Roadmap (Operators)

With TensorFlow-DirectML, our goal is to have largely equivalent operator coverage to the existing GPU device. We prioritize ops primarily on how frequently they are used in models that we've encountered, but there are other considerations as well. The following aspects are considered when lowering the priority of an operator:

• Usage: we don't have enough evidence that an op is frequently used.
• Cost: the implementation cost (e.g. a new DirectML API) is relatively high.
• Performance: falling back to CPU for the op isn't expected to result in a significant performance penalty, or the penalty is amortized over a longer training session.
• Domain: the op belongs to a set of ops that target a domain we don't intend to support at the moment.
• Generality: the op is unique to TensorFlow, not suitable for a DirectML API, but not easily implementable with a composition of existing APIs.

This prioritization may change, but we hope it's useful in highlighting the specific functionality we intend to light up over the coming months. Not all ops are equally important to accelerate, and it's definitely not required to reach 100% parity with the CPU; even the GPU device only implements a little over half (~600) of the total number of ops (~1100). As of today, DirectML implements ~480 ops.

Operators

The table below comprises all of the TF operators that have at least one GPU kernel.

• The first column is the name of the op. You can find more info for most of these by looking up tf.raw_ops.<name> on TensorFlow's documentation (e.g. tf.raw_ops.Bitcast).
• The second column indicates the operator type, which is a rough grouping of ops based on functionality. See Operator Types for details.
• The third column indicates DirectML support:
  • ✅ = the op has at least one DirectML kernel registered
  • 🚧 = the op will be implemented in the next milestone (~3 months)
  • ❌ = the op may be implemented in the future

Op Name	Type	DirectML Support
_ParallelConcatStart	array	✅
_ParallelConcatUpdate	array	✅
BatchMatrixBandPart	array	✅
BatchMatrixDiag	array	✅
BatchMatrixDiagPart	array	✅
BatchMatrixSetDiag	array	✅
BatchToSpace	array	✅
BatchToSpaceND	array	✅
Bitcast	array	✅
BroadcastArgs	array	✅
BroadcastGradientArgs	array	✅
BroadcastTo	array	✅
CheckNumerics	array	✅
Concat	array	✅
ConcatOffset	array	✅
ConcatV2	array	✅
Const	array	✅
DeepCopy	array	✅
DepthToSpace	array	✅
Diag	array	✅
DiagPart	array	✅
Empty	array	✅
EnsureShape	array	✅
ExpandDims	array	✅
ExtractImagePatches	array	✅
ExtractVolumePatches	array	✅
Fill	array	✅
Gather	array	✅
GatherNd	array	✅
GatherV2	array	✅
Identity	array	✅
InplaceAdd	array	❌ (Future: Usage, Cost)
InplaceSub	array	❌ (Future: Usage, Cost)
InplaceUpdate	array	❌ (Future: Usage, Cost)
InvertPermutation	array	✅
LowerBound	array	❌ (Future: Usage, Cost)
MatrixBandPart	array	✅
MatrixDiag	array	✅
MatrixDiagPart	array	✅
MatrixDiagPartV2	array	✅
MatrixDiagV2	array	✅
MatrixSetDiag	array	✅
MatrixSetDiagV2	array	✅
MirrorPad	array	✅
MirrorPadGrad	array	✅
OneHot	array	✅
OnesLike	array	✅
Pack	array	✅
Pad	array	✅
PadV2	array	✅
ParallelConcat	array	✅
PlaceholderWithDefault	array	✅
PreventGradient	array	✅
Rank	array	✅
RefIdentity	array	✅
Reshape	array	✅
ResourceStridedSliceAssign	array	✅
Reverse	array	✅
ReverseSequence	array	✅
ReverseV2	array	✅
Roll	array	✅
Shape	array	✅
ShapeN	array	✅
Size	array	✅
Slice	array	✅
Snapshot	array	✅
SpaceToBatch	array	✅
SpaceToBatchND	array	✅
SpaceToDepth	array	✅
Split	array	✅
SplitV	array	✅
Squeeze	array	✅
StopGradient	array	✅
StridedSlice	array	✅
StridedSliceAssign	array	✅
StridedSliceGrad	array	✅
TensorScatterAdd	array	✅
TensorScatterSub	array	✅
TensorScatterUpdate	array	✅
TensorStridedSliceUpdate	array	✅
Tile	array	✅
Transpose	array	✅
Unique	array	✅
Unpack	array	✅
UpperBound	array	❌ (Future: Usage, Cost)
Where	array	✅
ZerosLike	array	✅
BitwiseAnd	bitwise	✅
BitwiseOr	bitwise	✅
BitwiseXor	bitwise	✅
Invert	bitwise	✅
LeftShift	bitwise	✅
PopulationCount	bitwise	✅
RightShift	bitwise	✅
Angle	complex	❌ (Future: Domain (DirectML does not support complex data type))
Complex	complex	❌ (Future: Domain (DirectML does not support complex data type))
ComplexAbs	complex	❌ (Future: Domain (DirectML does not support complex data type))
Conj	complex	❌ (Future: Domain (DirectML does not support complex data type))
ConjugateTranspose	complex	❌ (Future: Domain (DirectML does not support complex data type))
Imag	complex	❌ (Future: Domain (DirectML does not support complex data type))
Real	complex	❌ (Future: Domain (DirectML does not support complex data type))
CudnnRNN	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNBackprop	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNBackpropV2	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNBackpropV3	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNCanonicalToParams	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNCanonicalToParamsV2	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNParamsSize	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNParamsToCanonical	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNParamsToCanonicalV2	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNV2	cudnn	❌ (Future: Generality (CUDA-only op))
CudnnRNNV3	cudnn	❌ (Future: Generality (CUDA-only op))
AnonymousIterator	dataset	✅
AnonymousIteratorV2	dataset	✅
DeleteIterator	dataset	✅
ExperimentalMapDataset	dataset	✅
ExperimentalSleepDataset	dataset	✅
GeneratorDataset	dataset	✅
IteratorFromStringHandleV2	dataset	✅
IteratorGetNext	dataset	✅
IteratorGetNextAsOptional	dataset	✅
IteratorGetNextSync	dataset	✅
IteratorToStringHandle	dataset	✅
IteratorV2	dataset	✅
MakeIterator	dataset	✅
OptionalFromValue	dataset	✅
OptionalGetValue	dataset	✅
OptionalHasValue	dataset	✅
OptionalNone	dataset	✅
PrefetchDataset	dataset	✅
SleepDataset	dataset	✅
UnwrapDatasetVariant	dataset	✅
WrapDatasetVariant	dataset	✅
Copy	debug	✅
CopyHost	debug	✅
DebugIdentity	debug	✅
DebugNanCount	debug	✅
DebugNumericSummary	debug	✅
CollectiveBcastRecv	distributed	❌ (Future: Domain (distributed execution), Cost)
CollectiveBcastSend	distributed	❌ (Future: Domain (distributed execution), Cost)
CollectiveGather	distributed	❌ (Future: Domain (distributed execution), Cost)
CollectiveReduce	distributed	❌ (Future: Domain (distributed execution), Cost)
ApplyKerasMomentum	directml	✅
DmlNonzeroCoordinates	directml	✅
_Arg	functional	✅
_ArrayToList	functional	✅
_DeviceArg	functional	✅
_DeviceRetval	functional	✅
_ListToArray	functional	✅
_Retval	functional	✅
_SwitchN	functional	✅
_While	functional	✅
Case	functional	✅
DeleteSessionTensor	functional	✅
DynamicPartition	functional	❌ (Future: Cost, Generality)
DynamicStitch	functional	✅
EagerPyFunc	functional	✅
EmptyTensorList	functional	✅
Enter	functional	✅
Exit	functional	✅
FakeParam	functional	✅
For	functional	✅
GetSessionHandle	functional	✅
GetSessionHandleV2	functional	✅
GetSessionTensor	functional	✅
If	functional	✅
MapClear	functional	✅
MapIncompleteSize	functional	✅
MapPeek	functional	✅
MapSize	functional	✅
MapStage	functional	✅
MapUnstage	functional	✅
MapUnstageNoKey	functional	✅
Merge	functional	✅
NextIteration	functional	✅
OrderedMapClear	functional	✅
OrderedMapIncompleteSize	functional	✅
OrderedMapPeek	functional	✅
OrderedMapSize	functional	✅
OrderedMapStage	functional	✅
OrderedMapUnstage	functional	✅
OrderedMapUnstageNoKey	functional	✅
ParallelDynamicStitch	functional	✅
PartitionedCall	functional	✅
RefEnter	functional	✅
RefExit	functional	✅
RefMerge	functional	✅
RefNextIteration	functional	✅
RefSwitch	functional	✅
RemoteCall	functional	✅
Stage	functional	✅
StageClear	functional	✅
StagePeek	functional	✅
StageSize	functional	✅
StatefulPartitionedCall	functional	✅
StatelessIf	functional	✅
StatelessWhile	functional	✅
Switch	functional	✅
SymbolicGradient	functional	✅
TensorArray	functional	✅
TensorArrayClose	functional	✅
TensorArrayCloseV2	functional	✅
TensorArrayCloseV3	functional	✅
TensorArrayConcat	functional	✅
TensorArrayConcatV2	functional	✅
TensorArrayConcatV3	functional	✅
TensorArrayGather	functional	✅
TensorArrayGatherV2	functional	✅
TensorArrayGatherV3	functional	✅
TensorArrayGrad	functional	✅
TensorArrayGradV2	functional	✅
TensorArrayGradV3	functional	✅
TensorArrayGradWithShape	functional	✅
TensorArrayPack	functional	✅
TensorArrayRead	functional	✅
TensorArrayReadV2	functional	✅
TensorArrayReadV3	functional	✅
TensorArrayScatter	functional	✅
TensorArrayScatterV2	functional	✅
TensorArrayScatterV3	functional	✅
TensorArraySize	functional	✅
TensorArraySizeV2	functional	✅
TensorArraySizeV3	functional	✅
TensorArraySplit	functional	✅
TensorArraySplitV2	functional	✅
TensorArraySplitV3	functional	✅
TensorArrayUnpack	functional	✅
TensorArrayV2	functional	✅
TensorArrayV3	functional	✅
TensorArrayWrite	functional	✅
TensorArrayWriteV2	functional	✅
TensorArrayWriteV3	functional	✅
TensorListConcat	functional	✅
TensorListConcatLists	functional	✅
TensorListConcatV2	functional	✅
TensorListElementShape	functional	✅
TensorListFromTensor	functional	✅
TensorListGather	functional	✅
TensorListGetItem	functional	✅
TensorListLength	functional	✅
TensorListPopBack	functional	✅
TensorListPushBack	functional	✅
TensorListPushBackBatch	functional	✅
TensorListReserve	functional	✅
TensorListResize	functional	✅
TensorListScatter	functional	✅
TensorListScatterIntoExistingList	functional	✅
TensorListScatterV2	functional	✅
TensorListSetItem	functional	✅
TensorListSplit	functional	✅
TensorListStack	functional	✅
Unstage	functional	✅
While	functional	✅
AdjustContrast	image	✅
AdjustContrastv2	image	✅
AdjustHue	image	✅
AdjustSaturation	image	✅
CropAndResize	image	✅
CropAndResizeGradBoxes	image	✅
CropAndResizeGradImage	image	✅
HSVToRGB	image	✅
NonMaxSuppressionV2	image	❌ (Future: Usage, Cost)
NonMaxSuppressionV3	image	❌ (Future: Usage, Cost)
ResizeBilinear	image	✅
ResizeBilinearGrad	image	✅
ResizeNearestNeighbor	image	✅
ResizeNearestNeighborGrad	image	✅
RGBToHSV	image	✅
_CopyFromGpuToHost	internal	✅
_CopyFromHostToGpu	internal	✅
_If	internal	✅
_ScopedAllocator	internal	❌ (Future: Usage)
_ScopedAllocatorConcat	internal	❌ (Future: Usage)
_ScopedAllocatorSplit	internal	❌ (Future: Usage)
BatchMatrixTriangularSolve	linalg	❌ (Future: Domain (linear algebra), Cost, Usage)
BatchSvd	linalg	❌ (Future: Domain (linear algebra), Cost)
Cholesky	linalg	❌ (Future: Domain (linear algebra), Cost)
Einsum	linalg	❌ (Future: Domain (linear algebra), Cost)
LogMatrixDeterminant	linalg	❌ (Future: Domain (linear algebra), Cost, Usage)
Lu	linalg	❌ (Future: Domain (linear algebra), Cost, Usage)
MatrixDeterminant	linalg	❌ (Future: Domain (linear algebra), Cost)
MatrixInverse	linalg	❌ (Future: Domain (linear algebra), Cost)
MatrixSolve	linalg	❌ (Future: Domain (linear algebra), Cost)
MatrixTriangularSolve	linalg	❌ (Future: Domain (linear algebra), Cost)
Qr	linalg	❌ (Future: Domain (linear algebra), Cost, Usage)
SelfAdjointEigV2	linalg	❌ (Future: Domain (linear algebra), Cost)
Svd	linalg	❌ (Future: Domain (linear algebra), Cost)
TridiagonalMatMul	linalg	❌ (Future: Domain (linear algebra), Cost, Usage)
TridiagonalSolve	linalg	❌ (Future: Domain (linear algebra), Cost, Usage)
Abs	math	✅
Acos	math	✅
Acosh	math	✅
Add	math	✅
AddN	math	✅
AddV2	math	✅
All	math	✅
Any	math	✅
ApproximateEqual	math	✅
ArgMax	math	✅
ArgMin	math	✅
Asin	math	✅
Asinh	math	✅
Atan	math	✅
Atan2	math	✅
Atanh	math	✅
BatchMatMul	math	✅
BatchMatMulV2	math	✅
Bincount	math	❌ (Future: Usage, Cost)
Bucketize	math	❌ (Future: Usage, Cost)
Cast	math	✅
Ceil	math	✅
ClipByValue	math	✅
CompareAndBitpack	math	✅
Cos	math	✅
Cosh	math	✅
Cross	math	✅
Cumprod	math	✅
Cumsum	math	✅
CumulativeLogsumexp	math	❌ (Future: Usage, Cost)
Div	math	✅
DivNoNan	math	✅
Equal	math	✅
Erf	math	✅
Erfc	math	✅
EuclideanNorm	math	✅
Exp	math	✅
Expm1	math	✅
Floor	math	✅
FloorDiv	math	✅
FloorMod	math	✅
Greater	math	✅
GreaterEqual	math	✅
HistogramFixedWidth	math	❌ (Future: Usage, Cost)
Inv	math	✅
InvGrad	math	✅
IsFinite	math	✅
IsInf	math	✅
IsNan	math	✅
Less	math	✅
LessEqual	math	✅
LinSpace	math	✅
Log	math	✅
Log1p	math	✅
LogicalAnd	math	✅
LogicalNot	math	✅
LogicalOr	math	✅
MatMul	math	✅
Max	math	✅
Maximum	math	✅
Mean	math	✅
Min	math	✅
Minimum	math	✅
Mod	math	✅
Mul	math	✅
MulNoNan	math	✅
Neg	math	✅
NextAfter	math	❌ (Future: Usage, Cost)
NotEqual	math	✅
Pow	math	✅
Prod	math	✅
Range	math	✅
RealDiv	math	✅
Reciprocal	math	✅
ReciprocalGrad	math	✅
Rint	math	✅
Round	math	✅
Rsqrt	math	✅
RsqrtGrad	math	✅
SegmentSum	math	❌ (Future: Generality, Cost)
Select	math	✅
SelectV2	math	✅
Sigmoid	math	✅
SigmoidGrad	math	✅
Sign	math	✅
Sin	math	✅
Sinh	math	✅
Sqrt	math	✅
SqrtGrad	math	✅
Square	math	✅
SquaredDifference	math	✅
Sub	math	✅
Sum	math	✅
Tan	math	✅
Tanh	math	✅
TanhGrad	math	✅
TruncateDiv	math	✅
TruncateMod	math	✅
UnsortedSegmentMax	math	❌ (Future: Generality, Cost)
UnsortedSegmentMin	math	❌ (Future: Generality, Cost, Usage)
UnsortedSegmentProd	math	❌ (Future: Generality, Cost, Usage)
UnsortedSegmentSum	math	❌ (Future: Generality, Cost)
Xdivy	math	✅
Xlogy	math	✅
_FusedBatchNormEx	nn	✅
_FusedConv2D	nn	✅
AvgPool	nn	✅
AvgPool3D	nn	✅
AvgPool3DGrad	nn	✅
AvgPoolGrad	nn	✅
BatchNormWithGlobalNormalization	nn	✅
BatchNormWithGlobalNormalizationGrad	nn	✅
BiasAdd	nn	✅
BiasAddGrad	nn	✅
BiasAddV1	nn	✅
Conv2D	nn	✅
Conv2DBackpropFilter	nn	✅
Conv2DBackpropInput	nn	✅
Conv3D	nn	✅
Conv3DBackpropFilter	nn	✅
Conv3DBackpropFilterV2	nn	✅
Conv3DBackpropInput	nn	✅
Conv3DBackpropInputV2	nn	✅
DataFormatDimMap	nn	✅
DataFormatVecPermute	nn	✅
DepthwiseConv2dNative	nn	✅
DepthwiseConv2dNativeBackpropFilter	nn	✅
DepthwiseConv2dNativeBackpropInput	nn	✅
Dilation2D	nn	❌ (Future: Usage, Cost)
Dilation2DBackpropFilter	nn	❌ (Future: Usage, Cost)
Dilation2DBackpropInput	nn	❌ (Future: Usage, Cost)
Elu	nn	✅
EluGrad	nn	✅
FusedBatchNorm	nn	✅
FusedBatchNormGrad	nn	✅
FusedBatchNormGradV2	nn	✅
FusedBatchNormGradV3	nn	✅
FusedBatchNormV2	nn	✅
FusedBatchNormV3	nn	✅
InTopKV2	nn	❌ (Future: Usage, Cost)
L2Loss	nn	✅
LeakyRelu	nn	✅
LeakyReluGrad	nn	✅
LogSoftmax	nn	✅
LRN	nn	✅
LRNGrad	nn	✅
MaxPool	nn	✅
MaxPool3D	nn	✅
MaxPool3DGrad	nn	✅
MaxPool3DGradGrad	nn	❌ (Future: Usage, Cost)
MaxPoolGrad	nn	✅
MaxPoolGradGrad	nn	❌ (Future: Usage, Cost)
MaxPoolGradGradV2	nn	❌ (Future: Usage, Cost)
MaxPoolGradGradWithArgmax	nn	❌ (Future: Usage, Cost)
MaxPoolGradV2	nn	✅
MaxPoolGradWithArgmax	nn	❌ (Future: Usage, Cost)
MaxPoolV2	nn	✅
MaxPoolWithArgmax	nn	❌ (Future: Usage, Cost)
Relu	nn	✅
Relu6	nn	✅
Relu6Grad	nn	✅
ReluGrad	nn	✅
Selu	nn	✅
SeluGrad	nn	✅
Softmax	nn	✅
SoftmaxCrossEntropyWithLogits	nn	✅
Softplus	nn	✅
SoftplusGrad	nn	✅
Softsign	nn	✅
SoftsignGrad	nn	✅
SparseSoftmaxCrossEntropyWithLogits	nn	✅
TopK	nn	✅
TopKV2	nn	✅
FakeQuantWithMinMaxArgs	quantization	❌ (Future: Domain (quantization), Cost)
FakeQuantWithMinMaxArgsGradient	quantization	❌ (Future: Domain (quantization), Cost)
FakeQuantWithMinMaxVars	quantization	❌ (Future: Domain (quantization), Cost)
FakeQuantWithMinMaxVarsGradient	quantization	❌ (Future: Domain (quantization), Cost)
FakeQuantWithMinMaxVarsPerChannel	quantization	❌ (Future: Domain (quantization), Cost)
FakeQuantWithMinMaxVarsPerChannelGradient	quantization	❌ (Future: Domain (quantization), Cost)
QuantizeAndDequantize	quantization	❌ (Future: Domain (quantization), Cost)
QuantizeAndDequantizeV2	quantization	❌ (Future: Domain (quantization), Cost)
QuantizeAndDequantizeV3	quantization	❌ (Future: Domain (quantization), Cost)
Multinomial	random	❌ (Future: Cost, Performance (amortizable)
ParameterizedTruncatedNormal	random	❌ (Future: Cost, Performance (amortizable)
RandomGammaGrad	random	❌ (Future: Cost, Performance (amortizable)
RandomStandardNormal	random	❌ (Future: Cost, Performance (amortizable)
RandomUniform	random	✅
RandomUniformInt	random	✅
RngSkip	random	❌ (Future: Cost, Performance (amortizable))
StatefulStandardNormalV2	random	❌ (Future: Cost, Performance (amortizable))
StatefulTruncatedNormal	random	❌ (Future: Cost, Performance (amortizable))
StatefulUniform	random	❌ (Future: Cost, Performance (amortizable))
StatefulUniformFullInt	random	❌ (Future: Cost, Performance (amortizable))
StatefulUniformInt	random	❌ (Future: Cost, Performance (amortizable))
StatelessMultinomial	random	❌ (Future: Cost, Performance (amortizable))
StatelessRandomNormal	random	❌ (Future: Cost, Performance (amortizable))
StatelessRandomUniform	random	✅
StatelessRandomUniformInt	random	✅
StatelessTruncatedNormal	random	❌ (Future: Cost, Performance (amortizable))
TruncatedNormal	random	❌ (Future: Cost, Performance (amortizable))
BlockLSTM	rnn	✅
BlockLSTMGrad	rnn	✅
BlockLSTMGradV2	rnn	✅
BlockLSTMV2	rnn	✅
GRUBlockCell	rnn	✅
GRUBlockCellGrad	rnn	✅
LSTMBlockCell	rnn	✅
LSTMBlockCellGrad	rnn	✅
BatchFFT	signal	❌ (Future: Domain (signal processing), Cost)
BatchFFT2D	signal	❌ (Future: Domain (signal processing), Cost)
BatchFFT3D	signal	❌ (Future: Domain (signal processing), Cost)
BatchIFFT	signal	❌ (Future: Domain (signal processing), Cost)
BatchIFFT2D	signal	❌ (Future: Domain (signal processing), Cost)
BatchIFFT3D	signal	❌ (Future: Domain (signal processing), Cost)
FFT	signal	❌ (Future: Domain (signal processing), Cost)
FFT2D	signal	❌ (Future: Domain (signal processing), Cost)
FFT3D	signal	❌ (Future: Domain (signal processing), Cost)
IFFT	signal	❌ (Future: Domain (signal processing), Cost)
IFFT2D	signal	❌ (Future: Domain (signal processing), Cost)
IFFT3D	signal	❌ (Future: Domain (signal processing), Cost)
IRFFT	signal	❌ (Future: Domain (signal processing), Cost)
IRFFT2D	signal	❌ (Future: Domain (signal processing), Cost)
IRFFT3D	signal	❌ (Future: Domain (signal processing), Cost)
RFFT	signal	❌ (Future: Domain (signal processing), Cost)
RFFT2D	signal	❌ (Future: Domain (signal processing), Cost)
RFFT3D	signal	❌ (Future: Domain (signal processing), Cost)
SparseTensorDenseMatMul	sparse	❌ (Future: Domain (sparse tensors), Usage)
BesselI0e	special	❌ (Future: Domain (special functions), Cost, Usage)
BesselI1e	special	❌ (Future: Domain (special functions), Cost, Usage)
Betainc	special	❌ (Future: Domain (special functions), Cost, Usage)
Digamma	special	❌ (Future: Domain (special functions), Cost, Usage)
Igamma	special	❌ (Future: Domain (special functions), Cost, Usage)
Igammac	special	❌ (Future: Domain (special functions), Cost, Usage)
IgammaGradA	special	❌ (Future: Domain (special functions), Cost, Usage)
Lgamma	special	❌ (Future: Domain (special functions), Cost)
Polygamma	special	❌ (Future: Domain (special functions), Cost, Usage)
Zeta	special	❌ (Future: Domain (special functions), Cost, Usage)
_ReadVariablesOp	state	✅
_VarHandlesOp	state	✅
Assign	state	✅
AssignAdd	state	✅
AssignAddVariableOp	state	✅
AssignSub	state	✅
AssignSubVariableOp	state	✅
AssignVariableOp	state	✅
ConsumeMutexLock	state	✅
DestroyResourceOp	state	✅
DestroyTemporaryVariable	state	✅
IsVariableInitialized	state	✅
MutexLock	state	✅
MutexV2	state	✅
ReadVariableOp	state	✅
ResourceGather	state	✅
ResourceGatherNd	state	✅
ResourceScatterAdd	state	✅
ResourceScatterDiv	state	✅
ResourceScatterMax	state	✅
ResourceScatterMin	state	✅
ResourceScatterMul	state	✅
ResourceScatterNdAdd	state	✅
ResourceScatterNdSub	state	✅
ResourceScatterNdUpdate	state	✅
ResourceScatterSub	state	✅
ResourceScatterUpdate	state	✅
ScatterAdd	state	✅
ScatterDiv	state	✅
ScatterMax	state	✅
ScatterMin	state	✅
ScatterMul	state	✅
ScatterNd	state	✅
ScatterNdAdd	state	✅
ScatterNdNonAliasingAdd	state	✅
ScatterNdSub	state	✅
ScatterNdUpdate	state	✅
ScatterSub	state	✅
ScatterUpdate	state	✅
TemporaryVariable	state	✅
VarHandleOp	state	✅
Variable	state	✅
VariableShape	state	✅
VariableV2	state	✅
VarIsInitializedOp	state	✅
ApplyAdadelta	training	✅
ApplyAdagrad	training	✅
ApplyAdagradV2	training	✅
ApplyAdam	training	✅
ApplyAdaMax	training	✅
ApplyAddSign	training	✅
ApplyCenteredRMSProp	training	✅
ApplyGradientDescent	training	✅
ApplyMomentum	training	✅
ApplyPowerSign	training	✅
ApplyRMSProp	training	✅
ResourceApplyAdadelta	training	✅
ResourceApplyAdagrad	training	✅
ResourceApplyAdagradV2	training	✅
ResourceApplyAdam	training	✅
ResourceApplyAdaMax	training	✅
ResourceApplyAdamWithAmsgrad	training	✅
ResourceApplyAddSign	training	✅
ResourceApplyCenteredRMSProp	training	✅
ResourceApplyGradientDescent	training	✅
ResourceApplyKerasMomentum	training	✅
ResourceApplyMomentum	training	✅
ResourceApplyPowerSign	training	✅
ResourceApplyRMSProp	training	✅

Operator Types

The operator type in the table above is a rough grouping of ops based on functionality. These labels are not official and only for organization purposes. Many of the names are derived from the kernel source file names.

Op Type	Purpose
array	Tensor transforms such as tile, reshape, slice, join, split, broadcast, generate, etc.
bitwise	Classic unary and binary bitwise ops: and, xor, not, bit count, etc.
complex	Op to support tensors that store complex numbers.
cudnn	Ops specific to CuDNN (no CPU implementation).
dataset	Training/inference data pipeline primitives for efficient storage, retrieval, and batching.
debug	Helpers for working with tfdbg (TensorFlow debugger).
distributed	Distributed/remote execution, including multi-device and multi-node operation.
directml	Ops specific to DirectML (no CPU implementation).
functional	Constructing dynamic graphs with higher-level primitives (tf.function, while_loop, AutoGraph).
image	Encoding, decoding, and working with image formats. Includes color space transforms, cropping, etc.
internal	Internal-only memory transfer optimization ops.
linalg	Classic linear algebra routines: determinants, decomposition, solving systems of equations, etc.
math	General-purpose compute ops: unary, binary, trig functions, reductions, norms, etc.
nn	Neural-network ops: pooling, conv, normalization, activation, etc.
quantization	Ops to support quantization.
random	Random number generation and distributions: uniform, normal, truncated normal, binomial, etc.
rnn	Recurrent neural-network ops, including GRU and LSTM.
signal	Signal processing ops for converting between time and frequency domains (FFTs).
special	Special functions.
state	Graph execution state: variables, resources, mutexes, etc.
training	Optimizers (e.g. SGD, Adam) and training-specific support ops.