Question

I’m trying to code a neon version of Sum of Absolute Difference of a 16 uint8_t sized inputs:

inline static int f_sad_16(const uint8_t* a, const uint8_t* b)
{
  int sad = 0;
  for (int i = 0; i < 16; i++) {
    sad += abs(static_cast<int>(a[i]) - static_cast<int>(b[i]));
  }
  return sad;
}

The neon code I wrote:

inline static int f_sad_16_neon(const uint8_t* a, const uint8_t* b)
{
  int32_t r[4] = { 0, 0, 0, 0 };
  uint8x16_t va, vb, vr;

  va = vld1q_u8(a);
  vb = vld1q_u8(b);

  vr = vabdq_u8(va, vb);

  uint16x8_t vr1 = vpaddlq_u8 (vr );
  uint32x4_t vr2 = vpaddlq_u16(vr1);
  uint64x2_t vr3 = vpaddlq_u32(vr2);

  vst1q_u64 (reinterpret_cast<uint64_t*>(r), vr3);

  return r[0] + r[2];
}

For some reason I get a bus error. My test program uses the __attribute__ ((aligned (16))) gcc directive for the two inputs, and I can see by the memory addresses assigned to them that the inputs are 16-byte aligned.

What can be the source of the problem?

Answer 1

You need to make sure that r is properly aligned – change:

int32_t r[4] = { 0, 0, 0, 0 };

to:

int32_t r[4] = { 0, 0, 0, 0 } __attribute__ ((aligned(16)));

Note however that it’s unlikely that this routine will be beneficial, as it uses a significant number of both scalar and NEON instructions, which means it will most likely have performance similar to that of your original scalar implementation.