我正在尝试运行oneAPI的hello-world DPC++示例,该示例在CPU和GPU上添加了两个一维阵列,并验证结果。代码如下所示:
/*
DataParallel Addition of two Vectors
*/
#include <CL/sycl.hpp>
#include <array>
#include <iostream>
using namespace sycl;
constexpr size_t array_size = 100000;
typedef std::array<int, array_size> IntArray;
// Initialize array with the same value as its index
void InitializeArray(IntArray& a) { for (size_t i = 0; i < a.size(); i++) a[i] = i; }
/*
Create an asynchronous Exception Handler for sycl
*/
static auto exception_handler = [](cl::sycl::exception_list eList) {
for (std::exception_ptr const& e : eList) {
try {
std::rethrow_exception(e);
}
catch (std::exception const& e) {
std::cout << "Failure" << std::endl;
std::terminate();
}
}
};
void VectorAddParallel(queue &q, const IntArray& x, const IntArray& y, IntArray& parallel_sum) {
range<1> num_items{ x.size() };
buffer x_buf(x);
buffer y_buf(y);
buffer sum_buf(parallel_sum.data(), num_items);
/*
Submit a command group to the queue by a lambda
which contains data access permissions and device computation
*/
q.submit([&](handler& h) {
auto xa = x_buf.get_access<access::mode::read>(h);
auto ya = y_buf.get_access<access::mode::read>(h);
auto sa = sum_buf.get_access<access::mode::write>(h);
std::cout << "Adding on GPU (Parallel)n";
h.parallel_for(num_items, [=](id<1> i) { sa[i] = xa[i] + ya[i]; });
std::cout << "Done on GPU (Parallel)n";
});
/*
queue runs the kernel asynchronously. Once beyond the scope,
buffers' data is copied back to the host.
*/
}
int main() {
default_selector d_selector;
IntArray a, b, sequential, parallel;
InitializeArray(a);
InitializeArray(b);
try {
// Queue needs: Device and Exception handler
queue q(d_selector, exception_handler);
std::cout << "Accelerator: "
<< q.get_device().get_info<info::device::name>() << "n";
std::cout << "Vector size: " << a.size() << "n";
VectorAddParallel(q, a, b, parallel);
}
catch (std::exception const& e) {
std::cout << "Exception while creating Queue. Terminating...n";
std::terminate();
}
/*
Do the sequential, which is supposed to be slow
*/
std::cout << "Adding on CPU (Scalar)n";
for (size_t i = 0; i < sequential.size(); i++) {
sequential[i] = a[i] + b[i];
}
std::cout << "Done on CPU (Scalar)n";
/*
Verify results, the old-school way
*/
for (size_t i = 0; i < parallel.size(); i++) {
if (parallel[i] != sequential[i]) {
std::cout << "Fail: " << parallel[i] << " != " << sequential[i] << std::endl;
std::cout << "Failed. Results do not match.n";
return -1;
}
}
std::cout << "Success!n";
return 0;
}
对于相对较小的array_size
(我测试了100-50k个元素(,计算结果很好。样本输出:
Accelerator: Intel(R) Gen9
Vector size: 50000
Adding on GPU (Parallel)
Done on GPU (Parallel)
Adding on CPU (Scalar)
Done on CPU (Scalar)
Success!
可以注意到,在CPU和GPU上完成计算只需一秒钟。但当我增加array_size
,也就是说,100000
时,我得到了一个看似毫无头绪的错误:
C:Usersmyusersourcereposdpcpp-iotasx64Debugdpcpp-iotas.exe (process 24472) exited with code -1073741571.
虽然我不确定错误开始发生的确切值,但我似乎确信它发生在70000
左右。我似乎不知道为什么会发生这种事,也不知道哪里出了问题?
事实证明,这是由于VS增强了堆栈大小。元素过多的连续数组导致堆栈溢出。
正如@user4581301所提到的,十六进制的错误代码-107374171
给出了C00000FD
,这是Visual Studio中"堆栈耗尽/溢出"的有符号表示。
解决方法:
- 在"项目属性"中将
/STACK
保留增加到高于1MB的值(这是默认值(>链接器>系统>堆栈保留/提交值 - 使用二进制编辑器(editbin.exe和dumpbin.exe(编辑
/STACK:reserve
- 改为使用
std::vector
,这允许动态分配(由@Retired Ninja建议(
我在oneAPI中找不到更改/STACK
的选项,这是Linker属性中的正常方式,如图所示。
我决定采用动态分配。
相关:https://stackoverflow.com/a/26311584/9230398
当我对大型应用程序进行编程时,我总是进行
ulimit -s unlimited
向shell解释我长大了,我真的想在堆栈上留出一些空间。
这是bash
语法,但您显然可以适应其他一些shell。
我想可能有一个非UNIX操作系统的等价物?