Jacinto 7 TIDL Release Notes
Version: 01.02.00
Contents
- Introduction
- Licensing
- What's New
- Documentation
- Upgrade and Compatibility Information
- Device Support
- Validation Information
- Fixed Issues
- Known Issues
- Technical Support
- Package Versioning
Introduction
The package consists of the Texas Instruments J7 TIDL Software
Licensing
The licensing information of this library, as well as any third-party components included which are made available under a number of other open-source licenses are enumerated as part of the manifest. Note : Below Links would work only after installing the package
Documentation
| 1 | User's Guide | [HTML] |
| 3 | Software Manifest | [HTML] |
What's New
Here are a few of the new features supported in this release:
- Optimized performace for below layers
- Resize layer
- Concatenate layer
- Batch Norm Layer
- Element-wise Sum Layer
- Depth-wise Deconvolution Layer
- Optimized networks with Large feature Map size
- Improved Post Training Quantization
Upgrade and Compatibility Information
L1, L2 and L3 Memory size variables has been removed from the TIDL_CreateParams Structure. Now these Parameters are directly used from the model Please remove the usage of these variables from the application code
Device Support
| Target SoC | Target Plaform | Build environment (OS) | Target environment (OS) |
| J721E EVM | x86_64 Linux PC | x86_64 (Linux) | NA |
| J721E EVM | x86_64 Windows PC | x86_64 Windows | NA |
Validation Information
This release was built and validated using the following tools:
Dependencies (included in Processor SDK RTOS Automotive):
- ivision - 01_00
- MMALIB - 01_02_00_03
- PDK - CORE SDK 7.0
- C7x CGT - ti-cgt-c7000_1.3.0.STS
Dependencies (NOT included in Processor SDK RTOS Automotive):
- Microsoft Visual Studio Verison 14.2
- GCC version 5.4.0 (In Ubuntu 18.04)
- OpenCV: 3.1.0 (for Test bench Only)
- Protobuf: 3.5.1 (for Import tool only)
- flatbuffers : 1.11.0(for Import tool only)
Fixed Issues
| ID | Description | Module | Affected Versions | Affected Platforms |
|---|---|---|---|---|
| TIDL-839 | 2x2 stride 2 deconvolution layer FAILED on host emulation | COMPUTE | 01.01.00 | J721E |
| TIDL-796 | [UG DOC] Improve Documentation on TF SSD model optimizations steps | Common | 01.01.00 | J721E |
| TIDL-808 | Deconvolution layer FAILED on host emulation | COMPUTE | 01.01.00 | J721E |
| TIDL-807 | Convolution FAILED on host emulation REFCOMP | COMPUTE | 01.01.00 | J721E |
| TIDL-878 | Import tool crashes when 1-D max pool is present in the model to import | COMPUTE | 01.01.00 | J721E |
| TIDL-707 | Pooling layer returns error for asymetric average pooling | COMPUTE | 01.00.00 | J721E |
| TIDL-861 | Stats collection tool crashes for more than one frames for given model | TOOL | 01.01.00 | J721E |
| TIDL-744 | 16-bits convolution layer in col flow hangs in DMA wait on EVM | COMPUTE | 01.01.00 | J721E |
| TIDL-822 | Sign of various layers are coming wrong with inDataNorm = 1 | TOOL | 01.01.00 | J721E |
| TIDL-869 | MeanIOU computation becomes wrong for large number of images | COMPUTE | 01.01.00 | J721E |
| TIDL-753 | Resize layer throws an error "DDR alloc failed for resize" | COMPUTE | 01.01.00 | J721E |
| TIDL-742 | [Resize] TIDL degrades the accuracy by 25% as compared to max expected drop of 5% | COMPUTE | 01.01.00, 01.00.01 | J721E |
| TIDL-806 | Convolution Layer got stuck on EVM in release mode | COMPUTE | 01.01.00 | J721E |
| TIDL-701 | [Network Graph Compiler] Convolution weights are taken from MSMC instead of L2 when Ni/No is lower | TOOL | 01.00.00 | J721E |
| TIDL-736 | Update custom layer interface to include sysMemHandle | Common | 01.00.01 | J721E |
| TIDL-799 | Few C7x kernels are not setting state2 correctly | COMPUTE | 01.00.01 | J721E |
| TIDL-738 | [Batchnorm] Update batchnorm kernel to support different output/input padding | COMPUTE | 01.01.00, 01.00.01 | J721E |
| TIDL-743 | Implement workaround for MMALIB-291 - 4x4 deconvolution kernel fails in CI in last call when last call's validColsIn > validColsIn of previous calls | COMPUTE | 01.01.00 | J721E |
| TIDL-833 | [Perfsim] Weight address is assigned a big number when analyzeNetwrokIter is run in a loop | TOOL | 01.00.01 | J721E |
| TIDL-810 | Batchnorm NATC kernel is not working properly for Resnet18 v2 16 bit test case | COMPUTE | 01.00.01 | J721E |
| TIDL-849 | ColumnFlow: Target and Ref mismatch when weight volume is greater than 64k | COMPUTE | 01.01.00 | J721E |
| TIDL-840 | Import tool gives error during algFree for MNv2 4 MP | TOOL | 01.00.01 | J721E |
| TIDL-741 | [Arg_Max] Arg Max fails when feature size is not aligned to 32/64 | COMPUTE | 01.00.01 | J721E |
| TIDL-805 | Object detection labels written out to txt file are wrong | COMPUTE | 01.01.00 | J721E |
| TIDL-845 | Concat layer is not working correctly when output is pushed to DDR | COMPUTE | 01.00.01 | J721E |
| TIDL-831 | SSD(Single Shot Object Detection) model with image size of 300x300 doesn't work. | COMPUTE | 01.01.00 | J721E |
| TIDL-832 | TIDL shouldn't allocate multiple buffers when output goes to DDR | COMPUTE | 01.00.01 | J721E |
| TIDL-734 | Implement workaround for convolution's predicate circularity issue in 16-bits: MMALIB-264 | COMPUTE | 01.00.01 | J721E |
| TIDL-737 | [Network Graph Compiler] Fixing co-execution layers constraint causes few networks to fail | TOOL | 01.00.01 | J721E |
Known Issues
| ID | Description | Module | Reported in Release | Affected Platforms | Occurrence | Workaround in this release |
|---|---|---|---|---|---|---|
| TIDL-510 | Faster RCNN second stage output is not matching with reference | COMPUTE | 00.09.00 | J721E | Always | Second Stage Can be executed as custom layer |
| TIDL-558 | TIDL Test applications don't work with SD card having capacity>4 GB | COMPUTE | 01.00.00 | J721E | Always | User need to format the card and create volume with size less that 4B even the card has large capacity |
| TIDL-705 | Max Pooling Output is not correct with signed input | COMPUTE | 01.00.01 | J721E | Rare | Use Relu to make the input unsigned to Max Pool Layer |
| TIDL-719 | For segmentation networks, dumping of multiple images doesn't always work on EVM | COMPUTE | 01.01.00 | J721E | Rare | User can Validate the application on Linux with DL vision apps |
| TIDL-887 | NC: Returning Error code ERR_HIGHER_INSTANCEMEM with few networks | TOOL | 01.01.00 | J721E | Rare | NA |
Technical Support
For technical support, please post your questions on TI E2E Forum for Automotive ADAS SoCs.
For additional assistance, contact local TI Field Application Engineer
Package Versioning
Each package version is composed of 4 period-delimited numbers - represented here by the letters M, m, p and b [M.m.p.b]. The table below provides a descriptive reference regarding package version numbering.
| Digit | Meaning | Description |
|---|---|---|
| 1 (M=Major) | Major revision | Incremented when the new version is substantially different from the previous For example, a new module added or an existing module's algorithm significantly altered. |
| 2 (m=minor) | Minor revision | Incremented when the new version has changed but not in a major way. For example, some minor changes in the API or feature set. |
| 3 (p=patch) | Patch number | Incremented for all other source code changes. This include any packaging support code. |
| 4 (b=build) | Build number | Incremented for each release delivery to CM. Reset for any change to M, m or p |
Copyright 2018, Texas Instruments Incorporated