Class DataProcessor

Data Processor class that processes operations. Currently supports CPU based operations.

Public Functions

inline DataProcessor()

Default Constructor.

~DataProcessor()

Default Destructor.

InferStatus initialize(const MultiMappings &process_operations, const Mappings &custom_kernels, bool use_cuda_graphs, const std::string config_path)

Checks the validity of supported operations.

Parameters

• process_operations – Map where tensor name is the key, and operations to perform on the tensor as vector of strings. Each value in the vector of strings is the supported operation.

• custom_kernels – Map of custom kernel identifier, mapped to related value as a string

• use_cuda_graphs – Flag to enable CUDA Graphs for custom kernel processing

• config_path – Path to the processing configuration settings

Returns

InferStatus with appropriate code and message

InferStatus process_operation(const std::string &operation, const std::vector<int> &in_dims, const void *in_data, std::vector<int64_t> &processed_dims, DataMap &processed_data_map, const std::vector<std::string> &output_tensors, const std::vector<std::string> &custom_strings, bool process_with_cuda, cudaStream_t cuda_stream)

Executes an operation via function callback.

Parameters

• operation – Operation to perform. Refer to user docs for a list of supported operations

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

• processed_dims – Dimension of the output tensor, is populated during the processing

• processed_data_map – Output data map, that will be populated

• output_tensors – Tensor names to be populated in the out_data_map

• custom_strings – Strings to display for custom print operations

• process_with_cuda – Flag defining if processing should be done with CUDA

• cuda_stream – CUDA stream to use when procseeing is done with CUDA

Returns

InferStatus with appropriate code and message

InferStatus process_transform(const std::string &transform, const std::string &key, const std::map<std::string, void*> &indata, const std::map<std::string, std::vector<int>> &indim, DataMap &processed_data, DimType &processed_dims)

Executes a transform via function callback. (Currently CPU based)

Parameters

• transform – Data transform operation to perform.

• key – String identifier for the transform

• indata – Map with key as tensor name and value as data buffer

• indim – Map with key as tensor name and value as dimension of the input tensor

• processed_data – Output data map, that will be populated

• processed_dims – Dimension of the output tensor, is populated during the processing

Returns

InferStatus with appropriate code and message

InferStatus compute_max_per_channel_scaled(const std::vector<int> &in_dims, const void *in_data, std::vector<int64_t> &out_dims, DataMap &out_data_map, const std::vector<std::string> &output_tensors, bool process_with_cuda, cudaStream_t cuda_stream)

Computes max per channel in input data and scales it to [0, 1]. (supports both GPU and CPU data)

Parameters

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

• out_dims – Dimension of the output tensor

• out_data_map – Output data buffer map

• output_tensors – Output tensor names, used to populate out_data_map

• process_with_cuda – Flag defining if processing should be done with CUDA

• cuda_stream – CUDA stream to use when procseeing is done with CUDA

InferStatus scale_intensity_cpu(const std::vector<int> &in_dims, const void *in_data, std::vector<int64_t> &out_dims, DataMap &out_data_map, const std::vector<std::string> &output_tensors)

Scales intensity using min-max values and histogram. (CPU based)

Parameters

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

• out_dims – Dimension of the output tensor

• out_data_map – Output data buffer map

• output_tensors – Output tensor names, used to populate out_data_map

InferStatus print_results(const std::vector<int> &in_dims, const void *in_data)

Print data in the input buffer in float32. Ideally to be used by classification models.

Parameters

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

InferStatus print_results_int32(const std::vector<int> &in_dims, const void *in_data)

Print data in the input buffer in int32 form. Ideally to be used by classification models.

Parameters

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

InferStatus print_custom_binary_classification(const std::vector<int> &in_dims, const void *in_data, const std::vector<std::string> &custom_strings)

Print custom text for binary classification results in the input buffer.

Parameters

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

• custom_strings – Strings to display for custom print operations

InferStatus export_binary_classification_to_csv(const std::vector<int> &in_dims, const void *in_data, const std::vector<std::string> &custom_strings)

Export binary classification results in the input buffer to CSV file using Data Exporter API.

Parameters

• in_dims – Dimension of the input tensor

• in_data – Input data buffer

• custom_strings – The comma separated list of strings containing information for the output CSV file. It should include application name as a first string required for the Data Exporter API and column names.

InferStatus launchCustomKernel(const std::vector<std::string> &ids, const std::vector<int> &dimensions, const void *input, std::vector<int64_t> &processed_dims, DataMap &processed_data_map, const std::vector<std::string> &output_tensors, bool process_with_cuda, cudaStream_t cuda_stream)

Launches custom kernel at runtime.

Parameters

• ids – Unique custom kernel ids

• dimensions – Dimensions of input buffer

• input – Input data buffer

• processed_dims – Dimension of the output tensor, is populated during the processing

• processed_data_map – Output data map, that will be populated

• output_tensors – Output tensor names, used to populate out_data_map

• process_with_cuda – Flag defining if processing should be done with CUDA

• cuda_stream – CUDA stream to use when procseeing is done with CUDA

InferStatus prepareCustomKernel()

Initialization and preparation of all custom cuda kernels.