CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习

GPU线程的理解 thread,block,grid

  • 一、从 cpu 多线程角度理解 gpu 多线程
    • 1、cpu 多线程并行加速
    • 2、gpu多线程并行加速
      • 2.1、cpu 线程与 gpu 线程的理解(核函数)
        • 2.1.1 、第一步:编写核函数
        • 2.1.2、第二步:调用核函数(使用内核函数)
        • 2.1.3、第三步:编写 CMakeLists & 编译代码
  • 二、重要概念 & 与线程索引的直观理解
    • 2.1、重要概念
    • 2.2、dim3与启动内核
    • 2.3、如何找到线程块的索引
    • 2.4、如何找到绝对线程索引
  • 三、参考代码(打印索引)
    • 3.1、打印一维索引
    • 3.2、打印二维索引
    • 3.2、扩展应用 (获取图片坐标)

一、从 cpu 多线程角度理解 gpu 多线程

1、cpu 多线程并行加速

在 cpu 中,用 openmp 并行计算,有限的线程数对 128 进行分组运算。

#pragma omp parallel for
for(int i =0;i<128;i++)
{
	a[i]=b[i]*c[i];
}

2、gpu多线程并行加速

在 gpu 中,可以直接开启 128 个线程对其进行计算。下面步骤和代码是演示如何开启 128个线程并打印

2.1、cpu 线程与 gpu 线程的理解(核函数)

2.1.1 、第一步:编写核函数
__global__ void some_kernel_func(int *a, int *b, int *c)
{
    // 初始化线程ID
    int i = (blockIdx.x * blockDim.x) + threadIdx.x;
    // 对数组元素进行乘法运算
    a[i] = b[i] * c[i];
    // 打印打前处理的进程ID
    // 可以看到blockIdx并非是按照顺序启动的,这也说明线程块启动的随机性
    printf("blockIdx.x = %d,blockDimx.x = %d,threadIdx.x = %d\n", blockIdx.x, blockDim.x, threadIdx.x);
}
2.1.2、第二步:调用核函数(使用内核函数)

#)调用语法
kernel_function<<>>(param1,param2,...)

  • num_blocks 线程块,至少保证一个线程块
  • num_threads 执行内核函数的线程数量

#)tips:

1、 some_kernel_func<<<1,128>>>(a,b,c); 调用 some_kernel_func 1*128 次

2、 some_kernel_func<<<2,128>>>(a,b,c); 调用 some_kernel_func 2*128 次

3、如果将 num_blocks1 改成 2 ,则表示 gpu 将启动两倍于之前的线程数量的线程,

在 blockIdx.x = 0 中,i = threadIdx.x
在 blockIdx.x = 1 中, blockDim.x 表示所要求每个线程块启动的线程数量,在这 = 128

2.1.3、第三步:编写 CMakeLists & 编译代码

CMakeLists.txt

cmake_minimum_required(VERSION 2.8 FATAL_ERROR)
project(demo)
add_definitions(-std=c++14)      
find_package(CUDA REQUIRED)

# add cuda
include_directories(${CUDA_INCLUDE_DIRS} ) 
message("CUDA_LIBRARIES:${CUDA_LIBRARIES}")
message("CUDA_INCLUDE_DIRS:${CUDA_INCLUDE_DIRS}")
cuda_add_executable(demo print_theardId.cu)

# link
target_link_libraries (demo ${CUDA_LIBRARIES})

print_theardId.cu

#include 
#include 

#include 
#include 

/*gpu 中的矩阵乘法*/
__global__ void some_kernel_func(int *a, int *b, int *c)
{
    // 初始化线程ID
    int i = (blockIdx.x * blockDim.x) + threadIdx.x;
    // 对数组元素进行乘法运算
    a[i] = b[i] * c[i];
    // 打印打前处理的进程ID
    // 可以看到blockIdx并非是按照顺序启动的,这也说明线程块启动的随机性
    printf("blockIdx.x = %d,blockDimx.x = %d,threadIdx.x = %d\n", blockIdx.x, blockDim.x, threadIdx.x);
}

int main(void)
{
    // 初始化指针元素
    int *a, *b, *c;
    // 初始化GPU指针元素
    int *gpu_a, *gpu_b, *gpu_c;
    // 初始化数组大小
    int size = 128 * sizeof(int);
    // 为CPU指针元素分配内存
    a = (int *)malloc(size);
    b = (int *)malloc(size);
    c = (int *)malloc(size);
    // 为GPU指针元素分配内存
    cudaMalloc((void **)&gpu_a, size);
    cudaMalloc((void **)&gpu_b, size);
    cudaMalloc((void **)&gpu_c, size);
    // 初始化数组元素
    for (int i = 0; i < 128; i++)
    {
        b[i] = i;
        c[i] = i;
    }
    // 将数组元素复制到GPU中
    cudaMemcpy(gpu_b, b, size, cudaMemcpyHostToDevice);
    cudaMemcpy(gpu_c, c, size, cudaMemcpyHostToDevice);
    // 执行GPU核函数
    some_kernel_func<<<4, 32>>>(gpu_a, gpu_b, gpu_c);
    // 将GPU中的结果复制到CPU中
    cudaMemcpy(a, gpu_a, size, cudaMemcpyDeviceToHost);
    // 释放GPU和CPU中的内存
    cudaFree(gpu_a);
    cudaFree(gpu_b);
    cudaFree(gpu_c);
    free(a);
    free(b);
    free(c);
    return 0;
}
mkdir build
cd build
cmake ..
make
./demo

部分结果:
CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第1张图片

可以看到,
1、gpu 可以直接调用很多个线程,
2、线程数量的多少是由线程块,线程,线程网格等决定的,
3、在核函数中编写单个线程的使用代码,再调用核函数,便可简单的达到 cpu 中 openmp 的多线程方式

二、重要概念 & 与线程索引的直观理解

2.1、重要概念

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第2张图片

gridDim.x – 线程网格X维度的线程块数目
gridDim.y – 线程网格Y维度的线程块数目

blockDim.x – 一个线程块X维度上的线程数量
blockDim.y – 一个线程块Y维度上的线程数量

theadIdx.x – 线程块X维度上的线程数量
theadIdx.y – 线程块Y维度上的线程数量

一般来说:
一个 kernel 对应一个 grid
一个 grid 可以有多个 block,一维~三维
一个 block 可以有多个 thread,一维~三维

2.2、dim3与启动内核

dim3 是CUDA中的特殊数据结构,可用来创建二维的线程块与线程网络
eg:4个线程块,128个线程

dim3 threads_rect(32,4)	// 每个线程块在X方向开启32个线程,Y方向开启4个线程
dim3 blocks_rect(1,4)	//在线程网格上,x方向1个线程块,Y方向4个线程
or
dim3 threads_square(16,8)
dim3 blocks_square(2,2)

以上两种方式线程数都是 324=128 , 168=128,只是线程块中线程的排布方式不一样

启动内核
1、 some_kernel_func<<>>(a,b,c);

2、 some_kernel_func<<>>(a,b,c);

2.3、如何找到线程块的索引

线程块的索引 x 线程块的大小 + 线程数量的起始点
参考核函数

// 定义ID查询函数
__global__ void what_is_my_id(
    unsigned int *const block,
    unsigned int *const thread,
    unsigned int *const warp,
    unsigned int *const calc_thread)
{
    /*线程ID是线程块的索引 x 线程块的大小 + 线程数量的起始点*/
    const unsigned int thread_idx = (blockIdx.x * blockDim.x) + threadIdx.x;

    block[thread_idx] = blockIdx.x;
    thread[thread_idx] = threadIdx.x;

    /*线程束 = 线程ID / 内置变量warpSize*/
    warp[thread_idx] = thread_idx / warpSize;

    calc_thread[thread_idx] = thread_idx;
}

来个.cu文件,体验一下这个核函数,// 编译方法同上

#include 
#include 

#include "cuda.h"
#include "cuda_runtime.h"

// 定义ID查询函数
__global__ void what_is_my_id(
    unsigned int *const block,
    unsigned int *const thread,
    unsigned int *const warp,
    unsigned int *const calc_thread)
{
    /*线程ID是线程块的索引 x 线程块的大小 + 线程数量的起始点*/
    const unsigned int thread_idx = (blockIdx.x * blockDim.x) + threadIdx.x;

    block[thread_idx] = blockIdx.x;
    thread[thread_idx] = threadIdx.x;

    /*线程束 = 线程ID / 内置变量warpSize*/
    warp[thread_idx] = thread_idx / warpSize;

    calc_thread[thread_idx] = thread_idx;
}

// 定义数组大小
#define ARRAY_SIZE 1024
// 定义数组字节大小
#define ARRAY_BYTES ARRAY_SIZE * sizeof(unsigned int)

// 声明主机下参数
unsigned int cpu_block[ARRAY_SIZE];
unsigned int cpu_thread[ARRAY_SIZE];
unsigned int cpu_warp[ARRAY_SIZE];
unsigned int cpu_calc_thread[ARRAY_SIZE];

// 定义主函数
int main(void)
{
    // 总线程数量为 2 x 64 = 128
    // 初始化线程块和线程数量
    const unsigned int num_blocks = 2;
    const unsigned int num_threads = 64;
    char ch;

    // 声明设备下参数
    unsigned int *gpu_block, *gpu_thread, *gpu_warp, *gpu_calc_thread;

    // 声明循环数量
    unsigned int i;

    // 为设备下参数分配内存
    cudaMalloc((void **)&gpu_block, ARRAY_BYTES);
    cudaMalloc((void **)&gpu_thread, ARRAY_BYTES);
    cudaMalloc((void **)&gpu_warp, ARRAY_BYTES);
    cudaMalloc((void **)&gpu_calc_thread, ARRAY_BYTES);

    // 调用核函数
    what_is_my_id<<<num_blocks, num_threads>>>(gpu_block, gpu_thread, gpu_warp, gpu_calc_thread);

    // 将设备下参数复制到主机下
    cudaMemcpy(cpu_block, gpu_block, ARRAY_BYTES, cudaMemcpyDeviceToHost);
    cudaMemcpy(cpu_thread, gpu_thread, ARRAY_BYTES, cudaMemcpyDeviceToHost);
    cudaMemcpy(cpu_warp, gpu_warp, ARRAY_BYTES, cudaMemcpyDeviceToHost);
    cudaMemcpy(cpu_calc_thread, gpu_calc_thread, ARRAY_BYTES, cudaMemcpyDeviceToHost);

    // 释放GPU内存
    cudaFree(gpu_block);
    cudaFree(gpu_thread);
    cudaFree(gpu_warp);
    cudaFree(gpu_calc_thread);

    // 循环打印结果
    for (i = 0; i < ARRAY_SIZE; i++)
    {
        printf("Calculated Thread: %d - Block: %d - Warp: %d - Thread: %d\n", cpu_calc_thread[i], cpu_block[i], cpu_warp[i], cpu_thread[i]);
    }
    return 0;
}

2.4、如何找到绝对线程索引

thread_idx = ( (gridDim.x * blockDim.x ) * idy ) + idx;

绝对线程索引 = 当前行索引 * 每行线程总数 + x方向的偏移
参考核函数

/*定义线程id计算函数*/
__global__ void what_is_my_id_2d_A(
    unsigned int *const block_x,
    unsigned int *const block_y,
    unsigned int *const thread,
    unsigned int *const calc_thread,
    unsigned int *const x_thread,
    unsigned int *const y_thread,
    unsigned int *const grid_dimx,
    unsigned int *const block_dimx,
    unsigned int *const grid_dimy,
    unsigned int *const block_dimy)
{
    /*获得线程索引*/
    const unsigned int idx = (blockIdx.x * blockDim.x) + threadIdx.x;
    const unsigned int idy = (blockIdx.y * blockDim.y) + threadIdx.y;
    /*
        计算线程id
        计算公式:线程ID = ((网格维度x * 块维度x) * 线程idy) + 线程idx(作为x维度上的偏移)
    */
    const unsigned int thread_idx = ((gridDim.x * blockDim.x) * idy) + idx;
    /*获取线程块的索引*/
    block_x[thread_idx] = blockIdx.x;
    block_y[thread_idx] = blockIdx.y;
    /*获取线程的索引*/
    thread[thread_idx] = threadIdx.x;
    /*计算线程id*/
    calc_thread[thread_idx] = thread_idx;
    /*获取线程的x维度索引*/
    x_thread[thread_idx] = idx;
    /*获取线程的y维度索引*/
    y_thread[thread_idx] = idy;
    /*获取网格维度的X,Y值*/
    grid_dimx[thread_idx] = gridDim.x;
    grid_dimy[thread_idx] = gridDim.y;
    /*获取block_dimy*/
    block_dimx[thread_idx] = blockDim.x;
}

来个.cu文件,体验一下这个核函数,// 编译方法同上

#include 
#include 
#include 
#include 

/*定义线程id计算函数*/
__global__ void what_is_my_id_2d_A(
    unsigned int *const block_x,
    unsigned int *const block_y,
    unsigned int *const thread,
    unsigned int *const calc_thread,
    unsigned int *const x_thread,
    unsigned int *const y_thread,
    unsigned int *const grid_dimx,
    unsigned int *const block_dimx,
    unsigned int *const grid_dimy,
    unsigned int *const block_dimy)
{
    /*获得线程索引*/
    const unsigned int idx = (blockIdx.x * blockDim.x) + threadIdx.x;
    const unsigned int idy = (blockIdx.y * blockDim.y) + threadIdx.y;
    /*
        计算线程id
        计算公式:线程ID = ((网格维度x * 块维度x) * 线程idy) + 线程idx(作为x维度上的偏移)
    */
    const unsigned int thread_idx = ((gridDim.x * blockDim.x) * idy) + idx;
    /*获取线程块的索引*/
    block_x[thread_idx] = blockIdx.x;
    block_y[thread_idx] = blockIdx.y;
    /*获取线程的索引*/
    thread[thread_idx] = threadIdx.x;
    /*计算线程id*/
    calc_thread[thread_idx] = thread_idx;
    /*获取线程的x维度索引*/
    x_thread[thread_idx] = idx;
    /*获取线程的y维度索引*/
    y_thread[thread_idx] = idy;
    /*获取网格维度的X,Y值*/
    grid_dimx[thread_idx] = gridDim.x;
    grid_dimy[thread_idx] = gridDim.y;
    /*获取block_dimy*/
    block_dimx[thread_idx] = blockDim.x;
}

/*定义矩阵宽度以及大小*/
#define ARRAY_SIZE_X 32
#define ARRAY_SIZE_Y 16
#define ARRAY_SIZE_IN_BYTES (ARRAY_SIZE_X * ARRAY_SIZE_Y * sizeof(unsigned int))

/*声明CPU端上的各项参数内存*/
unsigned int *cpu_block_x[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_block_y[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_thread[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_warp[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_calc_thread[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_x_thread[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_y_thread[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_grid_dimx[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_grid_dimy[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_block_dimx[ARRAY_SIZE_Y][ARRAY_SIZE_X];
unsigned int *cpu_block_dimy[ARRAY_SIZE_Y][ARRAY_SIZE_X];

int main(void)
{
   const dim3 thread_rect = (32, 4);
    /*注意这里的块的dim3值为1x4*/
    const dim3 block_rect = (1, 4);

    /*初始化矩形线程分布启动项*/
     const dim3 thread_square = (16, 8);
    /*注意这里的块的dim3值为2x2*/
    const dim3 block_square = (2, 2);

    /*定义一个临时指针用于打印信息*/
    char ch;

    /*定义GPU端上的各项参数内存*/
    unsigned int *gpu_block_x;
    unsigned int *gpu_block_y;
    unsigned int *gpu_thread;
    unsigned int *gpu_warp;
    unsigned int *gpu_calc_thread;
    unsigned int *gpu_x_thread;
    unsigned int *gpu_y_thread;
    unsigned int *gpu_grid_dimx;
    unsigned int *gpu_grid_dimy;
    unsigned int *gpu_block_dimx;

    /*分配GPU端上的各项参数内存*/
    cudaMalloc((void **)&gpu_block_x, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_block_y, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_thread, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_warp, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_calc_thread, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_x_thread, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_y_thread, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_grid_dimx, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_grid_dimy, ARRAY_SIZE_IN_BYTES);
    cudaMalloc((void **)&gpu_block_dimx, ARRAY_SIZE_IN_BYTES);

    /*调用核函数*/
    for (int kernel = 0; kernel < 2; kernel++)
    {
        switch (kernel)
        {
        case 0:
            /*执行矩形配置核函数*/
            what_is_my_id_2d_A<<<block_rect, thread_rect>>>(gpu_block_x, gpu_block_y, gpu_thread, gpu_warp, gpu_calc_thread, gpu_x_thread, gpu_y_thread, gpu_grid_dimx, gpu_grid_dimy, gpu_block_dimx);
            break;
        case 1:
            /*执行方形配置核函数*/
            what_is_my_id_2d_A<<<block_square, thread_square>>>(gpu_block_x, gpu_block_y, gpu_thread, gpu_warp, gpu_calc_thread, gpu_x_thread, gpu_y_thread, gpu_grid_dimx, gpu_grid_dimy, gpu_block_dimx);
            break;
        default:
            exit(1);
            break;
        }

        /*将GPU端上的各项参数内存拷贝到CPU端上*/
        cudaMemcpy(cpu_block_x, gpu_block_x, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_block_y, gpu_block_y, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_thread, gpu_thread, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_warp, gpu_warp, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_calc_thread, gpu_calc_thread, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_x_thread, gpu_x_thread, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_y_thread, gpu_y_thread, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_grid_dimx, gpu_grid_dimx, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_grid_dimy, gpu_grid_dimy, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);
        cudaMemcpy(cpu_block_dimx, gpu_block_dimx, ARRAY_SIZE_IN_BYTES, cudaMemcpyDeviceToHost);

        printf("\n kernel %d\n", kernel);
        /*打印结果*/
        for (int y = 0; y < ARRAY_SIZE_Y; y++)
        {
            for (int x = 0; x < ARRAY_SIZE_X; x++)
            {
                printf("CT: %2u Bkx: %1u TID: %2u YTID: %2u XTID: %2u GDX: %1u BDX: %1u GDY: %1u BDY:%1U\n", cpu_calc_thread[y * ARRAY_SIZE_X + x], cpu_block_x[y * ARRAY_SIZE_X + x], cpu_thread[y * ARRAY_SIZE_X + x], cpu_y_thread[y * ARRAY_SIZE_X + x], cpu_x_thread[y * ARRAY_SIZE_X + x], cpu_grid_dimx[y * ARRAY_SIZE_X + x], cpu_block_dimx[y * ARRAY_SIZE_X + x], cpu_grid_dimy[y * ARRAY_SIZE_X + x], cpu_block_y[y * ARRAY_SIZE_X + x]);
            }
            /*每行打印完后按任意键继续*/
            ch = getchar();
        }
        printf("Press any key to continue\n");
        ch = getchar();
    }
    /*释放GPU端上的各项参数内存*/
    cudaFree(gpu_block_x);
    cudaFree(gpu_block_y);
    cudaFree(gpu_thread);
    cudaFree(gpu_warp);
    cudaFree(gpu_calc_thread);
    cudaFree(gpu_x_thread);
    cudaFree(gpu_y_thread);
    cudaFree(gpu_grid_dimx);
    cudaFree(gpu_grid_dimy);
    cudaFree(gpu_block_dimx);
}

其中有个代码片段

   const dim3 thread_rect = (32, 4);
    /*注意这里的块的dim3值为1x4*/
    const dim3 block_rect = (1, 4);
    
     const dim3 thread_square = (16, 8);
    /*注意这里的块的dim3值为2x2*/
    const dim3 block_square = (2, 2);

如图理解,都是 2x2 / 1x4 = 四个线程块;每一块 32x4 / 16x8 =128个线程。这是两种不同的线程块布局方式。

但是一般会选择长方形的布局方式。
1、要以行的方式进行连续访问内存,而不是列的方式
2、同一个线程块可以通过共享内存进行通信
3、同一个线程束中的线程存储访问合并在一起了,长方形布局只需要一次访问操作就可以获得连续的内存数据 // 正方形要两次访问

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第3张图片

三、参考代码(打印索引)

3.1、打印一维索引

场景:
一个数组有 8 个数据,要开 8 个线程去访问。
我们想切成 2 个block 访问,所以一个 blcok 就有 4 个线程

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第4张图片
所以 线程设置如下:一个 block里面4个线程,一个grid里面2个block

一维索引的设置如下:

    dim3 block(4);// 一个 block 里面 4 个线程
    dim3 grid(2);//	一个 grid 里面 2 个 block	
#include 
#include 

__global__ void print_idx_kernel(){
    printf("block idx: (%3d, %3d, %3d), thread idx: (%3d, %3d, %3d)\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         threadIdx.z, threadIdx.y, threadIdx.x);
}

__global__ void print_dim_kernel(){
    printf("grid dimension: (%3d, %3d, %3d), block dimension: (%3d, %3d, %3d)\n",
         gridDim.z, gridDim.y, gridDim.x,
         blockDim.z, blockDim.y, blockDim.x);
}

__global__ void print_thread_idx_per_block_kernel(){
    int index = threadIdx.z * blockDim.x * blockDim.y + \
              threadIdx.y * blockDim.x + \
              threadIdx.x;

    printf("block idx: (%3d, %3d, %3d), thread idx: %3d\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         index);
}

__global__ void print_thread_idx_per_grid_kernel(){
    int bSize  = blockDim.z * blockDim.y * blockDim.x;

    int bIndex = blockIdx.z * gridDim.x * gridDim.y + \
               blockIdx.y * gridDim.x + \
               blockIdx.x;

    int tIndex = threadIdx.z * blockDim.x * blockDim.y + \
               threadIdx.y * blockDim.x + \
               threadIdx.x;

    int index  = bIndex * bSize + tIndex;

    printf("block idx: %3d, thread idx in block: %3d, thread idx: %3d\n", 
         bIndex, tIndex, index);
}

void print_one_dim(){
    int inputSize = 8;
    int blockDim = 4;
    int gridDim = inputSize / blockDim;

    dim3 block(blockDim);//4
    dim3 grid(gridDim);//2

    print_idx_kernel<<<grid, block>>>();
    //print_dim_kernel<<>>();
    //print_thread_idx_per_block_kernel<<>>();
    //print_thread_idx_per_grid_kernel<<>>();

    cudaDeviceSynchronize();	//用于同步
}

int main() {
    print_one_dim();
    return 0;
}

核函数及其结果:

  • 8个线程,8个输出;
  • 索引都是从 z到y到x的;

1、线程块与线程

__global__ void print_idx_kernel(){
    printf("block idx: (%3d, %3d, %3d), thread idx: (%3d, %3d, %3d)\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         threadIdx.z, threadIdx.y, threadIdx.x);
}

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第5张图片

2、线程网格与线程块

__global__ void print_dim_kernel(){
    printf("grid dimension: (%3d, %3d, %3d), block dimension: (%3d, %3d, %3d)\n",
         gridDim.z, gridDim.y, gridDim.x,
         blockDim.z, blockDim.y, blockDim.x);
}

1x1x2=2
1x1x4=4

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第6张图片

3、在 block 里面寻找每个线程的索引

__global__ void print_thread_idx_per_block_kernel(){
    int index = threadIdx.z * blockDim.x * blockDim.y + \
              threadIdx.y * blockDim.x + \
              threadIdx.x;

    printf("block idx: (%3d, %3d, %3d), thread idx: %3d\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         index);
}

可以根据下面的图来理解访问顺序:

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第7张图片

结果:
CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第8张图片
4、在 grid 里面寻找每个线程索引

__global__ void print_thread_idx_per_grid_kernel(){
    int bSize  = blockDim.z * blockDim.y * blockDim.x;	// block 的线程大小

    int bIndex = blockIdx.z * gridDim.x * gridDim.y + \
               blockIdx.y * gridDim.x + \
               blockIdx.x;

    int tIndex = threadIdx.z * blockDim.x * blockDim.y + \
               threadIdx.y * blockDim.x + \
               threadIdx.x;

    int index  = bIndex * bSize + tIndex;

    printf("block idx: %3d, thread idx in block: %3d, thread idx: %3d\n", 
         bIndex, tIndex, index);
}

可以根据下面的图来理解访问顺序:实际上就是从一堆方块里面找到那个红点

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第9张图片
结果:(thread 从 0 ~ 7 )
CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第10张图片

3.2、打印二维索引

#include 
#include 


__global__ void print_idx_kernel(){
    printf("block idx: (%3d, %3d, %3d), thread idx: (%3d, %3d, %3d)\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         threadIdx.z, threadIdx.y, threadIdx.x);
}

__global__ void print_dim_kernel(){
    printf("grid dimension: (%3d, %3d, %3d), block dimension: (%3d, %3d, %3d)\n",
         gridDim.z, gridDim.y, gridDim.x,
         blockDim.z, blockDim.y, blockDim.x);
}

__global__ void print_thread_idx_per_block_kernel(){
    int index = threadIdx.z * blockDim.x * blockDim.y + \
              threadIdx.y * blockDim.x + \
              threadIdx.x;

    printf("block idx: (%3d, %3d, %3d), thread idx: %3d\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         index);
}

__global__ void print_thread_idx_per_grid_kernel(){
    int bSize  = blockDim.z * blockDim.y * blockDim.x;

    int bIndex = blockIdx.z * gridDim.x * gridDim.y + \
               blockIdx.y * gridDim.x + \
               blockIdx.x;

    int tIndex = threadIdx.z * blockDim.x * blockDim.y + \
               threadIdx.y * blockDim.x + \
               threadIdx.x;

    int index  = bIndex * bSize + tIndex;

    printf("block idx: %3d, thread idx in block: %3d, thread idx: %3d\n", 
         bIndex, tIndex, index);
}


void print_two_dim(){
    int inputWidth = 4;

    int blockDim = 2;
    int gridDim = inputWidth / blockDim;


    dim3 block(blockDim, blockDim);// 2 , 2
    dim3 grid(gridDim, gridDim);    //2,2


    print_idx_kernel<<<grid, block>>>();
    // print_dim_kernel<<>>();
    // print_thread_idx_per_block_kernel<<>>();
    //print_thread_idx_per_grid_kernel<<>>();

    cudaDeviceSynchronize();
}

int main() {
     print_two_dim();
    return 0;
}

3.2、扩展应用 (获取图片坐标)

原理其实就是同上面(在 grid 里面寻找每个线程索引)一样,这里为了方便看,再次贴一次图。

CUDA编程- - GPU线程的理解 thread,block,grid - 再次学习_第11张图片
重写一个核函数,比之前的方便看

__global__ void print_cord_kernel(){
    int index = threadIdx.z * blockDim.x * blockDim.y + \
              threadIdx.y * blockDim.x + \
              threadIdx.x;

    int x  = blockIdx.x * blockDim.x + threadIdx.x;
    int y  = blockIdx.y * blockDim.y + threadIdx.y;

    printf("block idx: (%3d, %3d, %3d), thread idx: %3d, cord: (%3d, %3d)\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         index, x, y);
}

完整的 .cu 文件如下:

#include 
#include 

__global__ void print_cord_kernel(){
    int index = threadIdx.z * blockDim.x * blockDim.y + \
              threadIdx.y * blockDim.x + \
              threadIdx.x;

    int x  = blockIdx.x * blockDim.x + threadIdx.x;
    int y  = blockIdx.y * blockDim.y + threadIdx.y;

    printf("block idx: (%3d, %3d, %3d), thread idx: %3d, cord: (%3d, %3d)\n",
         blockIdx.z, blockIdx.y, blockIdx.x,
         index, x, y);
}

void print_cord(){
    int inputWidth = 4;

    int blockDim = 2;
    int gridDim = inputWidth / blockDim;

    dim3 block(blockDim, blockDim);
    dim3 grid(gridDim, gridDim);

    print_cord_kernel<<<grid, block>>>();

    cudaDeviceSynchronize();
}

int main() {
     print_cord();
    return 0;
}

你可能感兴趣的:(Cuda与TensorRT,学习,算法,人工智能)