Method Run

executor/operator/arm32/sigmoid_float.cpp:93–142

    }

    bool Run(Node* node)
    {
        Tensor* input_tensor = node->GetInputTensor(0);
        Tensor* output_tensor = node->GetOutputTensor(0);
        const TShape& shape = input_tensor->GetShape();
        int elem_num = shape.GetSize();
        float* input = ( float* )get_tensor_mem(input_tensor);
        float* output = ( float* )get_tensor_mem(output_tensor);
        /// void* data = get_tensor_mem(input_tensor);
        /// int ret = kernel_run(data, elem_num, &op_param);
#define SIGMOID_MAX(a, b) ((a) > (b) ? (a) : (b))
#define SIGMOID_MIN(a, b) ((a) < (b) ? (a) : (b))
        /*
        for(int i = 0; i < elem_num; i++)
        {
            input[i] = SIGMOID_MIN(input[i], 30.0f);
            input[i] = SIGMOID_MAX(input[i], -30.0f);
            std::cout<< input[i] << " ";
            float tmp_exp = exp8mul(-input[i]);
            output[i] = 1 / (1 + tmp_exp);
            std::cout<< tmp_exp << " " <<output[i] << "\n";
        }
        */

        float32x4_t min = vdupq_n_f32(-30.0f);
        float32x4_t max = vdupq_n_f32(30.0f);
        float32x4_t tmp_vec = vdupq_n_f32(1);
        for(int i = 0; i < (elem_num & -4); i += 4)
        {
            float32x4_t _input = vld1q_f32(input + i);
            _input = vmaxq_f32(_input, min);
            _input = vminq_f32(_input, max);
            float32x4_t tmp_exp = vaddq_f32(tmp_vec, vexpq10_f32(vmulq_n_f32(_input, -1.0f)));
            /// float32x4_t tmp_exp = vaddq_f32(tmp_vec, vexpq_f32(vmulq_n_f32(_input, -1.0f)));
            float32x4_t out = vrecpeq_f32(tmp_exp);
            out = vmulq_f32(vrecpsq_f32(tmp_exp, out), out);
            out = vmulq_f32(vrecpsq_f32(tmp_exp, out), out);
            vst1q_f32(output, out);
            output += 4;
        }
        for(int i = elem_num & ~3; i < elem_num; i++)
        {
            input[i] = SIGMOID_MIN(input[i], 30.0f);
            input[i] = SIGMOID_MAX(input[i], -30.0f);
            float tmp_exp = exp10_f32(-input[i]);
            *output++ = 1 / (1 + tmp_exp);
        }

        return true;
    }
};

NodeOps* SelectFunc(const CPUInfo* info, Node* node)