public void initialize(ComputeContext computeContext, int kernelPositionsLength, Misc.Vector2<int> inputMapSize)
{
ErrorCode errorCode;
OpenCL.Net.Event eventWriteBufferCompletedKernel;
this.kernelPositionsLength = kernelPositionsLength;
kernelResults = new float[kernelPositionsLength];
bufferForPositions = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, 2 * kernelPositionsLength, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForInputMap = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
OpenCL.Net.IMem<float> bufferForKernel = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, (kernelWidth * kernelWidth), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForKernelResults = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, kernelPositionsLength, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.EnqueueWriteBuffer<float>(computeContext.commandQueue, bufferForKernel, OpenCL.Net.Bool.True, this.kernelArray, 0, null, out eventWriteBufferCompletedKernel);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
program = Cl.CreateProgramWithSource(computeContext.context, 1, new[] { getProgramSource(inputMapSize.x, kernelPositionsLength) }, null, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.BuildProgram(program, 1, new[] { computeContext.chosenDevice }, "", null, IntPtr.Zero);
if (errorCode != ErrorCode.Success)
{
OpenCL.Net.InfoBuffer logInfoBuffer = Cl.GetProgramBuildInfo(program, computeContext.chosenDevice, ProgramBuildInfo.Log, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
throw new ComputeContext.OpenClError();
}
kernel = Cl.CreateKernel(program, "kernel0", out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<float>(kernel, 0, bufferForInputMap);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<float>(kernel, 1, bufferForKernel);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<int>(kernel, 2, bufferForPositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<float>(kernel, 3, bufferForKernelResults);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
Cl.ReleaseEvent(eventWriteBufferCompletedKernel);
}
public void calculate(ComputeContext computeContext)
{
ErrorCode errorCode;
int lengthOfPositionsForOpenCl;
int numberOfPositionsForOpenCl;
OpenCL.Net.Event eventWriteBufferInputPositions;
OpenCL.Net.Event eventWriteBufferForInputMap;
OpenCL.Net.Event eventExecutedKernelNearestPoint;
OpenCL.Net.Event eventReadBufferForResultPositions;
OpenCL.Net.Event eventReadBufferForFoundNewPosition;
numberOfPositionsForOpenCl = inputPositions.Length;
numberOfPositionsForOpenCl = numberOfPositionsForOpenCl + 32 + (32 - (numberOfPositionsForOpenCl % 32));
lengthOfPositionsForOpenCl = numberOfPositionsForOpenCl * 2;
if (lengthOfPositionsForOpenCl > numberOfAllocatedInputAndOutputPositions*2)
{
// we need to reallocate the buffer to the right size
numberOfAllocatedInputAndOutputPositions = lengthOfPositionsForOpenCl;
Cl.ReleaseMemObject(bufferForInputPositions);
Cl.ReleaseMemObject(bufferForOutputPositions);
Cl.ReleaseMemObject(bufferForFoundNewPosition);
bufferForInputPositions = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, numberOfAllocatedInputAndOutputPositions * 2, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForOutputPositions = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, numberOfAllocatedInputAndOutputPositions * 2, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForFoundNewPosition = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, numberOfAllocatedInputAndOutputPositions, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
}
// copy the positions
int[] inputPositionsForOpenCl;
inputPositionsForOpenCl = new int[lengthOfPositionsForOpenCl];
int i;
for (i = 0; i < inputPositions.Length; i++ )
{
inputPositionsForOpenCl[i * 2 + 0] = inputPositions[i].x;
inputPositionsForOpenCl[i * 2 + 1] = inputPositions[i].y;
}
errorCode = Cl.EnqueueWriteBuffer<int>(computeContext.commandQueue, bufferForInputPositions, OpenCL.Net.Bool.False, inputPositionsForOpenCl, 0, new Event[] { }, out eventWriteBufferInputPositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
// copy input map
int[] translatedInputMap;
translatedInputMap = new int[inputMap.getWidth() * inputMap.getLength()];
for (i = 0; i < inputMap.getWidth() * inputMap.getLength(); i++ )
{
if( inputMap.unsafeGetValues()[i] )
{
translatedInputMap[i] = 1;
}
}
errorCode = Cl.EnqueueWriteBuffer<int>(computeContext.commandQueue, bufferForInputMap, OpenCL.Net.Bool.False, translatedInputMap, 0, new Event[] { }, out eventWriteBufferForInputMap);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
//errorCode = Cl.SetKernelArg(kernelNearestPoint, 5, (IntPtr)4, /*lengthOfPositionsForOpenCl*/inputPositions.Length); // number of relative positions
//ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
// call kernel
IntPtr[] globalWorkSize = new IntPtr[] { (IntPtr)32, (IntPtr)(numberOfPositionsForOpenCl/32) };
IntPtr[] localWorkSize = new IntPtr[] { (IntPtr)32, (IntPtr)1 };
errorCode = Cl.EnqueueNDRangeKernel(computeContext.commandQueue, kernelNearestPoint, 2, null, globalWorkSize, localWorkSize, 2, new Event[] { eventWriteBufferInputPositions, eventWriteBufferForInputMap }, out eventExecutedKernelNearestPoint);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
int[] resultPositionsFromOpenCl;
int[] resultfoundNewPositionsFromOpenCl;
resultPositionsFromOpenCl = new int[lengthOfPositionsForOpenCl];
resultfoundNewPositionsFromOpenCl = new int[lengthOfPositionsForOpenCl / 2];
// read results back and copy into arrays
errorCode = Cl.EnqueueReadBuffer(computeContext.commandQueue, bufferForOutputPositions, OpenCL.Net.Bool.False, resultPositionsFromOpenCl, 1, new[] { eventExecutedKernelNearestPoint }, out eventReadBufferForResultPositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.EnqueueReadBuffer(computeContext.commandQueue, bufferForFoundNewPosition, OpenCL.Net.Bool.False, resultfoundNewPositionsFromOpenCl, 1, new[] { eventExecutedKernelNearestPoint }, out eventReadBufferForFoundNewPosition);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.Flush(computeContext.commandQueue);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.WaitForEvents(2, new Event[] { eventReadBufferForResultPositions, eventReadBufferForFoundNewPosition });
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
// copy back
outputPositions = new Vector2<int>[inputPositions.Length];
foundNewPositions = new bool[inputPositions.Length];
for( i = 0; i < inputPositions.Length; i++ )
{
// check output position
// NOTE< works only if kernel is prepared >
{
Vector2<int> inputPosition = inputPositions[i];
Vector2<int> outputPosition2;
outputPosition2 = new Vector2<int>();
outputPosition2.x = resultPositionsFromOpenCl[i * 2 + 0];
outputPosition2.y = resultPositionsFromOpenCl[i * 2 + 1];
if (System.Math.Abs(outputPosition2.x - inputPosition.x) > 20 || System.Math.Abs(outputPosition2.y - inputPosition.y) > 20)
{
int here0 = 0;
}
}
Vector2<int> outputPosition;
outputPosition = new Vector2<int>();
outputPosition.x = resultPositionsFromOpenCl[i * 2 + 0];
outputPosition.y = resultPositionsFromOpenCl[i * 2 + 1];
outputPositions[i] = outputPosition;
if( resultfoundNewPositionsFromOpenCl[i] == 1 )
{
foundNewPositions[i] = true;
}
}
Cl.ReleaseEvent(eventWriteBufferInputPositions);
Cl.ReleaseEvent(eventWriteBufferForInputMap);
Cl.ReleaseEvent(eventExecutedKernelNearestPoint);
Cl.ReleaseEvent(eventReadBufferForResultPositions);
Cl.ReleaseEvent(eventReadBufferForFoundNewPosition);
}
public void initialize(ComputeContext computeContext, int kernelRadius, Misc.Vector2<int> inputMapSize)
{
ErrorCode errorCode;
float[] kernelArray;
OpenCL.Net.Event eventWriteBufferCompletedKernel;
bufferForInputMap = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForTemporary = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForOutputMap = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
kernelArray = calculateKernel(kernelRadius);
bufferForKernelArray = Cl.CreateBuffer<float>(computeContext.context, MemFlags.AllocHostPtr, kernelArray.Length, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
// copy kernel into buffer
errorCode = Cl.EnqueueWriteBuffer<float>(computeContext.commandQueue, bufferForKernelArray, OpenCL.Net.Bool.True, kernelArray, 0, null, out eventWriteBufferCompletedKernel);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
string programLocation = Assembly.GetEntryAssembly().Location;
string pathToLoad = Path.Combine(Path.GetDirectoryName(programLocation), "..\\..\\", "ComputationBackend\\OpenCl\\src\\Blur.cl");
string openClSource = File.ReadAllText(pathToLoad);
program = Cl.CreateProgramWithSource(computeContext.context, 1, new[] { openClSource }, null, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.BuildProgram(program, 1, new[] { computeContext.chosenDevice }, "", null, IntPtr.Zero);
if (errorCode != ErrorCode.Success)
{
OpenCL.Net.InfoBuffer logInfoBuffer = Cl.GetProgramBuildInfo(program, computeContext.chosenDevice, ProgramBuildInfo.Log, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
throw new ComputeContext.OpenClError();
}
kernelBlurX = Cl.CreateKernel(program, "blurX", out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
//errorCode = Cl.SetKernelArg<float>(kernelBlurX, 0, bufferForInputMap);
//ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<float>(kernelBlurX, 1, bufferForKernelArray);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
//errorCode = Cl.SetKernelArg<float>(kernelBlurX, 2, bufferForTemporary);
//ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelBlurX, 3, (IntPtr)4, kernelRadius);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelBlurX, 4, (IntPtr)4, inputMapSize.x);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
kernelBlurY = Cl.CreateKernel(program, "blurY", out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
//errorCode = Cl.SetKernelArg<float>(kernelBlurY, 0, bufferForTemporary);
//ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<float>(kernelBlurY, 1, bufferForKernelArray);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
//errorCode = Cl.SetKernelArg<float>(kernelBlurY, 2, bufferForOutputMap);
//ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelBlurY, 3, (IntPtr)4, kernelRadius);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelBlurY, 4, (IntPtr)4, inputMapSize.x);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelBlurY, 5, (IntPtr)4, inputMapSize.y);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
Cl.ReleaseEvent(eventWriteBufferCompletedKernel);
}
public void initialize(ComputeContext computeContext, int searchRadius, Misc.Vector2<int> inputMapSize)
{
List<Vector2<int>> relativePositions;
ErrorCode errorCode;
int[] relativePositionsArray;
OpenCL.Net.Event eventWriteBufferForRelativePositions;
relativePositions = calculateRelativePositionsForRadius(searchRadius);
numberOfAllocatedInputAndOutputPositions = 50000;
bufferForInputMap = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForRelativePositions = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, relativePositions.Count * 2, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
relativePositionsArray = convertRelativePositionsToArray(relativePositions);
// copy relative positions into buffer
errorCode = Cl.EnqueueWriteBuffer<int>(computeContext.commandQueue, bufferForRelativePositions, OpenCL.Net.Bool.True, relativePositionsArray, 0, new Event[] { }, out eventWriteBufferForRelativePositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForInputPositions = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, numberOfAllocatedInputAndOutputPositions * 2, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForOutputPositions = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, numberOfAllocatedInputAndOutputPositions * 2, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForFoundNewPosition = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, numberOfAllocatedInputAndOutputPositions, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
program = Cl.CreateProgramWithSource(computeContext.context, 1, new[] { getOpenClSource() }, null, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.BuildProgram(program, 1, new[] { computeContext.chosenDevice }, "", null, IntPtr.Zero);
if (errorCode != ErrorCode.Success)
{
OpenCL.Net.InfoBuffer logInfoBuffer = Cl.GetProgramBuildInfo(program, computeContext.chosenDevice, ProgramBuildInfo.Log, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
throw new ComputeContext.OpenClError();
}
kernelNearestPoint = Cl.CreateKernel(program, "findNearestPoint", out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<int>(kernelNearestPoint, 0, bufferForInputMap);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<int>(kernelNearestPoint, 1, bufferForRelativePositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<int>(kernelNearestPoint, 2, bufferForInputPositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<int>(kernelNearestPoint, 3, bufferForOutputPositions);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg<int>(kernelNearestPoint, 4, bufferForFoundNewPosition);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelNearestPoint, 5, (IntPtr)4, relativePositions.Count); // number of relative positions
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelNearestPoint, 6, (IntPtr)4, inputMapSize.x);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelNearestPoint, 7, (IntPtr)4, inputMapSize.y);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
Cl.ReleaseEvent(eventWriteBufferForRelativePositions);
}
public void initialize(ComputeContext computeContext, Misc.Vector2<int> inputMapSize)
{
ErrorCode errorCode;
bufferForCounterMap = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForCounterOutputMap = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, (int)(inputMapSize.x * inputMapSize.y), out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
bufferForChangeMade = Cl.CreateBuffer<int>(computeContext.context, MemFlags.AllocHostPtr, 1, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
program = Cl.CreateProgramWithSource(computeContext.context, 1, new[] { getOpenClSource() }, null, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.BuildProgram(program, 1, new[] { computeContext.chosenDevice }, "", null, IntPtr.Zero);
if (errorCode != ErrorCode.Success)
{
OpenCL.Net.InfoBuffer logInfoBuffer = Cl.GetProgramBuildInfo(program, computeContext.chosenDevice, ProgramBuildInfo.Log, out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
throw new ComputeContext.OpenClError();
}
kernelNarrow = Cl.CreateKernel(program, "narrow", out errorCode);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelNarrow, 4, (IntPtr)4, inputMapSize.x);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
errorCode = Cl.SetKernelArg(kernelNarrow, 5, (IntPtr)4, inputMapSize.y);
ComputeContext.throwErrorIfNotSuccessfull(errorCode, "");
}
