gna_config.hpp

 // Copyright (C) 2018-2019 Intel Corporation


 // SPDX-License-Identifier: Apache-2.0


 //


 


 /**


  * @brief A header that defines advanced related properties for VPU plugins.


  * These properties should be used in SetConfig() and LoadNetwork() methods of plugins


  *


  * @file vpu_plugin_config.hpp


  */


 

 #pragma once


 

 #include <string>


 #include "ie_plugin_config.hpp"


 

 namespace InferenceEngine {

 

 namespace GNAConfigParams {

 

 #define GNA_CONFIG_KEY(name) InferenceEngine::GNAConfigParams::_CONFIG_KEY(GNA_##name)


 #define GNA_CONFIG_VALUE(name) InferenceEngine::GNAConfigParams::GNA_##name


 

 #define DECLARE_GNA_CONFIG_KEY(name) DECLARE_CONFIG_KEY(GNA_##name)


 #define DECLARE_GNA_CONFIG_VALUE(name) DECLARE_CONFIG_VALUE(GNA_##name)


 


 /**


 * @brief Scale factor that is calculated by user, in order to use static quantisation feature


 * This option should be used with floating point value serialized to string with decimal separator equals to . (dot)


 * @details For multiple input case, individual scale factors can be passed, using KEY_GNA_SCALE_FACTOR[_input_layer_name]


 * where input_layer can be obtained from from CNNNetwork::GetInputsInfo


 */


 DECLARE_GNA_CONFIG_KEY(SCALE_FACTOR);

 


 /**


 * @brief By default gna api work in Int16 precision, however this can be adjusted if necessary,


 * currently supported values are I16, I8


 */


 DECLARE_GNA_CONFIG_KEY(PRECISION);

 

 


 /**


 * @brief if turned on, dump GNA firmware model into specified file


 */


 DECLARE_GNA_CONFIG_KEY(FIRMWARE_MODEL_IMAGE);

 


 /**


 * @brief GNA proc_type setting that should be one of GNA_AUTO, GNA_HW, GNA_SW, GNA_SW_EXACT


 */


 DECLARE_GNA_CONFIG_KEY(DEVICE_MODE);

 

 DECLARE_GNA_CONFIG_VALUE(AUTO);

 DECLARE_GNA_CONFIG_VALUE(HW);

 DECLARE_GNA_CONFIG_VALUE(SW);

 DECLARE_GNA_CONFIG_VALUE(SW_EXACT);

 


 /**


 * @brief if enabled produced minimum memory footprint for loaded network in GNA memory, default value is YES


 */


 DECLARE_GNA_CONFIG_KEY(COMPACT_MODE);

 


 /**


 * @brief The option to enable/disable uniformly distributed PWL algorithm.


 * By default (in case of NO value set) the optimized algorithm called "Recursive Descent Algorithm for Finding


 * the Optimal Minimax Piecewise Linear Approximation of Convex Functions is used.


 * If value is YES then simple uniform distribution used to create PWL approximation of activation functions


 * Uniform distribution usually gives poor approximation with same number of segments


 */


 DECLARE_GNA_CONFIG_KEY(PWL_UNIFORM_DESIGN);

 


 /**


 * @brief By default, the GNA plugin uses one worker thread for inference computations.


 * This parameter allows you to create up to 127 threads for software modes.


 *


 * Note that multithreading mode does not guarantee the same computation order as order


 * of issuing. Additionally, in this case, software modes do not implement any serializations.


 */


 DECLARE_GNA_CONFIG_KEY(LIB_N_THREADS);

 }  // namespace GNAConfigParams


 }  // namespace InferenceEngine