CANN/asc-devkit寄存器加载存储掩码示例

ld_st_reg_mask Example

【免费下载链接】asc-devkit 本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言，原生支持C和C++标准规范，主要由类库和语言扩展层构成，提供多层级API，满足多维场景算子开发诉求。 项目地址: https://gitcode.com/cann/asc-devkit

Overview

This example implements UB (Unified Buffer) load/store operations for MaskReg (mask register) based on the Reg programming interface, as well as masked store operations using mask. It supports multiple scenarios. Select a scenario through the CMake build parameter SCENARIO_NUM.

Supported Products and CANN Versions

Directory Structure

├── reg_load_store_mask
│   ├── scripts
│   │   └── gen_data.py                // Input data and ground truth generation script
│   ├── figures                        // Illustrations
│   ├── CMakeLists.txt                 // Build configuration file
│   ├── data_utils.h                   // Data read/write functions
│   ├── README.md                      // Example description
│   ├── ld_st_reg_mask_scenario_1.asc  // Scenario 1: Ascend C implementation for basic transfer scenario
│   └── ld_st_reg_mask_scenario_2.asc  // Scenario 2: Ascend C implementation for composite computation scenario

Example Description

Scenario 1: Basic Transfer Scenario

• Example functionality:
  • Data load: Takes the first 256 bits of data from the input matrix and calls the LoadAlign API to load data from UB into MaskReg.
  • Data store: Sets all bits of MaskReg to 1 and calls the StoreAlign API to store data from MaskReg to UB.
• Example specifications:

  Example Type (OpType)	AIV Example
  Example Input	name	shape	data type
  	x	[1, 1024]	uint8_t
  Example Output	y	[1, 1024]	uint8_t
  Kernel Function Name	ld_st_reg_mask_scenario_1

• Example implementation:

1. In the CopyVF function, call the LoadAlign API to load 256 bits (32*uint8_t) of data from UB into MaskReg to achieve dynamic mask setting. In this example, the 32 uint8_t values are set to 1,0,...,1, meaning the first and last values are 1 (b'00000001). Since the chip reads numbers from the low bit each time, these 32 numbers are ultimately filled into MaskReg as b'1000...1...000.
2. Call the Duplicate API for data filling. MaskReg can indicate which elements participate in computation. From step 1, the 1st bit and the 249th bit in MaskReg are 1. Using this mask, only the 1st and 249th numbers in RegTensor are filled with the value 2.
3. Use the StoreAlign API to save the results from RegTensor to UB.
4. Set all bits in MaskReg to 1, and save the data from MaskReg to UB (address = save address from step 3 + 256B) through the StoreAlign API, implementing the function of storing MaskReg data on UB. Corresponding to 32 uint8_t values, each bit of each value is 1, so each value is 255 (0xFF).

Scenario 2: Composite Computation Scenario

• Example functionality: Selects values from the xReg or yReg vector at corresponding positions based on the bits in maskReg. When a mask bit is 1, the corresponding element from src0 is selected; when a mask bit is 0, the corresponding element from src1 is selected.

• Example specifications:

  Example Type (OpType)	AIV Example
  Example Input	name	shape	data type
  	x	[1, 256]	float
  	y	[1, 256]	float
  	mask	[1, 32]	uint8_t
  Example Output	z	[1, 256]	float
  Kernel Function Name	ld_st_reg_mask_scenario_2

• Example implementation: In the SelectVF function, call the LoadAlign API to load mask data from UB into MaskReg, then pass it to the Select API for computation, writing the result back to UB.

• Invocation implementation Uses the kernel invocation syntax <<<>>> to call the kernel function.

Build and Run

Run the following steps in the root directory of this example to build and run it.

• Configure environment variables Configure environment variables based on the installation method of the CANN development kit on the current environment.

  source ${install_path}/cann/set_env.sh
  

  Note: ${install_path} is the CANN package installation directory. When no installation directory is specified, the default installation path is /usr/local/Ascend.

• Run the example

  Run the following commands in the example directory.

  SCENARIO_NUM=1                                                                # Set the scenario number
  mkdir -p build && cd build;                                                   # Create and enter the build directory
  cmake -DCMAKE_ASC_ARCHITECTURES=dav-3510 -DSCENARIO_NUM=$SCENARIO_NUM ..;make -j; # Build the project (default npu mode)
  python3 ../scripts/gen_data.py -scenarioNum=$SCENARIO_NUM                     # Generate test input data
  ./demo                                                                        # Run the compiled executable to execute the example
  

  To use CPU debug or NPU simulation mode, add the -DCMAKE_ASC_RUN_MODE=cpu or -DCMAKE_ASC_RUN_MODE=sim parameter.

  Examples:

  cmake -DCMAKE_ASC_RUN_MODE=cpu -DCMAKE_ASC_ARCHITECTURES=dav-3510 -DSCENARIO_NUM=$SCENARIO_NUM ..;make -j; # CPU debug mode
  cmake -DCMAKE_ASC_RUN_MODE=sim -DCMAKE_ASC_ARCHITECTURES=dav-3510 -DSCENARIO_NUM=$SCENARIO_NUM ..;make -j; # NPU simulation mode
  

  Notice: Clear the cmake cache before switching build modes. Run rm CMakeCache.txt in the build directory and re-run cmake.

• Build option description

  Option	Values	Description
  CMAKE_ASC_RUN_MODE	npu (default), cpu, sim	Run mode: NPU execution, CPU debug, NPU simulation
  CMAKE_ASC_ARCHITECTURES	dav-3510	NPU architecture: dav-3510 corresponds to Ascend 950PR/Ascend 950DT
  SCENARIO_NUM	1, 2	Scenario number: 1=basic transfer scenario, 2=composite computation scenario

• Execution result The following execution result indicates that the precision comparison is successful.

  test pass!
  

【免费下载链接】asc-devkit 本项目是CANN 推出的昇腾AI处理器专用的算子程序开发语言，原生支持C和C++标准规范，主要由类库和语言扩展层构成，提供多层级API，满足多维场景算子开发诉求。 项目地址: https://gitcode.com/cann/asc-devkit