Class AbsvalOps

executor/operator/arm64/absval_float.cpp:45–116

const int default_prio = 300;

struct AbsvalOps : public MTNodeOps
{
    AbsvalOps()
    {
        name_ = "arm_abs_fp32";
    }

    bool OnBind(Node* node)
    {
        inplace_t io_map;

        io_map[0] = 0;

        node->SetAttr(ATTR_INPLACE, io_map);
        return true;
    }

    bool Prerun(Node* node)
    {
        return true;
    }

    bool Run(Node* node)
    {
        Tensor* input_tensor = node->GetInputTensor(0);
        const TShape& shape = input_tensor->GetShape();
        
        float* data = ( float* )get_tensor_mem(input_tensor);
        
        int channel_num = shape.GetC();
        int batch_number = shape.GetN();
        int channel_size = shape.GetW() * shape.GetH();
        
        for(int c = 0; c < channel_num * batch_number; c++)
        {
            for(int i = 0; i < (channel_size & -4); i += 4)
            {
                float32x4_t _p = vld1q_f32(data);
                _p = vabsq_f32(_p);
                vst1q_f32(data, _p);

                data += 4;
            }
            for(int i = channel_size & ~3; i < channel_size; i++)
            {
                if (*data < 0)
                    *data = -*data;
                data++;
            }
        }
        
        /*
        int elem_num = shape.GetSize();
        for(int i = 0; i < (elem_num & -4); i += 4)
        {
	        float32x4_t _p = vld1q_f32(data);
            _p = vabsq_f32(_p);
            vst1q_f32(data, _p);