File QnnTensor.h

↰ Parent directory (include/QNN)

Tensor component API.

Definition (include/QNN/QnnTensor.h)

Detailed Description

Requires Backend to be initialized. Tensors have either Context or Graph scope. Tensors created with Context scope can be used within Graphs that belong to same Context, but not vice versa. Tensors hold either operation’s static/constant data or input/output activation data.

Includes

Included By

Full File Listing

Tensor component API. 

    Requires Backend to be initialized.
    Tensors have either Context or Graph scope. Tensors created with
    Context scope can be used within Graphs that belong to same Context,
    but not vice versa. Tensors hold either operation's static/constant
    data or input/output activation data.

Enums

enum QnnTensor_Error_t

QNN Tensor API result / error codes.

Values:

enumerator QNN_TENSOR_MIN_ERROR = 7000
enumerator QNN_TENSOR_NO_ERROR = 0

Success.

enumerator QNN_TENSOR_ERROR_INVALID_HANDLE = 7000 + 1

Invalid context/graph handle in creating tensor.

enumerator QNN_TENSOR_ERROR_DOES_NOT_EXIST = 7000 + 2

Tensor with specified credentials not registered with a context/graph.

enumerator QNN_TENSOR_ERROR_ALREADY_EXISTS = 7000 + 3

(deprecated) Tensor has already been registered with backend.

enumerator QNN_TENSOR_ERROR_INVALID_TENSOR_PARAM = 7000 + 4

Invalid tensor param.

enumerator QNN_TENSOR_ERROR_UNSUPPORTED_TENSOR_PARAM = 7000 + 5

This tensor param is currently unsupported.

enumerator QNN_TENSOR_ERROR_NAME_HASH_COLLISION = 7000 + 6

(deprecated) A hash collision has occurred with a previously registered tensor’s name.

enumerator QNN_TENSOR_ERROR_INCOMPATIBLE_TENSOR_UPDATE = 7000 + 7

Tensor provided for update is invalid.

enumerator QNN_TENSOR_ERROR_UNSUPPORTED_FEATURE = QNN_COMMON_ERROR_NOT_SUPPORTED

There is optional API component that is not supported yet. See QnnProperty.

enumerator QNN_TENSOR_MAX_ERROR = 7999
enumerator QNN_TENSOR_ERROR_UNDEFINED = 0x7FFFFFFF

Functions

Qnn_ErrorHandle_t QnnTensor_createContextTensor(Qnn_ContextHandle_t context, Qnn_Tensor_t *tensor)

A function to create a new tensor on Qnn_ContextHandle_t. 

   This call may or may not allocate memory, depending on the Qnn_TensorType_t
   value specified in tensor and the accelerator implementation.
   Optionally it may be initialized with data provided in the tensor if present.

Parameters

  • context – [in] The context in which the tensor would be created.

  • tensor – [inout] Pointer to a user-allocated struct containing information on the tensor (type, name, data format, dimensions, data, etc). For tensors containing static data (such as weights or biases), the tensor type is expected to be QNN_TENSOR_TYPE_STATIC. Valid data must be presented in the tensor object at creation. This data will be copied, and may be safely de-allocated after this call returns. Other tensor types (e.g: APP_READ, APP_WRITE, APP_READWRITE, NULL) must have the data pointer set to NULL at the time of creation. Any preset value in id will be overwritten by the backend as part of this call. Subsequent usage of the tensor must reference this id. Creating a tensor with a name that duplicates a previously created tensor name in the context and all child graphs results in undefined behaviour. The dimensions are treated as the maximum dimensions during tensor creation.

Returns

Error code:

  • QNN_SUCCESS: Successfully created a context tensor

  • QNN_TENSOR_ERROR_INVALID_HANDLE: Provided context handle is invalid

  • QNN_TENSOR_ERROR_INVALID_TENSOR_PARAM: One or more tensor parameters is invalid

  • QNN_TENSOR_ERROR_UNSUPPORTED_TENSOR_PARAM: One or more tensor parameters are unsupported

  • QNN_COMMON_ERROR_MEM_ALLOC: Failure in creating tensor due to issues with memory allocation

  • QNN_TENSOR_ERROR_UNSUPPORTED_FEATURE: some API feature is not supported yet

Note

Use corresponding API through QnnInterface_t.

Warning

Context tensors cannot be of type QNN_TENSOR_TYPE_NATIVE. Native tensors connect nodes within a single graph.

Warning

Context tensors cannot be of datatype QNN_DATATYPE_STRING.

Qnn_ErrorHandle_t QnnTensor_createGraphTensor(Qnn_GraphHandle_t graph, Qnn_Tensor_t *tensor)

A function to create a new tensor on Qnn_GraphHandle_t. 

   This call may or may not allocate memory, depending on the Qnn_TensorType_t
   value specified in tensor and the accelerator implementation.
   Optionally it may be initialized with data provided in the tensor if present.

Parameters

  • graph – [in] The graph or sub-graph in which the tensor would be created.

  • tensor – [inout] Pointer to a user-allocated struct containing information on the tensor (type, name, data format, dimensions, data, etc). For tensors containing static data (such as weights or biases), the tensor type is expected to be QNN_TENSOR_TYPE_STATIC. Valid data must be presented in the tensor object at creation. This data will be copied, and may be safely de-allocated after this call returns. Other tensor types (e.g: NATIVE, APP_READ, APP_WRITE, NULL) must have the data pointer set to NULL at the time of creation. Any preset value in id will be overwritten by the backend as part of this call. Subsequent usage of the tensor must reference this id. Creating a tensor with a name that duplicates a previously created tensor name in the graph or parent context results in undefined behaviour. The dimensions are treated as the maximum dimensions during tensor creation.

Returns

Error code:

  • QNN_SUCCESS: Successfully created a graph tensor

  • QNN_TENSOR_ERROR_INVALID_HANDLE: Provided graph handle is invalid

  • QNN_TENSOR_ERROR_INVALID_TENSOR_PARAM: One or more tensor parameters is invalid

  • QNN_TENSOR_ERROR_UNSUPPORTED_TENSOR_PARAM: One or more tensor parameters are unsupported

  • QNN_COMMON_ERROR_MEM_ALLOC: Failure in creating tensor due to issues with memory allocation

  • QNN_TENSOR_ERROR_UNSUPPORTED_FEATURE: some API feature is not supported yet

Note

Use corresponding API through QnnInterface_t.

Warning

Graph tensors cannot be of type QNN_TENSOR_TYPE_APP_READWRITE. R/W tensors connect multiple graphs.

Warning

Graph tensors cannot be of datatype QNN_DATATYPE_STRING.

Qnn_ErrorHandle_t QnnTensor_updateGraphTensors(Qnn_GraphHandle_t graph, const Qnn_Tensor_t **tensor, uint64_t numTensors)

Update a graph tensor with the new provided tensor information. Tensors provided here are associated with the tensor in the backend through the ID field. Valid fields to update are: data and quantization parameters for UPDATEABLE_STATIC tensors, quantization parameters for UPDATEABLE_NATIVE, UPDATEABLE_APP_READ, UPDATEABLE_APP_WRITE, and UPDATEABLE_APP_READWRITE tensors. Multiple calls to QnnTensor_updateGraphTensors() can be made, but the updates will not take effect until QnnGraph_finalize() is called. Backends may support a subset of updateable tensor types.

Returns

Error code:

  • QNN_SUCCESS: Successfully updated the graph tensors

  • QNN_TENSOR_ERROR_INVALID_HANDLE: Provided graph handle is invalid

  • QNN_TENSOR_ERROR_INVALID_TENSOR_PARAM: One or more tensor parameters is invalid

  • QNN_TENSOR_ERROR_UNSUPPORTED_TENSOR_PARAM: One or more tensor parameters are unsupported

  • QNN_GRAPH_ERROR_GRAPH_NOT_FINALIZED: graph needs to be finalized before updating graph tensors.

  • QNN_COMMON_ERROR_MEM_ALLOC: Failure in creating tensor due to issues with memory allocation

  • QNN_TENSOR_ERROR_INCOMPATIBLE_TENSOR_UPDATE: provided tensor is invalid and cannot be applied as an update.

  • QNN_TENSOR_ERROR_UNSUPPORTED_FEATURE: some API feature is not supported yet

Qnn_ErrorHandle_t QnnTensor_updateContextTensors(Qnn_ContextHandle_t context, const Qnn_Tensor_t **tensor, uint64_t numTensors)

Update a context tensor with the new provided tensor information. Tensors provided here are associated with the tensor in the backend through the ID field. Valid fields to update are: data and quantization parameters for UPDATEABLE_STATIC tensors, quantization parameters for UPDATEABLE_NATIVE, UPDATEABLE_APP_READ, UPDATEABLE_APP_WRITE, and UPDATEABLE_APP_READWRITE tensors. Multiple calls to QnnTensor_updateContextTensors() can be made, but the updates will not take effect until QnnGraph_finalize() is called for one or more of the graphs to which the context tensors are associated. Backends may support a subset of updateable tensor types.

Returns

Error code:

  • QNN_SUCCESS: Successfully updated the context tensor

  • QNN_TENSOR_ERROR_INVALID_HANDLE: Provided context handle is invalid

  • QNN_TENSOR_ERROR_INVALID_TENSOR_PARAM: One or more tensor parameters is invalid

  • QNN_TENSOR_ERROR_UNSUPPORTED_TENSOR_PARAM: One or more tensor parameters are unsupported

  • QNN_COMMON_ERROR_MEM_ALLOC: Failure in creating tensor due to issues with memory allocation

  • QNN_TENSOR_ERROR_INCOMPATIBLE_TENSOR_UPDATE: provided tensor is invalid and cannot be applied as an update.

  • QNN_TENSOR_ERROR_UNSUPPORTED_FEATURE: some API feature is not supported yet