QNN LPAI Op Support Revision History

Introduced in QNN SDK Version

Description

2.36.0

  • Added support for RmsNorm Op

  • Added support Buffer Op Padding Param

2.35.0

  • Added support for StatelessLSTM/GRU Op

2.34.0

  • Added support for BatchToSpace,SpaceToBatch,DepthToSpace,SpaceToDepth,Quantize,Dequantize

2.32.0

  • Added support for Channel_Shuffle op

2.26.1

  • Added Default Support of antialias param for ResizeBilinear Op

2.24.1

  • Add Power op support

  • Add LAYERNORM op support

  • Add SPLIT op support

2.23.1
  • Added support of 8-bit and 16-bit Ops:

    • Cast, ElementWisePower, ReduceProd, ResizeNearestNeighbor, Split, Layernorm

2.16.1
  • Added Support of 8-bit and 16-bit ElementWiseUnary Op with sub operations:

    • ElementWiseExp, ElementWiseLog, ElementWiseNeg, ElementWiseSquareRoot

  • Added Support of 8-bit and 16-bit ElementWiseBinary Op with sub operations:

    • ElementWiseAdd, ElementWiseDivide, ElementWiseMultiply, ElementWiseSubtract

  • Added Support of 8-bit and 16-bit ElementWiseNeuron Op with sub operations:

    • Gelu, HardSwish, Relu, ReluMinMax, Sigmoid, Tanh

2.13.1

  • Added support for Resize and ElementWiseMinimum (8bit/16bit)

2.12.1

  • Added ROI support for Slice and Concat

  • Added support for Convert (8bit/16bit)

  • Added support for Pool to floor rounding mode

2.11.1

  • Removed Gather op limitation

  • Added support deconv with both input and output padding

  • Added 4-bit weights support for Conv2d, DepthWiseConv2d, TransposeConv2d

  • Added 8-bit support for HardSwish and Gelu

  • Added 16-bit support for Argmax, Argmin, Batchnorm, Conv2d, Concat, DepthWiseConv2d, HardSwish, ElementWiseMaximum, ElementWiseAdd, ElementWiseDivide, ElementWiseExp, ElementWiseLog, ElementWiseMultiply, ElementWiseSquareRoot, ElementWiseSubtract, FullyConnected, Gather, Gelu, LogSoftmax, MatMul, Pad, PoolAvg2d, PoolMax2d, Prelu, ReduceMax, ReduceMean, ReduceMin, ReluMinMax, Reshape, ReduceSum, Relu, Softmax, StridedSlice, Sigmoid, Transpose, Tanh, and TransposeConv2d.

2.10.1

  • Fix for overflow for clip in global AvgPool

  • Added 8-bit signed/unsigned support for ReduceMax, ReduceMin, ReduceSum, StridedSlice, Gather, and LogSoftMax.

  • Added AvgPool, MaxPool, Pad, TransposeConv2d, and StridedSlice validators.