Skip to content

Latest commit

 

History

History
1315 lines (910 loc) · 14.8 KB

OpList.md

File metadata and controls

1315 lines (910 loc) · 14.8 KB
Raw

MNN Op manual

The specific parameters of Op can be found in *.fbs, and the special case is specified in the corresponding Op below.

[TOC]

Concat

Inputs:

  • input0: float32(NHWC|NC4HW4)
  • input1: float32(NHWC|NC4HW4)
  • inputn: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Concat

Tensorflow op:

  • Concat
  • ConcatV2

Convolution

Inputs

  • input: float32|int8(NC4HW4)

Outputs:

  • output: float32|int8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Convolution
  • CuDNNGroupedConvolutionForward

Tensorflow op:

  • Conv2D

Pooling

Inputs:

  • input: float32(NC4HW4)

Outpus:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Pooling

Tensorflow op:

  • MaxPool
  • AvgPool

ReLU

Inputs:

  • inputs: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Relu

Tensorflow op:

  • Relu

Softmax

Inputs:

  • input: float32(NC4HW4|NHWC)

Outputs:

  • outpus: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Softmax

Tensorflow op:

  • Softmax

PReLU

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • outpus: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • PReLU

Tensorflow op:

  • PRelu

LRN

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • LRN
  • CuDNNLRNCrossChannel

Tensorflow op:

  • LRN

Scale

Inputs:

  • input: float32(NC4HW4|NHWC)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op

  • Scale
  • BatchNorm
  • CuDNNBatchNorm

Tensorflow op:

  • FusedBatchNorm

Eltwise

Inputs:

  • input0: float32(NC4HW4|NHWC)
  • input1: float32(NC4HW4|NHWC)
  • intputn: float32(NC4HW4|NHWC)

Outputs:

  • output: float32

Backend:

    • CPU: support Product, Sum, Max
    • Metal: support Product, Sum, Max
    • Vulkan: supports Product, Sum
    • OpenCL: supports Product, Sum, Max

Caffe op:

  • Eltwise

ConvolutionDepthwise

Inputs:

  • input: float32|int8(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • CuDNNGroupedConvolutionForward
  • Convolution(group)

Tensorflow op:

  • DepthwiseConv2dNative

ArgMax

Inputs:

  • input: float32(NC4HW4)

Outpus:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • ArgMax

Reshape

Inputs:

  • input: float32(NC4HW4|NHWC)
  • shape: int32(optional)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Flatten
  • Reshape

Tensorflow:

  • Reshape

Const

Inputs:

  • 无

Outputs:

  • output: float32|int32|int8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Const

Binary

Inputs:

  • input0: float32|int32
  • input1: float32|int32

Outputs:

  • output: float32|int32

Backend:

    • CPU: supports Max, Min, Mul, Add, Sub, ReadDiv, Greater
    • Metal
    • Vulkan: supports float32 only; supports Mul, Add, Sub
    • OpenCL: supports float32 only; supports Mul, Add, Sub, Max

Caffe op:

Tensorflow op:

  • Mul
  • Sub
  • Add
  • Maximum
  • RealDiv
  • Minimum
  • Greater
  • BiasAdd

Deconvolution

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Deconvolution

Tensorflow:

  • Conv2DBackpropInput

Crop

Inputs:

  • input0: float32(NC4HW4)
  • input1: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU: supports Crop Spatial only
    • Metal
    • Vulkan: supports Crop Spatial only
    • OpenCL: supports Crop Spatial only

Limit:

  • axis >= 2

Caffe op:

  • Crop

DetectionOutput

Inputs:

  • input: float32(NC4HW4)
  • confidence: float32(NC4HW4)
  • priorbox: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • DetectionOutput

ExpandDims

Inputs:

  • input: float32(NHWC)
  • dims: int32

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • ExpandDims

Interp

Inputs:

  • input: float32(NC4HW4)
  • shape: int32(optional)

Outputs:

  • output: float32

Backend:

    • CPU: supports Nearstneighbor, bilinear, cubic
    • Metal
    • Vulkan: supports bilinear
    • OpenCL: supports bilinear

Caffe op:

  • Interp

Tensorflow:

  • ResizeBilinear
  • ResizeNearestNeighbor

LSTM

Inputs:

  • input0: float32(NC4HW4)
  • input1: float32(NC4HW4, optional)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan: support one input only
    • OpenCL

Caffe op:

  • LSTM

Normalize

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU: unsupport acrossSpatial and channelShared
    • Metal
    • Vulkan: unsupport acrossSpatial and channelShared
    • OpenCL

Caffe op:

  • Normalize

Permute

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Permute

PriorBox

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • PriorBox

Proposal

Inputs:

  • input: float32(NC4HW4)
  • bounding_box: float32
  • image_size: float32

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Reduction

Inputs:

  • input: float32|int32(NHWC)

Outputs:

  • output: float32|int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Min
  • Max
  • Mean
  • Sum
  • Prod

ROIPooling

Inputs:

  • input: float32(NC4HW4)
  • roi: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • ROIPooling

Sigmoid

Inputs:

  • input: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • Sigmoid

Tensorflow op:

  • Sigmoid

Slice

Inputs:

  • input: float32(NHWC|NC4HW4)

Outputs:

  • output0: float32
  • output1: float32
  • outputn: float32

Backend:

    • CPU
    • Metal
    • Vulkan: supports NC4HW4 input with axis = 1 only
    • OpenCL

Caffe op:

  • Slice

Tensorflow op:

  • Split

Squeeze

Inputs:

  • input: float32|int8|int32(NHWC)

Outputs:

  • output: float32|int8|int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Squeeze

TanH

Inputs:

  • input: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • TanH

Tensorflow op:

  • TanH

Tile

Inputs:

  • input: float32(NHWC)
  • multipliers: int32(NHWC)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Tile

UnaryOp

Inputs:

  • input: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU: supports Rsqrt, Square, Neg, Exp, Sqrt, Abs, Ceil
    • Metal: supports Rsqrt, Square, Neg, Exp, Sqrt, Abs, Ceil
    • Vulkan: supports Rsqrt, Exp, Sqrt, Abs
    • OpenCL: supports Rsqrt, Exp, Sqrt, Abs

Caffe op:

  • UnaryOp

Tensorflow op:

  • Rsqrt
  • Square
  • Exp
  • Neg
  • Abs
  • Ceil
  • Sqrt

SpatialProduct

Inputs:

  • input0: float32(NC4HW4)
  • input1: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Caffe op:

  • SpatialProduct

DeconvolutionDepthwise

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Resize

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

AsString

Inputs:

  • input: float32|int32(NHWC)

Outputs:

  • output: float32|int32

Backend:

    • CPU: supports Bilinear
    • Metal
    • Vulkan: supports Bilinear
    • OpenCL

Limit:

  • DataType only support float and bool

Tensorflow op:

  • AsString

Cast

Inputs:

  • input: float32|int8|int32(NHWC|NC4HW4)

Outputs:

  • output: float32|int8|int32

Backend:

    • CPU
    • Metal: half->int32, int32->half, uint8->half
    • Vulkan
    • OpenCL

Tensorflow op:

  • Cast

StridedSlice

Inputs:

  • input: float32|int32(NHWC)
  • begin: int32(NHWC)
  • end: int32(NHWC)
  • strided: int32(NHWC)

Outputs:

  • output: float32|int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • StridedSlice

ReduceJoin

Inputs:

  • input: int8(NHWC)
  • axis: int32(NHWC)

Outputs:

  • output: int8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • ReduceJoin

Pack

Inputs:

  • input0: int32(NHWC)
  • input1: float32|int32(NHWC)
  • inputn: float32|int32(NHWC)

Outputs:

  • output: float32|int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Pack

ReLU6

Inputs:

  • input: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • ReLU6

NonMaxSuppressionV2

Inputs:

  • boxes: float32(NHWC)
  • scores: float32(NHWC)
  • max_output_size: int32(NHWC)
  • iou_threshold: float32(NHWC)

Outputs:

  • output: int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • NonMaxSuppressionV2
  • NonMaxSuppressionV3

SpaceToBatchND

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • SpaceToBatchND

BatchToSpaceND

Inputs:

  • input: float32(NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • BatchToSpaceND

Shape

Inputs:

  • input: float32|int32|int8

Outputs:

  • output: int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Shape

Selu

Inputs:

  • input: float32(NHWC|NC4HW4)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Selu

tfQuantizedConv2D

Inputs:

  • intput: uint8(NC4HW4)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedConv2D

Gather

Inputs:

  • params: float32(NHWC)
  • indices: int32(NHWC)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Gather

GatherV2

Inputs:

  • params: float32(NHWC)
  • indices: int32(NHWC)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • GatherV2

QuantizedReshape

Inputs:

  • intput: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedReshape

QuantizedMaxPool

Inputs:

  • intput: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedMaxPool

QuantizedAvgPool

Inputs:

  • intput: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedMaxPool

TopKV2

Inputs:

  • intput: float32
  • k: int

Outputs:

  • output: float32
  • indices: int

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • TopKV2

CropAndResize

Inputs:

  • input: float32(NHWC)
  • boxes: float32(NHWC)
  • box_index: int32(NHWC)
  • crop_size: int32(NHWC)

Outputs:

  • output: int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • CropAndResize

Fill

Inputs:

  • dims: int32(NHWC)
  • value: int32(NHWC)

Outputs:

  • output: int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Fill

Range

Inputs:

  • start: int32|int64|float32|double(NHWC)
  • limit: int32|int64|float32|double(NHWC)
  • delta: int32|int64|float32|double(NHWC)

Outputs:

  • output: int32|int64|float32|double

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Range

Rank

Inputs:

  • input: int32(NHWC)

Outputs:

  • output: int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Rank

Size

Inputs:

  • input: int32|int64|uint8|float32(NHWC)

Outputs:

  • output: int32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Size

SilceTf

Inputs:

  • input: float32|int32(NHWC)
  • begin: int32(NHWC)
  • size: int32(NHWC)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Slice

Transpose

Inputs:

  • input: float32(NHWC)
  • perm: int32|int64(NHWC)

Outputs:

  • output: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Transpose

Where

Inputs:

  • input: int32(NHWC)

Outputs:

  • output: int64

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Where

quantizedSoftmax

Inputs:

  • intput: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedSoftmax

QuantizedDepthwiseConv2D

Inputs:

  • intput: uint8(NC4HW4)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizeDepthwise

QuantizedAdd

Inputs:

  • input0: uint8(NHWC)
  • input1: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedAdd

QuantizedLogistic

Inputs:

  • intput: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedLogistic

Unpack

Inputs:

  • intput: float32(NHWC)

Outputs:

  • output0: float32
  • outputN: float32

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow op:

  • Unpack

QuantizedConcat

Inputs:

  • intput0: uint8(NHWC)
  • intput1: uint8(NHWC)
  • intputn: uint8(NHWC)

Outputs:

  • output: uint8

Backend:

    • CPU
    • Metal
    • Vulkan
    • OpenCL

Tensorflow Lite op:

  • QuantizedConcat