Skip to content

Latest commit

 

History

History
106 lines (98 loc) · 5.27 KB

supported-ops-and-platforms.md

File metadata and controls

106 lines (98 loc) · 5.27 KB
Raw

Supported precision

CUDA only supports FP16 precision and int8 precision on Turing Devices. If you want to use int8 percision, you should maintain a json file like sample quant file

Supported operators and opsets

  • ONNX
Op Type Op Set Linux/Windows CUDA
Add 7~12 ✓
And 7~16 ✓
ArgMax 1~11 ✓
AveragePool 1~16 ✓
BatchNormalization 9~13 ✓
Cast 9~12 ✓
Ceil 6~12 ✓
Clip 6~13 ✓
Concat 4~12 ✓
ConstantOfShape 9~16 ✓
Conv 1~16 ✓
ConvTranspose 1~16 ✓
Cos 7~16 ✓
CumSum 11~16 ✓
DepthToSpace 1~12 ✓
Div 7~12 ✓
Equal 7~16 ✓
Erf 9~16 ✓
Exp 6~12 ✓
Expand 8~12 ✓
Flatten 1~12 ✓
Floor 6~16 ✓
Gather 1~16 ✓
GatherND 11 ✓
Gemm 11~12 ✓
GlobalAveragePool 1~16 ✓
GlobalMaxPool 1~16 ✓
Greater 9~16 ✓
GreaterOrEqual 9~16 ✓
Identity 1~12 ✓
If 1~12 ✓
InstanceNormalization 6~13 ✓
LeakyRelu 6~16 ✓
Less 9~16 ✓
Log 6~12 ✓
Loop 1~12 ✓
LSTM 7~13 ✓
MatMul 9~12 ✓
Max 8~11 ✓
MaxPool 1~16 ✓
MaxUnpool 9~16 ✓
Min 8~11 ✓
Mul 7~12 ✓
NonMaxSuppression 10~16 ✓
NonZero 9~12 ✓
Not 1~16 ✓
Pad 2~12 ✓
Pow 7~11 ✓
Range 11~16 ✓
ReduceL2 1~16 ✓
ReduceMax 1~16 ✓
ReduceMean 1~16 ✓
ReduceMin 1~16 ✓
ReduceProd 1~16 ✓
ReduceSum 1~16 ✓
Relu 6~12 ✓
Reshape 5~12 ✓
Resize 11~12 ✓
RoiAlign 10~15 ✓
ScatterElements 11~12 ✓
ScatterND 11~12 ✓
SequenceAt 11~16 ✓
Shape 1~12 ✓
Sigmoid 6~12 ✓
Sin 1~16 ✓
Slice 1~12 ✓
Softmax 1~12 ✓
Split 2~12 ✓
SplitToSequence 11~16 ✓
Sqrt 6~12 ✓
Squeeze 1~12 ✓
Sub 7~12 ✓
Tanh 6~12 ✓
Tile 6~12 ✓
TopK 11~16 ✓
Transpose 1~12 ✓
Unsqueeze 1~12 ✓
Where 9~15 ✓
  • MMCV
Op Type Op Set Linux/Windows CUDA
grid_sample 1 ✓
ModulatedDeformConv2d 1 ✓
NonMaxSuppression 1 ✓
RoiAlign 1 ✓
  • PPL
Op Type Op Set Linux/Windows CUDA
ChannelShuffle 1 ✓
Reduce 1 ✓
ShapeOperation 1 ✓