Supported Operators
DirectML provides a comprehensive set of operators to enable hardware-accelerated machine learning inference on Windows. This page details the operators currently supported by DirectML, categorized for clarity.
Core Operators
These are fundamental operations used across many machine learning models.
| Operator | Description | Category | Supported Types |
|---|---|---|---|
Add |
Element-wise addition of two tensors. | Arithmetic | FP32, FP16, INT32 |
Sub |
Element-wise subtraction of two tensors. | Arithmetic | FP32, FP16, INT32 |
Mul |
Element-wise multiplication of two tensors. | Arithmetic | FP32, FP16, INT32 |
Div |
Element-wise division of two tensors. | Arithmetic | FP32, FP16 |
Relu |
Rectified Linear Unit activation (max(0, x)). | Activation | FP32, FP16 |
Sigmoid |
Sigmoid activation function. | Activation | FP32, FP16 |
Tanh |
Hyperbolic tangent activation function. | Activation | FP32, FP16 |
Conv2D |
2D Convolution operation. | Convolution | FP32, FP16 |
MaxPool2D |
2D Max Pooling operation. | Pooling | FP32, FP16 |
AvgPool2D |
2D Average Pooling operation. | Pooling | FP32, FP16 |
Gemm |
General Matrix Multiply (often used for Fully Connected layers). | Linear Algebra | FP32, FP16 |
Softmax |
Softmax activation function, typically for classification. | Activation | FP32, FP16 |
BatchNormalization |
Normalizes activations across the batch dimension. | Normalization | FP32, FP16 |
Commonly Used Operators
These operators are frequently found in popular deep learning models.
| Operator | Description | Category | Supported Types |
|---|---|---|---|
Concat |
Concatenates tensors along a specified axis. | Tensor Manipulation | FP32, FP16, INT32 |
Reshape |
Changes the shape of a tensor without changing its data. | Tensor Manipulation | FP32, FP16, INT32 |
Slice |
Extracts a slice from a tensor. | Tensor Manipulation | FP32, FP16, INT32 |
ReduceMean |
Computes the mean of elements across dimensions of a tensor. | Reduction | FP32, FP16 |
ReduceSum |
Computes the sum of elements across dimensions of a tensor. | Reduction | FP32, FP16, INT32 |
MatMul |
Matrix multiplication of two tensors. | Linear Algebra | FP32, FP16 |
Advanced Operators
Support for more specialized or complex operations.
| Operator | Description | Category | Supported Types |
|---|---|---|---|
LayerNormalization |
Normalizes activations across the feature dimension. | Normalization | FP32, FP16 |
LSTM |
Long Short-Term Memory recurrent layer. | Recurrent | FP32, FP16 |
GRU |
Gated Recurrent Unit recurrent layer. | Recurrent | FP32, FP16 |
LSTMCell |
Single LSTM cell operation. | Recurrent | FP32, FP16 |
Note: Support for specific operator versions, data types (e.g., INT8, UINT8), and combinations may vary depending on your DirectML driver version and hardware capabilities. Always refer to the latest DirectML documentation or use the DirectML capabilities query API for the most up-to-date information.
Operator Support Matrix
The table below provides a high-level overview of operator support across different DirectML feature levels. Check the DirectML Release Notes for detailed version compatibility.
| Operator | Feature Level 1.0 | Feature Level 1.1 | Feature Level 1.2 | Feature Level 1.3 |
|---|---|---|---|---|
Add |
✓ | ✓ | ✓ | ✓ |
Sub |
✓ | ✓ | ✓ | ✓ |
Mul |
✓ | ✓ | ✓ | ✓ |
Div |
✗ | ✓ | ✓ | ✓ |
Conv2D |
✓ | ✓ | ✓ | ✓ |
Relu |
✓ | ✓ | ✓ | ✓ |
MaxPool2D |
✓ | ✓ | ✓ | ✓ |
Gemm |
✓ | ✓ | ✓ | ✓ |
Softmax |
✓ | ✓ | ✓ | ✓ |
LSTM |
✗ | ✗ | ✓ | ✓ |
✓ Supported, ✗ Not supported
Contributing to Operator Support
We are continuously expanding DirectML operator support. If you need support for an operator not listed here, please visit the DirectML GitHub repository to submit a feature request or contribute.