剛開始學習CUDA的時候，對kernel加載的計算idx一直很模糊，threadIdx.x,blockx.x，blockDim,gridDim等一直分不清。經過查閱各方資料，特在此做個整理，表述一下個人理解。
1. Grid,Block,Thread三關係
從圖中我們可以看出，一個Grid裏可以包含多個Block，一個Block裏包含多個Thread。這三者的組成方式都可以是一維、二維、三維的。在CUDA程序中每個線程的ThreadIdx在任何時刻都是唯一的。
2. 維度
啓動kernel時，需要制定gridsize和blocksize
dim3 gridsize（x,y,z)
dim3 blocksize(x,y,z)
blockDim.x，blockDim.y，blockDim.z分別代表Block 在x,y,z三個方向的深度。Dim 數從1開始標，線程數Idx從0開始標。
3. 1D、2D、3D模式
3.1 1D模式
grid 1D，Block 1D（grid劃分成1維，block劃分成1維）
加載方式 int idx = blockIdx.x *blockDim.x + threadIdx.x;
Kernel<<< numBlock,threadsPerBlock>>>(argv)
grid 1D，Block 2D （grid劃分成1維，block劃分成2維）
int idx = blockIdx.x * blockDim.x * blockDim.y + threadIdx.y * blockDim.x + threadIdx.x;
dim3 dimBlock(x,y)
Kernel<<< numBlock,dimBlock>>>(argv)
以此爲例，因爲grid是一維的，所以blockIdx.x（從0開始標號）就是一個grid中含有的Block的數目-1；blockDim.x是一個block中x方向的線程數目，blockDim.y是一個block中y方向的線程數目，blockDim.x*blockDim.y就是一個Block中所含有的線程數， blockIdx.x * blockDim.x * blockDim.y 就是一個grid中所有滿線程的Block中所含有的線程總數。接下來我們看最後一個Block的情況，因爲Block是二維的，所以threadIdx.y * blockDim.x就是滿x的線程數，threadIdx.x是最後一行的線程數。三者相加就是所有線程數。
grid 1D，Block 3D
int idx = blockIdx.x * blockDim.x * blockDim.y * blockDim.z + threadIdx.z * blockDim.y * blockDim.x + threadIdx.y * blockDim.x + threadIdx.x;
dim3 dimBlock(x,y,z)
Kernel<<< numBlock,dimBlock>>>(argv)
grid 2D，Block 1D （grid劃分成2維，block劃分成1維）
int blockId = blockIdx.y * gridDim.x + blockIdx.x;
int Idx = blockId * blockDim.x + threadIdx.x;
dim3 dimGrid(x,y)；
Kernel<<< dimGrid,threadsPerBlock>>>(argv)；
grid 2D，Block 2D
int blockId = blockIdx.y * gridDim.x + blockIdx.x;
int Idx = blockId * (blockDim.x * blockDim.y) + (threadIdx.y * blockDim.x) + threadIdx.x;
dim3 dimGrid(x1,y1),dimBlock(x2,y2)；
Kernel<<< dimGrid,dimBlock>>>(argv)；
grid 2D，Block 3D
int blockId = blockIdx.y * gridDim.x + blockIdx.x;
int Idx = blockId * (blockDim.x * blockDim.y * blockDim.z) + (threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x)+ threadIdx.x;
dim3 dimGrid(x1,y1),dimBlock(x2,y2,z2)；
Kernel<<< dimGrid,dimBlock>>>(argv)；
grid 3D，Block 1D
int blockId = blockIdx.x+ blockIdx.y * gridDim.x+ gridDim.x * gridDim.y * blockIdx.z;
int Idx = blockId * blockDim.x + threadIdx.x;
dim3 dimGrid(x,y,z)；
Kernel<<< dimGrid,threadsPerBlock>>>(argv)；
grid 3D，Block 2D
int blockId = blockIdx.x+ blockIdx.y * gridDim.x + gridDim.x * gridDim.y * blockIdx.z;
int Idx = blockId * (blockDim.x * blockDim.y)+ (threadIdx.y * blockDim.x) + threadIdx.x;
dim3 dimGrid(x1,y1,z1),dimBlock(x2,y2)；
Kernel<<< dimGrid,dimBlock>>>(argv)；
grid 3D,block 3D
int blockId = blockIdx.x+ blockIdx.y * gridDim.x+ gridDim.x * gridDim.y * blockIdx.z;
int Idx = blockId * (blockDim.x * blockDim.y * blockDim.z) + (threadIdx.z * (blockDim.x * blockDim.y)) + (threadIdx.y * blockDim.x)+ threadIdx.x;
dim3 dimGrid(x1,y1),dimBlock(x2,y2,z2)；
Kernel<<< dimGrid,dimBlock>>>(argv)；