Jacinto 7 TIDL Release Notes

Version: 08.04.00, Date: Aug 23, 2022

Contents

  1. Introduction
  2. Licensing
  3. What's New
  4. Documentation
  5. Upgrade and Compatibility Information
  6. Device Support
  7. Validation Information
  8. Fixed Issues
  9. Known Issues
  10. Change Request
  11. Technical Support
  12. Package Versioning

Introduction

This package contains Texas Instrument's Deep learning solution for J721E SoCs from the Jacinto 7 family

Licensing

The licensing information of this product, 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]
2 Software Manifest [HTML]

What's New

Here are a few of the new features supported in this release:

  1. Grouped convolution performance improvements
  2. OSRT RTs (TFlite, ONNX runtime) validated with Ubuntu 18.04 and Ubuntu 20.04 Docker containers on EVM

Upgrade and Compatibility Information

This release is interface compatible with previous release. Note : Model's imported/compiled with previous release versions will not work with this release. User is expected to re-import/re-compile models with this release before using it

Device Support

Target SoC Target Plaform Build environment (OS) Target environment (OS)
J721E EVM x86_64 Linux PC x86_64 (Linux) NA

Validation Information

This release was built and validated using the following tools:

Dependencies (included in Processor SDK RTOS):

Dependencies (NOT included in Processor SDK RTOS):

Refer User's Guide for instructions to install and setup above dependencies.

Fixed Issues

ID Description Module Affected Versions Affected Platforms
[TIDL-605] Spatial pooling + inner-product will not generate error TOOL 08.02.00, 08.01.00, 08.00.01, 08.00.00, 02.00.00, 01.04.00, 01.03.00, 01.02.00, 01.01.00, 01.00.00 J721E
[TIDL-708] Network compiler should detect invalid layers that have inW=inH=outW=outH=No=Ni=0 and return an error TOOL 08.02.00, 08.01.00,08.00.01, 08.00.00, 02.00.00, 01.04.00, 01.03.00, 01.02.00, 01.01.00 J721E
[TIDL-864] DNN with Eltwise or Concat layers may have functional issue under some remote cases COMPUTE 08.02.00, 08.01.00,08.00.01, 08.00.00, 02.00.00, 01.04.00, 01.03.00, 01.02.00 J721E
[TIDL-1350] Accuracy of ONNX RT is poor than TVM (NEO AI DLR) for same model for some models COMPUTE 08.02.00, 08.01.00,08.00.01, 08.00.00, 02.00.00 J721E
[TIDL-1816] Model cannot be executed if number of input dimensions of the network/network subgraphs is not equal to 4 TOOL 08.02.00, 08.01.00,08.00.01 J721E
[TIDL-1874] Documentation update : Clearly indicate that post processing output dumped during import pass is not representative of inference output Documentation 08.02.00, 08.01.00 J721E
[TIDL-1893] Auto pad attribute "SAME LOWER" (for convolution and pooling) not imported correctly as part of ONNX import TOOL 08.02.00, 08.01.00 J721E
[TIDL-1915] Convolution layer with pad = 0 may result in a hang/crash COMPUTE 08.02.00, 08.01.00 J721E
[TIDL-2360] SSD Network with more than 8 Priorbox is not working COMPUTE 08.02.00, 08.01.00, 08.00.01 J721E
[TIDL-2385] ONNX Split layer not getting imported correctly TOOL 08.02.00 J721E
[TIDL-2391] ONNX import : Pooling layer with stride = 2 behaves incorrectly when padL < padR or padT < padB TOOL 08.02.00 J721E
[TIDL-2520] Global average pooling misbehaves with 16 bit and output feature map placed in DDR COMPUTE 08.02.00 J721E
[TIDL-2528] Convolution (non fully grouped case) layer may have functional issue with 16 bit during some specific cases COMPUTE 08.02.00, 08.01.00,08.00.01, 08.00.00 J721E
[TIDL-2529] Clip activation merged with EltWise layer does not work as expected COMPUTE 08.02.00 J721E
[TIDL-2556] Custom Layer : DMA abstraction API's datatype is different from documentation COMPUTE 08.02.00 J721E
[TIDL-2559] Global Average pooling with 1x1 spatial input resolution gives incorrect output on EVM COMPUTE 08.02.00 J721E

Known Issues

ID Description Module Reported in Release Affected Platforms Occurrence Workaround in this release
TIDL-1197 Model execution order for a branch is not optimal COMPUTE 01.02.00 J721E Rare No fucntional issue, Only Performance drop would be obsrved
TIDL-1530 Processing on EVM gets stuck in case of "GEMM" layer in onnx model OSRT 02.00.00 J721E Rare set reservedCtrl=1 of TIDL_CreateParams
TIDL-1872 Import failure for deeplabv3 (mobilenetv2 backbone) based OD+SEG Model COMPUTE 08.01.00 J721E Rare
TIDL-2593 OSRT : Softmax layer with non-flattened input gets delegated to C7X-MMA instead of ARM OSRT 08.02.00 J721E Rare Introduce a flatten layer before softmax or set softmax layer in deny_list runtime option

For latest status of the issues and any new issues found post this release, please refer TI External software Incident Report. Issues on this product can be filtered with search query as project = EXT_EP AND issuetype in (Bug, Enhancement) AND Product ~ "TI Deep Learning Library"

Change Request

ID Description Original Fix Version New Fix Version

Technical Support

For technical support, please post your questions on TI E2E Forum for Processors.

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 1995-2021, Texas Instruments Incorporated. All rights reserved.
Trademarks | Privacy policy | Terms of use | Terms of sale