This RKNN2 V2.5.0 platform reference helps ArmSoM developers select and adapt models. It does not establish support on every ArmSoM board: verify the board SoC, actual Toolkit2, Runtime, NPU driver and BSP versions. The source specifies no universal Linux kernel version. No ArmSoM hardware validation was performed.
Source: the supplied Chinese and English RKNN2 Supported Operators and Constraints Reference V2.5.0, 374 and 384 pages respectively. Chinese follows the Chinese edition; English, Russian and German follow the English edition. Platform constraints remain separate. The English PDF itself contains Chinese I/O figures and output tables; their provenance is retained with editable localized tables. Source spellings such as ConvTranpose and the original LSTM/GRU equations are preserved; these equations are not corrections to SDK implementations.
Source pages: CN 353–366; EN 363–376.
Operator reference series
RKNN2 V2.5.0 · RK3566/RK3568：Basics and convolution
RKNN2 V2.5.0 · RK3566/RK3568：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RK3566/RK3568：Shapes, slicing and transpose
RKNN2 V2.5.0 · RK3566/RK3568：Activations, normalization and extensions
RKNN2 V2.5.0 · RK3588：Basics and convolution
RKNN2 V2.5.0 · RK3588：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RK3588：Shapes, slicing and transpose
RKNN2 V2.5.0 · RK3588：Activations, normalization and extensions
RKNN2 V2.5.0 · RK3562：Basics and convolution
RKNN2 V2.5.0 · RK3562：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RK3562：Shapes, slicing and transpose
RKNN2 V2.5.0 · RK3562：Activations, normalization and extensions
RKNN2 V2.5.0 · RV1103/RV1106：Basics and convolution
RKNN2 V2.5.0 · RV1103/RV1106：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RV1103/RV1106：Shapes, slicing and transpose
RKNN2 V2.5.0 · RV1103/RV1106：Activations, normalization and extensions
RKNN2 V2.5.0 · RK3576/RV1126B：Basics and convolution
RKNN2 V2.5.0 · RK3576/RV1126B：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RK3576/RV1126B：Shapes, slicing and transpose
RKNN2 V2.5.0 · RK3576/RV1126B：Activations, normalization and extensions
RKNN2 V2.5.0 · RK2118：Basics and convolution
RKNN2 V2.5.0 · RK2118：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RK2118：Shapes, slicing and transpose
RKNN2 V2.5.0 · RK2118：Activations, normalization and extensions
RKNN2 V2.5.0 · RV1103B/RV1106B：Basics and convolution
RKNN2 V2.5.0 · RV1103B/RV1106B：Recurrent networks, arithmetic and pooling
RKNN2 V2.5.0 · RV1103B/RV1106B：Shapes, slicing and transpose
RKNN2 V2.5.0 · RV1103B/RV1106B：Activations, normalization and extensions
RKNN2 V2.5.0 · CPU operator support and constraints
CPU OP Specifications
Note:
- RK2118 does not support all CPU operators
- RV1106/RV1103/RV1106B/RV1103B only support some CPU operators as follows
- Concat, DataConvert, Div, exLayerNorm, Exp, Gather, ReduceL2, ReduceMax, Reshape, ScatterND,Softmax,Pow
Add
- Description: Addition operation
- Specification Constraints: No limitation
AveragePool
- Description: Average pooling
- Specification Constraints: No limitation
ArgMin
- Description: Returns the index of the minimum value
- Specification Constraints: No limitation
ArgMax
- Description: Returns the index of the maximum value
- Specification Constraints: No limitation
BatchNormalization
- Description: Batch normalization
- Specification Constraints: No limitation
Cast
- Description: Data type conversion
- Specification Constraints: SRC supports:float32/bool/int8/float16/int32/int64 DST supports:float32/int8/int32/float16
Clip
- Description: Clipping activation layer
- Specification Constraints: No limitation
Concat
- Description: Concatenation operation
- Specification Constraints: axis only supports {0,1,2,3}
Convolution
- Description: Convolution operation
- Specification Constraints: No limitation
ConvTranspose/Deconvolution
- Description: Transposed convolution
- Specification Constraints: No limitation
ConvTransposePad
- Description: Combined operator
- Specification Constraints: No limitation
Cos
- Description: Cosine function
- Specification Constraints: No limitation
DataConvert
- Description: Data type conversion
- Specification Constraints: only bool/int8/float type conversion is supported
DepthToSpace
- Description: Channel-to-spatial direction conversion
- Specification Constraints: No limitation
Div
- Description: Division operation
- Specification Constraints: No limitation
Equal
- Description: Equal to
- Specification Constraints: No limitation
Exp
- Description: Exponential function
- Specification Constraints: No limitation
Flatten
- Description: Flatten operation
- Specification Constraints: No limitation
Gather
- Description: Gather operation
- Specification Constraints: No limitation
Greater
- Description: Greater than
- Specification Constraints: No limitation
GreaterOrEqual
- Description: Greater than or equal to
- Specification Constraints: No limitation
GRU
- Description: Gated Recurrent Unit
- Specification Constraints: No limitation
exGRU
- Description: Gated Recurrent Unit
- Specification Constraints: No limitation
- Note: ONNX extension operator
exHardSwish
- Description: Activation function
- Specification Constraints: No limitation
- Note: ONNX extension operator
InstanceNormalization
- Description: Instance normalization
- Specification Constraints: No limitation
exLayerNorm
- Description: Layer normalization
- Specification Constraints: No limitation
- Note: ONNX extension operator
Less
- Description: Less than
- Specification Constraints: No limitation
LessOrEqual
- Description: Less than or equal to
- Specification Constraints: No limitation
LogSoftmax
- Description: Activation function
- Specification Constraints: batchsize only supports 1
LpNormalization
- Description: Lp normalization
- Specification Constraints: No limitation
exLRN
- Description: Local response normalization
- Specification Constraints: No limitation
- Note: ONNX extension operator
MatMul
- Description: Multi-dimensional matrix multiplication
- Specification Constraints: No limitation (4D x 4D and 4D x 3D computation are supported)
Max
- Description: Maximum
- Specification Constraints: No limitation
MaxPool
- Description: Max pooling
- Specification Constraints: No limitation
MaxRoiPool
- Description: Region Max pooling
- Specification Constraints: No limitation
MaxUnpool
- Description: Reverse Max pooling
- Specification Constraints: No limitation
exMish
- Description: Activation function
- Specification Constraints: No limitation
- Note: ONNX extension operator
Min
- Description: Minimum value
- Specification Constraints: No limitation
Mul
- Description: Multiplication
- Specification Constraints: No limitation
Pad
- Description: Padding
- Specification Constraints: No limitation
Pow
- Description: Exponentiation
- Specification Constraints: No limitation
exProposal
- Description: Region proposal network
- Specification Constraints: batchsize only supports 1
- Note: ONNX extension operator
ReduceMax
- Description: Compute Max along the specified dimension
- Specification Constraints: the output dimension cannot exceed 4 dimensions
ReduceMean
- Description: Compute Mean along the specified dimension
- Specification Constraints: the output dimension cannot exceed 4 dimensions
ReduceSum
- Description: Compute Sum along the specified dimension
- Specification Constraints: the output dimension cannot exceed 4 dimensions
ReduceMin
- Description: Compute Min along the specified dimension
- Specification Constraints: the output dimension cannot exceed 4 dimensions
Reorg
- Description: Data reorganization
- Specification Constraints: No limitation
Reshape
- Description: Data shape change
- Specification Constraints: No limitation
Resize
- Description: Data width/height direction scaling
- Specification Constraints: supported interpolation methods are bilinear; nearest2d
ReverseSequence
- Description: Sequence reversal
- Specification Constraints: No limitation
exRMSNorm
- Description: Root mean square normalization
- Specification Constraints: No limitation
- Note: ONNX extension operator
RoiAlign
- Description: Region align pooling
- Specification Constraints: only AvgPoolMode is supported, batchsize only supports 1
ScatterND
- Description: N-dimensional index gathering
- Specification Constraints: No limitation
Sin
- Description: Sine function
- Specification Constraints: No limitation
Slice
- Description: Slice operation
- Specification Constraints: batchsize only supports 1
Softmax
- Description: Activation function
- Specification Constraints: batchsize only supports 1
- Note: consistent with the ONNX OPSET11 specification
exSoftmax13
- Description: Activation function
- Specification Constraints: batchsize only supports 1
- Note: ONNX extension operator, consistent with the ONNX OPSET13 specification
SpaceToDetph
- Description: Spatial-to-channel direction conversion
- Specification Constraints: No limitation
Split
- Description: Split data
- Specification Constraints: No limitation
Sqrt
- Description: Square root
- Specification Constraints: No limitation
Squeeze
- Description: Squeeze data dimensions
- Specification Constraints: No limitation
Sub
- Description: Subtraction
- Specification Constraints: No limitation
Tanh
- Description: Hyperbolic tangent function
- Specification Constraints: No limitation
Tile
- Description: Expand and copy data
- Specification Constraints: batchsize only supports 1, broadcasting is not supported
Transpose
- Description: Transpose computation
- Specification Constraints: No limitation
Upsample
- Description: Upsample
- Specification Constraints: supported interpolation methods are bilinear; nearest2d
Not
- Description: Element-wise negation
- Specification Constraints: No limitation
where
- Description: Gather by mask
- Specification Constraints: No limitation
Erf
- Description: Error function (Erf)
- Specification Constraints: No limitation
Floor
- Description: Floor operation function
- Specification Constraints: No limitation
Mod
- Description: Modulo function
- Specification Constraints: No limitation
exMeanVarianceNormalization
- Description: Mean variance normalization function
- Specification Constraints: No limitation
And
- Description: AND function
- Specification Constraints: No limitation
GatherElements
- Description: Element gathering function
- Specification Constraints: No limitation
Log
- Description: Log function
- Specification Constraints: No limitation
exDataConvert
- Description: Precision conversion
- Specification Constraints: No limitation
SpaceToDepth
- Description: Dimension transformation
- Specification Constraints: No limitation
SoftmaxMask
- Description: where-fused softmax activation function
- Specification Constraints: No limitation
TransposeReshape
- Description: Combined operator
- Specification Constraints: No limitation
ReshapeTranspose
- Description: Combined operator
- Specification Constraints: No limitation
ReduceL2
- Description: L2 norm along an axis
- Specification Constraints: No limitation
LayerNormalization
- Description: Layer normalization
- Specification Constraints: No limitation
PassThrough
- Description: Direct connection
- Specification Constraints: No limitation
Expand
- Description: Tensor data copy or expansion
- Specification Constraints: No limitation
Sigmoid
- Description: Activation function
- Specification Constraints: No limitation
Tanh
- Description: Trigonometric functions
- Specification Constraints: No limitation
Cos
- Description: Trigonometric functions
- Specification Constraints: No limitation
Sin
- Description: Trigonometric functions
- Specification Constraints: No limitation
exSwish
- Description: Activation function
- Specification Constraints: No limitation
exGlu
- Description: Activation function
- Specification Constraints: No limitation
exGelu
- Description: Activation function
- Specification Constraints: No limitation
exWindow
- Description: exWindow originates from the window_partition and window_reverse operations of Swin-Transformer
- Specification Constraints: No limitation
exNorm
- Description: When mode is set to ‘pixel’, exNorm will use PixelNorm for normalization. When mode is omitted, exNorm uses LayerNorm for normalization by default.
- Specification Constraints: No limitation
Round
- Description: Data saturation
- Specification Constraints: No limitation
TopK
- Description: Get the top K ranked values
- Specification Constraints: No limitation
exSwooshR
- Description: Activation function
- Specification Constraints: No limitation
exSwooshL
- Description: Activation function
- Specification Constraints: No limitation
OneHot
- Description: One-hot encoding
- Specification Constraints: No limitation
Cumsum
- Description: Cumulative sum along the specified dimension
- Specification Constraints: No limitation