[cann document express issue 05] let you know what operators are

What is an operator

Deep learning algorithm is composed of computing units , We call these computational units operators （Operator, abbreviation OP）. In the network model , The operator corresponds to the computational logic in the layer , for example ： Convolution layer （Convolution Layer） It's an operator ; Fully connected layer （Fully-connected Layer, FC layer） The weight summation process in , It's an operator .

Another example is ：tanh、ReLU etc. , Is an operator used as an activation function in the network model .

The name of the operator （Name） And type （Type）

• The name of the operator ： Identify an operator in a network , The names of operators in the same network should be unique .
• The type of operator ： Each operator in the network implements logical matching according to the operator type , There may be more than one operator of the same type in a network .

As shown in the figure below ,Conv1、Pool1、Conv2 Are the names of operators in this network , among Conv1 And Conv2 The types of operators are Convolution, It means doing convolution calculation once respectively .

tensor （Tensor）

• Tensors are containers for operators to compute data , Including input data and output data .
• Tensor descriptor （TensorDesc） It is a description of input data and output data , It mainly contains the following attributes ：

Next, we will introduce the shape and data layout format in the lower tensor descriptor in detail .

shape （Shape）

The shape of the tensor , Like shape (3,4) Indicates that the first dimension has 3 Elements , The second dimension is 4 Elements , It's a 3 That's ok 4 Matrix array of columns . How many numbers are there in the shape , It means how many dimensions this tensor has . The first element of the shape depends on how many elements are in the outermost bracket of the tensor , The second element of the shape depends on how many elements are in the second bracket starting from the left in the tensor , And so on . for example ：

Let's take a look at the physical meaning of shape , hypothesis shape=(4, 20, 20, 3).

Suppose there is 4 A picture , namely shape in 4 The meaning of , The width and height of each photo are 20, That is to say 20*20=400 Pixel , Each pixel is marked by red / green / blue 3 Color composition , namely shape Inside 3 The meaning of , This is it. shape=(4, 20, 20, 3) The physical meaning of . 

Data layout format

In the field of deep learning , Multidimensional data is stored in multidimensional arrays , For example, the characteristic graph of convolutional neural network （Feature Map） Usually saved in a four-dimensional array , namely 4D Format ：

• N：Batch Number , For example, the number of images .
• H：Height, Feature map height , That is, the number of pixels in the vertical height direction .
• W：Width, Width of feature map , That is, the number of pixels in the horizontal width direction .
• C：Channels, Signature channel , For example, color RGB Graphic Channels by 3.

Because data can only be stored linearly , So these four dimensions have corresponding order . Different deep learning frameworks will store feature map data in different order , such as Caffe, The order is [Batch, Channels, Height, Width], namely NCHW.TensorFlow in , The order is [Batch, Height, Width, Channels], namely NHWC.

In one format RGB For example ,NCHW in ,C On the outside , What's actually stored is “RRRRRRGGGGGGBBBBBB”, That is, all pixel values of the same channel are stored together in sequence ; and NHWC in C On the innermost layer , What is actually stored is “RGBRGBRGBRGBRGBRGB”, That is, the pixel values at the same position of different channels are sequentially stored together .

Although the stored data is the same , But different storage order will lead to inconsistent data access characteristics , So even if you do the same operation , The corresponding computing performance will be different .
It's rising AI In the processor , In order to improve the efficiency of data access , Tensor data are used NC1HWC0 Five dimensional format of . among C0 Strongly related to microarchitecture , be equal to AI Core Size of matrix calculation unit in , This part of data needs to be stored continuously ;C1 Yes, it will C Dimensions are in accordance with C0 Number after splitting , namely C1=C/C0. If you don't divide , The last piece of data needs to be completed to align C0.

More about

Learn more about , You can log in to shengteng community https://www.hiascend.com/, Read the relevant documents ：

Rise CANN The document center is dedicated to providing developers with better content and more convenient development experience , help CANN Developers co build AI ecology . Any comments and suggestions can be fed back in shengteng community , Your every attention is our driving force .