AthCUDA::Details::Stream Class Reference

Helper class for managing a CUDA stream in memory. More...

#include <Stream.h>

Collaboration diagram for AthCUDA::Details::Stream:

Public Member Functions
	Stream ()
	Constructor, creating the CUDA stream.
	~Stream ()
	Destructor, destroying the CUDA stream.
cudaStream_t	stream () const
	Get the CUDA stream.
std::string	name () const
	Get the name of the device associated with the stream.

Private Attributes
cudaStream_t	m_stream {nullptr}
	The CUDA stream to use for asynchronous copies.

Detailed Description

Helper class for managing a CUDA stream in memory.

Definition at line 16 of file Stream.h.

Constructor & Destructor Documentation

Stream()

AthCUDA::Details::Stream::Stream

(

)

Constructor, creating the CUDA stream.

Definition at line 28 of file Stream.cxx.

               {
  CUDA_ERROR_CHECK(cudaStreamCreate(&m_stream));
}

~Stream()

AthCUDA::Details::Stream::~Stream

(

)

Destructor, destroying the CUDA stream.

Let's not check for errors here.

Since that would introduce a struct that may throw in its destructor. Which we don't want.

Definition at line 34 of file Stream.cxx.

                {
  cudaStreamDestroy(m_stream);
}

Member Function Documentation

name()

std::string AthCUDA::Details::Stream::name

(

)

const

Get the name of the device associated with the stream.

Definition at line 48 of file Stream.cxx.

                             {
 
  // Get the device's properties.
  int device = -1;
  CUDA_ERROR_CHECK(cudaStreamGetDevice(stream(), &device));
  cudaDeviceProp props;
  CUDA_ERROR_CHECK(cudaGetDeviceProperties(&props, device));
 
  // Construct a unique name out of those properties.
  return std::format("{} [id: {}, bus: {}, device: {}]", props.name, device,
                     props.pciBusID, props.pciDeviceID);
}

stream()

cudaStream_t AthCUDA::Details::Stream::stream

(

)

const

Get the CUDA stream.

Definition at line 38 of file Stream.cxx.

                                  {
 
  // The CUDA runtime promises thread safety for handling streams in parallel
  // from different CPU threads. Returning a non-const pointer of course allows
  // us to cause harm. But as long as user code is not trying to actively break
  // things, we should be fine.
  cudaStream_t result ATLAS_THREAD_SAFE = m_stream;
  return result;
}

Member Data Documentation

m_stream

cudaStream_t AthCUDA::Details::Stream::m_stream {nullptr}

private

The CUDA stream to use for asynchronous copies.

Definition at line 32 of file Stream.h.

{nullptr};

---

The documentation for this class was generated from the following files:

• Stream.h
• Stream.cxx