OpenACC 笔记

贡献者： addis

预备知识　C++ 基础

OpenACC 可以像 OpenMP 那样，通过添加少量的 #pragma（预编译指令）给已有的程序进行 GPU 加速支持在 Linux 上运行的 C++ 和 Fortran 程序。本文以 C++ 为例。它由于易用性，它常用于高性能科学计算中。
和 OpenMP 一样，OpenACC 需要编译器支持。一般使用英伟达显卡专用的 NVIDIA HPC SDK（包含 CUDA Toolkit）（202301：暂不支持 Windows）或者 GCC。
【ubuntu22.04 原装 g++ 测试失败】运行中显示 Not executing on GPU.，应该需要安装 OpenACC 版本的 GCC，或者直接用 NVIDIA SDK。

先确保 NVIDIA GPU 驱动安好，nvidia-smi 能用，然后安装好 NVIDIA SDK 以后，要在 .bashrc 最后添加以下几行设置环境（见这个文档）

NVARCH=`uname -s`_`uname -m`; export NVARCH
NVCOMPILERS=/opt/nvidia/hpc_sdk; export NVCOMPILERS
MANPATH=$MANPATH:$NVCOMPILERS/$NVARCH/23.11/compilers/man; export MANPATH
PATH=$NVCOMPILERS/$NVARCH/23.11/compilers/bin:$PATH; export PATH

source ~/.bashrc 以后，就可以试试 nvc --version，nvcc --version 了。
nvc/nvc++ 编译器支持 OpenACC 和 OpenMP，而 nvcc 则用于 CUDA 编程。

1. 入门例程

【装了 GPU 驱动仍然编译失败】在 GCC 中，直接加上 -fopenacc 选项即可！例如 g++ -fopenacc test1.cpp -o test1.x。在编译时可以加上 -fopt-info 选项输出更多信息。
【测试成功】用 Nvidia 编译器如 nvc -acc -gpu=managed -Minfo=accel test1.cpp -o test1.x

代码 1：test1.cpp

#include <iostream>
#include <vector>
#include <openacc.h>

// Function to initialize the vectors with values
void initialize(std::vector<double>& a, std::vector<double>& b, int n) {
	for(int i = 0; i < n; ++i) {
		a[i] = static_cast<double>(i);
		b[i] = static_cast<double>(2 * i);
	}
}

// detect if GPU is actually running
void detect_gpu()
{
	double a[100], b[100];
	#pragma acc parallel loop
	for (int i = 0; i < 100; ++i) {
		if (i == 10) {
			if (acc_on_device(acc_device_not_host))
				printf("Executing on GPU.\n");
			else
				printf("Not executing on GPU.\n");
		}
		a[i] += b[i];
	}
}

int main() {
	const int n = 1000000; // Size of the vectors
	std::vector<double> a(n), b(n), c(n);
	double *pa = a.data(), *pb = b.data(), *pc = c.data();

	// Initialize vectors a and b
	initialize(a, b, n);

	detect_gpu();

	// Using OpenACC to offload the following computation to an accelerator
	// and explicitly handle data movement
#pragma acc data copyin(pa[0:n], pb[0:n]) copyout(pc[0:n])
	{
#pragma acc parallel loop
		for(int i = 0; i < n; ++i)
			pc[i] = pa[i] + pb[i];
	}

	// Display the first 10 results
	for(int i = 0; i < 10; ++i) {
		std::cout << "c[" << i << "] = " << c[i] << std::endl;
	}
}

2. 编译器选项

-ta=tesla: Compiler option to target NVIDIA Tesla GPUs.
-Minfo=accel: Provides feedback about the code generated by the compiler.

3. 常用命令

循环

#pragma acc parallel: GPU 并行运算
#pragma acc kernels: Identifies a code block for parallelization, allowing the compiler to automatically manage parallelism.
#pragma acc loop: Used within parallel or kernels regions to indicate loops that should be parallelized.

函数和变量

#pragma acc routine: 让一个函数可以在 GPU 代码中被调用（也可以在 CPU 代码调用）。
#pragma acc declare: Used for declaring variables or creating a data region.

数据传输

#pragma acc data: Manages data movement to and from the GPU.
#pragma acc enter data: Specifies data that should be moved to the GPU.
#pragma acc exit data: Specifies data to be moved back from the GPU.
#pragma acc update: Synchronizes data between the host and the GPU.
copy, copyin, copyout, create, present: Clauses for data construct to define how data is handled (e.g., whether it's copied to/from the GPU or just created there).

线程精细控制

gang, worker, vector: Used with loop directive to control how loop iterations are distributed over parallel execution units.
collapse(n): Collapses nested loops to enhance parallelism.
reduction(operator:list): Performs a reduction operation (like sum, max) across parallel elements.