Cuda C权威指南学习笔记（二）

复习

---

CUDA的抽象层次：内存结构，线程结构，障碍同步。

线程管理

---

CUDA线程管理是通过一个两层的线程层次结构完成的，也就是网格和线程块。通过以下代码可以看到这个结构是如何运行的

#include<cuda_runtime.h>
#include<stdio.h>
#include<iostream>
using namespace std;
__global__ void checkIndex(void){
    printf("threadIdx: %d %d %d, blockIdx: %d %d %d,threadDim: %d %d %d, blockDim: %d %d %d\n",threadIdx.x,threadIdx.y,threadIdx.z,blockIdx.x,blockIdx.y,blockIdx.z,blockDim.x,blockDim.y,blockDim.z,gridDim.x,gridDim.y,gridDim.z);
}
int main(){
    checkIndex<<<4,8>>>();
    cudaDeviceReset();
    return 0;
}

输出如下

threadIdx: 0 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 1 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 2 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 3 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 4 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 5 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 6 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 7 0 0, blockIdx: 1 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 0 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 1 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 2 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 3 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 4 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 5 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 6 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 7 0 0, blockIdx: 0 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 0 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 1 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 2 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 3 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 4 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 5 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 6 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 7 0 0, blockIdx: 2 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 0 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 1 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 2 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 3 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 4 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 5 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 6 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1
threadIdx: 7 0 0, blockIdx: 3 0 0,threadDim: 8 1 1, blockDim: 4 1 1

编写核函数

---

限定符	执行	调用	备注
global	在设备端执行	可以在主机端调用，计算能力需大于等于3	返回类型必须为void
device	在设备端执行	只能在设备端调用
host	在主机端执行	只能在主机端调用	可以省略

cuda核函数限制：

• 只能访问设备内存
• 必须拥有void返回类型
• 不支持可变数量的参数
• 不支持静态变量
• 显示异步行为