Detailed Description

This structure define the parameters CNN/Deep learning net in TIDL.

Data Fields
int32_t	netVersion

int32_t	deviceName

int32_t	numLayers

int32_t	weightsElementSize

int32_t	slopeElementSize

int32_t	biasElementSize

int32_t	dataElementSize

int32_t	interElementSize

int32_t	quantizationStyle

int32_t	calibrationOption

sTIDL_CalibParams_t	calibrationParams

int32_t	dataFlowInfo

int32_t	isQuantStatsAvailable

float32_tidl	preEmptWorstCaseDelay

int32_t	reserved

sTIDL_Layer_t	TIDLLayers [TIDL_NUM_MAX_LAYERS]

Field Documentation

◆ netVersion

int32_t sTIDL_Network_t::netVersion

Version Number for TIDl net format

◆ deviceName

int32_t sTIDL_Network_t::deviceName

Device for which the network has been compiled

◆ numLayers

int32_t sTIDL_Network_t::numLayers

Number of layers in the network including the input and output data Layers

◆ weightsElementSize

int32_t sTIDL_Network_t::weightsElementSize

Size of compute layer weight parameters in bytes

◆ slopeElementSize

int32_t sTIDL_Network_t::slopeElementSize

Size of PRelU layer weight/slope parameters in bytes

◆ biasElementSize

int32_t sTIDL_Network_t::biasElementSize

Size of compute layer Bias parameters in bytes

◆ dataElementSize

int32_t sTIDL_Network_t::dataElementSize

Size of compute layer input and data buffers in bytes

◆ interElementSize

int32_t sTIDL_Network_t::interElementSize

Size of compute layer intermediate data in bytes

◆ quantizationStyle

int32_t sTIDL_Network_t::quantizationStyle

Variable to indicate different types of quantization Styles

◆ calibrationOption

int32_t sTIDL_Network_t::calibrationOption

Variable to indicate different types of calibration options. Please refer eTIDL_CalibOption for various options available. Even though user can try multiple options following options are recommended : 0 : Simple calibration ( Default) 1 : Histogram based activation range collection 7 : Advanced bias calibration

◆ calibrationParams

sTIDL_CalibParams_t sTIDL_Network_t::calibrationParams

Calibration specific parameters, parameters are specific to calibration method and are only applicable if the corresponding option is set by user as part of eTIDL_CalibOption

◆ dataFlowInfo

int32_t sTIDL_Network_t::dataFlowInfo

dataFlow Info from offline Analysis

◆ isQuantStatsAvailable

int32_t sTIDL_Network_t::isQuantStatsAvailable

Quantization statistics availability - Updated with tidl_quant_stats_tool

◆ preEmptWorstCaseDelay

float32_tidl sTIDL_Network_t::preEmptWorstCaseDelay

It is the WorstCaseDelay in milliSecond for pre-emption for the given network

◆ reserved

int32_t sTIDL_Network_t::reserved

Reserved Parameters

◆ TIDLLayers

sTIDL_Layer_t sTIDL_Network_t::TIDLLayers[TIDL_NUM_MAX_LAYERS]

Layer Parameter for each layer