CANN/asc-devkit DeepNorm归一化样例

DeepNorm 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 uses the DeepNorm high-level API provided by Ascend C in a deep neural network training scenario to implement DeepNorm normalization. This API performs LayerNorm normalization while improving training stability of deep Transformer networks by scaling the residual connection (alpha coefficient).

Supported Products and CANN Software Versions

Directory Structure

├── deepnorm
│   ├── scripts
│   │   ├── gen_data.py         // Script for generating input data and ground truth data
│   ├── CMakeLists.txt          // Build project file
│   ├── data_utils.h            // Data read/write functions
│   ├── deepnorm.asc            // Ascend C example implementation & invocation example
│   └── README.md               // Example documentation

Example Description

• Example function:
  This example implements DeepNorm normalization for input data with shape [B, S, H]. The computation formula is: $$ DeepNorm(x) = LayerNorm(\alpha * X + SubLayer(X)) $$

• Example specifications:

  Example Type (OpType)	deepnorm
  Example Input
  	name	shape	data type	format
  	inputX	[4, 16, 64]	float	ND
  	inputGx	[4, 16, 64]	float	ND
  	beta	[1, 64]	float	ND
  	gamma	[1, 64]	float	ND
  Example Output
  	output	[4, 16, 64]	float	ND
  	outputMean	[4, 16]	float	ND
  	outputVariance	[4, 16]	float	ND
  Kernel Function Name	deepnorm_custom

• Example implementation:

  • Kernel implementation

    The computation logic is:
    Use the DeepNorm high-level API to complete the deepnorm computation, obtain the final result, and transfer it to external storage. For detailed API information, refer to DeepNorm.

  • Tiling implementation

    The tiling implementation process is as follows:

    1. Call AscendC::GetDeepNormMaxMinTmpSize to obtain and use the minimum temporary space size required by the DeepNorm API.
    2. Call AscendC::GetDeepNormTilingInfo to obtain the tiling parameters required by the kernel-side API based on the input shape and workspace size.

  • Invocation implementation
    Use the kernel invocation operator <<<>>> to call the kernel function.

Build and Run

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

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

  mkdir -p build && cd build;   # Create and enter the build directory
  cmake ..;make -j;             # Build the project (NPU mode by default)
  python3 ../scripts/gen_data.py   # Generate test input data
  ./demo                        # Run the compiled executable to execute the example
  

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

  For example:

  cmake -DCMAKE_ASC_RUN_MODE=cpu -DCMAKE_ASC_ARCHITECTURES=dav-2201 ..;make -j; # CPU debugging mode
  cmake -DCMAKE_ASC_RUN_MODE=sim -DCMAKE_ASC_ARCHITECTURES=dav-2201 ..;make -j; # NPU simulation mode
  

  Notice: Before switching the build mode, clear the cmake cache by running rm CMakeCache.txt in the build directory and then re-run cmake.

• Build option description

  Option	Values	Description
  CMAKE_ASC_RUN_MODE	npu (default), cpu, sim	Run mode: NPU run, CPU debug, NPU simulation
  CMAKE_ASC_ARCHITECTURES	dav-2201 (default), dav-3510	NPU architecture: dav-2201 corresponds to Atlas A2 Training Series Products/Atlas A2 Inference Series Products and Atlas A3 Training Series Products/Atlas A3 Inference Series Products; dav-3510 corresponds to Ascend 950PR/Ascend 950DT

  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