CANN/asc-devkit Subnormal示例

Subnormal 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 demonstrates the behavioral differences of the Div operation in Ascend C Reg vector computation when Subnormal mode is enabled or disabled. Through two different Div precision algorithm configurations (PRECISION_1ULP_FTZ_FALSE and PRECISION_1ULP_FTZ_TRUE), it shows the impact of Subnormal floating-point number support on computation results.

Supported Products and CANN Versions

Directory Structure

subnormal
│   ├── scripts/             // Test script directory
│   │   └── gen_data.py      // Generate test input and golden data
│   ├── CMakeLists.txt       // Build configuration file
│   ├── data_utils.h         // Data read/write utility functions
│   ├── subnormal.asc        // Ascend C operator implementation & invocation example
│   └── README.md            // Example description document

Example Description

• Example functionality:

  The example computes float type data, using PRECISION_1ULP_FTZ_FALSE and PRECISION_1ULP_FTZ_TRUE to configure DivSpecificMode as the precision algorithm parameter for the Div operation, executing division. The computation formula is as follows: $$z_i = \frac{x_i}{y_i}$$

• Example specifications

  Example Type (OpType)	AIV Example
  Example Input	name	shape	data type
  	x	[1024]	float
  	y	[1024]	float
  Example Output	z	[1024]	float
  Kernel Function Name	subnormal

• Example implementation

  • Implementation flow:

    1. Allocate memory on the Host side and initialize input data
    2. Transfer data from Host memory to Device global memory
    3. Call the kernel function to execute computation on the Vector core
    4. Inside the kernel function: transfer input data from Global Memory to Unified Buffer via DataCopy
    5. Call the VF (Vector Function) function via asc_vf_call on the Unified Buffer
    6. Inside the VF function, use the DivSpecificMode template parameter to perform LoadAlign, Div, and StoreAlign operations
    7. Transfer computation results from Unified Buffer back to Global Memory
    8. Perform precision verification and golden data comparison on the Host side

  • Subnormal and FTZ concept description:

    • Subnormal: Floating-point numbers where the exponent is all zeros and the mantissa is not all zeros, representing extremely small values close to 0. The IEEE 754 standard supports Subnormal numbers, providing better numerical precision and gradual underflow characteristics.
    • FTZ: A mode that forces Subnormal numbers to 0. Although it simplifies hardware design and improves performance, it sacrifices numerical precision.

  • Multi-scenario description:

    • Scenario 1 (SCENARIO_NUM=1): Use PRECISION_1ULP_FTZ_FALSE as the precision algorithm parameter for Div

      • Implementation function: SubnormalFTZFalseVF
      • Implementation:

        static constexpr AscendC::Reg::DivSpecificMode mode = {ZEROING, false, PRECISION_1ULP_FTZ_FALSE};
        AscendC::Reg::Div<T, &mode>(zReg, xReg, yReg, mask);
        

      • Description: Supports Subnormal data computation, uses a single instruction to compute the result, with a maximum precision error of 1 ulp. When the division result is a Subnormal number, the value is preserved.
      • Application scenarios: Scientific computing, numerical simulation, and high-precision computation scenarios that require precise handling of Subnormal floating-point numbers.

    • Scenario 2 (SCENARIO_NUM=2): Use PRECISION_1ULP_FTZ_TRUE as the precision algorithm parameter for Div

      • Implementation function: SubnormalFTZTrueVF
      • Implementation:

        static constexpr AscendC::Reg::DivSpecificMode mode = {ZEROING, false, PRECISION_1ULP_FTZ_TRUE};
        AscendC::Reg::Div<T, &mode>(zReg, xReg, yReg, mask);
        

      • Description: Does not support Subnormal data computation. FTZ mode uses a single instruction to compute the result, with a maximum precision error of 1 ulp. When the division result is a Subnormal number, it is forced to 0.
      • Application scenarios: Deep learning, image processing, and real-time inference scenarios where Subnormal precision requirements are not high, simplifying hardware implementation and improving computation performance.

    • Computation data scenario comparison:

      No.	Div Combination	Dividend	Divisor	Scenario 1 (golden match)	Analysis	Scenario 2 Output (precision loss)	Analysis
      1	N / N = S	$2^{-126}$	4.0	0x00200000	Result is subnormal $2^{-128}$	0x00000000	Result falls in Subnormal range, forced to zero
      2	N / S = N	$2^{-125}$	$2^{-127}$	0x40800000	Result is normal $2^{2}$	0x7f800000	Divisor S treated as 0, any non-zero positive divided by 0 is infinity
      3	S / N = S	$2^{-130}$	2.0	0x00040000	Result is subnormal $2^{-131}$	0x00000000	Dividend S treated as 0 on input, result is 0
      4	S / N = N	$2^{-130}$	$2^{-20}$	0x08800000	Result is normal $2^{110}$	0x00000000	Dividend S treated as 0 on input, result is 0
      5	S / S = N	$2^{-140}$	$2^{-145}$	0x42000000	Result is normal $2^{5}$	0x7fffffff	Both dividend S and divisor S treated as 0, $0 \div 0$, result is NaN

      • N: normal data, normal number.
      • S: Subnormal data, subnormal number.

  • Constraints:

    • Input data length must be a multiple of GetVecLen()
    • The divisor cannot be 0. In practical applications, a check for zero divisors is required
    • The example currently only supports the float data type

Build and Run

• 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 2: Use PRECISION_1ULP_FTZ_TRUE (FTZ mode)
  SCENARIO_NUM=2
  mkdir -p build && cd build;                                               # Create and enter the build directory
  cmake .. -DSCENARIO_NUM=$SCENARIO_NUM; make -j;                           # Build the project
  python3 ../scripts/gen_data.py;                                           # Generate test input data
  ./demo                                                                    # Run 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 ..;make -j; # CPU debug mode
  cmake -DCMAKE_ASC_RUN_MODE=sim -DCMAKE_ASC_ARCHITECTURES=dav-3510 ..;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
  SCENARIO_NUM	1 (default), 2	1: Supports subnormal computation; 2: Does not support subnormal computation;
  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

• 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