/// <summary>
/// Added by Hybrid DSP
/// </summary>
/// <typeparam name="T"></typeparam>
/// <param name="destinationHandle">The destination.</param>
/// <param name="blocking">if set to <c>true</c> [blocking].</param>
/// <param name="destinationOffset">The destination offset.</param>
/// <param name="region">The region.</param>
/// <param name="source">The source.</param>
/// <param name="events">The events.</param>
public void WriteEx <T>(CLMemoryHandle destinationHandle, bool blocking, long destinationOffset, long region, IntPtr source, ICollection <ComputeEventBase> events) where T : struct
{
int sizeofT = HDSPUtils.SizeOf(typeof(T));
int eventWaitListSize;
CLEventHandle[] eventHandles = ComputeTools.ExtractHandles(events, out eventWaitListSize);
bool eventsWritable = (events != null && !events.IsReadOnly);
CLEventHandle[] newEventHandle = (eventsWritable) ? new CLEventHandle[1] : null;
ComputeErrorCode error = CL10.EnqueueWriteBuffer(Handle, destinationHandle, blocking, new IntPtr(destinationOffset * sizeofT), new IntPtr(region * sizeofT), source, eventWaitListSize, eventHandles, newEventHandle);
ComputeException.ThrowOnError(error);
if (eventsWritable)
{
events.Add(new ComputeEvent(newEventHandle[0], this));
}
}
/// <summary>
/// Enqueues a command to write data to a buffer.
/// </summary>
/// <param name="destination"> The buffer to write to. </param>
/// <param name="blocking"> The mode of operation of this command. If <c>true</c> this call will not return until the command has finished execution. </param>
/// <param name="destinationOffset"> The <paramref name="destination"/> element position where writing starts. </param>
/// <param name="region"> The region of elements to write. </param>
/// <param name="source"> The data written to the buffer. </param>
/// <param name="events"> A collection of events that need to complete before this particular command can be executed. If <paramref name="events"/> is not <c>null</c> or read-only a new <see cref="OpenCLEvent"/> identifying this command is created and attached to the end of the collection. </param>
/// <remarks> If <paramref name="blocking"/> is <c>true</c> this method will not return until the command completes. If <paramref name="blocking"/> is <c>false</c> this method will return immediately after the command is enqueued. </remarks>
public void Write(OpenCLBufferBase destination, bool blocking, long destinationOffset, long region, IntPtr source, IReadOnlyList <OpenCLEventBase> events = null, IList <OpenCLEventBase> newEvents = null)
{
int sizeofT = Marshal.SizeOf(destination.ElementType);
int eventWaitListSize;
CLEventHandle[] eventHandles = OpenCLTools.ExtractHandles(events, out eventWaitListSize);
CLEventHandle[] newEventHandle = (newEvents != null) ? new CLEventHandle[1] : null;
OpenCLErrorCode error = CL10.EnqueueWriteBuffer(Handle, destination.Handle, blocking, new IntPtr(destinationOffset * sizeofT), new IntPtr(region * sizeofT), source, eventWaitListSize, eventHandles, newEventHandle);
OpenCLException.ThrowOnError(error);
if (newEvents != null)
{
lock (newEvents)
{
newEvents.Add(new OpenCLEvent(newEventHandle[0], this));
}
}
}
public void sortKeysValue(CLMemoryHandle key, CLMemoryHandle value,
int numElements)
{
debugRead = new int[Math.Max(numElements, numCounters)];
ComputeErrorCode error;
ComputeEvent eve;
/*
* error = CL10.EnqueueReadBuffer(cqCommandQueue, input, Bool.True, IntPtr.Zero, (IntPtr)(numElements * 4),
* debugRead, 0, null, out eve);
* CheckErr(error, "CL10.EnqueueReadBuffer");
*/
mCounters = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, gpuConstants.numGroupsPerBlock * gpuConstants.numRadices * gpuConstants.numBlocks * sizeof(int),
out error);
CheckErr(error, "CL10.CreateBuffer");
mRadixPrefixes = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, gpuConstants.numRadices * sizeof(int), out error);
CheckErr(error, "CL10.CreateBuffer");
CLMemoryHandle outputValue = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, (IntPtr)(8 * numElements),
out error);
CheckErr(error, "CL10.CreateBuffer");
CLMemoryHandle outputKey = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, (IntPtr)(4 * numElements),
out error);
CheckErr(error, "CL10.CreateBuffer");
gpuConstants.numElementsPerGroup = (numElements / (gpuConstants.numBlocks * gpuConstants.numGroupsPerBlock)) + 1;
gpuConstants.numTotalElements = numElements;
int i;
for (i = 0; i < 8; i++)
{
error = CL10.EnqueueWriteBuffer(cqCommandQueue, mCounters, Bool.True, IntPtr.Zero, (IntPtr)(numCounters * 4),
counters, 0, null, out eve);
CheckErr(error, "CL10.EnqueueWriteBuffer Counter initialize");
if (i % 2 == 0)
{
SetupAndCount(key, 4 * i);
SumIt(key, 4 * i);
ReorderingKeyValue(key, outputKey, value, outputValue, 4 * i);
}
else
{
SetupAndCount(outputKey, 4 * i);
SumIt(outputKey, 4 * i);
ReorderingKeyValue(outputKey, key, outputValue, value, 4 * i);
}
}
if (i % 2 == 0)
{
error = CL10.EnqueueCopyBuffer(cqCommandQueue, outputKey, key, IntPtr.Zero, IntPtr.Zero, (IntPtr)(numElements * 4), 0, null, out eve);
CheckErr(error, "CL10.EnqueueCopyBuffer");
error = CL10.Finish(cqCommandQueue);
CheckErr(error, "CL10.Finish Copybuffer");
error = CL10.EnqueueCopyBuffer(cqCommandQueue, outputValue, value, IntPtr.Zero, IntPtr.Zero, (IntPtr)(numElements * 8), 0, null, out eve);
CheckErr(error, "CL10.EnqueueCopyBuffer");
error = CL10.Finish(cqCommandQueue);
CheckErr(error, "CL10.Finish Copybuffer");
}
error = CL10.ReleaseMemObject(outputKey);
CheckErr(error, "CL10.ReleaseMemObj");
error = CL10.ReleaseMemObject(outputValue);
CheckErr(error, "CL10.ReleaseMemObj");
error = CL10.ReleaseMemObject(mRadixPrefixes);
CheckErr(error, "CL10.ReleaseMemObj");
error = CL10.ReleaseMemObject(mCounters);
CheckErr(error, "CL10.ReleaseMemObj");
Log_Idx++;
}
public void sortKeysOnly(CLMemoryHandle input, CLMemoryHandle output,
int numElements)
{
debugRead = new int[Math.Max(numElements, numCounters)];
ComputeErrorCode error;
Compute ComputeEvent eve;
mCounters = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, gpuConstants.numGroupsPerBlock * gpuConstants.numRadices * gpuConstants.numBlocks * sizeof(int),
out error);
CheckErr(error, "CL10.CreateBuffer");
mRadixPrefixes = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, gpuConstants.numRadices * sizeof(int), out error);
CheckErr(error, "CL10.CreateBuffer");
gpuConstants.numElementsPerGroup = (numElements / (gpuConstants.numBlocks * gpuConstants.numGroupsPerBlock)) + 1;
gpuConstants.numTotalElements = numElements;
if (DEBUG)
{
CL10.EnqueueReadBuffer(cqCommandQueue, input, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintAsArray(debugRead, gpuConstants.numTotalElements);
}
int i;
for (i = 0; i < 8; i++)
{
error = CL10.EnqueueWriteBuffer(cqCommandQueue, mCounters, true, IntPtr.Zero, (IntPtr)(numCounters * 4),
counters, 0, null, out eve);
CheckErr(error, "CL10.EnqueueWriteBuffer Counter initialize");
if (i % 2 == 0)
{
DateTime before = DateTime.Now;
SetupAndCount(input, 4 * i);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine("Setup and Count =" + (DateTime.Now - before).TotalMilliseconds);
}
before = DateTime.Now;
SumIt(input, 4 * i);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine("SumIt =" + (DateTime.Now - before).TotalMilliseconds);
}
before = DateTime.Now;
ReorderingKeysOnly(input, output, 4 * i);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine("Reorder =" + (DateTime.Now - before).TotalMilliseconds);
}
}
else
{
SetupAndCount(output, 4 * i);
SumIt(output, 4 * i);
ReorderingKeysOnly(output, input, 4 * i);
}
}
if (i % 2 != 0)
{
error = CL10.EnqueueCopyBuffer(cqCommandQueue, input, output, IntPtr.Zero, IntPtr.Zero, (IntPtr)(numElements * 4), 0, null, out eve);
CheckErr(error, "CL10.EnqueueCopyBuffer");
error = CL10.Finish(cqCommandQueue);
CheckErr(error, "CL10.Finish Copybuffer");
}
error = CL10.ReleaseMemObject(mRadixPrefixes);
CheckErr(error, "CL10.ReleaseMemObj");
error = CL10.ReleaseMemObject(mCounters);
CheckErr(error, "CL10.ReleaseMemObj");
Log_Idx++;
}