', A);
   for i = 1:numel(A)
       C(i) = (A(i) .* B(i)) + 
                   (A(i) .* B(i));
   end
end

loop_ub = A_size[0] * A_size[1];
for (i = 0; i < loop_ub; i++) {
  C_data_tmp = A_data[i] * B_data[i];
  C_data[i] = C_data_tmp + C_data_tmp;
}

loop_ub = A_size[0] * A_size[1];
scalarLB = loop_ub & -4;
vectorUB = scalarLB - 4;
for (i = 0; i <= vectorUB; i += 4) {
   r = _mm_mul_ps(_mm_loadu_ps(&A_data[i]), 
       _mm_loadu_ps(&B_data[i]));
       _mm_storeu_ps(&C_data[i],
       _mm_add_ps(r, r));
}

for (i = scalarLB; i < loop_ub; i++) {
  C_data_tmp = A_data[i] * B_data[i];
  C_data[i] = C_data_tmp + C_data_tmp;
}

For a list of a Intel intrinsic functions for supported MATLAB functions, see https://software.intel.com/sites/landingpage/IntrinsicsGuide/. For a list of Inlined ARM NEON Intrinsic functions, see https://developer.arm.com/architectures/instruction-sets/simd-isas/neon/intrinsics.

Limitations

The generated code is not optimized through SIMD when the MATLAB code contains: