the code doesn't speed up while using Intel Intrinsics - c++

Question

I am using intrinsics to accelerate the running openCV code. But after i replaced the code with Intrinsics, the runtime cost of the code is almost the same or maybe even worse. i cannot figure out what and why this is happening. I have been searching this issue for quite long time but noting change. It is appreciated if someone can help me out. Thank you very much! Here is my code
// if useSSE is true,run the code with intrinsics and takes 1.45ms in my computer
// and if not run the general code and takes the same time.
cv::Mat<float> results(shape.rows,2);
if (useSSE) {
float* pshape = (float*)shape.data;
results = shape.clone();
float* presults = (float*)results.data;
// use SSE
__m128 xyxy_center = _mm_set_ps(bbox.center_y, bbox.center_x, bbox.center_y, bbox.center_x);
float bbox_width = bbox.width/2;
float bbox_height = bbox.height/2;
__m128 xyxy_size = _mm_set_ps(bbox_height, bbox_width, bbox_height, bbox_width);
gettimeofday(&start, NULL); // this is for counting time
int shape_size = shape.rows*shape.cols;
for (int i=0; i<shape_size; i +=4) {
__m128 a = _mm_loadu_ps(pshape+i);
__m128 result = _mm_div_ps(_mm_sub_ps(a, xyxy_center), xyxy_size);
_mm_storeu_ps(presults+i, result);
}
}else {
//SSE TO BE DONE
for (int i = 0; i < shape.rows; i++){
results(i, 0) = (shape(i, 0) - bbox.center_x) / (bbox.width / 2.0);
results(i, 1) = (shape(i, 1) - bbox.center_y) / (bbox.height / 2.0);
}
}
gettimeofday(&end, NULL);
diff = 1000000*(end.tv_sec-start.tv_sec)+end.tv_sec-start.tv_usec;
std::cout<<diff<<"-----"<<std::endl;
return results;

Answer 1

Your SSE optimization will corrupt memory near results variable, if shape.rows % 2 == 1
Try avoiding using i variable in the loop, use pointers directly. Compiler may optimize additional plus operation, or it may not.
Use multiplication instead of division:
float bbox_width_inv = 2./bbox.width;
float bbox_height_inv = 2./bbox.height;
__m128 xyxy_size = _mm_set_ps(bbox_height, bbox_width, bbox_height, bbox_width);
float* p_shape_end = p_shape + shape.rows*shape.cols;
float* p_shape_end_batch = p_shape + shape.rows*shape.cols & (~3);
for (; p_shape<p_shape_end_batch; p_shape+=4, presults+=4) {
__m128 a = _mm_loadu_ps(pshape);
__m128 result = _mm_mul_ps(_mm_sub_ps(a, xyxy_center), xyxy_size_inv);
_mm_storeu_ps(presults, result);
}
while (p_shape < p_shape_end) {
presults++ = (p_shape++ - bbox.center_x) * bbox_width_inv;
presults++ = (p_shape++ - bbox.center_y) * bbox_height_inv;
}
Try to disassemble code generated from intrinsics, and make sure there is enough registers to perform your operations, and it doesn't store temporary results into RAM