Skip to content

Commit

Permalink
Update ARM_NEON_CNN编程.md
Browse files Browse the repository at this point in the history
  • Loading branch information
Ewenwan authored May 13, 2019
1 parent d1bcf74 commit bc67b3f
Showing 1 changed file with 30 additions and 0 deletions.
30 changes: 30 additions & 0 deletions CNN/HighPerformanceComputing/ARM_NEON_CNN编程.md
Original file line number Diff line number Diff line change
Expand Up @@ -1682,7 +1682,37 @@ int dot(char* A,char* B,int K)
return sum;
}
```
### 示例8:3x3 pool 池化代码 最大值/均值池化

```c
// 先分别读取三列
constexpr const int pool_size = 3;
const float32x4_t top_data = vld1q_f32(reinterpret_cast<const float *>(input_top_ptr + input.offset()));
const float32x4_t middle_data = vld1q_f32(reinterpret_cast<const float *>(input_middle_ptr + input.offset()));
const float32x4_t bottom_data = vld1q_f32(reinterpret_cast<const float *>(input_bottom_ptr + input.offset()));

float32x2_t res = {};
if(pooling_type == PoolingType::AVG)
{// 均值池化=============
// Calculate scale
float scale = calculate_avg_scale(id, pool_size, upper_bound_w, upper_bound_h, pool_pad_x, pool_pad_y, pool_stride_x, pool_stride_y);
const float32x2_t scale_v = vdup_n_f32(scale);// 寄存器 初始化为 scale 2个32位

// Perform pooling
const float32x4_t sum_data = vaddq_f32(vaddq_f32(top_data, bottom_data), middle_data);
res = vpadd_f32(vget_high_f32(vsetq_lane_f32(0.f, sum_data, 3)), vget_low_f32(sum_data));
res = vmul_f32(vpadd_f32(res, res), scale_v);// 得到4个最大的float
}
else
{// 最大值池化
const float32x4_t max_data = vmaxq_f32(vmaxq_f32(top_data, bottom_data), middle_data);
res = vpmax_f32(vget_high_f32(vsetq_lane_f32(-std::numeric_limits<float>::max(), max_data, 3)), vget_low_f32(max_data));
res = vpmax_f32(res, res);
}

*(reinterpret_cast<float *>(output.ptr())) = vget_lane_f32(res, 0);

```
## 4. NEON assembly

采用汇编语言进行NEON(**NEON 汇编(assembly)**)的最底层优化,可以使优化性能最大化,但汇编语言比较灵活,手写汇编程序对开发人员来说具有较大挑战,如果使用不恰当,反而会影响优化性能。
Expand Down

0 comments on commit bc67b3f

Please sign in to comment.