| 50 | name_ = "arm_hardswish_fp32"; |
| 51 | } |
| 52 | bool Run(Node* node) |
| 53 | { |
| 54 | Tensor* input_tensor = node->GetInputTensor(0); |
| 55 | Tensor* output_tensor = node->GetOutputTensor(0); |
| 56 | Hardswish* Hardswish_op = dynamic_cast<Hardswish*>(node->GetOp()); |
| 57 | HardswishParam* param_ = Hardswish_op->GetParam(); |
| 58 | float alpha = param_->alpha; |
| 59 | float beta = param_->beta; |
| 60 | float lower = -beta / alpha; |
| 61 | float upper = (1.f / alpha) + lower; |
| 62 | |
| 63 | int elem_num = input_tensor->GetShape().GetSize(); |
| 64 | |
| 65 | float* data = ( float* )get_tensor_mem(input_tensor); |
| 66 | float* out_data = ( float* )get_tensor_mem(output_tensor); |
| 67 | |
| 68 | float32x4_t _zero = vdupq_n_f32(0.f); |
| 69 | float32x4_t _one = vdupq_n_f32(1.f); |
| 70 | for(int i = 0; i < (elem_num & -4); i += 4) |
| 71 | { |
| 72 | float32x4_t _p = vld1q_f32(data+i); |
| 73 | float32x4_t _ans = vdupq_n_f32(beta); |
| 74 | _ans = vmlaq_n_f32(_ans, _p, alpha); |
| 75 | _ans = vmaxq_f32(_ans, _zero); |
| 76 | _ans = vminq_f32(_ans, _one); |
| 77 | _ans = vmulq_f32(_ans, _p); |
| 78 | vst1q_f32(out_data+i, _ans); |
| 79 | } |
| 80 | for(int i = elem_num & ~3; i < elem_num; i++) |
| 81 | { |
| 82 | if (data[i] < lower) |
| 83 | out_data[i] = 0.f; |
| 84 | else if (data[i] > upper) out_data[i] = data[i]; |
| 85 | else |
| 86 | out_data[i] = data[i] * (data[i] * alpha + beta); |
| 87 | } |
| 88 | /* |
| 89 | for(int i = 0; i < elem_num; i++) |
| 90 | { |
| 91 | if (data[i] < lower) |
| 92 | out_data[i] = 0.f; |
| 93 | else if (data[i] > upper) out_data[i] = data[i]; |
| 94 | else |
| 95 | out_data[i] = data[i] * (data[i] * alpha + beta); |
| 96 | } |
| 97 | */ |
| 98 | |
| 99 | return true; |
| 100 | |
| 101 | } |
| 102 | }; |
| 103 | |
| 104 | NodeOps* SelectFunc(const CPUInfo* cpu_info, Node* node) |
nothing calls this directly
no test coverage detected