Porting analysis

A Sobel filter implementation is used as the example application as it is an applicable embedded computer vision workload. The Sobel SIMD OpenCV repo is implemented in three different ways which makes it a great example candidate to show different aspects of porting.

It is implemented in the following ways:

  • Non-SIMD
    • pure C++ code
  • SIMD
    • x86_64 intrinsics
  • OpenCV
    • a computer vision library

The application builds and runs on an x86_64 machine. The application runs on CPU only (no hardware acceleration).

You will follow the porting methodology and gather information about the application.

Programming languageC++
OSUbuntu22.04 LTS
Build toolsCMake3.22.1
External librariesOpenCV4.5.4

This table is the starting point for the porting analysis.

Using the original software and tool versions when porting an application isn’t a requirement, however it is recommended as it will make the porting smoother. By looking at the Sobel filter code and with the questions from the previous section in mind, you can start the porting analysis.

In src/CMakeLists.txt#L12 :


        # Enable SIMD instructions for Intel Intrinsics
# https://software.intel.com/sites/landingpage/IntrinsicsGuide/
if(NOT WIN32)


The flag -mavx used with GCC is architecture-specific. It is only available on x86_64 and will prevent the application from compiling. Even though the application won’t compile, no changes to the source code for the non-SIMD and OpenCV implementations are necessary.

In src/main.cpp#L26 :


        /* GCC-compatible compiler, targeting x86/x86-64 */
#include <x86intrin.h>


The header file x86intrin.h isn’t supported on aarch64.

In src/main.cpp#L253 :


        p1 = _mm_loadu_si128((__m128i *)(inputPointer + i * width + j));
p2 = _mm_loadu_si128((__m128i *)(inputPointer + i * width + j + 1));
p3 = _mm_loadu_si128((__m128i *)(inputPointer + i * width + j + 2));


The lines of code above is just a snippet from the function SobelSimd which has intrinsics prefixed with _mm_. These aren’t supported on aarch64 and will need to be ported for the application to compile on aarch64.

The table below summarizes the migration analysis.

versionavailable on ArmComment
Ubuntu22.04 LTSYesUbuntu for Arm
Compiler option -mavxN/ANox86-specific
AVX intrinsicsN/ANox86-specific

You can draw the following conclusions:

  • The compiler options need to be modified
  • The AVX intrinsics need to ported to utilize Arm SIMD intrinsics
    • Arm has three SIMD technologies
      • NEON
      • Scalable Vector Extension ( SVE )
      • Scalable Vector Extension version 2 ( SVE2 )