TfLite Delegate

About the delegate

'armnnDelegate' is a library for accelerating certain TensorFlow Lite (TfLite) operators on Arm hardware. It can be integrated in TfLite using its delegation mechanism. TfLite will then delegate the execution of operators supported by Arm NN to Arm NN.

The main difference to our Arm NN Tf Lite Parser is the amount of operators you can run with it. If none of the active backends support an operation in your model you won't be able to execute it with our parser. In contrast to that, TfLite only delegates operations to the armnnDelegate if it does support them and otherwise executes them itself. In other words, every TfLite model can be executed and every operation in your model that we can accelerate will be accelerated. That is the reason why the armnnDelegate is our recommended way to accelerate TfLite models.

If you need help building the armnnDelegate, please take a look at our build guide.

Supported Operators

This reference guide provides a list of TensorFlow Lite operators the Arm NN SDK currently supports.

Fully supported

The Arm NN SDK TensorFlow Lite delegate currently supports the following operators:

• ABS
• ADD
• ARG_MAX
• ARG_MIN
• AVERAGE_POOL_2D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• AVERAGE_POOL_3D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, SIGN_BIT, TANH, NONE
• BATCH_MATMUL
• BATCH_TO_SPACE_ND
• BROADCAST_TO
• CAST
• CEIL
• CONCATENATION, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• CONV_2D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• CONV_3D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• DEPTH_TO_SPACE
• DEPTHWISE_CONV_2D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• DEQUANTIZE
• DIV
• EQUAL
• ELU
• EXP
• EXPAND_DIMS
• FILL
• FLOOR
• FLOOR_DIV
• FULLY_CONNECTED, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• GATHER
• GATHER_ND
• GELU
• GREATER
• GREATER_EQUAL
• HARD_SWISH
• L2_NORMALIZATION
• L2_POOL_2D
• LESS
• LESS_EQUAL
• LOCAL_RESPONSE_NORMALIZATION
• LOG
• LOGICAL_AND
• LOGICAL_NOT
• LOGICAL_OR
• LOGISTIC
• LOG_SOFTMAX
• LSTM
• MAXIMUM
• MAX_POOL_2D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, TANH, NONE
• MAX_POOL_3D, Supported Fused Activation: RELU, RELU6, RELU_N1_TO_1, SIGMOID, SIGN_BIT, TANH, NONE
• MEAN
• MINIMUM
• MIRROR_PAD
• MUL
• NEG
• NOT_EQUAL
• PACK
• PAD
• PADV2
• POW
• PRELU
• QUANTIZE
• RANK
• REDUCE_MAX
• REDUCE_MIN
• REDUCE_PROD
• RELU
• RELU6
• RELU_N1_TO_1
• RESHAPE
• RESIZE_BILINEAR
• RESIZE_NEAREST_NEIGHBOR
• REVERSEV2
• RSQRT
• SCATTERND
• SHAPE
• SIN
• SOFTMAX
• SPACE_TO_BATCH_ND
• SPACE_TO_DEPTH
• SPLIT
• SPLIT_V
• SQRT
• SQUARE_DIFFERENCE
• SQUEEZE
• SLICE
• STRIDED_SLICE
• SUB
• SUM
• TANH
• TILE
• TRANSPOSE
• TRANSPOSE_CONV
• UNIDIRECTIONAL_SEQUENCE_LSTM
• UNPACK

More machine learning operators will be supported in future releases.

Delegate Options

The general list of runtime options are described in Runtime options

Opaque Delegate Options

In Opaque Delegate, delegate options are passed via ArmNNSettings which is a FlatBuffer of tflite::TFLiteSettings.

Arm NN Settings	Possible Values	Description
backends	["GpuAcc"/"CpuAcc"]	A comma separated list without whitespaces of backends which should be used for execution. Falls back to next backend in list if previous does not provide support for operation.
fastmath	[true/false]	Allows the use of optimisation techniques e.g. Winograd that will reduce execution time with the possibility of a drop in accuracy.
additional_parameters	JSON string of additional Arm NN delegate options	JSON string of additional Arm NN delegate options. The general list of runtime options are described in Runtime options.