RKNN2 V2.5.0 · RK3576/RV1126B：Shapes, slicing and transpose

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 250–263; EN 258–271.

Rockchip RKNN2

Relu

Sets negative values of the input tensor to zero, keeping positive values unchanged

Inputs

• input:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
  • Broadcasting support
    • None

Outputs

• output:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Other Support

• Multi-core: Not yet supported

Clip/ReLU6

Clips the values of the input tensor to the range [0,6]

Inputs

• input:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
  • Broadcasting support
    • None

Outputs

• output:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Other Support

• Multi-core: Not yet supported

PRelu

Scales negative values of the input tensor by a learnable parameter, keeping positive values unchanged

Inputs

• input:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer/per-channel
  • Broadcasting support
    • None

Outputs

• output:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer/per-channel

Attributes

• slope: PRelu coefficient
  • Only a single scalar or C-dimension coefficients are supported

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Other Support

• Multi-core: Not yet supported

LeakyRelu

Scales negative values of the input tensor by a fixed ratio, keeping positive values unchanged

Inputs

• input:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
  • Broadcasting support
    • None

Outputs

• output:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Other Support

• Multi-core: Not yet supported

Reshape

Returns a tensor with the specified shape without changing the input data or the number of elements.

Inputs

• input:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Outputs

• output:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Attributes

• shape:int
  • Specifies the shape of the output tensor

Other Support

• Multi-core: Not yet supported

Resize

Resizes the input tensor. In general, it computes each value in the output tensor as a weighted average of the input tensor values in the neighborhood (i.e., sampling locations) of the corresponding position

Inputs

• X:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
• roi(optional):T2
  • Currently not supported
• scales(optional):T2
  • The array of scale factors along each dimension
• sizes(optional):T3
  • The target size of the output tensor

Outputs

• Y:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)
• T2:float(Tensor)
• T3:int64(Tensor)

Attributes

• antialias: int (default value 0)
  • Currently not supported to set
• axes:int[]
  • Currently not supported to set
• coordinate_transformation_mode: strings (default value “half_pixel”)
  • Currently only half_pixel, pytorch_half_pixel, and align_corners are supported
• cubic_coeff_a: float (default -0.75)
  • Currently not supported to set
• exclude_outside: int (default value 0)
  • Currently not supported to set
• extrapolation_value: float (default 0)
  • Currently not supported to set
• keep_aspect_ratio_policy: strings (default value “stretch”)
  • Currently not supported to set
• mode: strings (default value “nearest”)
  • Currently only nearest and linear are configurable
• nearest_mode: strings (default value “round_prefer_floor”)
  • Currently not supported to set Support Note Currently only integer scaling in width/height direction is supported, and the scale factor must be an integer multiple of 2-8. If the input is (1,c,1,1), it can be scaled up to (1,c,h,w)
• Multi-core: Not supported

Slice

Obtains a slice of the current vector, with the same rules as numpy slicing

Inputs

• input_tensor:T
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Outputs

• output:T
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T:int8(Tensor),float16(Tensor)

Attributes

• starts:int(Tensor)
  • The start position of the split: No limitation
• ends:int(Tensor)
  • The end position of the split: No limitation
• axes:int(Tensor)
  • The axis to split: any axis 0~3 is supported, and multiple axes can be selected simultaneously
• steps:int(Tensor)
  • The step size of the corresponding axis for splitting
  • When height*width==1&&channel%subc==0&&starts[1]%subc==0&&ends[1]% when subc==0&&steps[1]<=subc, steps can be other than 1; otherwise, Slice OP can only run on the NPU when steps==[1,1,1,1]

Softmax

This operator computes the normalized exponential value of the given input:

Softmax(input, axis) = exp(input − max(input, axis)) / Σ_axis exp(input − max(input, axis))

The “axis” attribute indicates the dimension along which Softmax is performed. The output tensor has the same shape and contains the Softmax values of the corresponding input.

Inputs

• input:T1
  • shape:[batch,channel,height,width]
    • Subject to the transpose specification constraints

Outputs

• output:T1
  • shape:[batch,channel,height,width]

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Attributes

• axis:int64
  • The axis for softmax: 1,3, i.e., channel and width directions

Other Support

• Multi-core: Not yet supported

exSoftmax13

This operator computes the normalized exponential value of the given input:

exSoftmax13(input, axis) = exp(input − max(input, axis)) / Σ_axis exp(input − max(input, axis))

The “axis” attribute indicates the dimension along which Softmax is performed. The output tensor has the same shape and contains the Softmax values of the corresponding input.

Inputs

• input:T1
  • shape:[batch,channel,height,width]
    • Subject to the transpose specification constraints

Outputs

• output:T1
  • shape:[batch,channel,height,width]

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Attributes

• axis:int64
  • The axis for softmax: 1,3, i.e., channel and width directions

Other Support

• Multi-core: Not yet supported

exSoftmaxMask

This operator computes the normalized exponential value of the given input:

input(x, mask_t, mask_value) =
    x + (mask_t - 1) * inf       if mask_value = 0
    x + mask_t * (-inf)          elif mask_value = 1

exSoftmaxMask(input, axis) = exp(input - max(input, axis)) / Σ_axis(exp(input - max(input, axis)))

The “axis” attribute indicates the dimension along which Softmax is performed. The output tensor has the same shape and contains the Softmax values of the corresponding input.

Inputs

• input:T1
  • shape:[batch,channel,height,width]
    • Subject to the transpose specification constraints
• mask(optional):T1
  • shape:[1,channel,height,width]
    • channel∈(0,8192]
    • width: · axis=1,No limitation; · axis=3/-1,width∈(0,8192]

Outputs

• output:T1
  • shape:[batch,channel,height,width]

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)

Attributes

• axis:int64
  • The axis for softmax: 1,3, i.e., channel and width directions
• mask_value:int64
  • The values to be masked: 0,1

Other Support

• Multi-core: Not yet supported

Split

Splits the vector into num_outputs vectors along the axis direction. split specifies the size of each split vector along the axis

Inputs

• input_tensor:T
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Outputs

• output:T
  • One or more output vectors, determined by the num_outputs attribute
  • shape: determined by the splitting situation
  • Quantization support
    • per-layer

Data Type Constraints

• T:int8(Tensor),float16(Tensor)

Attributes

• axis:int
  • axis∈{0,1,2,3}
• num_outputs:int
• split:int(Tensor)

Sub

• Performs element-wise binary subtraction (C=A-B), supporting multi-directional NumPy-style broadcasting

Inputs

• A:T1,T2
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
  • Broadcasting support
    • [b,c,1,1]
    • [scalar]
• B:T1,T2
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
  • Broadcasting support
    • [b,c,1,1]
    • [scalar]

Outputs

• C:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)
• T2:float(constTensor)

Other Support

• Multi-core: Not supported

Tile

Constructs a tensor by tiling a given tensor. This is the same as the tile function in Numpy, but without broadcasting

Inputs

• input:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer
• repeats:T2
  • shape:[batch,channel,height,width]

Outputs

• output:T1
  • shape:[batch,channel,height,width]
  • Quantization support
    • per-layer

Data Type Constraints

• T1:int8(Tensor),float16(Tensor)
• T2:int64(Tensor)

Other Support

• Multi-core: Not supported

Transpose

Transposes the input tensor

Inputs

• X:T1
  • shape:[n1,c1,h1,w1]
    • See Attributes for detailed constraint rules
  • Quantization support
    • per-layer

Outputs

• Y:T1,T2
  • shape:[n2,c2,h2,w2]
    • See Attributes for detailed constraint rules
  • Quantization support
    • per-layer

Attributes

• The axis order of the perm transpose: The perm parameter restrictions are as follows: (where p=268435456, q=65536)
  • perm=[0,1,2,3]
    • No limitation
  • perm=[0,1,3,2]
    • T1:h1%16=0,h1w1<16p,h1w1<2q
    • T2:h1%8=0,h1w1<8p,h1w1<2q
  • perm=[0,2,3,1]
    • T1:c1<8192,h1w1%8=0,w1c1<p*16
    • T2:c1<8192,h1w1%8=0,w1c1<p*8
  • perm=[0,2,1,3]
    • T1:h1%16=0,h1w1<16p,h1w1<2q,c1*w1<p
    • T2:h1%8=0,h1w1<8p,h1w1<2q,c1*w1<p
  • perm=[0,3,1,2]
    • T1,T2:h1w1<p,w1<8192,h1c1<p
  • perm=[0,3,2,1]
    • T1:h1%16=0,h1w1<16p,h1w1<2q,c1<p*16
    • T2:h1%8=0,h1w1<8p,h1w1<2q,c1<p*8
  • perm=[1,0,2,3]
    • T1:h1w1<16p,n1<8192,h1*w1<p
    • T2:h1w1<8p,n1<8192,h1*w1<p
  • perm=[1,0,3,2]
    • T1:h1%16=0,c1h1w1<16p,h1w1<2*q
    • T2:h1%8=0,c1h1w1<8p,h1w1<2*q
  • perm=[1,2,0,3]
    • T1:n1w1<p16,c1h1w1<p,w1%16=0
    • T2:n1w1<p8,c1h1w1p,w1%8=0
  • perm=[1,2,3,0]
    • T1:h1w1<p,n1<p16,c1h1w1<p16,n1w1<p*16
    • T2:h1w1<p,n1<p8,c1h1w1<p8,n1w1<p*8
  • perm=[1,3,0,2]
    • T1:c1w1<8192,n1h1<p,w1%16=0
    • T2:c1w1<8192,n1h1<p,w1%8=0
  • perm=[1,3,2,0]
    • T1:c1h1w1<p,n1*h1<p,w1<8192,w1%16=0
    • T2:c1h1w1<p,n1*h1<p,w1<8192,w1%8=0
  • perm=[2,1,0,3]
    • T1,T2:c1<8192,c1h1w1<p,n1*w1<p
  • perm=[2,1,3,0]
    • T1:c1h1w1<p,n1w1<p16,c1<8192,n1*w1%4=0
    • T2:c1h1w1<p,n1w1<p8,c1<8192,n1*w1%4=0
  • perm=[2,0,3,1]
    • T1,T2:c1<8192,c1h1w1<p
  • perm=[2,0,1,3]
    • T1:n1c1h1w1<p,h1w1%16=0
    • T2:n1c1h1w1<p,h1w1%8=0
  • perm=[2,3,1,0]
    • T1:c1h1w1<p,n1c1<p,h1w1<8192,h1*w1%16=0
    • T2:c1h1w1<p,n1c1<p,h1w1<8192,h1*w1%8=0
  • perm=[2,3,0,1]
    • T1,T2:h1w1<8192,n1c1<p
  • perm=[3,0,2,1]
    • T1:h1w1<p,n1c1<8192,c1h1<p,n1w1<8192,n1*w1%16=0
    • T2:h1w1<p,n1c1<8192,c1h1<p,n1w1<8192,n1*w1%8=0
  • perm=[3,0,1,2]
    • T1:h1w1<p,n1c1<8192,c1h1<p,n1w1<8192,n1h1w1%16=0
    • T2:h1w1<p,n1c1<8192,c1h1<p,n1w1<8192,n1h1w1%8=0
  • perm=[3,2,0,1]
    • T1,T2:h1w1<p,n1c1<8192,n1c1w1<p
  • perm=[3,2,1,0]
    • T1:c1h1w1<p,h1w1<8192,2h1w1<q,n1c1<p*16,h1%16=0
    • T2:c1h1w1<p,h1w1<8192,2h1w1<q,n1c1<p*8,h1%8=0
  • perm=[3,1,0,2]
    • T1,T2:h1w1<p,c1w1<8192,n1*h1<p
  • perm=[3,1,2,0]
    • T1,T2:h1w1<p,n1c1<8192,n1h1<p,n1c1*w1<p

Data Type Constraints

• T1:int8(Tensor)
• T2:int16(Tensor),float16(Tensor)

Other Support

• Multi-core: Not yet supported