目录
1.线程的组成结构
具体的结构从上至下可以分为：Grid-> Block-> Warp ->thread。
Grid/Block/Warp/Thread都是软件的组织结构，并不是硬件的，因此理论上我们可以以任意的维度（一维、二维、三维）去排列Thread；在硬件上就是一个个的SP，并没有维度这一说，只是软件上抽象成了具有维度的概念。
Warp是结构中的第三个层次，32个线程为一组。
2.确定自己显卡的配置
在进行程序设计前要注意显卡所支持的thread数量和block数量是有限制的。因此首先要对显卡的配置进行确认。否则一旦程序超过了显卡的硬件限制，便会停止工作。
打印显卡配置的cuda例程：
/***
*打印显卡属性
*2017-11-8
* ***/
#include<iostream>
#include <cuda_runtime.h>
#include<device_launch_parameters.h>
using namespace std;
int main()
{
int deviceCount;
cudaGetDeviceCount(&deviceCount); //Returns in *deviceCount the number of devices
cout << "deviceCount: " << deviceCount << "\n\n";
if (deviceCount == 0)
{
cout << "error: no devices supporting CUDA.\n";
exit(EXIT_FAILURE);
}
int dev = 0;
cudaSetDevice(dev); //Sets dev=0 device as the current device for the calling host thread.
cudaDeviceProp devProps;
cudaGetDeviceProperties(&devProps, dev);
cout << "name: " << devProps.name << "\n";
cout << "totalGlobalMem: " << devProps.totalGlobalMem << "\n";
cout << "regsPerBlock: " << devProps.regsPerBlock << "\n";
cout << "warpSize: " << devProps.warpSize << "\n";
cout << "memPitch: " << devProps.memPitch << "\n\n";
cout << "一个线程块中可使用的最大共享内存\n";
cout << "devProps.sharedMemPerBlock: " << devProps.sharedMemPerBlock << "\n\n";
cout << "一个线程块中可包含的最大线程数量\n";
cout << "maxThreadsPerBlock: " << devProps.maxThreadsPerBlock << "\n\n";
cout << "多维线程块数组中每一维可包含的最大线程数量\n";
cout << "maxThreadsDim[0]: " << devProps.maxThreadsDim[0] << "\n";
cout << "maxThreadsDim[1]: " << devProps.maxThreadsDim[1] << "\n";
cout << "maxThreadsDim[2]: " << devProps.maxThreadsDim[2] << "\n\n";
cout << "一个线程格中每一维可包含的最大线程块数量\n";
cout << "maxGridSize[0]: " << devProps.maxGridSize[0] << "\n";
cout << "maxGridSize[1]: " << devProps.maxGridSize[1] << "\n";
cout << "maxGridSize[2]: " << devProps.maxGridSize[2] << "\n\n";
cout << "clockRate: " << devProps.clockRate << "\n";
cout << "totalConstMem: " << devProps.totalConstMem << "\n";
cout << "textureAlignment: " << devProps.textureAlignment << "\n\n";
cout << "计算能力：" << devProps.major << "." << devProps.minor << "\n\n";
cout << "minor: " << devProps.minor << "\n";
cout << "texturePitchAlignment: " << devProps.texturePitchAlignment << "\n";
cout << "deviceOverlap: " << devProps.deviceOverlap << "\n";
cout << "multiProcessorCount: " << devProps.multiProcessorCount << "\n";
cout << "kernelExecTimeoutEnabled: " << devProps.kernelExecTimeoutEnabled << "\n";
cout << "integrated: " << devProps.integrated << "\n";
cout << "canMapHostMemory: " << devProps.canMapHostMemory << "\n";
cout << "computeMode: " << devProps.computeMode << "\n";
cout << "maxTexture1D: " << devProps.maxTexture1D << "\n";
cout << "maxTexture1DMipmap: " << devProps.maxTexture1DMipmap << "\n";
cout << "maxTexture1DLinear: " << devProps.maxTexture1DLinear << "\n";
cout << "maxTexture2D: " << devProps.maxTexture2D << "\n";
cout << "maxTexture2DMipmap: " << devProps.maxTexture2DMipmap << "\n";
cout << "maxTexture2DLinear: " << devProps.maxTexture2DLinear << "\n";
cout << "maxTexture2DGather: " << devProps.maxTexture2DGather << "\n";
cout << "maxTexture3D: " << devProps.maxTexture3D << "\n";
cout << "maxTexture3DAlt: " << devProps.maxTexture3DAlt << "\n";
cout << "maxTextureCubemap: " << devProps.maxTextureCubemap << "\n";
cout << "maxTexture1DLayered: " << devProps.maxTexture1DLayered << "\n";
cout << "maxTexture2DLayered: " << devProps.maxTexture2DLayered << "\n";
cout << "maxTextureCubemapLayered: " << devProps.maxTextureCubemapLayered << "\n";
cout << "maxSurface1D: " << devProps.maxSurface1D << "\n";
cout << "maxSurface2D: " << devProps.maxSurface2D << "\n";
cout << "maxSurface3D: " << devProps.maxSurface3D << "\n";
cout << "maxSurface1DLayered: " << devProps.maxSurface1DLayered << "\n";
cout << "maxSurface2DLayered: " << devProps.maxSurface2DLayered << "\n";
cout << "maxSurfaceCubemap: " << devProps.maxSurfaceCubemap << "\n";
cout << "maxSurfaceCubemapLayered: " << devProps.maxSurfaceCubemapLayered << "\n";
cout << "surfaceAlignment: " << devProps.surfaceAlignment << "\n";
cout << "concurrentKernels: " << devProps.concurrentKernels << "\n";
cout << "ECCEnabled: " << devProps.ECCEnabled << "\n";
cout << "pciBusID: " << devProps.pciBusID << "\n";
cout << "pciDeviceID: " << devProps.pciDeviceID << "\n";
cout << "pciDomainID: " << devProps.pciDomainID << "\n";
cout << "tccDriver: " << devProps.tccDriver << "\n";
cout << "asyncEngineCount: " << devProps.asyncEngineCount << "\n";
cout << "unifiedAddressing: " << devProps.unifiedAddressing << "\n";
cout << "memoryClockRate: " << devProps.memoryClockRate << "\n";
cout << "memoryBusWidth: " << devProps.memoryBusWidth << "\n";
cout << "l2CacheSize: " << devProps.l2CacheSize << "\n";
cout << "maxThreadsPerMultiProcessor: " << devProps.maxThreadsPerMultiProcessor << "\n";
cout << "streamPrioritiesSupported: " << devProps.streamPrioritiesSupported << "\n";
cout << "globalL1CacheSupported: " << devProps.globalL1CacheSupported << "\n";
cout << "localL1CacheSupported: " << devProps.localL1CacheSupported << "\n";
cout << "sharedMemPerMultiprocessor: " << devProps.sharedMemPerMultiprocessor << "\n";
cout << "regsPerMultiprocessor: " << devProps.regsPerMultiprocessor << "\n";
cout << "isMultiGpuBoard: " << devProps.isMultiGpuBoard << "\n";
cout << "multiGpuBoardGroupID: " << devProps.multiGpuBoardGroupID << "\n";
cout << "singleToDoublePrecisionPerfRatio: " << devProps.singleToDoublePrecisionPerfRatio << "\n";
cout << "pageableMemoryAccess: " << devProps.pageableMemoryAccess << "\n";
cout << "concurrentManagedAccess: " << devProps.concurrentManagedAccess << "\n";
return 0;
}
3.CUDA核函数的参数（Kernel Function)
要对GPU编程中线程执行的任务进行了解。
/*
kernel函数的样例
*/
// CUDA核函数的定义
__global__ void addKernel(int *c, const int *a, const int *b)
{
int i = threadIdx.x;
c[i] = a[i] + b[i];
}
// CUDA核函数调用
addKernel<<<Dg,Db, Ns, S>>>(c, a, b);
需要注意以下几点：
1.函数前的标识“global”表示该函数在GPU执行。
2.在函数调用时，要通过<<< >>>来确定线程的数量。
3.调用kernel函数时各个参数的含义：
Dg：int型或者dim3类型(x,y,z)，用于定义一个Grid中Block是如何组织的，如果是int型，则表示一维组织结构
Db：int型或者dim3类型(x,y,z)，用于定义一个Block中Thread是如何组织的，如果是int型，则表示一维组织结构
Ns：size_t类型，可缺省，默认为0； 用于设置每个block除了静态分配的共享内存外，最多能动态分配的共享内存大小，单位为byte。 0表示不需要动态分配。
S：cudaStream_t类型，可缺省，默认为0。 表示该核函数位于哪个流。
3.1 1d Grid 和 1d Block
cuda的调用模式：
下图为1个grid有4个block块（编号：0-3），1个block块有8个线程（编号：0-7）。
kernel_name<<<4, 8>>>(...)
在cuda核函数中线程索引计算方式如下：
int threadId = blockIdx.x * blockDim.x + threadIdx.x
3.2 2d Grid 和 2d Block
对于高维度计算索引的方式来说类似于计算多维数组中数据的数量。需要计算出Thread前面的所有Block中的Thread数量，在求出该Thread在本Block中的序号相加即可。
调用形式：
dim3 grid(4,1), block(2,2);
kernel_name<<<grid, block>>>(...)
索引计算方式：
int x = threadIdx.x + blockIdx.x * blockDim.x;
int y = threadIdx.y + blockIdx.y * blockDim.y;
int offset = x + y * blockDim.x * gridDim.x;
或者
int blockId = blockIdx.x + blockId.y * gridDim.x;
int threadId = blockId * (blockDim.x * blockDim.y) + (threadIdx.y * blockDim.x) + threadIdx.x;
注：在二维排序中，Thread(0,1)表示第1行第0列的Thread，这跟我们传统中理解的横坐标和纵坐标不太一样；我们定义grid(4,2)表示第一维度有4个索引值，第二个维度有2个索引值，即2行4列。
注：dim3是NVIDIA的CUDA编程中一种自定义的整型向量类型，基于用于指定维度的uint3。
例如：
dim3 grid（num1，num2，num3）；
dim3类型最终设置的是一个三维向量，三维参数分别为x,y,z;
3.3 3d Grid 和 3d Block
3d的情况相对较少，它的线程计算公式如下：
int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
int threadIc = blockId * (blockDim.x * blockDim.y * blockDim.z)
+ (threadIdx.z * (blockDim.x * blockDim.y))
+ (threadIdx.y * blockDim.x) + threadIdx.x;
3.4 计算公式汇总
一维Grid 一维Block
blockId = blockIdx.x
threadId = blockIdx.x *blockDim.x + threadIdx.x
一维Grid 二维Block
blockId = blockIdx.x
threadId = blockIdx.x * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x
一维Grid 三维Block
blockId = blockIdx.x
threadId = blockIdx.x * blockDim.x * blockDim.y * blockDim.z
+ threadIdx.z * blockDim.y * blockDim.x
+ threadIdx.y * blockDim.x + threadIdx.x
二维Grid 一维Block
int blockId = blockIdx.y * gridDim.x + blockIdx.x;
int threadId = blockId * blockDim.x + threadIdx.x;
二维Grid 二维Block
int blockId = blockIdx.x + blockIdx.y * gridDim.x;
int threadId = blockId * (blockDim.x * blockDim.y)
+ (threadIdx.y * blockDim.x) + threadIdx.x;
二维Grid 三维Block
int blockId = blockIdx.x + blockIdx.y * gridDim.x;
int threadId = blockId * (blockDim.x * blockDim.y * blockDim.z)
+ (threadIdx.z * (blockDim.x * blockDim.y))
+ (threadIdx.y * blockDim.x) + threadIdx.x;
三维Grid 一维Block
int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
int threadId = blockId * blockDim.x + threadIdx.x;
三维Grid 二维Block
int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
int threadId = blockId * (blockDim.x * blockDim.y)
+ (threadIdx.y * blockDim.x) + threadIdx.x;
三维Grid 三维Block
int blockId = blockIdx.x + blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
int threadId = blockId * (blockDim.x * blockDim.y * blockDim.z)
+ (threadIdx.z * (blockDim.x * blockDim.y))
+ (threadIdx.y * blockDim.x) + threadIdx.x;