public Event EnqueueCopyBuffer(
Buffer source, ulong sourceOffset,
Buffer destination, ulong destinationOffset, ulong count, Event[] events)
{
ClHelper.ThrowNullException(Handle);
if (source == Buffer.Null)
{
throw new ArgumentNullException("source");
}
if (destination == Buffer.Null)
{
throw new ArgumentNullException("destination");
}
unsafe
{
int num_events_in_wait_list = events == null ? 0 : events.Length;
IntPtr *wait_list = stackalloc IntPtr[num_events_in_wait_list];
for (int i = 0; i < num_events_in_wait_list; ++i)
{
wait_list[i] = events[i].Handle;
}
if (events == null)
{
wait_list = null;
}
IntPtr event_ptr = IntPtr.Zero;
ClHelper.GetError(Cl.EnqueueCopyBuffer(Handle,
source.Handle, destination.Handle,
new UIntPtr(sourceOffset), new UIntPtr(sourceOffset), new UIntPtr(count),
(uint)num_events_in_wait_list, wait_list, &event_ptr));
return(new Event(event_ptr));
}
}
public void FeedData(DataSet dataSet, int[] iExamples)
{
#if TIMING_LAYERS
Utils.InputFeedTimer.Start();
#endif
int dataPointSize = dataSet.DataDimension;
for (int m = 0; m < outputNeurons.MiniBatchSize; m++)
{
#if OPENCL_ENABLED
int iDataPoint = iExamples[m];
OpenCLSpace.ClError = Cl.EnqueueCopyBuffer(OpenCLSpace.Queue,
dataSet.DataContainer[iDataPoint].Data, // source
outputNeurons.ActivationsGPU, // destination
(IntPtr)0, // source offset (in bytes)
(IntPtr)(sizeof(float) * m * dataPointSize), // destination offset (in bytes)
(IntPtr)(sizeof(float) * dataPointSize), // size of buffer to copy
0,
null,
out OpenCLSpace.ClEvent);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "InputLayer.FeedData Cl.EnqueueCopyBuffer inputData");
OpenCLSpace.ClError = Cl.ReleaseEvent(OpenCLSpace.ClEvent);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.ReleaseEvent");
// Dropout!
if (dropoutParameter < 1.0)
{
// Set kernel arguments
OpenCLSpace.ClError = Cl.SetKernelArg(OpenCLSpace.InputDropout, 0, outputNeurons.ActivationsGPU);
OpenCLSpace.ClError |= Cl.SetKernelArg(OpenCLSpace.InputDropout, 1, (IntPtr)sizeof(int), nOutputUnits * outputNeurons.MiniBatchSize);
OpenCLSpace.ClError |= Cl.SetKernelArg(OpenCLSpace.InputDropout, 2, (IntPtr)sizeof(float), (float)dropoutParameter);
OpenCLSpace.ClError |= Cl.SetKernelArg(OpenCLSpace.InputDropout, 3, (IntPtr)sizeof(ulong), (ulong)Guid.NewGuid().GetHashCode());
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "InputDropout: Cl.SetKernelArg");
// Run kernel
OpenCLSpace.ClError = Cl.EnqueueNDRangeKernel(OpenCLSpace.Queue,
OpenCLSpace.InputDropout,
1,
null,
dropoutGlobalWorkSizePtr,
dropoutLocalWorkSizePtr,
0,
null,
out OpenCLSpace.ClEvent);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "InputDropout: Cl.EnqueueNDRangeKernel");
OpenCLSpace.ClError = Cl.ReleaseEvent(OpenCLSpace.ClEvent);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.ReleaseEvent");
OpenCLSpace.ClError = Cl.Finish(OpenCLSpace.Queue);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.Finish");
}
#else
outputNeurons.SetHost(m, dataSet.Data[iExamples[m]]);
#endif
}
#if OPENCL_ENABLED
OpenCLSpace.ClError = Cl.Finish(OpenCLSpace.Queue);
OpenCLSpace.CheckErr(OpenCLSpace.ClError, "Cl.Finish");
#endif
#if TIMING_LAYERS
Utils.InputFeedTimer.Stop();
#endif
}