4. Supported Operators¶

TIDL accelerates the operators listed in this chapter on the C7™ NPU. Operators not listed here fall back to TVM-generated C7™ NPU code (c7x_codegen=1) or Arm (c7x_codegen=0).

ONNX 1.14.0 / ONNX Runtime 1.15.0 (OPSET-19, IR-8)

For detailed per-operator test reports including attribute and shape coverage, see the edgeai-tidl-tools operator reports.

4.1. Convolution¶

ONNX Operator	TIDL Layer	Key Constraints
`Conv`	TIDL_ConvolutionLayer	Stride must match horizontally and vertically. 3x3 kernel with stride 3 not supported on AM62A/AM67A. Kernel >7 with stride 2 not supported. Depthwise supported for 1x3s1, 3x3s1/s2, 5x5s1/s2, 7x7s1/s2. Stride 4 only with 11x11 kernel.
`ConvTranspose`	TIDL_Deconv2DLayer	Supported kernels: 4x4, 3x3, 2x2 with 2x2 stride. 16-bit not supported on AM62A/AM67A. Default dilation (1,1) only.
`DeformConv`	TIDL_DeformableConvLayer	3x3 stride-1 filter only.

4.2. Pooling¶

ONNX Operator	TIDL Layer	Key Constraints
`AveragePool` / `GlobalAveragePool` / `MaxPool`	TIDL_PoolingLayer	GlobalAveragePool: plane size (H×W) must be ≤ 1024. Validated kernel sizes: 3x3, 2x2, 1x1 with stride 1 or 2 (except 2x2 stride 1). Default dilation (1,1) only.

4.3. Normalization¶

ONNX Operator	TIDL Layer	Key Constraints
`BatchNormalization`	TIDL_BatchNormLayer	training_mode=1 not supported. Scale, bias, mean, and variance must be constant 1-D tensors.
`InstanceNormalization`	TIDL_InstanceNormLayer	Input must be ≥ 3D. Scale and bias must be 1-D channel vectors.
`LayerNormalization`	TIDL_LayerNormLayer	Width axis only. Scale and bias must be [1,N] or [N].

4.4. Activations¶

ONNX Operator	TIDL Layer	Key Constraints
`Relu`	TIDL_ReLULayer
`PRelu`	TIDL_PReLULayer	Variable slope not supported. Slope must be broadcast-able to the channel dimension.
`LeakyRelu`	TIDL_LeakyReluLayer
`Sigmoid` / `Logistic`	TIDL_SigmoidLayer
`Tanh`	TIDL_TanhLayer
`HardSigmoid`	TIDL_HardSigmoidLayer
`HardSwish`	TIDL_HardSwishLayer
`Elu`	TIDL_ELULayer
`Mish`	TIDL_MishLayer
`Clip`	TIDL_ClipLayer	min ≤ 0 and max > 0 only.
`Softmax`	TIDL_SoftMaxLayer	Width and height axes only.

4.5. Element-wise Arithmetic¶

ONNX Operator	TIDL Layer	Key Constraints
`Add` / `Sub` / `Mul` / `Div` / `Max` / `Min` / `Sum`	TIDL_EltWiseLayer	Two inputs only. Inputs must have matching dimensions or be broadcast-able.
`Abs`	TIDL_AbsLayer
`Neg`	TIDL_NegLayer
`Sqrt`	TIDL_SqrtLayer
`Pow`	TIDL_PowLayer	Exponent must be a constant scalar tensor.
`Exp`	TIDL_ExpLayer
`Log`	TIDL_LogLayer
`Floor`	TIDL_FloorLayer

4.6. Trigonometric¶

ONNX Operator	TIDL Layer
`Sin`	TIDL_SinLayer
`Cos`	TIDL_CosLayer
`Tan`	TIDL_TanLayer
`Asin`	TIDL_AsinLayer
`Acos`	TIDL_AcosLayer
`Atan`	TIDL_AtanLayer
`Sinh`	TIDL_SinhLayer
`Cosh`	TIDL_CoshLayer
`Asinh`	TIDL_AsinhLayer

4.7. Linear and Matrix¶

ONNX Operator	TIDL Layer	Key Constraints
`Gemm` / `MatMul`	TIDL_InnerProductLayer	Gemm: transA=0, alpha=1.0, beta=1.0. Bias must be [1,N] or [N]. MatMul with signed/unsigned mixed inputs not supported on TDA4VM.
`TopK`	TIDL_TopKLayer	sorted=0 not supported. K must be a model initializer. Width and height axes only.

4.8. Shape Manipulation¶

ONNX Operator	TIDL Layer	Key Constraints
`Reshape`	TIDL_ReshapeLayer	Variable shape not supported. Input and output volume must match.
`Flatten`	TIDL_FlattenLayer
`Squeeze`	TIDL_SqueezeLayer
`Unsqueeze`	TIDL_UnsqueezeLayer	Output dimensions must be ≤ 6.
`Transpose`	TIDL_TransposeLayer	For >4D inputs, permutations mapping the width dimension to output position 0 or 1 are not supported.
`Concat`	TIDL_ConcatLayer	Axis values -3, -2, -1 only. Batch dimension not supported.
`Slice` / `Split`	TIDL_SliceLayer	4D input only. Batch size=1. Non-singular strides only supported as part of the Patch Merging fusion pattern.
`Pad`	TIDL_PadLayer	Constant pad mode with constant_value=0 only. Width/height axes only.
`Expand`	TIDL_ExpandLayer	Shape tensor must be constant.

4.9. Spatial and Image¶

ONNX Operator	TIDL Layer	Key Constraints
`Upsample` / `Resize`	TIDL_ResizeLayer	Modes: nearest or linear. Width and height axes only. Scales < 1 not supported. Width and height scales must match. Power-of-2 scales only.
`DepthToSpace`	TIDL_DepthToSpaceLayer	4D input. Channel depth must be multiple of blocksize². DCR and CRD modes supported.
`SpaceToDepth`	TIDL_SpaceToDepthLayer	4D input. H and W must be multiples of blocksize.
`GridSample`	TIDL_GridSampleLayer	Nearest and bilinear modes. Zero padding only. 2D grid only. Nearest-neighbour rounding may differ from ONNX reference.

4.10. Reduction¶

ONNX Operator	TIDL Layer	Key Constraints
`ReduceMean`	TIDL_ReduceMeanLayer
`ReduceSum`	TIDL_ReduceSumLayer
`ReduceMin` / `ReduceMax`	TIDL_ReduceLayer	Height axis only. keepdims=1 only.
`ArgMin` / `ArgMax`	TIDL_ArgOpLayer	keepdims=1 only. axis=-3 only.

4.11. Gather and Scatter¶

ONNX Operator	TIDL Layer	Key Constraints
`Gather`	TIDL_GatherLayer	1D indices only. Input rank must be > 1. Data cannot be constant; only indices can be constant.
`ScatterND` / `ScatterElements`	TIDL_ScatterElementsLayer	ScatterElements: ‘none’ reduction only. ScatterND: ‘mul’ reduction not supported. Scatter along width axis only. Input data must be a zero tensor.

4.12. Quantization¶

ONNX Operator	TIDL Layer	Key Constraints
`Cast`	TIDL_CastLayer	Only at terminal nodes (network input/output).
`DequantizeLinear`	TIDL_DequantizeLayer	axis=1 only. QDQ models only.
`QuantizeLinear`	TIDL_QuantizeLayer	axis=1 only. QDQ models only.

4.13. Fusion Patterns Only¶

The following operators are not supported in isolation but are accelerated when they appear as part of a recognised fusion pattern.

ONNX Operator	TIDL Layer	Fusion Pattern
`Erf` / `Identity`	TIDL_IdentityLayer	GELU activation
`DropOut`	TIDL_DropOutLayer	Not supported standalone