【C++】CUDA线程在全局索引中的计算方式

1. 一维网格一维线程块

定义grid与block尺寸：

dim3 grid_size(4);
dim3 block_size(8);

调用核函数：

kernal_fun<<<grid_size, block_size>>>(...);

具体索引方式如下图所示，$blockIdx.x$ 从0~3，$threadIdx.x$ 从0~7

计算方式：

$$intid=blockIdx.x\ast blockDim.x+threadIdx.x$$

2. 二维网格二维线程块

定义grid与block尺寸

dim3 grid_size(2,2);
dim3 block_size(4,4);

调用核函数：

kernal_fun<<<grid_size, block_size>>>(...);

具体线程索引方式如下图所示， $blockIdx.x$ 和 $blockIdx.y$ 从0到1，$threadIdx.x$ 和 $threadIdx.y$从0到3:

计算方式：
$$\begin{array}{rl}& intblockId=blockIdx.x+blockId.y\ast girdDim.x\\ & intthreadId=threadIdx.y\ast blockDim.x+threadIdx.x\\ & intid=blockId\ast (blockDim.x\ast blockDim.y)+threadId\end{array}$$

3. 三维网格三维线程块

定义grid和block尺寸：

dim3 grid_size(2,2,2);
dim3 block_size(4,4,2);

调用核函数：

kernal_fun<<<grid_size, block_size>>>(...);

具体线程索引方式如图所示：

• $blockIdx.x$ 从0到1
• $blockIdx.y$ 从0到1
• $blockIdx.z$ 从0到1
• $threadIdx.x$ 从0到3
• $threadIdx.y$ 从0到3
• $threadIdx.z$ 从0到1

计算方式：
$$\begin{array}{rl}& intblockId=blockIdx.x+blockIdx.y\ast girdDim.x+gridDim.x\ast gridDim.y\ast blockIdx.z\\ & intthreadId=threadIdx.z\ast blockDim.x\ast blockDim.y+threadIdx.y\ast blockDim.x+threadIdx.x\\ & intid=blockId\ast (blockDim.x\ast blockDim.y\ast blockDim.z)+threadId\end{array}$$

4. 不同组合形式

4.1 一维网格一维线程块

$$\begin{array}{rl}& intblockId=blockIdx.x\\ & intid=blockIdx.x\ast blockDim.x+threadIdx.x\end{array}$$

4.2 一维网格二维线程块

$$\begin{array}{rl}& intblockId=blockIdx.x\\ & intid=blockIdx.x\ast blockDim.x\ast blockDim.y+threadIdx.y\ast blockDim.x+threadIdx.x\end{array}$$

4.3 一维网格三维线程块

$$\begin{gathered} intblockId=blockIdx.x \\ intid=blockIdx.x\ast blockDim.x\ast blockDim.y\ast blockDim.z+threadIdx.z\ast blockDim.y\ast blockDim.x+threadIdx.y\ast blockDim.x+threadIdx.x \end{gathered}$$

4.4 二维网格一维线程块

​
$$\begin{gathered} intblockId=blockIdx.x+blockIdx.y\ast gridDim.x \\ intid=blockId\ast blockDim.x+threadIdx.x \end{gathered}$$

4.5 二维网格二维线程块

$$\begin{gathered} intblockId=blockIdx.x+blockIdx.y\ast gridDim.x \\ intid=blockId\ast blockDim.x\ast blockDim.y+threadIdx.y\ast blockDim.x+threadIdx.x \end{gathered}$$

4.6 二维网格三维线程块

$$\begin{gathered} intblockId=blockIdx.x+blockIdx.y\ast gridDim.x \\ intid=blockId\ast blockDim.x\ast blockDim.y\ast blockDim.z+threadIdx.z\ast blockDim.x\ast blockDim.y+threadIdx.y\ast blockDim.x+threadIdx.x \end{gathered}$$

4.7 三维网格一维线程块

$$\begin{gathered} intblockId=blockIdx.x+blockIdx.y\ast gridDim.x+blockIdx.z\ast gridDim.x\ast gridDim.y \\ intid=blockId\ast blockDim.x+threadIdx.x \end{gathered}$$

4.8 三维网格二维线程块

$$\begin{gathered} intblockId=blockIdx.x+blockIdx.y\ast gridDim.x+blockIdx.z\ast gridDim.x\ast gridDim.y \\ intid=blockId\ast blockDim.x\ast blockDim.y+threadIdx.y\ast blockDim.x+threadIdx.x \end{gathered}$$

4.9 三维网格三维线程块

$$\begin{gathered} intblockId=blockIdx.x+blockIdx.y\ast gridDim.x+blockIdx.z\ast gridDim.x\ast gridDim.y \\ intid=blockId\ast blockDim.x\ast blockDim.y\ast blockDim.z+threadIdx.z\ast blockDim.x\ast blockDim.y+threadIdx.y\ast blockDim.x+threadIdx.x \end{gathered}$$