TfLiteDelegate

#include <common.h>

WARNING: This is an experimental interface that is subject to change.

Summary

Public attributes

CopyFromBufferHandle)(TfLiteContext *context, struct TfLiteDelegate *delegate, TfLiteBufferHandle buffer_handle, TfLiteTensor *tensor)
TfLiteStatus(*
Copy the data from delegate buffer handle into raw memory of the given tensor.
CopyToBufferHandle)(TfLiteContext *context, struct TfLiteDelegate *delegate, TfLiteBufferHandle buffer_handle, TfLiteTensor *tensor)
TfLiteStatus(*
Copy the data from raw memory of the given tensor to delegate buffer handle.
FreeBufferHandle)(TfLiteContext *context, struct TfLiteDelegate *delegate, TfLiteBufferHandle *handle)
void(*
Free the Delegate Buffer Handle.
Prepare)(TfLiteContext *context, struct TfLiteDelegate *delegate)
TfLiteStatus(*
Invoked by ModifyGraphWithDelegate.
data_
void *
Data that delegate needs to identify itself.
flags
int64_t
Bitmask flags. See the comments in TfLiteDelegateFlags.
opaque_delegate_builder
The opaque delegate builder associated with this object.

Public attributes

CopyFromBufferHandle

TfLiteStatus(* TfLiteDelegate::CopyFromBufferHandle)(TfLiteContext *context, struct TfLiteDelegate *delegate, TfLiteBufferHandle buffer_handle, TfLiteTensor *tensor)

Copy the data from delegate buffer handle into raw memory of the given tensor.

Note that the delegate is allowed to allocate the raw bytes as long as it follows the rules for kTfLiteDynamic tensors, in which case this cannot be null.

CopyToBufferHandle

TfLiteStatus(* TfLiteDelegate::CopyToBufferHandle)(TfLiteContext *context, struct TfLiteDelegate *delegate, TfLiteBufferHandle buffer_handle, TfLiteTensor *tensor)

Copy the data from raw memory of the given tensor to delegate buffer handle.

This can be null if the delegate doesn't use its own buffer.

FreeBufferHandle

void(* TfLiteDelegate::FreeBufferHandle)(TfLiteContext *context, struct TfLiteDelegate *delegate, TfLiteBufferHandle *handle)

Free the Delegate Buffer Handle.

Note: This only frees the handle, but this doesn't release the underlying resource (e.g. textures). The resources are either owned by application layer or the delegate. This can be null if the delegate doesn't use its own buffer.

Prepare

TfLiteStatus(* TfLiteDelegate::Prepare)(TfLiteContext *context, struct TfLiteDelegate *delegate)

Invoked by ModifyGraphWithDelegate.

This prepare is called, giving the delegate a view of the current graph through TfLiteContext*. It typically will look at the nodes and call ReplaceNodeSubsetsWithDelegateKernels() to ask the TensorFlow lite runtime to create macro-nodes to represent delegated subgraphs of the original graph.

data_

void * TfLiteDelegate::data_

Data that delegate needs to identify itself.

This data is owned by the delegate. The delegate is owned in the user code, so the delegate is responsible for deallocating this when it is destroyed.

flags

int64_t TfLiteDelegate::flags

Bitmask flags. See the comments in TfLiteDelegateFlags.

opaque_delegate_builder

struct TfLiteOpaqueDelegateBuilder * TfLiteDelegate::opaque_delegate_builder

The opaque delegate builder associated with this object.

If set then the TF Lite runtime will give precedence to this field. E.g. instead of invoking Prepare via the function pointer inside the TfLiteDelegate object, the runtime will first check if the corresponding function pointer inside opaque_delegate_builder is set and if so invoke that.

If this field is non-null, then the Prepare field (of the TfLiteDelegate) should be null.