Arm loop code optimization for calculating mean standard deviation - objective-c

Question

I want to improve some code which is using 25% of my app CPU, the code is the next:
for (int i=0; i<8; i++) {
unsigned f = *p++;
sum += f;
sqsum += f*f;
}
I made some arm code but it is not working, even not compiling, which is the next:
void loop(uint8_t * p , int * sum ,int * qsum)
{
__asm__ volatile("vld4.8 {d0}, [%0]! \n"
"mov r4, #0 \n"
"vmlal.u8 [%1]!, [%1]!, d0 \n"
"vmull.u8 r4, d0 , d0 \n"
"vmlal.u8 [%2]!, [%2]!, r4\n"
:
: "r"(p), "r"(sum), "r"(qsum)
: "r4"
);
}
Any help?
Here is the my function to improve:
void calculateMeanStDev8x8(cv::Mat* patch, int sx, int sy, int& mean, float& stdev)
{
unsigned sum=0;
unsigned sqsum=0;
for (int j=0; j< 8; j++) {
const unsigned char* p = (const unsigned char*)(patch->data + (j+sy)*patch->step + sx); //Apuntador al inicio de la matrix
//The code to improve
for (int i=0; i<8; i++) {
unsigned f = *p++;
sum += f;
sqsum += f*f;
}
}
mean = sum >> 6;
int r = (sum*sum) >> 6;
stdev = sqrtf(sqsum - r);
if (stdev < .1) {
stdev=0;
}
}

Answer 1

That loop is a perfect candidate for NEON optimization. You can fit your 8 unsigned integers into a single NEON register. There is no "sum all elements of a vector" instruction, but you can use the pairwise add to compute the sum of the 8 elements in 3 steps. Since we can't see the rest of your application, it's hard to know what the big picture is, but NEON is your best bet for improving the speed. All recent Apple products support NEON instructions and in XCode you can use the NEON intrinsics mixed with your C++ code.