SUPPORT MATRIX FOR TENSORRT
SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | August 2019
Support Guide
TABLE OF CONTENTS
Chapter 1. Features For Platforms And Software......................................................... 1
Chapter 2. Layers And Features.............................................................................. 2
Chapter 3. Layers And Precision..............................................................................5
Chapter 4. Hardware And Precision..........................................................................8
Chapter 5. Software Versions Per Platform................................................................ 9
Chapter 6. Supported Ops.................................................................................... 10
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | ii
Chapter 1.
FEATURES FOR PLATFORMS AND
SOFTWARE
Table 1 List of supported features per platform.
Linux Linux
Linux x86-64 Windows x64 ppc64le AArch64 QNX AArch64
Supported ‣ 10.2 ‣ 10.1 10.1 ‣ 10.2 10.2
CUDA versions
‣ 10.1 ‣ 10.0 ‣ 10.0
‣ 10.0 ‣ 9.0
‣ 9.0
Supported 7.6.3 7.6.3 7.6.3 7.6.3 7.6.3
cuDNN versions
1
TensorRT Yes No Yes Yes No
Python API
NvUffParser Yes Yes Yes Yes Yes
NvOnnxParser Yes Yes Yes Yes Yes
Serialized engines are not portable across platforms or TensorRT versions.
1
Python is not supported on automotive platforms.
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 1
Chapter 2.
LAYERS AND FEATURES
Table 2 List of supported features per TensorRT layer.
Does the
operation
apply to Supports
Dimensions Dimensions only the Supports broadcast
of input of output innermost 3 broadcast across batch
Layer tensor tensor dimensions? (see Note 1) (see Note 2)
IActivationLayer 0-7 dimensions 0-7 dimensions No No No
IConcatenationLayer
1-7 dimensions 1-7 dimensions No No No
IConstantLayer 0-7 dimensions 0-7 dimensions No No Always
IConvolutionLayer3 or more 3 or more Yes No No
> 2D dimensions dimensions
Convolution
IConvolutionLayer4 or more 4 or more No No No
> 3D dimensions dimensions
Convolution
IDeconvolutionLayer
3 or more 3 or more Yes No No
> 2D dimensions dimensions
Deconvolution
IDeconvolutionLayer
4 or more 4 or more No No No
> 3D dimensions dimensions
Deconvolution
IElementWiseLayer0-7 dimensions 0-7 dimensions No Yes Yes
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 2
Layers And Features
Does the
operation
apply to Supports
Dimensions Dimensions only the Supports broadcast
of input of output innermost 3 broadcast across batch
Layer tensor tensor dimensions? (see Note 1) (see Note 2)
IFullyConnectedLayer
3 or more 3 or more Yes No No
dimensions dimensions
IGatherLayer ‣ Input1: 0-7 dimensions No No Yes
1-7
dimensions
‣ Input2:
0-7
dimensions
IIdentityLayer 0-7 dimensions 0-7 dimensions No No No
ILRNLayer 3 or more 3 or more Yes No No
dimensions dimensions
IMatrixMultiplyLayer
2 or more 2 or more No Yes Yes
dimensions dimensions
IPaddingLayer 3 or more 3 or more Yes No No
dimensions dimensions
IPluginLayer User defined User defined User defined User defined User defined
IPluginV2Layer User defined User defined User defined User defined User defined
IPoolingLayer > 3 or more 3 or more Yes Yes Yes
2D Pooling dimensions dimensions
IPoolingLayer > 4 or more 4 or more No Yes Yes
3D Pooling dimensions dimensions
IRaggedSoftMaxLayer
‣ Input: 2 2 or more No No Yes
dimensions dimensions
‣ Bounds: 2
dimensions
IReduceLayer 1-7 dimensions 0-7 dimensions No No No
IResizeLayer 1-7 dimensions 1-7 dimensions No No No
IRNNLayer 3 dimensions 3 dimensions No No No
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 3
Layers And Features
Does the
operation
apply to Supports
Dimensions Dimensions only the Supports broadcast
of input of output innermost 3 broadcast across batch
Layer tensor tensor dimensions? (see Note 1) (see Note 2)
IRNNv2Layer ‣ Data/ Data/Hidden/ No No No
Hidden/ Cell: 2 or more
Cell: 2 dimensions
or more
dimensions
‣ Seqlen: 0
or more
dimensions
IScaleLayer 3 or more 3 or more Yes No No
dimensions dimensions
IShapeLayer 1 or more 1 dimension No No No
dimensions
IShuffleLayer 0-7 dimensions 0-7 dimensions No No No
ISliceLayer 1-7 dimensions 1-7 dimensions No No Yes
ISoftMaxLayer 1-7 dimensions 1-7 dimensions No No Yes
ITopKLayer 1-7 dimensions ‣ Output1: Yes No Yes
1-7
dimensions
‣ Output2:
1-7
dimensions
IUnaryLayer 0-7 dimensions 0-7 dimensions No No No
1. Indicates support for broadcast in this layer. This layer allows its two input
tensors to be of dimensions [1, 5, 4, 3] and [1, 5, 1, 1], and its output out be [1,
5, 4, 3]. Note: The second input tensor has been broadcast in the innermost 2
dimensions.
2. Indicates support for broadcast across the batch dimension.
For more information about each of the TensorRT layers, see TensorRT Layers.
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 4
Chapter 3.
LAYERS AND PRECISION
The following table lists the TensorRT layers and the precision modes that each layer
supports. It also lists the ability of the layer to run on Deep Learning Accelerator (DLA).
For more information about additional constraints, see DLA Supported Layers.
For more information about each of the TensorRT layers, see TensorRT Layers. To view a
list of the specific attributes that are supported by each layer, refer to the TensorRT API
documentation.
Table 3 List of supported precision mode per TensorRT layer.
Layer FP32 FP16 INT8 INT32 DLA FP16 DLA INT8
2 3
IActivationLayerYes Yes Yes No Yes Yes
4 3
IConcatenationLayer
Yes Yes Yes Yes Yes Yes
IConstantLayer Yes Yes Yes Yes No No
5
IConvolutionLayer
Yes Yes Yes No Yes Yes
> 2D
Convolution
IConvolutionLayer
Yes Yes No No No No
> 3D
Convolution
6
IDeconvolutionLayer
Yes Yes Yes No Yes Yes
> 2D
Deconvolution
2
Partial support. Yes for ReLU, sigmoid and TanH activation types only.
3
Partial support. Yes for ReLU activation type only.
4
Partial support. Yes for concatenation across C dimension only.
5
Partial support. Yes for ungrouped convolutions and No for grouped.
6
Partial support. Yes for ungrouped deconvolutions and No for grouped.
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 5
Layers And Precision
Layer FP32 FP16 INT8 INT32 DLA FP16 DLA INT8
IDeconvolutionLayer
Yes Yes No No No No
> 3D
Deconvolution
7 8
IElementWiseLayer
Yes Yes No Yes Yes Yes
IFullyConnectedLayer
Yes Yes Yes No Yes Yes
IGatherLayer Yes Yes No Yes No No
IIdentityLayer Yes Yes Yes Yes No No
IPluginV2Layer Yes Yes Yes No No No
ILRNLayer Yes Yes Yes No Yes No
IMatrixMultiplyLayer
Yes Yes No No No No
IPaddingLayer Yes Yes Yes No No No
IPluginLayer Yes Yes No No No No
9 8
IPoolingLayer Yes Yes Yes No Yes Yes
> 2D Pooling
IPoolingLayer Yes Yes No No No No
> 3D Pooling
IRaggedSoftMaxLayer
Yes No No No No No
IReduceLayer Yes Yes No No No No
IResizeLayer Yes Yes No No No No
IRNNLayer Yes Yes No No No No
IRNNv2Layer Yes Yes No No No No
10 9
IScaleLayer Yes Yes Yes No Yes Yes
11
IShapeLayer Yes Yes Yes Yes No No
IShuffleLayer Yes Yes Yes Yes No No
12
ISliceLayer Yes Yes No Yes No No
ISoftMaxLayer Yes Yes No No No No
7
Partial support. Yes for sum, sub, prod, min and max elementwise operations only.
8
Partial support. Yes for sum elementwise operation only.
9
Partial support. Yes for max and average pooling type only.
10
Partial support. DLA does not support power on scale layer.
11
Output is always INT32.
12
Partial support. Yes for unstrided Slice and No for strided.
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 6
Layers And Precision
Layer FP32 FP16 INT8 INT32 DLA FP16 DLA INT8
ITopKLayer Yes Yes No No No No
IUnaryLayer Yes Yes No No No No
DLA with FP16/INT8 precision with some restrictions on layer parameters.
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 7
Chapter 4.
HARDWARE AND PRECISION
The following table lists NVIDIA hardware and which precision modes each hardware
supports. It also lists availability of Deep Learning Accelerator (DLA) on these
hardware. TensorRT supports all NVIDIA hardware with capability SM 3.0 or higher.
Table 4 List of supported precision mode per hardware.
CUDA FP16 INT8
Compute Example Tensor Tensor
Capability Device FP32 FP16 INT8 Cores Cores DLA
7.5 Tesla T4 Yes Yes Yes Yes Yes No
7.2 Jetson AGX Yes Yes Yes Yes Yes Yes
Xavier
7.0 Tesla V100 Yes Yes Yes Yes No No
6.2 Jetson TX2 Yes Yes No No No No
6.1 Tesla P4 Yes No Yes No No No
6.0 Tesla P100 Yes Yes No No No No
5.3 Jetson TX1 Yes Yes No No No No
5.2 Tesla M4 Yes No No No No No
5.0 Quadro Yes No No No No No
K2200
3.7 Tesla K80 Yes No No No No No
3.5 Tesla K40 Yes No No No No No
3.0 Tesla K10 Yes No No No No No
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 8
Chapter 5.
SOFTWARE VERSIONS PER PLATFORM
Table 5 List of supported platforms per software version.
Compiler version Python version
Ubuntu 14.04 x86-64 gcc 4.8.4 2.7, 3.4
Ubuntu 16.04 x86-64 gcc 5.4.0 2.7, 3.5
Ubuntu 18.04 x86-64 gcc 7.4.0 2.7, 3.6
CentOS 7.5 x86-64 gcc 4.8.5 2.7, 3.6
Windows 10 x64 CUDA 10.0, 10.1
MSVC 2017u5
CUDA 9.0
MSVC 2017u3
Ubuntu 18.04 ppc64le gcc 7.4.0 2.7, 3.6
CentOS 7.5 ppc64le gcc 4.8.5 2.7, 3.6
Ubuntu 18.04 AArch64 gcc 7.4.0 2.7, 3.6
QNX AArch64 gcc 5.4.0
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 9
Chapter 6.
SUPPORTED OPS
The following lists describe the operations that are supported in a Caffe or TensorFlow
framework and in the ONNX TensorRT parser:
Caffe
These are the operations that are supported in a Caffe framework:
‣ BatchNormalization
‣ BNLL
‣ Clip13
‣ Concatenation
‣ Convolution
‣ Crop
‣ Deconvolution
‣ Dropout
‣ ElementWise
‣ ELU
‣ InnerProduct
‣ Input
‣ LeakyReLU
‣ LRN
‣ Permute
‣ Pooling
‣ Power
‣ Reduction
‣ ReLU, TanH, and Sigmoid
‣ Reshape
13
When using the Clip operation, Caffe users must serialize their layers using ditcaffe.pb.h instead of caffe.pb.h in
order to import the layer into TensorRT.
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 10
Supported Ops
‣ SoftMax
‣ Scale
TensorFlow
These are the operations that are supported in a TensorFlow framework:
‣ Add, Sub, Mul, Div, Minimum and Maximum
‣ ArgMax
‣ ArgMin
‣ AvgPool
‣ BiasAdd
‣ Clip
‣ ConcatV2
‣ Const
‣ Conv2D
‣ ConvTranspose2D
‣ DepthwiseConv2dNative
‣ Elu
‣ ExpandDims
‣ FusedBatchNorm
‣ Identity
‣ LeakyReLU
‣ MaxPool
‣ Mean
‣ Negative, Abs, Sqrt, Recip, Rsqrt, Pow, Exp and Log
‣ Pad is supported if followed by one of these TensorFlow layers: Conv2D,
DepthwiseConv2dNative, MaxPool, and AvgPool.
‣ Placeholder
‣ ReLU, TanH, and Sigmoid
‣ Relu6
‣ Reshape
‣ ResizeBilinear, ResizeNearestNeighbor
‣ Sin, Cos, Tan, Asin, Acos, Atan, Sinh, Cosh, Asinh, Acosh, Atanh,
Ceil and Floor
‣ Selu
‣ Slice
‣ SoftMax
If the input to a TensorFlow SoftMax op is not NHWC, TensorFlow will
automatically insert a transpose layer with a non-constant permutation, causing
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 11
Supported Ops
the UFF converter to fail. It is therefore advisable to manually transpose SoftMax
inputs to NHWC using a constant permutation.
‣ Softplus
‣ Softsign
‣ Transpose
ONNX
Since the ONNX parser is an open source project, the most up-to-date information
regarding the supported operations can be found in GitHub: ONNX TensorRT.
These are the operations that are supported in the ONNX framework:
‣ Abs
‣ Add
‣ ArgMax
‣ ArgMin
‣ AveragePool
‣ BatchNormalization
‣ Cast
‣ Ceil
‣ Clip
‣ Concat
‣ Constant
‣ Conv
‣ ConvTranspose
‣ DepthToSpace
‣ Div
‣ Dropout
‣ Elu
‣ Exp
‣ Flatten
‣ Floor
‣ Gather
‣ Gemm
‣ GlobalAveragePool
‣ GlobalMaxPool
‣ HardSigmoid
‣ Identity
‣ ImageScaler
‣ InstanceNormalization
‣ LRN
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 12
Supported Ops
‣ LeakyRelU
‣ Log
‣ LogSoftmax
‣ MatMul
‣ Max
‣ MaxPool
‣ Mean
‣ Min
‣ Mul
‣ Neg
‣ Pad
‣ ParametricSoftplus
‣ Pow
‣ Reciprocal
‣ ReduceL1
‣ ReduceL2
‣ ReduceLogSum
‣ ReduceLogSumExp
‣ ReduceMax
‣ ReduceMean
‣ ReduceMin
‣ ReduceProd
‣ ReduceSum
‣ ReduceSumSquare
‣ Relu
‣ Reshape
‣ Resize
‣ ScaledTanh
‣ Selu
‣ Shape
‣ Sigmoid
‣ Sin, Cos, Tan, Asin, Acos, Atan, Sinh, Cosh, Asinh, Acosh, and
Atanh
‣ Size
‣ Slice
‣ Softmax
‣ Softplus
‣ Softsign
‣ SpaceToDepth
‣ Split
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 13
Supported Ops
‣ Squeeze
‣ Sub
‣ Sum
‣ Tanh
‣ ThresholdedRelu
‣ TopK
‣ Transpose
‣ Unsqueeze
‣ Upsample
www.nvidia.com
Support Matrix For TensorRT SWE-SWDOCTRT-001-SPMT _vTensorRT 6.0.1 | 14
Notice
THE INFORMATION IN THIS GUIDE AND ALL OTHER INFORMATION CONTAINED IN NVIDIA DOCUMENTATION
REFERENCED IN THIS GUIDE IS PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED,
STATUTORY, OR OTHERWISE WITH RESPECT TO THE INFORMATION FOR THE PRODUCT, AND EXPRESSLY
DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A
PARTICULAR PURPOSE. Notwithstanding any damages that customer might incur for any reason whatsoever,
NVIDIA’s aggregate and cumulative liability towards customer for the product described in this guide shall
be limited in accordance with the NVIDIA terms and conditions of sale for the product.
THE NVIDIA PRODUCT DESCRIBED IN THIS GUIDE IS NOT FAULT TOLERANT AND IS NOT DESIGNED,
MANUFACTURED OR INTENDED FOR USE IN CONNECTION WITH THE DESIGN, CONSTRUCTION, MAINTENANCE,
AND/OR OPERATION OF ANY SYSTEM WHERE THE USE OR A FAILURE OF SUCH SYSTEM COULD RESULT IN A
SITUATION THAT THREATENS THE SAFETY OF HUMAN LIFE OR SEVERE PHYSICAL HARM OR PROPERTY DAMAGE
(INCLUDING, FOR EXAMPLE, USE IN CONNECTION WITH ANY NUCLEAR, AVIONICS, LIFE SUPPORT OR OTHER
LIFE CRITICAL APPLICATION). NVIDIA EXPRESSLY DISCLAIMS ANY EXPRESS OR IMPLIED WARRANTY OF FITNESS
FOR SUCH HIGH RISK USES. NVIDIA SHALL NOT BE LIABLE TO CUSTOMER OR ANY THIRD PARTY, IN WHOLE OR
IN PART, FOR ANY CLAIMS OR DAMAGES ARISING FROM SUCH HIGH RISK USES.
NVIDIA makes no representation or warranty that the product described in this guide will be suitable for
any specified use without further testing or modification. Testing of all parameters of each product is not
necessarily performed by NVIDIA. It is customer’s sole responsibility to ensure the product is suitable and
fit for the application planned by customer and to do the necessary testing for the application in order
to avoid a default of the application or the product. Weaknesses in customer’s product designs may affect
the quality and reliability of the NVIDIA product and may result in additional or different conditions and/
or requirements beyond those contained in this guide. NVIDIA does not accept any liability related to any
default, damage, costs or problem which may be based on or attributable to: (i) the use of the NVIDIA
product in any manner that is contrary to this guide, or (ii) customer product designs.
Other than the right for customer to use the information in this guide with the product, no other license,
either expressed or implied, is hereby granted by NVIDIA under this guide. Reproduction of information
in this guide is permissible only if reproduction is approved by NVIDIA in writing, is reproduced without
alteration, and is accompanied by all associated conditions, limitations, and notices.
Trademarks
NVIDIA, the NVIDIA logo, and cuBLAS, CUDA, cuDNN, cuFFT, cuSPARSE, DALI, DIGITS, DGX, DGX-1, Jetson,
Kepler, NVIDIA Maxwell, NCCL, NVLink, Pascal, Tegra, TensorRT, and Tesla are trademarks and/or registered
trademarks of NVIDIA Corporation in the Unites States and other countries. Other company and product
names may be trademarks of the respective companies with which they are associated.
Copyright
© 2019 NVIDIA Corporation. All rights reserved.
www.nvidia.com
