private void SetupAndCount(CLMemoryHandle input, int bitOffset)
{
ComputeErrorCode error;
IntPtr agentPtrSize = (IntPtr)0;
agentPtrSize = (IntPtr)Marshal.SizeOf(typeof(IntPtr));
var ptrSize = (IntPtr)Marshal.SizeOf(typeof(Mem));
int globalWorkSize = gpuConstants.numThreadsPerBlock * gpuConstants.numBlocks;
int localWorkSize = gpuConstants.numThreadsPerBlock;
IntPtr[] workGroupSizePtr = new IntPtr[] { (IntPtr)globalWorkSize };
IntPtr[] localWorkGroupSizePtr = new IntPtr[] { (IntPtr)localWorkSize };
ComputeEvent clevent;
error = CL10.SetKernelArg(ckSetupAndCount, 0, ptrSize, input);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckSetupAndCount, 1, ptrSize, mCounters);
CheckErr(error, "CL10.SetKernelArg");
//if(DEBUG_CONSOLE_OUTPUT) Console.WriteLine((Marshal.SizeOf(typeof(GPUConstants))));
error = CL10.SetKernelArg(ckSetupAndCount, 2, (IntPtr)(Marshal.SizeOf(typeof(GPUConstants))), gpuConstants);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckSetupAndCount, 3, (IntPtr)4, bitOffset);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.EnqueueNDRangeKernel(cqCommandQueue, ckSetupAndCount, 1, null, workGroupSizePtr, localWorkGroupSizePtr, 0, null, out clevent);
CheckErr(error, "CL10.EnqueueNDRangeKernel");
error = CL10.Finish(cqCommandQueue);
CheckErr(error, "CL10.Finish");
if (DEBUG)
{
ComputeEvent eve;
CL10.EnqueueReadBuffer(cqCommandQueue, input, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintElementBuffer(debugRead, gpuConstants.numTotalElements, "Setup and Count -> Input -> bitoffset = " + bitOffset);
CL10.EnqueueReadBuffer(cqCommandQueue, mCounters, Bool.True, IntPtr.Zero, (IntPtr)(numCounters * sizeof(int)), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintCounterBuffer(debugRead, "Setup and Count -> bitoffset = " + bitOffset);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine("Setup and Count -> bitoffset = " + bitOffset);
}
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine();
}
}
}
/// <summary>
/// Enqueues a command to read data from a buffer.
/// </summary>
/// <param name="source"> The buffer to read from. </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="offset"> The <paramref name="source"/> element position where reading starts. </param>
/// <param name="region"> The region of elements to read. </param>
/// <param name="destination"> A pointer to a preallocated memory area to read the data into. </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="ComputeEvent"/> 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 Read <T>(ComputeBufferBase <T> source, bool blocking, long offset, long region, IntPtr destination, 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.EnqueueReadBuffer(Handle, source.Handle, blocking, new IntPtr(offset * sizeofT), new IntPtr(region * sizeofT), destination, eventWaitListSize, eventHandles, newEventHandle);
ComputeException.ThrowOnError(error);
if (eventsWritable)
{
events.Add(new ComputeEvent(newEventHandle[0], this));
}
}
/// <summary>
/// Enqueues a command to read data from a buffer.
/// </summary>
/// <param name="source"> The buffer to read from. </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="offset"> The <paramref name="source"/> element position where reading starts. </param>
/// <param name="region"> The region of elements to read. </param>
/// <param name="destination"> A pointer to a preallocated memory area to read the data into. </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 Read(OpenCLBufferBase source, bool blocking, long offset, long region, IntPtr destination, IReadOnlyList <OpenCLEventBase> events = null, IList <OpenCLEventBase> newEvents = null)
{
int sizeofT = Marshal.SizeOf(source.ElementType);
int eventWaitListSize;
CLEventHandle[] eventHandles = OpenCLTools.ExtractHandles(events, out eventWaitListSize);
CLEventHandle[] newEventHandle = (newEvents != null) ? new CLEventHandle[1] : null;
OpenCLErrorCode error = CL10.EnqueueReadBuffer(Handle, source.Handle, blocking, new IntPtr(offset * sizeofT), new IntPtr(region * sizeofT), destination, eventWaitListSize, eventHandles, newEventHandle);
OpenCLException.ThrowOnError(error);
if (newEvents != null)
{
lock (newEvents)
{
newEvents.Add(new OpenCLEvent(newEventHandle[0], this));
}
}
}
private void ReorderingKeyValue(CLMemoryHandle inputKey, CLMemoryHandle outputKey, CLMemoryHandle inputValue, CLMemoryHandle outputValue, int bitOffset)
{
ComputeErrorCode error;
IntPtr agentPtrSize = (IntPtr)0;
agentPtrSize = (IntPtr)Marshal.SizeOf(typeof(IntPtr));
var ptrSize = (IntPtr)Marshal.SizeOf(typeof(Mem));
int globalWorkSize = gpuConstants.numThreadsPerBlock * gpuConstants.numBlocks;
int localWorkSize = gpuConstants.numThreadsPerBlock;
IntPtr[] workGroupSizePtr = new IntPtr[] { (IntPtr)globalWorkSize };
IntPtr[] localWorkGroupSizePtr = new IntPtr[] { (IntPtr)localWorkSize };
ComputeEvent clevent;
error = CL10.SetKernelArg(ckReorderingKeyValue, 0, ptrSize, inputKey);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 1, ptrSize, outputKey);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 2, ptrSize, inputValue);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 3, ptrSize, outputValue);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 4, ptrSize, mCounters);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 5, ptrSize, mRadixPrefixes);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 6, (IntPtr)(gpuConstants.numGroupsPerBlock * gpuConstants.numBlocks * gpuConstants.numRadicesPerBlock * 4), null);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 7, (IntPtr)(gpuConstants.numRadices * 4), null);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 8, (IntPtr)(Marshal.SizeOf(typeof(GPUConstants))), gpuConstants);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.SetKernelArg(ckReorderingKeyValue, 9, (IntPtr)4, bitOffset);
CheckErr(error, "CL10.SetKernelArg");
error = CL10.EnqueueNDRangeKernel(cqCommandQueue, ckReorderingKeyValue, 1, null, workGroupSizePtr, localWorkGroupSizePtr, 0, null, out clevent);
CheckErr(error, "CL10.EnqueueNDRangeKernel");
error = CL10.Finish(cqCommandQueue);
CheckErr(error, "CL10.Finish");
if (DEBUG)
{
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine("-------------------------------Reordering-------------------------------------------------");
}
ComputeEvent eve;
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine(" Input ");
}
CL10.EnqueueReadBuffer(cqCommandQueue, inputKey, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintElementBuffer(debugRead, gpuConstants.numTotalElements, "Reordering -> Input -> bitoffset = " + bitOffset);
CL10.EnqueueReadBuffer(cqCommandQueue, inputValue, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintElementBuffer(debugRead, gpuConstants.numTotalElements, "Reordering -> InputValues -> bitoffset = " + bitOffset);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine(" Counters ");
}
CL10.EnqueueReadBuffer(cqCommandQueue, mCounters, Bool.True, IntPtr.Zero, (IntPtr)(numCounters * sizeof(int)), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintCounterBuffer(debugRead, "Reordering -> bitoffset = " + bitOffset);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine(" Counters ");
}
CL10.EnqueueReadBuffer(cqCommandQueue, mRadixPrefixes, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numRadices * sizeof(int)), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintElementBuffer(debugRead, gpuConstants.numRadices, "Reordering -> RadixPrefixe -> bitoffset = " + bitOffset);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine(" Output ");
}
CL10.EnqueueReadBuffer(cqCommandQueue, outputKey, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintElementBuffer(debugRead, gpuConstants.numTotalElements, "Reordering -> Output -> bitoffset = " + bitOffset);
CL10.EnqueueReadBuffer(cqCommandQueue, outputValue, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintElementBuffer(debugRead, gpuConstants.numTotalElements, "Reordering -> OutputValue -> bitoffset = " + bitOffset);
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine("Reordering -> bitoffset = " + bitOffset);
}
if (DEBUG_CONSOLE_OUTPUT)
{
Console.WriteLine();
}
}
;
}
public void sortKeysValue(CLMemoryHandle inputKey, CLMemoryHandle outputKey, CLMemoryHandle inputValue, CLMemoryHandle outputValue,
int numElements)
{
debugRead = new int[Math.Max(numElements, numCounters)];
ComputeErrorCode error;
ComputeEvent eve;
mCounters = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, (IntPtr)(gpuConstants.numGroupsPerBlock * gpuConstants.numRadices * gpuConstants.numBlocks * sizeof(int)), debugRead,
out error);
CheckErr(error, "CL10.CreateBuffer");
mRadixPrefixes = CL10.CreateBuffer(cxGPUContext, ComputeMemoryFlags.ReadWrite, gpuConstants.numRadices * sizeof(int), out error);
CheckErr(error, "CL10.CreateBuffer");
/*
* error = CL10.EnqueueReadBuffer(cqCommandQueue, input, Bool.True, IntPtr.Zero, (IntPtr)(numElements * 4),
* debugRead, 0, null, out eve);
* CheckErr(error, "CL10.EnqueueReadBuffer");
*/
if (DEBUG)
{
CL10.EnqueueReadBuffer(cqCommandQueue, inputKey, Bool.True, IntPtr.Zero, (IntPtr)(gpuConstants.numTotalElements * 4), debugRead, 0,
null, out eve);
CheckErr(error, "CL10.EnqueueReadBuffer");
PrintAsArray(debugRead, gpuConstants.numTotalElements);
}
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(inputKey, 4 * i);
SumIt(inputKey, 4 * i);
ReorderingKeyValue(inputKey, outputKey, inputValue, outputValue, 4 * i);
}
else
{
SetupAndCount(outputKey, 4 * i);
SumIt(outputKey, 4 * i);
ReorderingKeyValue(outputKey, inputKey, outputValue, inputValue, 4 * i);
}
}
if (i % 2 != 0)
{
error = CL10.EnqueueCopyBuffer(cqCommandQueue, inputKey, outputKey, 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, inputValue, outputValue, 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(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++;
}