[英]How can I pass a C++ array of structs to a CUDA device?
I've spent 2 days trying to figure this out and getting nowhere. 我花了两天的时间试图解决这个问题并且无处可去。 Say I had a struct that looks like this:
假设我有一个看起来像这样的结构:
struct Thing {
bool is_solid;
double matrix[9];
}
I want to create an array of that struct called things
and then process that array on the GPU. 我想创建一个名为
things
结构数组,然后在GPU上处理该数组。 Something like: 就像是:
Thing *things;
int num_of_things = 100;
cudaMallocManaged((void **)&things, num_of_things * sizeof(Thing));
// Something missing here? Malloc individual structs? Everything I try doesn't work.
things[10].is_solid = true; // Segfaults
Is it even best practice to do it this way rather than pass a single struct with arrays that are num_of_things
large? 以这种方式执行此操作是否是最佳实践,而不是使用
num_of_things
较大的数组传递单个结构? It seem to me that can get pretty nasty especially when you have arrays already (like matrix
, which would need to be 9 * num_of_things
. 在我看来,可能会变得非常讨厌,尤其是当你已经有阵列时(比如
matrix
,需要9 * num_of_things
。
Any info would be much appreciated! 任何信息将不胜感激!
After some dialog in the comments, it seems that OP's posted code has no issues. 在评论中的一些对话框之后,似乎OP发布的代码没有问题。 I was able to successfully compile and run this test case built around that code, and so was OP:
我能够成功编译并运行围绕该代码构建的测试用例,OP也是如此:
$ cat t1005.cu
#include <iostream>
struct Thing {
bool is_solid;
double matrix[9];
};
int main(){
Thing *things;
int num_of_things = 100;
cudaError_t ret = cudaMallocManaged((void **)&things, num_of_things * sizeof(Thing));
if (ret != cudaSuccess) {
std::cout << cudaGetErrorString(ret) << std::endl;
return 1;}
else {
things[10].is_solid = true;
std::cout << "Success!" << std::endl;
return 0;}
}
$ nvcc -arch=sm_30 -o t1005 t1005.cu
$ ./t1005
Success!
$
Regarding this question: 关于这个问题:
Is it even best practice to do it this way rather than pass a single struct with arrays that are num_of_things large?
以这种方式执行此操作是否是最佳实践,而不是使用num_of_things较大的数组传递单个结构?
Yes, this is a sensible practice and is usable whether managed memory is being used or not. 是的,这是一种明智的做法,无论是否使用托管内存,都可以使用。 An array of more or less any structure that does not contain embedded pointers to dynamically allocated data elsewhere can be transferred to the GPU in a simple fashion using a single
cudaMemcpy
call (for example, if managed memory were not being used.) 可以使用单个
cudaMemcpy
调用以简单的方式将一个或多或少任何不包含嵌入式指针的结构数组转移到GPU(例如,如果未使用托管内存)。
To address the question about the 3rd ( flags
) parameter to cudaMallocManaged
: 要解决有关
cudaMallocManaged
的3rd( flags
)参数的问题:
cudaMemAttachGlobal
is provided. cudaMemAttachGlobal
的默认参数。 This can be confirmed by reviewing the cuda_runtime.h
file or else simply compiling/running the test code above. cuda_runtime.h
文件来确认,或者只是编译/运行上面的测试代码。 This particular point appears to be an oversight in the documentation, and I've filed an internal issue at NVIDIA to take a look at that. Finally, proper cuda error checking is always in order any time you are having trouble with a CUDA code, and the use of such may shed some light on any errors that are made. 最后,在您遇到CUDA代码时遇到问题时,总是按顺序进行正确的cuda错误检查 ,并且使用这些错误检查可能会对所发生的任何错误有所了解。 The seg fault that the OP reported in code comments was almost certainly due to the
cudaMallocManaged
call failing (perhaps because a zero parameter was supplied incorrectly) and as a result the pointer in question ( things
) had no actual allocation. OP在代码注释中报告的seg错误几乎肯定是由于
cudaMallocManaged
调用失败(可能是因为错误地提供了零参数),因此有问题的指针( things
)没有实际分配。 Subsequent usage of that pointer would lead to a seg fault. 随后使用该指针将导致seg错误。 My test code demonstrates how to avoid that seg fault, even if the
cudaMallocManaged
call fails for some reason, and the key is proper error checking. 我的测试代码演示了如何避免seg故障,即使
cudaMallocManaged
调用由于某种原因失败,并且密钥是正确的错误检查。
声明:本站的技术帖子网页,遵循CC BY-SA 4.0协议,如果您需要转载,请注明本站网址或者原文地址。任何问题请咨询:yoyou2525@163.com.