转：ARM NEON 优化举例

只看该作者 · 2012-1-28 18:35

ARM NEON 优化举例

累加数组内起始元素到制定个数元素的累加和：
uint32_t vector_add_of_n(uint32_t *ptr, uint32_t nitems)
{
uint32_t result, *i;
uint32x2_t vec64a, vec64b;
uint32x4_t vec128 = vdupq_n_u32(0);
for (i = ptr; i < (ptr + nitems); i += 4)
{
      uint32x4_t temp128 = vld1q_u32(i);
      vec128 = vaddq_u32(vec128, temp128);
}
vec64a = vget_low_u32(vec128);
vec64b = vget_high_u32(vec128);
vec64a = vadd_u32 (vec64a, vec64b);
result =  vget_lane_u32(vec64a, 0);
result += vget_lane_u32(vec64a, 1);
return result;
}

两个数组相关映射元素的乘积和：
void  fir(short * y,const short *x, const short *h,int n_out, int n_coefs)
{
int n, k;
int sum;
int16x4_t h_vec;
int16x4_t x_vec;
int32x4_t result_vec;
for (n = 0; n < n_out; n++)
{
sum = 0;
result_vec = vdupq_n_s32(0);
for(k = 0; k < n_coefs / 4; k++)
{
h_vec = vld1_s16(&h[k*4]);
x_vec = vld1_s16(&x[n - n_coefs + 1 + k*4]);
result_vec = vmlal_s16(result_vec, h_vec, x_vec);
}
sum += vgetq_lane_s32(result_vec, 0);
sum += vgetq_lane_s32(result_vec, 1);
sum += vgetq_lane_s32(result_vec, 2);
sum += vgetq_lane_s32(result_vec, 3);
if(n_coefs % 4)
{
for(k = n_coefs - (n_coefs % 4); k < n_coefs; k++)
sum += h[k] * x[n - n_coefs + 1 + k];
}
y[n] = ((sum>>15) + 1) >> 1;
}

只看该作者 · 2012-1-28 18:36

彩色转灰度：（C）
void reference_convert (uint8_t * __restrict dest, uint8_t * __restrict src, int n)
{
  int i;
  for (i=0; i<n; i++)
  {
int r = *src++;
int g = *src++;
int b = *src++;

   int y = (r*77)+(g*151)+(b*28);
*dest++ = (y>>8);
  }
}

void neon_convert (uint8_t * __restrict dest, uint8_t * __restrict src, int n)
{
  int i;
  uint8x8_t **c = vdup_n_u8 (77);
  uint8x8_t gfac = vdup_n_u8 (151);
  uint8x8_t bfac = vdup_n_u8 (28);
  n/=8;

  for (i=0; i<n; i++)
  {
uint16x8_t  temp;
uint8x8x3_t rgb  = vld3_u8 (src);
uint8x8_t result;

temp = vmull_u8 (rgb.val[0],    **c);
temp = vmlal_u8 (temp,rgb.val[1], gfac);
temp = vmlal_u8 (temp,rgb.val[2], bfac);

result = vshrn_n_u16 (temp, 8);
vst1_u8 (dest, result);
src  += 8*3;
dest += 8;
  }
}

转：ARM NEON 优化举例

相关下载

相关帖子