I want to accelerate my code using SSE or NEON, opencv has uniform interface for different device, so I'd like to use the SIMD feature of OpenCV. But I find the macro CV_SIMD128 is not defined outside opencv, so is there any way to use the v_int16x8 or v_reduce_sum structure or function safely?
The easiest way is to create OpenCV module. But you can also try to include intrin.hpp in your build (copy files if needed) and define some macros, e.g CV_SSE2.
// @file myapp.cpp
#define CV_SSE2
// #define CV_NEON // uncomment for ARM
#include <opencv2/core/hal/intrin.hpp>
...
And your app should have corresponding compiler options enabled, e.g. -mfpu=neon for GCC/Clang
Thank you for your help! Your solution is good, it works. But in this way I should detect whether the device support SSE or neon myself, I want to know if there is some way to get this information from OpenCV. I tried to include both <opencv2/core/cv_cpu_helper.h> and <opencv2/core/hal/intrin.hpp>, and I did not need to add any macros myself, it just worked. I want to know if this operation is normal.
If you wonn't declare any macro (e.g. CV_SSE2), default non-optimized implementation will be used: https://github.com/opencv/opencv/blob...
Universal intrinsics are header-only library, so they should not depend on OpenCV build options.
Asked: 2019-04-22 08:04:07 -0600
Seen: 2,286 times
Last updated: Apr 23 '19
