Skip to content

Latest commit

 

History

History
97 lines (88 loc) · 4.9 KB

supported-ops-and-platforms.md

File metadata and controls

97 lines (88 loc) · 4.9 KB

Supported precision

  1. CUDA only supports FP16 precision on Turing Devices.
  2. x86 only supports FP32 precision on AVX512/FMA.

Supported operators

  • ONNX
Op Type Op Set Linux X86-64 Linux CUDA
Add 11 ✓ ✓
And 11 ✓ ✓
ArgMax 11 ✓ ✓
AveragePool 11 ✓ ✓
BatchNormalization 11 ✓ ✓
Cast 11 ✓ ✓
Ceil 11 ✓ ✓
Clip 11 ✓ ✓
Concat 11 ✓ ✓
Constant 11 ✓
ConstantOfShape 11 ✓ ✓
Conv 11 ✓ ✓
ConvTranspose 11 ✓ ✓
DepthToSpace 11 ✓ ✓
Div 11 ✓ ✓
Equal 11 ✓ ✓
Exp 11 ✓ ✓
Expand 11 ✓ ✓
Flatten 11 ✓ ✓
Floor 11 ✓ ✓
Gather 11 ✓ ✓
GatherND 11 ✓ ✓
Gemm 11 ✓ ✓
Greater 11 ✓ ✓
Identity 11 ✓ ✓
If 13 ✓ ✓
LeakyRelu 11 ✓ ✓
Less 11 ✓ ✓
Log 11 ✓ ✓
Loop 13 ✓ ✓
MatMul 11 ✓
Max 11 ✓ ✓
MaxPool 11 ✓ ✓
MaxUnpool 11 ✓ ✓
Min 11 ✓ ✓
Mul 11 ✓ ✓
NonMaxSuppression 11 ✓ ✓
NonZero 11 ✓ ✓
Not 11 ✓ ✓
Pad 11 ✓ ✓
Pow 11 ✓ ✓
ReduceMax 11 ✓ ✓
ReduceMean 11 ✓ ✓
ReduceMin 11 ✓ ✓
ReduceProd 11 ✓ ✓
ReduceSum 11 ✓ ✓
Relu 11 ✓ ✓
Reshape 11 ✓ ✓
Resize 11 ✓ ✓
RoiAlign 11 ✓ ✓
ScatterElements 11 ✓ ✓
ScatterND 11 ✓ ✓
SequenceAt 13 ✓ ✓
Shape 11 ✓ ✓
Sigmoid 11 ✓ ✓
Size 11 ✓ ✓
Slice 11 ✓ ✓
Softmax 11 ✓ ✓
Split 11 ✓ ✓
SplitToSequence 13 ✓ ✓
Sqrt 11 ✓ ✓
Squeeze 11 ✓ ✓
Sub 11 ✓ ✓
Sum 11 ✓ ✓
Tanh 11 ✓ ✓
TopK 11 ✓ ✓
Transpose 11 ✓ ✓
Unsqueeze 11 ✓ ✓
Where 11 ✓ ✓
  • MMCV
Op Type Op Set Linux X86-64 Linux CUDA
NonMaxSuppression 1 ✓ ✓
RoiAlign 1 ✓ ✓
grid_sample 1 ✓ ✓
  • PPL
Op Type Op Set Linux X86-64 Linux CUDA
ChannelShuffle 1 ✓ ✓
Shape 1 ✓ ✓
Swish 1 ✓