Question

I’m working on translating the code below into Neon Assembly. Any help would be greatly appreciated.

void sum(int length, int *a, int *b, int *c, int *d, char *result)
{
   int i;

   for (i = 0; i < length; i++)
      {
          int sum = (a[i] + b[i] + c[i] + d[i])/4;
          if (sum > threshold)
             result[i] = 1;
          else
             result[i] = 0;
      }
}

The actual code is an image binarization algorithm. The above code is just to demonstrate the idea and not to make simple things more complicate.

Answer 1

Here’s a fairly straightforward implementation. Note that we convert the divide and threshold test into just a test against threshold * 4 (in order to eliminate the divide):

void sum(const int n, const int32_t *a, const int32_t *b, const int32_t *c, const int32_t *d, int32_t *result)
{
   const int32_t threshold4 = threshold * 4;
   const int32x4_t vthreshold4 = { threshold4, threshold4, threshold4, threshold4 };
   const uint32x4_t vk1 = { 1, 1, 1, 1 };
   int i;

   for (i = 0; i < n; i += 4)
   {
      int32x4_t va = vld1q_s32(&a[i]);    // load values from a, b, c, d
      int32x4_t vb = vld1q_s32(&b[i]);
      int32x4_t vc = vld1q_s32(&c[i]);
      int32x4_t vd = vld1q_s32(&d[i]);

      int32x4_t vsum = vaddq_s32(va, vb); // sum values form a, b, c, d
      vsum = vaddq_s32(vsum, vc);
      vsum = vaddq_s32(vsum, vd);

      uint32x4_t vcmp = vcgtq_s32(vsum, vthreshold4);
                                          // compare with threshold * 4
      int32x4_t vresult = (int32x4_t)vandq_u32(vcmp, vk1);
                                          // convert result to 0/1
      vst1q_s32(&result[i], vresult);     // store result
   }
}

Notes:

• completely untested code – may need further work
• result has been changed to int32_t * – it’s not hard to pack down to uint8_t but it adds a lot of complexity to this initial example so I thought I’d keep it simple for now
• a, b, c, d, result all need to be 16 byte aligned
• n needs to be a multiple of 4
• sum of a, b, c, d needs to fit within 32 bit signed int
• threshold * 4 needs to fit within 32 bit signed int