Method Run

executor/operator/common/lrn.cpp:84–154

        name_ = "com_lrn_fp32";
    }
    bool Run(Node* node)
    {
        Tensor* input_tensor = node->GetInputTensor(0);
        Tensor* output_tensor = node->GetOutputTensor(0);

        LRN* lrn_op = dynamic_cast<LRN*>(node->GetOp());
        LRNParam* param = lrn_op->GetParam();

        float* input = ( float* )get_tensor_mem(input_tensor);
        float* output = ( float* )get_tensor_mem(output_tensor);

        const TShape& shape = input_tensor->GetShape();
        const std::vector<int>& dims = shape.GetDim();

        int n = dims[0];
        int c = dims[1];
        int h = dims[2];
        int w = dims[3];

        int img_size = c * h * w;
        float alpha = param->alpha;
        float beta = param->beta;
        float bias = param->k;
        int local_size = param->local_size;
        float alpha_over_size = alpha / local_size;

        float* square = ( float* )(std::malloc(img_size * sizeof(float)));
        int cpu_number = cpu_info->GetCPUNumber();
        int num_task = c < cpu_number ? c : cpu_number;
        int step = c / num_task;

        for(int i = 0; i < n; i++)
        {
            /* get square value */

            float* in_base = input + i * img_size;
            float* out_base = output + i * img_size;

            if(param->norm_region != LRN_ACROSS_CHANNELS)
            {
                LOG_ERROR()<<"LRN Only support ACORSS_CHANNEL\n";
                return false;
            }
            else
            {
                for(int j = 0; j < img_size; j++)
                    square[j] = in_base[j] * in_base[j] + bias;
            }
            if(num_task == 1)
            {
                lrn_kernel(0, 0, &c, in_base, out_base, square, h, w, c, local_size, alpha_over_size, beta);
            }
            else
            {
                MULTI_THREAD_START(num_task, step, id, param)
                    lrn_kernel(0, id, param, in_base, out_base, square, h, w, c, local_size, alpha_over_size, beta);
                MULTI_THREAD_END();
            }