QNN HTP API Revision History

QNN HTP API Version

Introduced in QNN SDK Version

Description

5.39.0

2.39.0

5.38.0

2.38.0

  • QnnHtpContext.h

    • Added QNN_HTP_CONTEXT_CONFIG_OPTION_LORA_WEIGHT_SHARING_ENABLED config option

  • Please regenerate op package when moving to a new SDK version

5.37.0

2.37.0

  • QnnHtpCommon.h

    • Updated QNN_HTP_CONTEXT_BLOB_VERSION_MINOR for new blob data alignment

  • Please regenerate op package when moving to a new SDK version

5.36.0

2.36.0

  • QnnHtpContext.h

    • Introduced extended UDMA prepare time config

  • QnnHtpGraph.h

    • Added QNN_HTP_GRAPH_CONFIG_OPTION_SHARE_IO_BUFFER config option

  • QnnHtpDevice.h

    • Updated config for better supporting secure PD

  • Please regenerate op package when moving to a new SDK version

5.35.0

2.35.0

  • QnnHtpContext.h

    • Introduced VTCM backup buffer sharing

    • Added QNN_HTP_CONTEXT_CONFIG_OPTION_SKIP_VALIDATION_ON_BINARY_SECTION config option from QnnHtpGraph.h

    • Introduced QnnContext_CustomProperty_t for setting external weights and spill/fill buffer

    • Introduced QNN_HTP_CONTEXT_CONFIG_OPTION_USE_EXTENDED_UDMA config to support preparing graphs with extended UDMA utilization on supported architectures

  • QnnHtpGraph.h

    • Removed QNN_HTP_CONTEXT_CONFIG_OPTION_SKIP_VALIDATION_ON_BINARY_SECTION from custom graph config, moved to custom context config in QnnHtpContext.h

  • QnnHtpMem.h

    • Added QNN_HTP_MEM_WEIGHTS_BUFFER and QNN_HTP_MEM_SHARED_SPILLFILL_BUFFER memory type and corresponding config structures

  • Please regenerate op package when moving to a new SDK version

5.34.0

2.34.0

  • QnnHtpCommon.h

    • Updated DSP Context blob version

  • QnnHtpGraph.h

    • Introduced QNN_HTP_GRAPH_OPTIMIZATION_TYPE_ENABLE_SLC_ALLOCATOR config option

  • QnnHtpDevice.h

    • Introduced QNN_HTP_DEVICE_CONFIG_OPTION_CUSTOM enum to support SecurePD option

  • Please regenerate op package when moving to a new SDK version

5.33.0

2.33.0

5.32.0

2.32.0

  • QnnHtpContext.h

    • Enable concurrent deserialization on context create from binary without async api

  • Unified HTP API headers between auto and mobile SDK

  • Please regenerate op package when moving to a new SDK version

5.31.0

2.31.0

5.30.0

2.30.0

  • QnnHtpCommon.h

    • Updated DSP Context blob version to support Parallel Graph Execution

  • QnnHtpGraph.h

    • Introduced graph finalize config option

  • Please regenerate op package when moving to a new SDK version

5.29.0

2.29.0

  • QnnHtpContext.h

    • Introduced QNN_HTP_CONTEXT_CONFIG_OPTION_PREPARE_ONLY config option to enable model preparation only without mapping its content on the DSP side

  • QnnHtpGraph.h

    • Introduced QNN_HTP_GRAPH_OPTIMIZATION_TYPE_ENABLE_DLBC_WEIGHTS optimization type to enable DLBC weights compression

    • Introduced QNN_HTP_GRAPH_CONFIG_OPTION_ASSUME_SAME_QUANT config option

  • QnnHtpProfile.h

    • Introduced QNN_HTP_PROFILE_EVENTTYPE_GRAPH_APPLY_BINARY_SECTION_QNN, QNN_HTP_PROFILE_EVENTTYPE_GRAPH_APPLY_BINARY_SECTION_RPC, QNN_HTP_PROFILE_EVENTTYPE_GRAPH_APPLY_BINARY_SECTION_QNN_ACC, QNN_HTP_PROFILE_EVENTTYPE_GRAPH_APPLY_BINARY_SECTION_ACC profile event types, for detailed usages, please refer to the documentation in QnnHtpProfile.h

  • Please regenerate OpPackage when moving to a new SDK version. Custom OpPackages produced with previous releases may not be compatible, it is recommended to rebuild them with this release.

  • Context Binaries built since QNN 2.0 can used after this release, subject to the following restrictions:

    • It does not contain graphs that make use of custom op

    • It was not built on a 32-bit target prior to QNN 2.3

5.28.0

2.28.0

  • QnnHtpSystemContext.h

    • Retrieved HTP backend graph settings and device settings and stored them into QNN context binary info

  • QnnHtpPerfInfrastructure.h

    • Introduced adaptive polling time config option to allow clients set up the adaptive polling time in microseconds

  • Please regenerate op package when moving to a new SDK version

5.27.0

2.27.0

  • QnnHtpProfile.h

    • Introduced apply binary section event type to support applyBinarySection profiling

  • QnnHtpGraph.h

    • Introduced weights packing config option

  • Please regenerate op package when moving to a new SDK version

5.26.0

2.26.0

5.25.0

2.25.0

  • QnnHtpCommon.h

    • Updated DSP Context blob version

  • QnnHtpDevice.h

    • Introduced QNN_HTP_CORE_TYPE_MAX to QNN_HTP_CORE_TYPE enums

  • Please regenerate op package when moving to a new SDK version

5.24.0

2.24.0

  • QnnHtpContext.h

    • Introduced share resources config option to enable resource sharing across different context for the new QnnContext_createFromBinaryListAsync API through listConfig param

    • Introduced IO memory estimation config option to enable estimation of the total size of the IO tensors required by the context to ensure sufficient space on the PD before deserialization

  • QnnHtpDevice.h

    • Introduced QNN_HTP_CORE_TYPE enums to distinguish NSP and HPASS cores

  • QnnHtpGraph.h

    • Introduced VTCM size in mb config option, which currently behaves the same as QNN_HTP_GRAPH_CONFIG_OPTION_VTCM_SIZE

  • Please regenerate op package when moving to a new SDK version

5.23.0

2.23.0

  • QnnHtpContext.h

    • Introduced DSP memory profiling config option to enable memory profiling on DSP

  • Please regenerate op package when moving to a new SDK version

5.22.0

2.22.0

  • QnnHtpPerfInfrastructure.h

    • Introduced HMX frequency vote to allow clients to place a separate vote just for HMX frequency alongside HVX frequency

  • Please regenerate op package when moving to a new SDK version

5.21.0

2.21.0

5.20.0

2.20.0

  • QnnHtpCommon.h

    • Updated DSP Context blob version

  • QnnHtpGraph.h

    • Added new config macro to represent to use the maximum available number of resource

  • Introduced backward incompatible changes to HTP core API for custom op development

5.19.0

2.19.0

5.18.0

2.18.0

  • QnnHtpContext.h

    • Added new context config parameter to turn on the peak avoidance memcpy

  • QnnHtpSystemContext.h

    • Introduced QnnHtpSystemContext.h to supplement QnnSystemContext.h for HTP backend

    • Updated the spillFill variable type from size_t to uint64_t to prevent misalignment

  • Introduced backward incompatible changes to HTP core API for custom op development

5.17.0

2.17.0

  • QnnHtpProfile.h

    • Added true accelerator time event type for detailed profiling

  • Introduced backward incompatible changes to HTP core API for custom op development

5.16.0

2.16.0

  • QnnHtpContext.h

    • Added custom config for weight sharing

    • Added custom config for spill-fill buffer sharing

  • QnnHtpMem.h

    • Added custom memory type for multi tensor shared buffer

  • Introduced backward incompatible changes to HTP core API for custom op development

5.15.0

2.15.0

  • QnnHtpCommon.h

    • Updated DSP Context blob Version

  • Introduced backward incompatible changes to HTP core API for custom op development

5.14.0

2.14.0

  • QnnHtpGraph.h, QnnHtpDevice.h

    • Updated description of graph finalize optimization options

  • Introduced backward incompatible changes to HTP core API for custom op development

5.13.0

2.13.0

  • QnnHtpGraph.h

    • Updated description of graph finalize optimization options

  • Introduced backward incompatible changes to HTP core API for custom op development

5.12.0

2.12.0

  • Introduced backward incompatible changes to HTP core API for custom op development

5.11.0

2.11.0

  • Introduced backward incompatible changes to HTP core API for custom op development

5.10.0

2.10.0

  • Introduced backward incompatible changes to HTP core API for custom op development

5.9.0

2.9.0

  • QnnHtpCommon.h

    • Added library security versioning

  • QnnHtpPerfInfrastructure.h

    • Added support to configure Turbo L2 and L3 power levels

  • Introduced backward incompatible changes to HTP core API for custom op development

5.8.0

2.8.0

  • QnnHtpProfile.h

    • Added support to configure number of HVX threads in profiling

  • Introduced backward incompatible changes to HTP core API for custom op development

5.7.0

2.7.0

  • Introduced backward incompatible changes internal to HTP core opPackage and HTP core API

5.6.0

2.6.0

  • Introduced backward incompatible changes internal to HTP core API

    • shape.h - increased shape_reduce_map size to 7 for 6D tensor support

  • QnnHtpDevice.h

    • Deprecated PCIe device

  • QnnHtpGraph.h

    • Added support to configure number of HVX threads

5.5.0

2.5.0

  • Introduced backward incompatible changes internal to HTP core opPackage and HTP core API

    • optimize.h - conv_s2d_shape function added

    • tensor.h - CoreTensors type updated

5.4.0

2.4.0

  • Introduced backward incompatible changes internal to HTP core opPackage

5.3.0

2.3.0

  • Introduced backward incompatible changes internal to HTP core opPackage

5.2.0

2.2.0

  • Added QnnHtpMem.h to support VTCM sharing

  • Introduced backward incompatible changes internal to HTP core API

5.1.0

2.1.0

  • QnnHtpProfile.h

    • Node specific profile event types have been added.

  • Introduced backward incompatible changes internal to HTP core API

5.0.0

2.0.0

  • Branched from QNN SDK version 1.10.0