mirror of
https://gitlab.kitware.com/vtk/vtk-m
synced 2024-09-16 17:22:55 +00:00
ef58bd9c4a
We have been noticing occasional failures of UnitTestTimer on the dashboard for CUDA devices. The timer seems to be recording less time than is actually elapsed. The problem might be that the CUDA timer actually inserts fences to the CUDA stream rather than record the current time. Thus, the actual time might start after some pending operations complete. To attempt to match the UnitTestTimer measurements closer to wall-clock time, add a synchronize before starting the timer.
101 lines
3.6 KiB
C++
101 lines
3.6 KiB
C++
//============================================================================
|
|
// Copyright (c) Kitware, Inc.
|
|
// All rights reserved.
|
|
// See LICENSE.txt for details.
|
|
//
|
|
// This software is distributed WITHOUT ANY WARRANTY; without even
|
|
// the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR
|
|
// PURPOSE. See the above copyright notice for more information.
|
|
//============================================================================
|
|
#ifndef vtk_m_cont_Timer_h
|
|
#define vtk_m_cont_Timer_h
|
|
|
|
#include <vtkm/cont/DeviceAdapterTag.h>
|
|
|
|
#include <vtkm/cont/vtkm_cont_export.h>
|
|
|
|
#include <memory>
|
|
|
|
namespace vtkm
|
|
{
|
|
namespace cont
|
|
{
|
|
namespace detail
|
|
{
|
|
struct EnabledDeviceTimerImpls;
|
|
}
|
|
|
|
/// A class that can be used to time operations in VTK-m that might be occuring
|
|
/// in parallel. Users are recommended to provide a device adapter at construction
|
|
/// time which matches the one being used to execute algorithms to ensure that thread
|
|
/// synchronization is correct and accurate.
|
|
/// If no device adapter is provided at construction time, the maximum
|
|
/// elapsed time of all enabled deivces will be returned. Normally cuda is expected to
|
|
/// have the longest execution time if enabled.
|
|
/// Per device adapter time query is also supported. It's useful when users want to reuse
|
|
/// the same timer to measure the cuda kernal call as well as the cuda device execution.
|
|
/// It is also possible to change the device adapter after construction by calling the form
|
|
/// of the Reset method with a new DeviceAdapterId.
|
|
///
|
|
/// The there is no guaranteed resolution of the time but should generally be
|
|
/// good to about a millisecond.
|
|
///
|
|
class VTKM_CONT_EXPORT Timer
|
|
{
|
|
public:
|
|
VTKM_CONT
|
|
Timer();
|
|
|
|
VTKM_CONT Timer(vtkm::cont::DeviceAdapterId device);
|
|
|
|
VTKM_CONT ~Timer();
|
|
|
|
/// Resets the timer.
|
|
VTKM_CONT void Reset();
|
|
|
|
/// Resets the timer and changes the device to time on.
|
|
VTKM_CONT void Reset(vtkm::cont::DeviceAdapterId device);
|
|
|
|
VTKM_CONT void Start();
|
|
|
|
VTKM_CONT void Stop();
|
|
|
|
VTKM_CONT bool Started() const;
|
|
|
|
VTKM_CONT bool Stopped() const;
|
|
|
|
/// Used to check if Timer has finished the synchronization to get the result from the device.
|
|
VTKM_CONT bool Ready() const;
|
|
|
|
/// Get the elapsed time measured by the given device adapter. If no device is
|
|
/// specified, the max time of all device measurements will be returned.
|
|
VTKM_CONT
|
|
vtkm::Float64 GetElapsedTime() const;
|
|
|
|
/// Returns the device for which this timer is synchronized. If the device adapter has the same
|
|
/// id as `DeviceAdapterTagAny`, then the timer will synchronize all devices.
|
|
VTKM_CONT vtkm::cont::DeviceAdapterId GetDevice() const { return this->Device; }
|
|
|
|
/// Synchronize the device(s) that this timer is monitoring without starting or stopping the
|
|
/// timer. This is useful for ensuring that external events are synchronized to this timer.
|
|
///
|
|
/// Note that this method will allways block until the device(s) finish even if the
|
|
/// `Start`/`Stop` methods do not actually block. For example, the timer for CUDA does not
|
|
/// actually wait for asynchronous operations to finish. Rather, it inserts a fence and
|
|
/// records the time as fences are encounted. But regardless, this `Synchronize` method
|
|
/// will block for the CUDA device.
|
|
VTKM_CONT void Synchronize() const;
|
|
|
|
private:
|
|
/// Some timers are ill-defined when copied, so disallow that for all timers.
|
|
VTKM_CONT Timer(const Timer&) = delete;
|
|
VTKM_CONT void operator=(const Timer&) = delete;
|
|
|
|
vtkm::cont::DeviceAdapterId Device;
|
|
std::unique_ptr<detail::EnabledDeviceTimerImpls> Internal;
|
|
};
|
|
}
|
|
} // namespace vtkm::cont
|
|
|
|
#endif //vtk_m_cont_Timer_h
|