Note

The Deep Neural Network (DNN) component in Intel MKL is deprecated and will be removed in a future release. You can continue to use optimized functions for deep neural networks through Intel Math Kernel Library for Deep Neural Networks.

The type dnnPrimitive_t is a pointer to an opaque structure that describes DNN operations.

To perform a DNN operation, in your application call available Intel MKL DNN functions to do the following:

  1. Create a description of a DNN operation
  2. Perform the DNN operation
  3. Destroy the description of the operation

To create a description of a DNN operation, call a function:

  • dnnConversionCreate_<Data Type> for conversions
  • dnn<Operation>Create<Direction>[<Resource Name>]_<Data Type> for other operations

Here

<Operation> is Convolution, Pooling, Normalization, or ReLU,

<Direction> is Forward or Backward,

<Resource Name>, optional, is Bias, Data, Filter, or an empty string,

<Data Type> is F32 or F64.

To perform a DNN operation, call a dnnExecute_<Data Type> function and pass to it the description of the operation and an array of pointers to operation resources.

Because the dnnConversionExecute_<Data Type> function has a simpler interface than dnnExecute_<Data Type>, if a DNN operation is a conversion, you can alternatively call the dnnExecute_<Data Type> function and pass to it the description of the conversion and pointers to input and output arrays.

To destroy a description of a DNN operation, call a dnnDelete_<Data Type> function.

Notational Conventions

Descriptions of functions that create DNN operations use the following notational conventions:

x, dx
denote dimension-2 spatial indices, where dimension specifies the number of dimensions in the input and output
ic
denotes the index of an input channel
oc
denotes the index of an output channel
bc
denotes the index of an element in a batch
Для получения подробной информации о возможностях оптимизации компилятора обратитесь к нашему Уведомлению об оптимизации.
Выберите цвет&nbsp;фиксируемой кнопки: 
Orange (only for download buttons)