推理前处理neon加速实现neon_mean_scale

转自：https://github.com/PaddlePaddle/PaddleOCR/blob/release/2.1/deploy/android_demo/app/src/main/cpp/preprocess.cpp

// fill tensor with mean and scale and trans layout: nhwc -> nchw, neon speed up
void neon_mean_scale(const float *din, float *dout, int size,
                     const std::vector<float> &mean,
                     const std::vector<float> &scale) {
  if (mean.size() != 3 || scale.size() != 3) {
    LOGE("[ERROR] mean or scale size must equal to 3");
    return;
  }

  float32x4_t vmean0 = vdupq_n_f32(mean[0]);
  float32x4_t vmean1 = vdupq_n_f32(mean[1]);
  float32x4_t vmean2 = vdupq_n_f32(mean[2]);
  float32x4_t vscale0 = vdupq_n_f32(scale[0]);
  float32x4_t vscale1 = vdupq_n_f32(scale[1]);
  float32x4_t vscale2 = vdupq_n_f32(scale[2]);

  float *dout_c0 = dout;
  float *dout_c1 = dout + size;
  float *dout_c2 = dout + size * 2;

  int i = 0;
  for (; i < size - 3; i += 4) {
    float32x4x3_t vin3 = vld3q_f32(din);
    float32x4_t vsub0 = vsubq_f32(vin3.val[0], vmean0);
    float32x4_t vsub1 = vsubq_f32(vin3.val[1], vmean1);
    float32x4_t vsub2 = vsubq_f32(vin3.val[2], vmean2);
    float32x4_t vs0 = vmulq_f32(vsub0, vscale0);
    float32x4_t vs1 = vmulq_f32(vsub1, vscale1);
    float32x4_t vs2 = vmulq_f32(vsub2, vscale2);
    vst1q_f32(dout_c0, vs0);
    vst1q_f32(dout_c1, vs1);
    vst1q_f32(dout_c2, vs2);

    din += 12;
    dout_c0 += 4;
    dout_c1 += 4;
    dout_c2 += 4;
  }
  for (; i < size; i++) {
    *(dout_c0++) = (*(din++) - mean[0]) * scale[0];
    *(dout_c1++) = (*(din++) - mean[1]) * scale[1];
    *(dout_c2++) = (*(din++) - mean[2]) * scale[2];
  }
}