DirectML feature level history

  • Article

For general DirectML version history, see DirectML version history.

DML_FEATURE_LEVEL_6_4

Introduced in DirectML version 1.15.0.

Added the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_RESAMPLE3
  • DML_OPERATOR_FOLD
  • DML_OPERATOR_UNFOLD

Extended the following operators to accept the DML_PADDING_MODE_WRAP padding mode.

  • DML_OPERATOR_PADDING
  • DML_OPERATOR_PADDING1

Updated DML_OPERATOR_ACTIVATION_SOFTPLUS to allow Steepness < 1.

DML_FEATURE_LEVEL_6_3

Introduced in DirectML version 1.15.0

Added the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_MEAN_VARIANCE_NORMALIZATION2
  • DML_OPERATOR_MULTIHEAD_ATTENTION1
  • DML_OPERATOR_QUANTIZE
  • DML_OPERATOR_DEQUANTIZE

Introduced DML_TENSOR_DATA_TYPE_UINT4 and DML_TENSOR_DATA_TYPE_INT4 data types, currently supported by the following operators:

  • DML_OPERATOR_QUANTIZE
  • DML_OPERATOR_DEQUANTIZE

DML_FEATURE_LEVEL_6_2

Introduced in DirectML version 1.13.0.

Added the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ACTIVATION_HARD_SWISH
  • DML_OPERATOR_ACTIVATION_SWISH
  • DML_OPERATOR_AVERAGE_POOLING1
  • DML_OPERATOR_LP_POOLING1
  • DML_OPERATOR_MATRIX_MULTIPLY_INTEGER_TO_FLOAT
  • DML_OPERATOR_QUANTIZED_LINEAR_AVERAGE_POOLING

Extended data type support for the following operators, documented in DML_OPERATOR_TYPE. For details on the specific support added in DML_FEATURE_LEVEL_6_2, see each operator's structure topic.

  • DML_OPERATOR_RESAMPLE2

Made ZeroPointTensor optional for the following operators:

  • DML_OPERATOR_ELEMENT_WISE_DEQUANTIZE_LINEAR
  • DML_OPERATOR_ELEMENT_WISE_QUANTIZE_LINEAR

Added a new graph node type DML_GRAPH_NODE_TYPE_CONSTANT to enable compile-time optimizations that require content of small tensors.

DML_FEATURE_LEVEL_6_1

Introduced in DirectML version 1.12.0.

The operator types mentioned below are documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • Added DML_OPERATOR_MULTIHEAD_ATTENTION.
  • DML_OPERATOR_GEMM. FusedActivation now supports DML_OPERATOR_ACTIVATION_SOFTMAX and DML_OPERATOR_ACTIVATION_SOFTMAX1.

DML_FEATURE_LEVEL_6_0

Introduced in DirectML version 1.11.0.

The operator types mentioned below are documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • Added UINT64 and INT64 data type support for DML_OPERATOR_ELEMENT_WISE_DIVIDE, DML_OPERATOR_ELEMENT_WISE_MODULUS_FLOOR, and DML_OPERATOR_ELEMENT_WISE_MODULUS_TRUNCATE.
  • Added FLOAT16 data type support in ScaleTensor for DML_OPERATOR_ELEMENT_WISE_QUANTIZE_LINEAR.
  • Added FLOAT16 data type support in ScaleTensor and OutputTensor for DML_OPERATOR_ELEMENT_WISE_DEQUANTIZE_LINEAR.
  • Added DML_OPERATOR_ELEMENT_WISE_CLIP operator to the supported fused activation list.

DML_FEATURE_LEVEL_5_2

Introduced in DirectML version 1.10.0.

The operator types mentioned below are documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

The range of tensor dimension has been increased to 1 to 4 for the following parameters:

  • DML_OPERATOR_MATRIX_MULTIPLY_INTEGER, BZeroPointTensor parameter.
  • DML_OPERATOR_QUANTIZED_LINEAR_CONVOLUTION, FilterScaleTensor parameter.

ScaleTensor and BiasTensor can be null independent of each other for the following operators:

  • DML_OPERATOR_MEAN_VARIANCE_NORMALIZATION
  • DML_OPERATOR_MEAN_VARIANCE_NORMALIZATION1

DML_FEATURE_LEVEL_5_1

Introduced in DirectML version 1.9.0.

Added the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ACTIVATION_GELU
  • DML_OPERATOR_ACTIVATION_SOFTMAX1
  • DML_OPERATOR_ACTIVATION_LOG_SOFTMAX1
  • DML_OPERATOR_ACTIVATION_HARDMAX1
  • DML_OPERATOR_RESAMPLE2
  • DML_OPERATOR_RESAMPLE_GRAD1
  • DML_OPERATOR_DIAGONAL_MATRIX1

Extended data type support for the following operators, documented in DML_OPERATOR_TYPE. For details on the specific support added in DML_FEATURE_LEVEL_5_1, see each operator's structure topic.

  • DML_OPERATOR_ACTIVATION_RELU
  • DML_OPERATOR_ACTIVATION_RELU_GRAD
  • DML_OPERATOR_ACTIVATION_PARAMETERIZED_RELU
  • DML_OPERATOR_ELEMENT_WISE_ADD
  • DML_OPERATOR_ELEMENT_WISE_DIVIDE
  • DML_OPERATOR_ELEMENT_WISE_MULTIPLY
  • DML_OPERATOR_ELEMENT_WISE_SUBTRACT
  • DML_OPERATOR_DIAGONAL_MATRIX

DML_FEATURE_LEVEL_5_0

Introduced in DirectML version 1.8.0.

Added the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ELEMENT_WISE_CLIP1
  • DML_OPERATOR_ELEMENT_WISE_CLIP_GRAD1
  • DML_OPERATOR_ELEMENT_WISE_NEGATE
  • DML_OPERATOR_PADDING1

Extended data type support for the following operators, documented in DML_OPERATOR_TYPE. For details on the specific support added in DML_FEATURE_LEVEL_5_0, see each operator's structure topic.

  • DML_OPERATOR_CUMULATIVE_PRODUCT
  • DML_OPERATOR_CUMULATIVE_SUMMATION
  • DML_OPERATOR_DEPTH_TO_SPACE
  • DML_OPERATOR_DEPTH_TO_SPACE1
  • DML_OPERATOR_ELEMENT_WISE_CLIP
  • DML_OPERATOR_ELEMENT_WISE_CLIP_GRAD
  • DML_OPERATOR_ELEMENT_WISE_CLIP_GRAD1
  • DML_OPERATOR_ELEMENT_WISE_CLIP1
  • DML_OPERATOR_ELEMENT_WISE_IF
  • DML_OPERATOR_ELEMENT_WISE_MAX
  • DML_OPERATOR_ELEMENT_WISE_MIN
  • DML_OPERATOR_ELEMENT_WISE_NEGATE
  • DML_OPERATOR_FILL_VALUE_SEQUENCE
  • DML_OPERATOR_MAX_POOLING
  • DML_OPERATOR_MAX_POOLING1
  • DML_OPERATOR_MAX_POOLING2
  • DML_OPERATOR_MAX_UNPOOLING
  • DML_OPERATOR_PADDING
  • DML_OPERATOR_PADDING1
  • DML_OPERATOR_REDUCE, when using one of the following reduce functions.
    • DML_REDUCE_FUNCTION_L1
    • DML_REDUCE_FUNCTION_MAX
    • DML_REDUCE_FUNCTION_MIN
    • DML_REDUCE_FUNCTION_MULTIPLY
    • DML_REDUCE_FUNCTION_SUM
    • DML_REDUCE_FUNCTION_SUM_SQUARE
  • DML_OPERATOR_REVERSE_SUBSEQUENCES
  • DML_OPERATOR_ROI_ALIGN
  • DML_OPERATOR_ROI_ALIGN1
  • DML_OPERATOR_SPACE_TO_DEPTH
  • DML_OPERATOR_SPACE_TO_DEPTH1
  • DML_OPERATOR_TOP_K
  • DML_OPERATOR_TOP_K1

DML_FEATURE_LEVEL_4_1

Introduced in DirectML version 1.7.0.

Added the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ROI_ALIGN_GRAD
  • DML_OPERATOR_BATCH_NORMALIZATION_TRAINING
  • DML_OPERATOR_BATCH_NORMALIZATION_TRAINING_GRAD

Extended data type support for the following operators, documented in DML_OPERATOR_TYPE. For details on the specific support added in DML_FEATURE_LEVEL_4_1, see each operator's structure topic.

  • DML_OPERATOR_ELEMENT_WISE_IDENTITY
  • DML_OPERATOR_ELEMENT_WISE_ADD
  • DML_OPERATOR_ELEMENT_WISE_SUBTRACT
  • DML_OPERATOR_ELEMENT_WISE_MULTIPLY
  • DML_OPERATOR_ELEMENT_WISE_ABS
  • DML_OPERATOR_ELEMENT_WISE_SIGN
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_EQUALS
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_GREATER_THAN
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_LESS_THAN
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_GREATER_THAN_OR_EQUAL
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_LESS_THAN_OR_EQUAL
  • DML_OPERATOR_ELEMENT_WISE_BIT_SHIFT_LEFT
  • DML_OPERATOR_ELEMENT_WISE_BIT_SHIFT_RIGHT
  • DML_OPERATOR_ELEMENT_WISE_BIT_AND
  • DML_OPERATOR_ELEMENT_WISE_BIT_OR
  • DML_OPERATOR_ELEMENT_WISE_BIT_NOT
  • DML_OPERATOR_ELEMENT_WISE_BIT_XOR
  • DML_OPERATOR_ELEMENT_WISE_BIT_COUNT
  • DML_OPERATOR_ARGMIN
  • DML_OPERATOR_ARGMAX
  • DML_OPERATOR_CAST
  • DML_OPERATOR_SLICE
  • DML_OPERATOR_SLICE1
  • DML_OPERATOR_SLICE_GRAD
  • DML_OPERATOR_SPLIT
  • DML_OPERATOR_JOIN
  • DML_OPERATOR_GATHER
  • DML_OPERATOR_GATHER_ELEMENTS
  • DML_OPERATOR_GATHER_ND
  • DML_OPERATOR_GATHER_ND1
  • DML_OPERATOR_SCATTER
  • DML_OPERATOR_SCATTER_ND
  • DML_OPERATOR_FILL_VALUE_CONSTANT
  • DML_OPERATOR_TILE
  • DML_OPERATOR_ONE_HOT

DML_FEATURE_LEVEL_4_0

Introduced in DirectML version 1.6.0.

Added support for the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ELEMENT_WISE_QUANTIZED_LINEAR_ADD
  • DML_OPERATOR_DYNAMIC_QUANTIZE_LINEAR
  • DML_OPERATOR_ROI_ALIGN1

Extended data type and dimension count support for the following operators, documented in DML_OPERATOR_TYPE. For details on the specific support added in DML_FEATURE_LEVEL_4_0, see each operator's structure topic.

  • DML_OPERATOR_ACTIVATION_RELU_GRAD
  • DML_OPERATOR_ADAM_OPTIMIZER
  • DML_OPERATOR_CONVOLUTION
  • DML_OPERATOR_CONVOLUTION_INTEGER
  • DML_OPERATOR_CUMULATIVE_PRODUCT
  • DML_OPERATOR_CUMULATIVE_SUMMATION
  • DML_OPERATOR_DIAGONAL_MATRIX
  • DML_OPERATOR_FILL_VALUE_CONSTANT
  • DML_OPERATOR_FILL_VALUE_SEQUENCE
  • DML_OPERATOR_GEMM
  • DML_OPERATOR_MATRIX_MULTIPLY_INTEGER
  • DML_OPERATOR_MAX_POOLING_GRAD
  • DML_OPERATOR_NONZERO_COORDINATES
  • DML_OPERATOR_QUANTIZED_LINEAR_CONVOLUTION
  • DML_OPERATOR_QUANTIZED_LINEAR_MATRIX_MULTIPLY
  • DML_OPERATOR_RANDOM_GENERATOR
  • DML_OPERATOR_REVERSE_SUBSEQUENCES

DML_FEATURE_LEVEL_3_1

Introduced in DirectML version 1.5.0.

Added support for the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ELEMENT_WISE_ATAN_YX
  • DML_OPERATOR_ELEMENT_WISE_CLIP_GRAD
  • DML_OPERATOR_ELEMENT_WISE_DIFFERENCE_SQUARE
  • DML_OPERATOR_LOCAL_RESPONSE_NORMALIZATION_GRAD
  • DML_OPERATOR_CUMULATIVE_PRODUCT
  • DML_OPERATOR_BATCH_NORMALIZATION_GRAD

The maximum number of supported dimensions for the following operators has increased from 4 to 8.

  • DML_OPERATOR_BATCH_NORMALIZATION
  • DML_OPERATOR_CAST
  • DML_OPERATOR_JOIN
  • DML_OPERATOR_LP_NORMALIZATION
  • DML_OPERATOR_MEAN_VARIANCE_NORMALIZATION1
  • DML_OPERATOR_PADDING
  • DML_OPERATOR_ACTIVATION_RELU_GRAD
  • DML_OPERATOR_SLICE_GRAD
  • DML_OPERATOR_TILE
  • DML_OPERATOR_TOP_K
  • DML_OPERATOR_TOP_K1

DML_FEATURE_LEVEL_3_0

Introduced in DirectML version 1.4.0.

Added support for the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ELEMENT_WISE_BIT_AND
  • DML_OPERATOR_ELEMENT_WISE_BIT_OR
  • DML_OPERATOR_ELEMENT_WISE_BIT_XOR
  • DML_OPERATOR_ELEMENT_WISE_BIT_NOT
  • DML_OPERATOR_ELEMENT_WISE_BIT_COUNT
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_GREATER_THAN_OR_EQUAL
  • DML_OPERATOR_ELEMENT_WISE_LOGICAL_LESS_THAN_OR_EQUAL
  • DML_OPERATOR_ACTIVATION_CELU
  • DML_OPERATOR_ACTIVATION_RELU_GRAD
  • DML_OPERATOR_AVERAGE_POOLING_GRAD
  • DML_OPERATOR_MAX_POOLING_GRAD
  • DML_OPERATOR_RANDOM_GENERATOR
  • DML_OPERATOR_NONZERO_COORDINATES
  • DML_OPERATOR_RESAMPLE_GRAD
  • DML_OPERATOR_SLICE_GRAD
  • DML_OPERATOR_ADAM_OPTIMIZER
  • DML_OPERATOR_ARGMIN
  • DML_OPERATOR_ARGMAX
  • DML_OPERATOR_ROI_ALIGN
  • DML_OPERATOR_GATHER_ND1

Added the following enhancements.

  • The maximum number of tensor dimensions has been increased from 5 to 8. See DML_TENSOR_DIMENSION_COUNT_MAX1.
  • Additional support for integer datatypes has been added to the following operators.
    • DML_OPERATOR_ELEMENT_WISE_POW
    • DML_OPERATOR_ELEMENT_WISE_CONSTANT_POW
    • DML_OPERATOR_MAX_POOLING, DML_OPERATOR_MAX_POOLING1, and DML_OPERATOR_MAX_POOLING2
    • DML_OPERATOR_REDUCE, when using DML_REDUCE_FUNCTION_ARGMIN or DML_REDUCE_FUNCTION_ARGMAX
  • The following 64-bit data types have been added, and are supported by select operators.
    • DML_TENSOR_DATA_TYPE_FLOAT64
    • DML_TENSOR_DATA_TYPE_UINT64
    • DML_TENSOR_DATA_TYPE_INT64

Deprecated functionality.

  • DML_REDUCE_FUNCTION_ARGMAX and DML_REDUCE_FUNCTION_ARGMIN have been deprecated. You should prefer to use the standalone DML_OPERATOR_ARGMIN and DML_OPERATOR_ARGMAX operators in their place.

DML_FEATURE_LEVEL_2_1

Introduced in DirectML version 1.2.0.

Added the following APIs.

Added support for the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ELEMENT_WISE_BIT_SHIFT_LEFT
  • DML_OPERATOR_ELEMENT_WISE_BIT_SHIFT_RIGHT
  • DML_OPERATOR_ELEMENT_WISE_ROUND
  • DML_OPERATOR_ELEMENT_WISE_IS_INFINITY
  • DML_OPERATOR_ELEMENT_WISE_MODULUS_TRUNCATE
  • DML_OPERATOR_ELEMENT_WISE_MODULUS_FLOOR
  • DML_OPERATOR_FILL_VALUE_CONSTANT
  • DML_OPERATOR_FILL_VALUE_SEQUENCE
  • DML_OPERATOR_CUMULATIVE_SUMMATION
  • DML_OPERATOR_REVERSE_SUBSEQUENCES
  • DML_OPERATOR_GATHER_ELEMENTS
  • DML_OPERATOR_GATHER_ND
  • DML_OPERATOR_SCATTER_ND
  • DML_OPERATOR_MAX_POOLING2
  • DML_OPERATOR_SLICE1
  • DML_OPERATOR_TOP_K1
  • DML_OPERATOR_DEPTH_TO_SPACE1
  • DML_OPERATOR_SPACE_TO_DEPTH1
  • DML_OPERATOR_MEAN_VARIANCE_NORMALIZATION1
  • DML_OPERATOR_RESAMPLE1
  • DML_OPERATOR_MATRIX_MULTIPLY_INTEGER
  • DML_OPERATOR_QUANTIZED_LINEAR_MATRIX_MULTIPLY
  • DML_OPERATOR_CONVOLUTION_INTEGER
  • DML_OPERATOR_QUANTIZED_LINEAR_CONVOLUTION

Added the following enhancements.

  • Additional support for integer datatypes has been added to the following operators.
    • DML_OPERATOR_ELEMENT_WISE_IDENTITY
    • DML_OPERATOR_ELEMENT_WISE_ABS
    • DML_OPERATOR_ELEMENT_WISE_ADD
    • DML_OPERATOR_ELEMENT_WISE_CLIP
    • DML_OPERATOR_ELEMENT_WISE_DIVIDE
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_EQUALS
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_GREATER_THAN
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_LESS_THAN
    • DML_OPERATOR_ELEMENT_WISE_MAX
    • DML_OPERATOR_ELEMENT_WISE_MEAN
    • DML_OPERATOR_ELEMENT_WISE_MIN
    • DML_OPERATOR_ELEMENT_WISE_MULTIPLY
    • DML_OPERATOR_ELEMENT_WISE_SUBTRACT
    • DML_OPERATOR_ELEMENT_WISE_THRESHOLD
    • DML_OPERATOR_ELEMENT_WISE_QUANTIZE_LINEAR
    • DML_OPERATOR_ELEMENT_WISE_DEQUANTIZE_LINEAR
    • DML_OPERATOR_ELEMENT_WISE_SIGN
    • DML_OPERATOR_ELEMENT_WISE_IF
    • DML_OPERATOR_ACTIVATION_SHRINK
    • DML_OPERATOR_PADDING
    • DML_OPERATOR_GATHER
    • DML_OPERATOR_SCATTER
    • DML_OPERATOR_DEPTH_TO_SPACE
    • DML_OPERATOR_SPACE_TO_DEPTH
    • DML_OPERATOR_TILE
    • DML_OPERATOR_TOP_K and DML_OPERATOR_TOP_K1
    • DML_OPERATOR_ONE_HOT
    • DML_OPERATOR_REDUCE, when using one of the following reduce functions.
      • DML_REDUCE_FUNCTION_ARGMIN
      • DML_REDUCE_FUNCTION_ARGMAX
      • DML_REDUCE_FUNCTION_MAX
      • DML_REDUCE_FUNCTION_MIN
      • DML_REDUCE_FUNCTION_MULTIPLY
      • DML_REDUCE_FUNCTION_SUM
  • Relaxed tensor shape restrictions for DML_OPERATOR_GATHER

DML_FEATURE_LEVEL_2_0

Introduced in DirectML version 1.1.0.

Added the following APIs.

Added support for the following operator types, documented in DML_OPERATOR_TYPE. For each operator type constant, that topic provides a link to the corresponding structure.

  • DML_OPERATOR_ELEMENT_WISE_SIGN
  • DML_OPERATOR_ELEMENT_WISE_IS_NAN
  • DML_OPERATOR_ELEMENT_WISE_ERF
  • DML_OPERATOR_ELEMENT_WISE_SINH
  • DML_OPERATOR_ELEMENT_WISE_COSH
  • DML_OPERATOR_ELEMENT_WISE_TANH
  • DML_OPERATOR_ELEMENT_WISE_ASINH
  • DML_OPERATOR_ELEMENT_WISE_ACOSH
  • DML_OPERATOR_ELEMENT_WISE_ATANH
  • DML_OPERATOR_ELEMENT_WISE_IF
  • DML_OPERATOR_ELEMENT_WISE_ADD1
  • DML_OPERATOR_ACTIVATION_SHRINK
  • DML_OPERATOR_MAX_POOLING1
  • DML_OPERATOR_MAX_UNPOOLING
  • DML_OPERATOR_DIAGONAL_MATRIX
  • DML_OPERATOR_SCATTER_ELEMENTS
  • DML_OPERATOR_SCATTER
  • DML_OPERATOR_ONE_HOT
  • DML_OPERATOR_RESAMPLE

Added the following enhancements.

  • When binding an input resource for dispatch of an IDMLOperatorInitializer, it's now legal to provide a resource with D3D12_HEAP_TYPE_CUSTOM (in addition to D3D12_HEAP_TYPE_DEFAULT), as long as appropriate heap properties are also set. See Binding in DirectML.
  • The following logical boolean operators now support UINT8 output tensors, in addition to the existing support for UINT32.
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_AND
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_EQUALS
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_GREATER_THAN
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_LESS_THAN
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_NOT
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_OR
    • DML_OPERATOR_ELEMENT_WISE_LOGICAL_XOR
  • 5D activation functions now support the use of strides on their input and output tensors.

DML_FEATURE_LEVEL_1_0

The feature level in which DirectML was introduced.

See also