Abstract
NVIDIA TensorRT is a C++ library that facilitates high performance inference on NVIDIA GPUs. It is designed to work in connection with deep learning frameworks that are commonly used for training. TensorRT focuses specifically on running an already trained network quickly and efficiently on a GPU for the purpose of generating a result; also known as inferencing. These release notes describe the key features, software enhancements and improvements, and known issues for the TensorRT 8.5.10 product package.
1. TensorRT for DRIVE OS
1.1. DRIVE OS Linux "Standard"
1.2. DRIVE OS QNX "Standard"
DRIVE OS QNX for Safety
The safety package is available in the NVIDIA DRIVE OS 6.0.6.0 release. The safety package for NVIDIA DRIVE OS users of TensorRT, which is only available on QNX safety, contains the safety runtime, safety headers only, and the API documentation specific to the safety runtime.
1.3. DRIVE OS for Safety Proxy
- The TensorRT proxy runtime is a version of the safety runtime for platforms that are not safety certified. This includes NVIDIA DRIVE OS x86 SDK, NVIDIA DRIVE OS Linux SDK, NVIDIA DRIVE OS Linux PDK, NVIDIA DRIVE OS QNX SDK and NVIDIA DRIVE OS QNX PDK. The proxy runtime is part of the development flow for safety but it is not certified itself. The proxy runtime only supports engines with engine capability kSAFETY (safe engines).
- Headers allow applications to compile against the proxy runtime and the safety runtime.
- The safety runtime is also a library that allows applications to load serialized engine plans and perform inference. It is only available for QNX safety. The safety runtime only supports engines with engine capability kSAFETY (safe engines).
2. Release Highlights
2.1. Deprecations in this Release
| Summary | Impact |
|---|---|
| Two DLA Safety samples, dlaSafetyBuilder and dlaSafetyRuntime have been removed from the TensorRT 8.5.10 release. |
Module: TensorRT samples Action: Since TensorRT safety runtime does not support DLA, this removal does not have any impact for TensorRT safety users. You can check TensorRT standard runtime for the DLA usage. |
| ICudaEngine::getNbBindings |
Module: TensorRT runtime Action: Since the TensorRT standard runtime was updated to enqueueV3() in the TensorRT 8.5.1 release, the enqueueV3() in the TensorRT safety runtime reduces the API changes when migrating from the standard runtime to the safety runtime. Name-based functions have been added to safe::ICudaEngine. |
| ICudaEngine::getBindingIndex | |
| ICudaEngine::getBindingName | |
| ICudaEngine::bindingIsInput | |
| ICudaEngine::getBindingDimensions | |
| ICudaEngine::getBindingDataType | |
| ICudaEngine::getBindingBytesPerComponent | |
| ICudaEngine::getBindingComponentsPerElement | |
| ICudaEngine::getBindingFormat | |
| ICudaEngine::getBindingVectorizedDim | |
| IExecutionContext::getStrides |
Module: TensorRT safety execution context |
| IExecutionContext::enqueueV2 |
2.2. Planned Upcoming Changes
IGatherLayer Support
The TensorRT safety and proxy runtimes will add support for IGatherLayer in a future release. This is to complement the IGatherLayer functionality in the TensorRT standard runtime. IGatherLayer support in the TensorRT safety and proxy runtimes will be designed to support the Gather mode for the first DRIVE OS 6.0 safety cycle. Refer to the NVIDIA TensorRT 8.5.10 API Reference for DRIVE OS or the NVIDIA TensorRT Operator’s Reference documentation to get more information and limitations.
IMatrixMultiplyLayer Support
TensorRT will add support for IMatrixMultiplyLayer in a future release. This will allow users of TensorRT safety to use the ONNX GEMM or MatMul operators as inputs to the TensorRT builder. Refer to the NVIDIA TensorRT 8.5.10 API Reference for DRIVE OS or the NVIDIA TensorRT Operator’s Reference documentation to get more information and limitations.
Safe Plugin Registry Interface Updates
- nvinfer1::IPluginRegistry in NvInferRuntime.h, which will be used in TensorRT standard.
- nvinfer1::safe::IPluginRegistry in NvInferSafeRuntime.h, which will be used in TensorRT safety.
- The explicit use of nvinfer1::IPluginRegistry must be replaced by nvinfer1::safe::IPluginRegistry.
- Instead of directly including the NvInferRuntimeCommon.h header, the user code must include NvInferSafeRuntime.h instead.
- The use of getBuilderPluginRegistry() should be replaced by getBuilderSafePluginRegistry().
3. New Features and Enhancements
API Changes
| Interface | Impact |
|---|---|
| ICudaEngine::getTensorShape |
Affected: Binding index-based functions have been deprecated. Name-based functions have been added. Action: Refer to the NVIDIA TensorRT 8.5.10 API Reference for DRIVE OS. Use enqueueV3() for asynchronous inference execution. Use name-based functions. |
| ICudaEngine::getTensorDataType | |
| ICudaEngine::getTensorIOMode | |
| ICudaEngine::getTensorBytesPerComponent | |
| ICudaEngine::getTensorComponentsPerElement | |
| ICudaEngine::getTensorVectorizedDim | |
| ICudaEngine::getNbIOTensors | |
| ICudaEngine::getIOTensorName | |
| IExecutionContext::getTensorStrides | |
| IExecutionContext::setInputTensorAddress | |
| IExecutionContext::setOutputTensorAddress | |
| IExecutionContext::setInputConsumedEvent | |
| IExecutionContext::getInputConsumedEvent | |
| IExecutionContext::getInputTensorAddress | |
| IExecutionContext::getOutputTensorAddress | |
| IExecutionContext::enqueueV3 |
TensorRT Standard Build
The TensorRT 8.5 release includes changes to the standard builder and runtime that appear in TensorRT for DRIVE OS 6.0. For more information, refer to the TensorRT 8.5.1 Release Notes.
Documentation Changes
- The NVIDIA TensorRT 8.5.10 Developer Guide for DRIVE OS is based on the enterprise TensorRT 8.5.x release. We have modified the TensorRT 8.5.x Developer Guide documentation for DRIVE OS 6.0.6 accuracy. The TensorRT safety content has been removed.
- The TensorRT safety content is in the NVIDIA TensorRT 8.5.10 Safety Developer Guide Supplement for DRIVE OS. Refer to this PDF for all TensorRT safety specific documentation.
Safety Samples Update
New safety samples have been added to TensorRT 8.5.10. These samples focus on TensorRT safety features and functionality. Refer to the NVIDIA TensorRT 8.5.10 Safety Developer Guide Supplement for DRIVE OS for more information.
You can find the safety samples in the /usr/src/tensorrt/samples package directory. For more information on running samples, refer to the README.md file included with the sample.
Two DLA Safety samples, dlaSafetyBuilder and dlaSafetyRuntime have been removed from the TensorRT 8.5.10 release.
ISliceLayer Support
The TensorRT 8.5.10 release supports a new layer called ISliceLayer, which enables the slice operation of input tensors along all axes. Slices must be FP32, FP16, or INT8 precision and must meet the requirements listed in the NVIDIA TensorRT 8.5.10 Developer Guide for DRIVE OS. Refer to ONNX slice Op definition and the NVIDIA TensorRT 8.5.10 API Reference for DRIVE OS documentation for more information.
Equal Operation Support in IElementWiseLayer
The TensorRT 8.5.10 release extends IElementWiseLayer to support equal operation (ElementWiseOperation::kEQUAL). This is the first ElementWise logical operation that DLA supports, so there are several restrictions and requirements imposed when adopting this operation in DLA.
One such requirement is that you must explicitly set the device type of the ElementWise equal layer to DLA. trtexec now supports a flag --layerDeviceTypes to let you explicitly specify the device type for individual layers. Refer to the NVIDIA TensorRT 8.5.10 Developer Guide for DRIVE OS documentation for more information on the above changes.
Opportunistic TensorRT Safe Engine Version Forward Compatibility
Opportunistic TensorRT Safe Engine Version Forward Compatibility allows users to run safe engines generated by some older TensorRT versions with the current TensorRT safety runtime under specific conditions. This is only possible when the safety runtime does not change materially between releases, which would generally be limited to the safety and stabilization phase of a safety cycle leading up to safety assessment.
While we strive to ensure safe engine version forward compatibility opportunistically, safe engines generated from previous TensorRT versions are not forward compatible with the current TensorRT 8.5.10 safety runtime due to material changes in the runtime from new functionality. Similarly, safe engines generated from the current release will not be forward compatible with the TensorRT 8.6.10 runtime. Absent bugs and safety-related refactoring that would force us to do otherwise, our goal is to support safe engines generated from TensorRT 8.6.10+ for usage in later releases throughout the remainder of TensorRT 8 development.
Refer to the NVIDIA TensorRT 8.5.10 Safety Developer Guide Supplement for DRIVE OS and the Release Notes of each release for the supported forward compatible safe engine versions and limitations.
enqueueV3()
The TensorRT 8.5.10 release added a new function called enqueueV3() to support asynchronous inference execution. enqueueV3() adds support for constant input buffers, which is a safety requirement. Since the TensorRT standard runtime was updated to enqueueV3() in the TensorRT 8.5.1 release, the enqueueV3() in the TensorRT safety runtime reduces the API changes when migrating from the standard runtime to the safety runtime. Binding index-based functions have been deprecated and name-based functions have been added to safe::ICudaEngine. For more information regarding API changes, refer to the NVIDIA TensorRT 8.5.10 API Reference for DRIVE OS.
TensorRT Consistency Checker
The TensorRT 8.5.10 release of the consistency checker performs most checks to ensure that engines can be run in the safety runtime without invoking undefined or nondeterministic behavior. Operations within the safety scope are checked, tensor sizes and formats are checked, and inputs to each layer are analyzed to ensure no uninitialized values are read from memory. Some tactics require specialized kernels and internal data structures. Most, but not all, of these internal data structures are validated in the release.
4. Fixed Issues
| Feature | Module | Description |
|---|---|---|
| 3785919 | TensorRT safety package | When installing files from Debians on the same system, some files installed by NVIDIA DRIVE OS QNX safety TensorRT Debian would be in the same location as the NVIDIA DRIVE OS QNX proxy TensorRT Debian. This limitation has been fixed in this release. |
| 3698033 | DLA | Some networks would fail to build DLA INT8 loadable in DLA_STANDALONE mode with INT8 calibrator. This bug has been fixed in this release. |
| 3698054 | TensorRT builder |
In some cases, the TensorRT builder would allow input and output tensors in HWC16 format in FP16 precision. This format is outside the safety scope. HWC16 format has been removed from available formats in this release. |
| 3657753 | TensorRT runtime | There would sometimes be issues with large channel sizes with structured sparsity convolution kernels (seen at size 4096). This bug has been fixed in this release. |
| 3689094 | TensorRT builder | TensorRT would take some dense weights as sparse, if they match some special pattern. This bug has been fixed in this release. |
| 3448473 | TensorRT builder | The DLA compilation process in NVIDIA DRIVE OS 6.0.5.0 had a deep recursive call which required a lot of stack memory. On QNX, this may have exceeded the available stack space, leading to memory faults. This bug has been fixed in this release. |
5. Known Limitations
6. Known Issues
| Feature | Module | Description |
|---|---|---|
| 3494734 | DLA |
What is the issue? Some networks may produce incorrect outputs when run on DLA with large batch sizes. How does it impact the customer? Running networks on DLA with batch sizes larger than 32 may produce incorrect outputs. If there is a workaround, what is it? To work around this issue, use a batch size smaller than 32. When can we expect the fix? The issue will be fixed in a future DLA release. Is it for Standard/Safety, SDK/PDK? Standard, SDK |
| 3656116 | TensorRT runtime |
What is the issue? There is an up to 7% performance regression for the 3D-UNet networks compared to TensorRT 8.4 EA when running in INT8 precision on NVIDIA Orin due to a functionality fix. How does it impact the customer? When running 3D-UNet networks in INT8 precision, the latency will be up to 7% longer than in TensorRT 8.4 EA. If there is a workaround, what is it? To work around this issue, set the input type and format to kINT8 and kCHW32, respectively. When can we expect the fix? We do not plan to fix this performance regression since it was caused by a necessary fix for an accuracy issue. Is it for Standard/Safety, SDK/PDK? Standard, SDK |
| 3263411 | TensorRT builder |
What is the issue? For some networks, building and running an engine in the standard runtime will have better performance than the safety runtime. This can be due to various limitations in scope of the safety runtime including more limited tactics, tensor size limits, and operations supported in the safety scope. How does it impact the customer? Inference in the safety runtime may be significantly slower than in the standard runtime. If there is a workaround, what is it? Depending on the network, it may or may not be possible to reorganize operations into a more efficient form matching the safety runtime scope. What is the recommendation? It is recommended to work with NVIDIA and provide proxy networks as early as possible that demonstrate key performance metrics close to actual production networks. Future releases will target performance improvements for networks within the safety scope. Is it for Standard/Safety, SDK/PDK? Standard, SDK |
| 3827883 | Samples |
What is the issue? The trtexec binary shipped with TensorRT has an unnecessary dependency on deprecated NVMedia libraries. How does it impact the customer? The binary will not be usable if the deprecated NVMedia libraries are missing. If there is a workaround, what is it? Building trtexec from source will result in a binary without the extra dependency. Refer to the samples README for details on how to do so. When can we expect the fix? This issue is not expected to be fixed in a future release. Is it for Standard/Safety, SDK/PDK? Safety, PDK |
7. TensorRT Release Properties
| Linux x86-64 | Linux AArch64 | QNX AArch64 | ||
|---|---|---|---|---|
| QNX Safety | QNX Standard | |||
| Supported NVIDIA CUDA? versions | 11.4 | 11.4 | 11.4 | 11.4 |
| Supported NVIDIA cuDNN versions | 8.6.0 | 8.6.0 | No | 8.6.0 |
| TensorRT Python API | Yes | Yes | No | No |
| NvUffParser | Deprecated | Deprecated | No | Deprecated |
| NvOnnxParser | Yes | Yes | No | Yes |
7.1. Hardware Precision
| CUDA Compute Capability | Example Device | TF32 | FP32 | FP16 | INT8 | FP16 Tensor Cores | INT8 Tensor Cores | DLA |
|---|---|---|---|---|---|---|---|---|
| 8.7 | NVIDIA Orin |
No (TensorRT safe) Yes (TensorRT standard) |
Yes | Yes | Yes | Yes | Yes | Yes |
| 8.6 | NVIDIA A10 | Yes | Yes | Yes | Yes | Yes | Yes | No |
| 8.0 | NVIDIA PG199 | Yes | Yes | Yes | Yes | Yes | Yes | No |
7.2. Software Versions Per Platform
| Platform | Compiler Version | Python Version |
|---|---|---|
| Ubuntu 20.04 x86-64 | gcc 9.3.0 | 3.8 |
| Ubuntu 20.04 AArch64 | gcc 9.3.0 | 3.8 |
| QNX AArch64 | QNX 7.1.0 Q++ 8.3.0 | N/A |
7.3. Compatibility
- CUDA 11.4.20
- cuDNN 8.6.0
- TensorFlow 1.15.0
- PyTorch 1.9.0
- ONNX 1.9.0 and opset 13
- DLA 3.12
- ElementWise 2.5.0
Notice
Notice
This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA Corporation (“NVIDIA”) makes no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this document and assumes no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material (defined below), code, or functionality.
NVIDIA reserves the right to make corrections, modifications, enhancements, improvements, and any other changes to this document, at any time without notice.
Customer should obtain the latest relevant information before placing orders and should verify that such information is current and complete.
NVIDIA products are sold subject to the NVIDIA standard terms and conditions of sale supplied at the time of order acknowledgement, unless otherwise agreed in an individual sales agreement signed by authorized representatives of NVIDIA and customer (“Terms of Sale”). NVIDIA hereby expressly objects to applying any customer general terms and conditions with regards to the purchase of the NVIDIA product referenced in this document. No contractual obligations are formed either directly or indirectly by this document.
NVIDIA products are not designed, authorized, or warranted to be suitable for use in medical, military, aircraft, space, or life support equipment, nor in applications where failure or malfunction of the NVIDIA product can reasonably be expected to result in personal injury, death, or property or environmental damage. NVIDIA accepts no liability for inclusion and/or use of NVIDIA products in such equipment or applications and therefore such inclusion and/or use is at customer’s own risk.
NVIDIA makes no representation or warranty that products based on this document will be suitable for any specified use. Testing of all parameters of each product is not necessarily performed by NVIDIA. It is customer’s sole responsibility to evaluate and determine the applicability of any information contained in this document, ensure the product is suitable and fit for the application planned by customer, and perform 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 document. NVIDIA accepts no 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 document or (ii) customer product designs.
No license, either expressed or implied, is granted under any NVIDIA patent right, copyright, or other NVIDIA intellectual property right under this document. Information published by NVIDIA regarding third-party products or services does not constitute a license from NVIDIA to use such products or services or a warranty or endorsement thereof. Use of such information may require a license from a third party under the patents or other intellectual property rights of the third party, or a license from NVIDIA under the patents or other intellectual property rights of NVIDIA.
Reproduction of information in this document is permissible only if approved in advance by NVIDIA in writing, reproduced without alteration and in full compliance with all applicable export laws and regulations, and accompanied by all associated conditions, limitations, and notices.
THIS DOCUMENT AND ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. TO THE EXTENT NOT PROHIBITED BY LAW, IN NO EVENT WILL NVIDIA BE LIABLE FOR ANY DAMAGES, INCLUDING WITHOUT LIMITATION ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, PUNITIVE, OR CONSEQUENTIAL DAMAGES, HOWEVER CAUSED AND REGARDLESS OF THE THEORY OF LIABILITY, ARISING OUT OF ANY USE OF THIS DOCUMENT, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. Notwithstanding any damages that customer might incur for any reason whatsoever, NVIDIA’s aggregate and cumulative liability towards customer for the products described herein shall be limited in accordance with the Terms of Sale for the product.
Arm
Arm, AMBA and Arm Powered are registered trademarks of Arm Limited. Cortex, MPCore and Mali are trademarks of Arm Limited. "Arm" is used to represent Arm Holdings plc; its operating company Arm Limited; and the regional subsidiaries Arm Inc.; Arm KK; Arm Korea Limited.; Arm Taiwan Limited; Arm France SAS; Arm Consulting (Shanghai) Co. Ltd.; Arm Germany GmbH; Arm Embedded Technologies Pvt. Ltd.; Arm Norway, AS and Arm Sweden AB.
HDMI
HDMI, the HDMI logo, and High-Definition Multimedia Interface are trademarks or registered trademarks of HDMI Licensing LLC.
Blackberry/QNX
Copyright ? 2020 BlackBerry Limited. All rights reserved.
Trademarks, including but not limited to BLACKBERRY, EMBLEM Design, QNX, AVIAGE, MOMENTICS, NEUTRINO and QNX CAR are the trademarks or registered trademarks of BlackBerry Limited, used under license, and the exclusive rights to such trademarks are expressly reserved.
Trademarks
NVIDIA, the NVIDIA logo, and CUDA, DALI, DGX-1, DRIVE, JetPack, Orin, Pegasus, TensorRT, Triton and Xavier are trademarks and/or registered trademarks of NVIDIA Corporation in the United States and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.