TFLite Operations

NeuroPilot supports a wide variety of operations in TFLite. This allows most neural network models to run on the specialized compute cores available on MediaTek platforms.

Supported Data Types

The following table lists the supported data types of each hardware target.

Device	AsymU8	AsymI8	SymI8	SymI16	FP16	FP32	Bool8	Int32
MDLA 2.0	O	O	O	O	O
VPU	O	O			O		O	O
CPU	O	O			O	O	O	O

Supported NeuroPilot Operations

The following table lists the TFLite operations supported for each data type.

Note

For full details of each operation, see the corresponding hardware guidelines section.

OP Name	TFLite OP	Ncc-tflite	AsymU8	AsymI8	SymI8	SymI16	Fp16	Fp32	Bool8	Int32
Abs	kTfLiteBuiltinAbs	O	O	O	O	O	O	O
ArgMax	kTfLiteBuiltinArgMax	O	O	O			O	O
ArgMin	kTfLiteBuiltinArgMin	O	O	O			O	O
AvgPooling	kTfLiteBuiltinAveragePool2d	O	O	O	O	O	O	O
Cast	kTfLiteBuiltinCast	O	O	O						O
Concat	kTfLiteBuiltinConcatenation	O	O	O	O	O	O	O
Conv2D	kTfLiteBuiltinConv2d	O	O	O	O	O	O	O
DepthToSpace	kTfLiteBuiltinDepthToSpace	O	O	O	O	O	O	O
DepthwiseConv2D	kTfLiteBuiltinDepthwiseConv2d	O	O	O	O	O	O	O
Dequantize	kTfLiteBuiltinDequantize	O	O	O	O	O	O	O
ElementWiseAdd	kTfLiteBuiltinAdd	O	O	O	O	O	O	O
ElementWiseDiv	kTfLiteBuiltinDiv	O	O	O	O	O	O	O
ElementWiseMul	kTfLiteBuiltinMul	O	O	O	O	O	O	O
ElementWiseSub	kTfLiteBuiltinSub	O	O	O	O	O	O	O
Elu	kTfLiteBuiltinElu	O	O	O	O	O	O	O
Equal	kTfLiteBuiltinEqual	O	O	O			O	O	O	O
FullyConnected	kTfLiteBuiltinFullyConnected	O	O	O	O	O	O	O
Gather	kTfLiteBuiltinGather	O	O	O
Greater	kTfLiteBuiltinGreater	O	O	O			O	O	O	O
GreaterEqual	kTfLiteBuiltinGreaterEqual	O	O	O			O	O	O	O
HardSwish	kTfLiteBuiltinHardSwish	O	O	O	O	O	O	O
L2Norm	kTfLiteBuiltinL2Normalization	O	O	O				O
Less	kTfLiteBuiltinLess	O	O	O			O	O	O	O
LessEqual	kTfLiteBuiltinLessEqual	O	O	O			O	O	O	O
Maximum	kTfLiteBuiltinMaximum	O	O	O	O	O	O	O
MaxPooling	kTfLiteBuiltinMaxPool2d	O	O	O	O	O	O	O
Mean	kTfLiteBuiltinMean	O	O	O	O	O	O	O
Minimum	kTfLiteBuiltinMinimum	O	O	O	O	O	O	O
Neg	kTfLiteBuiltinNeg	O	O	O	O	O	O	O
NotEqual	kTfLiteBuiltinNotEqual	O	O	O			O	O	O	O
Pack	kTfLiteBuiltinPack	O	O	O	O	O	O
Pad	kTfLiteBuiltinPad	O	O	O	O	O	O	O
Pad	kTfLiteBuiltinPadv2	O	O	O	O	O	O	O
Pow	kTfLiteBuiltinPow	O	O	O	O	O	O
PRelu	kTfLiteBuiltinPrelu	O	O	O	O	O	O	O
PRelu	kTfLiteBuiltinLeakyRelu	O	O	O	O	O	O	O
Quantize	kTfLiteBuiltinQuantize	O	O	O	O	O	O	O
ReduceAny	kTfLiteBuiltinReduceAny	O	O	O
ReduceMax	kTfLiteBuiltinReduceMax	O	O	O
ReduceMin	kTfLiteBuiltinReduceMin	O	O	O
ReLU	kTfLiteBuiltinRelu	O	O	O	O	O	O	O
ReLU6	kTfLiteBuiltinRelu6	O	O	O	O	O	O	O
Reshape	kTfLiteBuiltinReshape	O	O	O	O	O	O	O
Reshape	kTfLiteBuiltinSqueeze	O	O	O	O	O	O	O
Resize::BILINEAR	kTfLiteBuiltinResizeBilinear	O	O	O	O	O	O	O
Resize::NEAREST	kTfLiteBuiltinResizeNearestNeighbor	O	O	O	O	O	O	O
RSqrt	kTfLiteBuiltinRsqrt	O	O	O	O	O	O	O
Sigmoid	kTfLiteBuiltinLogistic	O	O	O	O	O	O	O
Slice	kTfLiteBuiltinSlice	O	O	O	O	O	O
SoftMax	kTfLiteBuiltinSoftmax	O	O	O	O	O	O	O
SpaceToDepth	kTfLiteBuiltinSpaceToDepth	O	O	O	O	O	O	O
Split	kTfLiteBuiltinSplit	O	O	O	O	O	O	O
Sqrt	kTfLiteBuiltinSqrt	O	O	O	O	O	O	O
Square	kTfLiteBuiltinSquare	O	O	O	O	O	O
SquaredDifference	kTfLiteBuiltinSquaredDifference	O	O	O	O	O	O
StridedSlice	kTfLiteBuiltinStridedSlice	O	O	O	O	O	O	O
Tanh	kTfLiteBuiltinTanh	O	O	O	O	O	O	O
Tile	kTfLiteBuiltinTile	O	O	O
Transpose	kTfLiteBuiltinTranspose	O	O	O	O	O	O	O
TransposeConv2D	kTfLiteBuiltinTransposeConv	O	O	O	O	O	O	O

Supported Hardware Operations

The table below lists the supported TFLite operations (OPs) for hardware targets on different MediaTek platforms.

Note

• Each OP has both might have hardware and software constraints.

• For details on MDLA constrains, see MDLA 2.0 Guidelines.

• To check the supported TFLite OP versions, use the command --show-builtin-ops in ncc-tflite.

TFLite operations	MDLA 1.0	MDLA 1.5/1.7	MDLA 2.0	MDLA 3.0	VPU (asym8)	VPU (FP16)	MVPU 2.0
ABS	O	O	O	O
ADD	O	O	O	O	O	O	O
ARG_MAX					O		O
ARG_MIN					O		O
AVERAGE_POOL_2D	O	O	O	O	O		O
BATCH_TO_SPACE_ND	O	O	O	O	O
CAST		O	O	O	O		O
CEIL							O
CHANNEL_SHUFFLE							O
CONCATENATION	O	O	O	O	O
CONV_2D	O	O	O	O	O
DEPTH_TO_SPACE		O	O	O	O		O
DEPTHWISE_CONV_2D	O	O	O	O	O
DEQUANTIZE		O	O	O	O	O	O
DIV		O	O	O	O	O	O
ELU	O	O	O	O
EQUAL					O		O
EXP							O
EXPAND_DIMS	O	O	O	O	O
FILL							O
FLOOR							O
FULLY_CONNECTED	O	O	O	O	O	O
GATHER					O		O
GREATER					O		O
GREATER_EQUAL					O		O
HARD_SWISH		O	O	O	O		O
L2_NORMALIZATION					O		O
L2_POOL_2D	O	O	O	O
LEAKY_RELU							O
LESS					O		O
LESS_EQUAL					O		O
LOCAL_RESPONSE_NORMALIZATION
LOG							O
LOG_SOFTMAX							O
LOGICAL_AND							O
LOGICAL_NOT							O
LOGICAL_OR							O
LOGISTIC	O	O	O	O	O		O
LSTM (QSTM)		O	O	O	O		O
MAX_POOL_2D	O	O	O	O	O		O
MAXIMUM	O	O	O	O	O		O
MEAN	O	O	O	O	O		O
MINIMUM	O	O	O	O	O		O
MIRRORPAD		O	O	O
MUL	O	O	O	O	O	O	O
NEG		O	O	O
NOT_EQUAL					O		O
PACK		O	O	O	O
PAD	O	O	O	O	O
POW		O	O	O	O (SQRT)
PRELU	O	O	O	O	O		O
QUANTIZE		O	O	O	O	O	O
REDUCE_ANY					O		O
REDUCE_MAX		O	O	O	O		O
REDUCE_MIN		O	O	O	O		O
RELU	O	O	O	O	O		O
RELU_N1_TO_1	O	O	O	O	O		O
RELU6	O	O	O	O	O		O
RESHAPE	O	O	O	O	O		O
RESIZE_BILINEAR	O	O	O	O	O
RESIZE_NEAREAST		O	O	O	O
ROUND							O
RSQRT			O	O	O	O
SELECT							O
SLICE	O	O	O	O	O
SOFTMAX			O	O	O	O	O
SPACE_TO_BATCH_ND		O	O	O	O
SPACE_TO_DEPTH		O	O	O	O		O
SPLIT		O	O	O	O		O
SPLIT_V		O	O	O	O		O
SQRT			O	O	O	O
SQUARE		O	O	O	O	O
SQUARED_DIFFERENCE		O	O	O
SQUEEZE		O	O	O	O
STRIDED_SLICE		O	O	O	O
SUB	O	O	O	O	O		O
SUM							O
TANH	O	O	O	O	O		O
TILE					O
TOPK_V2					O		O
TRANSPOSE		O	O	O	O	O
TRANSPOSE_CONV	O	O	O	O	O
UNPACK		O	O	O