Basic Vector Operators

Operator Analysis

The steps are as follows:

1. Specify the mathematical expression and computation logic of an operator.

   The Add operator is used as an example.

   z = x + y

   The computation logic is as follows: The operation elements of the vector computation API provided by Ascend C are all LocalTensor. The input data needs to be moved from the external storage (Global Memory) to the on-chip storage (Unified Buffer), and then the computation API is called to add the two input parameters to obtain the final result. Then, the final result is moved to the external storage. The following figure shows the computation logic of the Ascend C Add operator.

   Figure 2 Operator computation logic
   

2. Specify the input and output.
   • The Add operator has two inputs, x and y, and outputs the result z.
   • In this sample, the supported input data type is half (float16), and so is the output data type.
   • The supported input shape is (1, 2048), and so is the output shape.
   • The supported input format is ND.

3. Define the kernel function name and parameters.
   • You can customize the kernel function name. In this example, the kernel function is named vec_add_custom.
   • Based on the analysis of the operator input and output, the kernel function has three parameters: x, y, and z. x and y indicate the memory address of the input in the global memory, and z indicates the memory address of the output in the global memory.

4. Specify the APIs required for operator implementation.
   • Move data between the local memory and global memory. For details about the data movement API, see the Ascend C API reference. Call DataCopy to move data.
   • This sample only involves the addition operation in vector computation. For details, see the vector computation API in the Ascend C API reference. Based on the preliminary analysis, the basic arithmetic Add API Add can be used to implement x+y.
   • The tensor data structure used in computation is managed by the queue, involving calling APIs such as EnQue and DeQue.

Based on the preceding analysis, the design specifications of the Ascend C Add operator are as follows.

• Operator type (OpType): Add
• Operator input and output:

  Table 1 Input and output specifications of the Add operator

  name	shape	data type	format
  x (input)	(1, 2048)	half	ND
  y (input)	(1, 2048)	half	ND
  z (output)	(1, 2048)	half	ND

• Kernel function name: vec_add_custom
• Main APIs:
  • DataCopy: data movement API
  • Add: vector basic arithmetic API
  • EnQue, DeQue, and others: queue management APIs
• Operator implementation file: vector_add.asc

Kernel Function Definition

Define kernel functions based on the rules described in Kernel Function.

1. Define the function prototype.

   In this sample, the function name is vector_add_custom (the kernel function name can be customized). Based on the analysis of the operator input and output in Operator Analysis, there are three parameters x, y, and z, where x and y indicate the input memory and z indicates the output memory. According to the rules described in Kernel Function, the function prototype is defined as follows: The __global__ function type qualifier is used to identify a kernel function that can be called by <<<>>>. The __vector__ function type qualifier is used to identify that the kernel function is executed on the Vector Core of the AI Core on the device. For convenience, the GM_ADDR macro is used to modify input parameters. For details about the definition of the GM_ADDR macro, see Kernel Function.

   1 2 3

   __global__ __vector__ void vector_add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z) { }

2. Call the Init and Process functions of the operator class.
   The Init function of the operator class completes memory initialization, and the Process function completes the core logic of operator implementation. For details, see Operator Class Implementation.

   1 2 3 4 5 6 7

   __global__ __vector__ void vector_add_custom(GM_ADDR x, GM_ADDR y, GM_ADDR z) { AscendC::TPipe pipe; KernelAdd op; op.Init(x, y, z, &pipe); op.Process(); }

3. According to Defining and Calling Kernel Functions, when the kernel function is called, in addition to the x, y, and z parameters, numBlocks (number of cores executed by the kernel function), nullptr (reserved parameter, set to nullptr), and stream (stream for maintaining the execution sequence of asynchronous operations in the application) also need to be input, to specify the execution configuration of the kernel function.

   1	vector_add_custom<<<numBlocks, nullptr, stream>>>(xDevice, yDevice, zDevice);

Operator Class Implementation

The kernel function calls the Init and Process functions of the operator class. This section describes how to implement the operator class based on the programming paradigm.

The following describes how to design the implementation process of the Add operator based on the vector programming paradigm. For details about the vector programming paradigm, see Vector Programming Paradigm. For details about the implementation process of the Add operator, see Figure 3 Add operator implementation process.

• The implementation process of the Add operator consists of three basic tasks: CopyIn, Compute, and CopyOut. The CopyIn task moves the input tensors xGm and yGm from the global memory to the local memory and stores them in xLocal and yLocal, respectively. The Compute task performs the addition operation on xLocal and yLocal, and stores the computation result in zLocal. The CopyOut task moves the output data from zLocal to the output tensor zGm in the global memory.
• CopyIn and Compute tasks synchronize with each other through the VECIN queues inQueueX and inQueueY. Compute and CopyOut tasks synchronize with each other through the VECOUT queue outQueueZ.
• The memory used for interaction between tasks and the memory used by temporary variables are managed using the pipe memory management object.

Figure 3 Add operator implementation process

The process is mainly implemented by the operator class, including the initialization function Init and core processing function Process. The Process function calls the three basic tasks. Some private members used in operator implementation are also included, such as the GlobalTensor (xGm, yGm, and zGm) and VECIN and VECOUT queues. Members of the KernelAdd operator class are as follows:

 1
 2
 3
 4
 5
 6
 7
 8
 9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25

class KernelAdd {
public:
    __aicore__ inline KernelAdd() {}
    // Initialization function, which is used to initialize memory.
    __aicore__ inline void Init(GM_ADDR x, GM_ADDR y, GM_ADDR z, AscendC::TPipe* pipeIn){}
    // Core processing function, which implements the operator logic and calls the private member functions CopyIn, Compute, and CopyOut to complete the three-level pipeline operation of the vector operator.
    __aicore__ inline void Process(){}

private:
    // Copy-in function, which is called by the core Process function to complete the processing in the CopyIn phase.
    __aicore__ inline void CopyIn(){}
    // Compute function, which is called by the core Process function to complete the processing in the Compute phase.
    __aicore__ inline void Compute(){}
    // Copy-out function, which is called by the core Process function to complete the processing in the CopyOut phase.
    __aicore__ inline void CopyOut(){}

private:
    AscendC::TPipe pipe;  // Pipe memory management object.
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueX;  // Queue management object of the input data. The value of TPosition is VECIN.
    AscendC::TQue<AscendC::TPosition::VECIN, 1> inQueueY;  // Queue management object of the input data. The value of TPosition is VECIN.
    AscendC::TQue<AscendC::TPosition::VECOUT, 1> outQueueZ;  // Queue management object of the output data. The value of TPosition is VECOUT.
    AscendC::GlobalTensor<half> xGm;  // Object for managing the global memory addresses of the input and output data. xGm and yGm are input, and zGm is output.
    AscendC::GlobalTensor<half> yGm;
    AscendC::GlobalTensor<half> zGm;
};