* the research papers on the package. Check out http://www.gromacs.org.
*/
/*! \libinternal \file
- * \brief Implements a GpuEventSynchronizer class for OpenCL.
+ * \brief Implements a DeviceEvent class for OpenCL.
*
* \author Aleksei Iupinov <a.yupinov@gmail.com>
+ * \author Andrey Alekseenko <al42and@gmail.com>
* \inlibraryapi
*/
-#ifndef GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_OCL_H
-#define GMX_GPU_UTILS_GPUEVENTSYNCHRONIZER_OCL_H
+#ifndef GMX_GPU_UTILS_DEVICE_EVENT_OCL_H
+#define GMX_GPU_UTILS_DEVICE_EVENT_OCL_H
-#ifndef DOXYGEN
+#include "gromacs/gpu_utils/gputraits_ocl.h"
+#include "gromacs/gpu_utils/oclutils.h"
+#include "gromacs/utility/exceptions.h"
+#include "gromacs/utility/gmxassert.h"
-# include "gromacs/gpu_utils/gputraits_ocl.h"
-# include "gromacs/gpu_utils/oclutils.h"
-# include "gromacs/utility/exceptions.h"
-# include "gromacs/utility/gmxassert.h"
+#ifndef DOXYGEN
-/*! \libinternal \brief
- * A class which allows for CPU thread to mark and wait for certain GPU stream execution point.
- * The event can be put into the stream with markEvent() and then later waited on with waitForEvent().
- * This can be repeated as necessary, but the current implementation does not allow waiting on
- * completed event more than once, expecting only exact pairs of markEvent(stream); waitForEvent().
- * The class generally attempts to track the correctness of its state transitions, but
- * please note that calling waitForEvent() right after the construction will fail with OpenCL but succeed with CUDA.
- *
- * Another possible mode of operation can be implemented if needed:
- * multiple calls to waitForEvent() after a single markEvent(). For this, clReleaseEvent() call
- * from waitForEvent() should instead happen conditionally at the beginning of markEvent(), replacing
- * the GMX_ASSERT(). This was tested to work both with CUDA and NVidia OpenCL, but not with AMD/Intel OpenCl.
- */
-class GpuEventSynchronizer
+class DeviceEvent
{
public:
//! A constructor
- GpuEventSynchronizer() : event_(nullptr) {}
+ DeviceEvent() : event_(sc_nullEvent) {}
+ DeviceEvent(cl_event event) : event_(event) {}
//! A destructor
- ~GpuEventSynchronizer()
+ ~DeviceEvent()
{
- // This additional code only prevents cl_event leak in an unlikely situation of destructor
- // being called after markEvent() but before waitForEvent() / enqueueWaitEvent().
- if (event_)
+ if (isMarked())
{
+ // Can not throw in destructor, so not checking for any error
clReleaseEvent(event_);
}
}
- //! No copying
- GpuEventSynchronizer(const GpuEventSynchronizer&) = delete;
- //! No assignment
- GpuEventSynchronizer& operator=(GpuEventSynchronizer&&) = delete;
- //! Moving is disabled but can be considered in the future if needed
- GpuEventSynchronizer(GpuEventSynchronizer&&) = delete;
+ // Disable copy, move, and assignment. Move can be allowed, but not needed yet.
+ DeviceEvent& operator=(const DeviceEvent&) = delete;
+ DeviceEvent(const DeviceEvent&) = delete;
+ DeviceEvent& operator=(DeviceEvent&&) = delete;
+ DeviceEvent(DeviceEvent&&) = delete;
/*! \brief Marks the synchronization point in the \p stream.
- * Should be called first and then followed by waitForEvent().
+ * Should be called first and then followed by wait().
*/
- inline void markEvent(const DeviceStream& deviceStream)
+ inline void mark(const DeviceStream& deviceStream)
{
- GMX_ASSERT(nullptr == event_, "Do not call markEvent more than once!");
+ reset();
cl_int clError = clEnqueueMarkerWithWaitList(deviceStream.stream(), 0, nullptr, &event_);
if (CL_SUCCESS != clError)
{
+ ocl_get_error_string(clError)));
}
}
+
/*! \brief Synchronizes the host thread on the marked event. */
- inline void waitForEvent()
+ inline void wait()
{
+ GMX_RELEASE_ASSERT(isMarked(), "Can not wait for an unmarked event");
cl_int clError = clWaitForEvents(1, &event_);
if (CL_SUCCESS != clError)
{
GMX_THROW(gmx::InternalError("Failed to synchronize on the GPU event: "
+ ocl_get_error_string(clError)));
}
+ }
- reset();
+ /*! \brief Enqueues a wait for the recorded event in stream \p stream. */
+ inline void enqueueWait(const DeviceStream& deviceStream)
+ {
+ GMX_RELEASE_ASSERT(isMarked(), "Can not enqueue an unmarked event");
+ cl_int clError = clEnqueueBarrierWithWaitList(deviceStream.stream(), 1, &event_, nullptr);
+ if (CL_SUCCESS != clError)
+ {
+ GMX_THROW(gmx::InternalError("Failed to enqueue device barrier for the GPU event: "
+ + ocl_get_error_string(clError)));
+ }
}
- /*! \brief Checks the completion of the underlying event and resets the object if it was. */
+
+ //! Checks the completion of the underlying event.
inline bool isReady()
{
+ GMX_RELEASE_ASSERT(isMarked(), "Can not check the status of unmarked event");
cl_int result;
cl_int clError = clGetEventInfo(
event_, CL_EVENT_COMMAND_EXECUTION_STATUS, sizeof(cl_int), &result, nullptr);
{
GMX_THROW(gmx::InternalError("Failed to retrieve event info: " + ocl_get_error_string(clError)));
}
- bool hasTriggered = (result == CL_COMPLETE);
- if (hasTriggered)
- {
- reset();
- }
- return hasTriggered;
+ return (result == CL_COMPLETE);
}
- /*! \brief Enqueues a wait for the recorded event in stream \p stream
- *
- * After enqueue, the associated event is released, so this method should
- * be only called once per markEvent() call.
- */
- inline void enqueueWaitEvent(const DeviceStream& deviceStream)
- {
- cl_int clError = clEnqueueBarrierWithWaitList(deviceStream.stream(), 1, &event_, nullptr);
- if (CL_SUCCESS != clError)
- {
- GMX_THROW(gmx::InternalError("Failed to enqueue device barrier for the GPU event: "
- + ocl_get_error_string(clError)));
- }
- reset();
- }
+ //! Checks whether this object encapsulates an underlying event.
+ inline bool isMarked() { return event_ != sc_nullEvent; }
//! Reset (release) the event to unmarked state.
inline void reset()
{
- cl_int clError = clReleaseEvent(event_);
- if (CL_SUCCESS != clError)
+ if (isMarked())
{
- GMX_THROW(gmx::InternalError("Failed to release the GPU event: "
- + ocl_get_error_string(clError)));
+ cl_int clError = clReleaseEvent(event_);
+ if (CL_SUCCESS != clError)
+ {
+ GMX_THROW(gmx::InternalError("Failed to release the GPU event: "
+ + ocl_get_error_string(clError)));
+ }
}
- event_ = nullptr;
+ event_ = sc_nullEvent;
}
private:
cl_event event_;
+
+ //! Magic value to indicate uninitialized state.
+ static constexpr cl_event sc_nullEvent = nullptr;
};
#endif