• English
  • 3.8.1

    • Release Notes

    Notes Before Use

    Attention

    This version is applicable to J6E/M/P/H/B, it is the official release of the algorithm toolchain, the version properties based on different chip platforms are as follows:

    • For J6E/M Linux platform: Commercial-ready version.

    • For J6E/M QNX platform: Restricted commercial version.

    • For J6P/H Linux platform: Commercial-ready version.

    • For J6B QNX platform: Restricted commercial version.

    Please make sure to read the “J6 Algorithm Toolchain License Agreement” carefully before using the toolchain. If you choose to continue using the toolchain, it means that you have fully understood, comprehended and accepted the “J6 Algorithm Toolchain License Agreement”.

    Release information

    Horizon OpenExplorer 3.8.1(release_date:202603)

    ModuleVersion
    HMCTv2.7.3
    HORIZON_TC_UIv3.5.11
    HORIZON_Plugin_Pytorchv3.2.4
    Horizon_Model_Train_Samplev3.2.6
    HBDK_Compilerv4.9.7
    UCP_Tutorialv3.14.7
    J6EM-Linux-BSPv3.1.0, v3.0.0, v2.1.0, v2.0.0
    J6EM-QNX-BSPv3.0.0
    J6PH-Linux-Matrix-BSPv3.0.0, v2.0.0, v1.0.0
    J6B-QNX-BSPv2.0.0, v1.0.0, v0.9.5

    Main Functional Updates

    1. The PTQ pipeline supports dual int16 input precision configuration for Conv/Gemm/MatMul operators. For details, refer to the introduction in the Configure the Computation Accuracy for Nodes with Certain Type section.

    2. The atan2 operator in the QAT pipeline is modified to a high-precision implementation by default, which only applies to the J6E/M/P/H platforms.

    3. The Python interface get_export_hbir_plugin_version is added to the QAT pipeline, which can be used to obtain the Plugin version information of the generated hbir model.

    4. Add the version method to the hbdk4 Module class to obtain the hbdk4 version number used during HBIR model generation (obtaining the hbdk4 version number of historical HBIR models is not supported).

    5. Optimize the initial memory resource usage of the inference library (UCP). The HBTL operator library is not loaded when inferring models without CPU operators, reducing memory usage.

    6. UCP adds version compatibility check for deployment-side dynamic link libraries.

    7. The hb_verifier tool adds the --remote-root parameter to support explicitly specifying the board-side directory.

    8. The hb_analyzer performance analysis tool is added. For usage instructions, refer to the introduction in the Performance Analysis Tool section.

    9. Officially support Torch version 2.8.

    10. J6B QNX adapts to QoS 804.

    11. The reference algorithms are streamlined and iterated. For the reference algorithm list, refer to the introduction in the Reference Algorithm - Sample Models section.

    12. Optimize the structure of the OE user guide to align with user workflow and improve reading experience.

    Compatibility Statement

    None.

    Use Restrictions

    The inference preemption function for models containing the TopK operator depends on the BSP version, which needs to meet: J6EM-Linux >= v3.1.0, J6PH-Linux-Matrix >= v3.0.0, J6B-QNX >= v2.0.0.

    Deprecated Functions

    None.

    Fix Issues

    1. Fix the issue where the Where operator abnormally falls back to the CPU.

    2. Fix the conflict issue between high consistency and the fp16 mean operator.

    3. Fix the incorrect automatic handling issue of fp16 Layernorm when data overflows.

    4. Fix the error reporting issue of QconfigSetter under certain corner cases.

    5. Fix the incompatibility issue between SyncBN and fp16 BN.

    Known Issues

    None.