Dynamic parallelism is a new feature of CUDA 5.0 for GPUs with compute capability 3.5, allowing to launch kernels directly from other kernels.
It promises to further speedup applications by better handling computing workloads at runtime directly on the GPU; that avoids CPU/GPU interactions with benefits to mechanisms like recursion.

To use dynamic parallelism in Visual Studio 2010, do the following:

1) View -> Property Pages 2) Configuration Properties -> CUDA C/C++ -> Common -> -> Generate Relocatable Device Code -> Yes (-rdc=true) 3) Configuration Properties -> CUDA C/C++ -> Device -> -> Code Generation -> compute_35, sm_35 4) Configuration Properties -> Linker -> Input -> -> Additional Dependencies -> cudadevrt.lib

Thrust:  https://github.com/thrust/thrust/wiki/Quick-Start-Guide 

            http://thrust.github.io/ 

            http://thrust.github.io/doc/modules.html

CUDPP: http://cudpp.github.io/ 

转自：http://download.csdn.net/download/anson2004110/5912747

VS2010 CUDA 5.5 Win7 64位配置

VS2010+CUDA+5.5+Win7+64位配置以及项目创建配置.docx

错误：thrust::system::system_error in transform_reduce

I'm using VS2010 and when it breaks at the errors it points to the following in the dbgheap.c file.

__finally {

   /* unlock the heap

    */

   _munlock(_HEAP_LOCK);

}

I forgot to adjust the Properties of the project to my CUDA card compute capability

Configuration Properties > CUDA CC++ > Device > Code Generation change compute_10,sm_10 to your GPU compute capability

For Nvidia card with 2.1 compute capability it will be compute_20,sm_21

CUDA Thread Indexing

Sample code

simpleIndexing.cu

1D grid of 1D blocks

1D grid of 2D blocks

__device__ int getGlobalIdx_2D_2D()
{
int blockId = blockIdx.x + blockIdx.y * gridDim.x;
int threadId = blockId * (blockDim.x * blockDim.y) + (threadIdx.y * blockDim.x) + threadIdx.x;
return threadId;
}

2D grid of 3D blocks

__device__ int getGlobalIdx_2D_3D()
{
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;
return threadId;
}

3D grid of 1D blocks

__device__ int getGlobalIdx_3D_1D()
{
int blockId = blockIdx.x
+ blockIdx.y * gridDim.x
+ gridDim.x * gridDim.y * blockIdx.z;
int threadId = blockId * blockDim.x + threadIdx.x;
return threadId;
}

3D grid of 2D blocks

__device__ int getGlobalIdx_3D_2D()
{
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;
return threadId;
}

3D grid of 3D blocks

__device__ int getGlobalIdx_3D_3D()
{
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;
return threadId;
}

转载本文请联系原作者获取授权，同时请注明本文来自梅钢科学网博客。
链接地址：https://blog.sciencenet.cn/blog-617277-725248.html


下一篇：博士论文