public float ElapsedTime(CUevent start, CUevent end)
{
float pMilliseconds = 0f;
this.LastError = CUDADriver.cuEventElapsedTime(ref pMilliseconds, start, end);
return(pMilliseconds);
}
public static void cuEventRecord(CUevent hEvent, CUstream hStream)
{
Wrap(() =>
{
try
{
var error = nativeEventRecord(hEvent, hStream);
if (error != CUresult.CUDA_SUCCESS)
{
throw new CudaException(error);
}
}
catch (CudaException)
{
throw;
}
catch (DllNotFoundException dnfe)
{
throw new CudaException(CudaError.NoDriver, dnfe);
}
catch (Exception e)
{
throw new CudaException(CudaError.Unknown, e);
}
});
}
public CUevent CreateEvent(CUEventFlags flags)
{
CUevent phEvent = new CUevent();
this.LastError = CUDADriver.cuEventCreate(ref phEvent, (uint)flags);
return(phEvent);
}
public static ElapsedTime cuEventElapsedTime(CUevent hStart, CUevent hEnd)
{
return(Wrap(() =>
{
try
{
float milliseconds;
var error = nativeEventElapsedTime(out milliseconds, hStart, hEnd);
if (error != CUresult.CUDA_SUCCESS)
{
throw new CudaException(error);
}
// note. cannot use TimeSpan here because it ain't work with fractions of milliseconds
return new ElapsedTime(milliseconds);
}
catch (CudaException)
{
throw;
}
catch (DllNotFoundException dnfe)
{
throw new CudaException(CudaError.NoDriver, dnfe);
}
catch (Exception e)
{
throw new CudaException(CudaError.Unknown, e);
}
}));
}
/// <summary>
/// Creates a new Event
/// </summary>
/// <param name="flags">Parameters for event creation</param>
public CudaEvent(CUEventFlags flags)
{
_event = new CUevent();
res = DriverAPINativeMethods.Events.cuEventCreate(ref _event, flags);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuEventCreate", res));
if (res != CUResult.Success) throw new CudaException(res);
}
public CudaTest(CUDA cuda)
{
_cuda = cuda;
//+ load frame module
LoadFrameModule(Cuda.CudaModle.BuiltinModules.Test);
//+ create cuda event handles
_cudaStartEvent = _cuda.CreateEvent();
_cudaStopEvent = _cuda.CreateEvent();
}
public void DestroyEvent(CUevent e)
{
if (_version >= 4000)
{
this.LastError = CUDADriver.cuEventDestroy_v2(e);
}
else
{
this.LastError = CUDADriver.cuEventDestroy(e);
}
}
/// <summary>
/// Make a compute stream wait on an event<para/>
/// Makes all future work submitted to the Stream wait until <c>hEvent</c>
/// reports completion before beginning execution. This synchronization
/// will be performed efficiently on the device.
/// <para/>
/// The stream will wait only for the completion of the most recent
/// host call to <see cref="CudaEvent.Record()"/> on <c>hEvent</c>. Once this call has returned,
/// any functions (including <see cref="CudaEvent.Record()"/> and <see cref="Dispose()"/> may be
/// called on <c>hEvent</c> again, and the subsequent calls will not have any
/// effect on this stream.
/// <para/>
/// If <c>hStream</c> is 0 (the NULL stream) any future work submitted in any stream
/// will wait for <c>hEvent</c> to complete before beginning execution. This
/// effectively creates a barrier for all future work submitted to the context.
/// <para/>
/// If <see cref="CudaEvent.Record()"/> has not been called on <c>hEvent</c>, this call acts as if
/// the record has already completed, and so is a functional no-op.
/// </summary>
/// <returns></returns>
public void WaitEvent(CUevent cuevent)
{
if (disposed)
{
throw new ObjectDisposedException(this.ToString());
}
res = DriverAPINativeMethods.Streams.cuStreamWaitEvent(_stream, cuevent, 0);
Debug.WriteLine(String.Format("{0:G}, {1}: {2}", DateTime.Now, "cuStreamWaitEvent", res));
if (res != CUResult.Success)
{
throw new CudaException(res);
}
}
public CUDAExecution(GASS.CUDA.CUDA cuda, string module, string function)
{
this.parameters = new List <Parameter>();
this.textures = new List <CUtexref>();
this.CUDAInstance = cuda;
if (!module.EndsWith("cubin"))
{
module = module + ".cubin";
}
FileInfo info = new FileInfo(module);
this.Module = info.FullName;
this.Function = function;
this.CUDAModule = cuda.LoadModule(this.module);
this.CUDAFunction = cuda.GetModuleFunction(this.function);
this.start = cuda.CreateEvent();
this.stop = cuda.CreateEvent();
}
// http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/online/group__CUEVENT_g7895332c94680b174ef41373af09d9ce.html private static extern CUresult nativeEventElapsedTime(out float pMilliseconds, CUevent hStart, CUevent hEnd);
public void RecordEvent(CUevent e)
{
this.RecordEvent(e, new CUstream());
}
static void Main(string[] args)
{
// Init and select 1st device.
CUDA cuda = new CUDA(0, true);
// load module
//cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "transpose_kernel.cubin"));
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "transpose_kernel.ptx"));
CUfunction transpose = cuda.GetModuleFunction("transpose");
CUfunction transpose_naive = cuda.GetModuleFunction("transpose_naive");
const int size_x = 4096;
const int size_y = 4096;
const int mem_size = sizeof(float) * size_x * size_y;
float[] h_idata = new float[size_x * size_y];
for (int i = 0; i < h_idata.Length; i++)
{
h_idata[i] = (float)i;
}
// allocate device memory
// copy host memory to device
CUdeviceptr d_idata = cuda.CopyHostToDevice <float>(h_idata);
CUdeviceptr d_odata = cuda.Allocate <float>(h_idata);
// setup execution parameters
cuda.SetFunctionBlockShape(transpose_naive, BLOCK_DIM, BLOCK_DIM, 1);
cuda.SetParameter(transpose_naive, 0, (uint)d_odata.Pointer);
cuda.SetParameter(transpose_naive, IntPtr.Size, (uint)d_idata.Pointer);
cuda.SetParameter(transpose_naive, IntPtr.Size * 2, (uint)size_x);
cuda.SetParameter(transpose_naive, IntPtr.Size * 2 + 4, (uint)size_y);
cuda.SetParameterSize(transpose_naive, (uint)(IntPtr.Size * 2 + 8));
cuda.SetFunctionBlockShape(transpose, BLOCK_DIM, BLOCK_DIM, 1);
cuda.SetParameter(transpose, 0, (uint)d_odata.Pointer);
cuda.SetParameter(transpose, IntPtr.Size, (uint)d_idata.Pointer);
cuda.SetParameter(transpose, IntPtr.Size * 2, (uint)size_x);
cuda.SetParameter(transpose, IntPtr.Size * 2 + 4, (uint)size_y);
cuda.SetParameterSize(transpose, (uint)(IntPtr.Size * 2 + 8));
// warmup so we don't time CUDA startup
cuda.Launch(transpose_naive, size_x / BLOCK_DIM, size_y / BLOCK_DIM);
cuda.Launch(transpose, size_x / BLOCK_DIM, size_y / BLOCK_DIM);
//System.Threading.Thread.Sleep(10);
int numIterations = 100;
Console.WriteLine("Transposing a {0} by {1} matrix of floats...", size_x, size_y);
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
cuda.RecordEvent(start);
for (int i = 0; i < numIterations; i++)
{
cuda.Launch(transpose_naive, size_x / BLOCK_DIM, size_y / BLOCK_DIM);
}
cuda.SynchronizeContext();
cuda.RecordEvent(end);
cuda.SynchronizeContext();
float naiveTime = cuda.ElapsedTime(start, end);
Console.WriteLine("Naive transpose average time: {0} ms\n", naiveTime / numIterations);
cuda.RecordEvent(start);
for (int i = 0; i < numIterations; i++)
{
cuda.Launch(transpose, size_x / BLOCK_DIM, size_y / BLOCK_DIM);
}
cuda.SynchronizeContext();
cuda.RecordEvent(end);
cuda.SynchronizeContext();
float optimizedTime = cuda.ElapsedTime(start, end);
Console.WriteLine("Optimized transpose average time: {0} ms\n", optimizedTime / numIterations);
float[] h_odata = new float[size_x * size_y];
cuda.CopyDeviceToHost <float>(d_odata, h_odata);
float[] reference = new float[size_x * size_y];
computeGold(reference, h_idata, size_x, size_y);
bool res = CompareF(reference, h_odata, size_x * size_y);
Console.WriteLine("Test {0}", res == true? "PASSED":"FAILED");
cuda.Free(d_idata);
cuda.Free(d_odata);
Console.ReadKey();
}
/// <summary>
/// implementation of sparese matrix product
/// </summary>
/// <param name="repetition">how many times kernel should be launch</param>
/// <param name="moduleFunction">cuda kenrel name</param>
/// <param name="blockSizeX">block size X</param>
/// <param name="blockSizeY">block size Y</param>
/// <param name="transposeGrid">indicate that grid dimensions should be
/// computed alternativly, if false than gridDimY- connected with rows
/// else gridDim.Y conected with cols</param>
/// <returns></returns>
public static float[] CRSSparseMM(int repetition, string moduleFunction,
int blockSizeX, int blockSizeY, bool transposeGrid)
{
//int blockSizeX = 4;
//int blockSizeY = 4;
CUDA cuda = new CUDA(0, true);
// load module
CUmodule module = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "matrixKernels.cubin"));
CUfunction cuFunc = cuda.GetModuleFunction(moduleFunction);
int maxRowSize = avgElements + stdElements - 1;
Console.WriteLine("------------------------------------");
Console.WriteLine("init Matrix");
Stopwatch t = Stopwatch.StartNew();
//values in CRS format
float[] AVals, BVals;
//indexes in Crs format
int[] AIdx, BIdx;
//Lenght of each row in CRS format
int[] ARowLen, BRowLen;
int maxIndex = 0;
MakeRandCrsSparseMatrix(Rows, maxRowSize, out AVals, out AIdx, out ARowLen, out maxIndex);
// DisplayCrsMatrix(AVals, AIdx, ARowLen,maxIndex);
MakeRandCrsSparseMatrix(Cols, maxRowSize, out BVals, out BIdx, out BRowLen, out maxIndex);
//DisplayCrsMatrix(BVals, BIdx, BRowLen, maxIndex);
Console.WriteLine("Init takes {0}", t.Elapsed);
t.Start();
CUdeviceptr AValsPtr = cuda.CopyHostToDevice(AVals);
CUdeviceptr AIdxPtr = cuda.CopyHostToDevice(AIdx);
CUdeviceptr ALenghtPtr = cuda.CopyHostToDevice(ARowLen);
CUdeviceptr BValsPtr = cuda.CopyHostToDevice(BVals);
CUdeviceptr BIdxPtr = cuda.CopyHostToDevice(BIdx);
CUdeviceptr BLenghtPtr = cuda.CopyHostToDevice(BRowLen);
int outputSize = Rows * Cols;
float[] output = new float[outputSize];
//CUdeviceptr dOutput = cuda.Allocate(output);
IntPtr outputPtr2 = cuda.HostAllocate((uint)(outputSize * sizeof(float)), CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
CUdeviceptr dOutput = cuda.GetHostDevicePointer(outputPtr2, 0);
Console.WriteLine("copy to device takes {0}", t.Elapsed);
#region set cuda parameters
int Aelements = AVals.Length;
int Belements = BVals.Length;
cuda.SetFunctionBlockShape(cuFunc, blockSizeX, blockSizeY, 1);
int offset = 0;
cuda.SetParameter(cuFunc, offset, AValsPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, AIdxPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, ALenghtPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, BValsPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, BIdxPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, BLenghtPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, (uint)Rows);
offset += sizeof(int);
cuda.SetParameter(cuFunc, offset, (uint)Cols);
offset += sizeof(int);
cuda.SetParameter(cuFunc, offset, (uint)Aelements);
offset += sizeof(int);
cuda.SetParameter(cuFunc, offset, (uint)Belements);
offset += sizeof(int);
cuda.SetParameterSize(cuFunc, (uint)offset);
#endregion
Console.WriteLine("start computation");
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
//CUtexref cuTexRef = cuda.GetModuleTexture(module, "texRef");
//cuda.SetTextureFlags(cuTexRef, 0);
int gridDimX = (int)Math.Ceiling((Cols + 0.0) / (blockSizeX));
int gridDimY = (int)Math.Ceiling((0.0 + Rows) / blockSizeY);
if (transposeGrid)
{
gridDimX = (int)Math.Ceiling((Rows + 0.0) / (blockSizeX));
gridDimY = (int)Math.Ceiling((0.0 + Cols) / blockSizeY);
}
Stopwatch timer = Stopwatch.StartNew();
cuda.RecordEvent(start);
for (int k = 0; k < repetition; k++)
{
cuda.Launch(cuFunc, gridDimX, gridDimY);
cuda.SynchronizeContext();
// cuda.CopyDeviceToHost(dOutput, output);
Marshal.Copy(outputPtr2, output, 0, outputSize);
}
cuda.RecordEvent(end);
cuda.SynchronizeContext();
timer.Stop();
float cudaTime = cuda.ElapsedTime(start, end);
Console.WriteLine("Matrix products with kernel {0}", moduleFunction);
Console.WriteLine(" takes {0} ms stopwatch time {1} ms", cudaTime, timer.Elapsed);
int lenght = displayCount;// Math.Min(displayCount, Rows);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
cuda.Free(AValsPtr);
cuda.Free(AIdxPtr);
cuda.Free(ALenghtPtr);
cuda.Free(BValsPtr);
cuda.Free(BIdxPtr);
cuda.Free(BLenghtPtr);
cuda.Free(dOutput);
cuda.DestroyEvent(start);
cuda.DestroyEvent(end);
return(output);
}
/// <summary>
/// Creates a new Event using <see cref="CUEventFlags.Default"/>
/// </summary>
internal CudaEvent(CUevent event_)
{
_event = event_;
}
public void RecordEvent(CUevent e, CUstream stream)
{
this.LastError = CUDADriver.cuEventRecord(e, stream);
}
public static float[] CRSSparseMMwithDenseVector(int repetition,
string moduleFunction, int blockSizeX, int blockSizeY)
{
CUDA cuda = new CUDA(0, true);
// load module
CUmodule module = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "matrixKernels.cubin"));
CUfunction cuFunc = cuda.GetModuleFunction(moduleFunction);
int maxRowSize = avgElements + stdElements - 1;
Console.WriteLine("------------------------------------");
Console.WriteLine("init Matrix");
Stopwatch t = Stopwatch.StartNew();
//values in CRS format
float[] AVals, BVals;
//indexes in Crs format
int[] AIdx, BIdx;
//Lenght of each row in CRS format
int[] ARowLen, BRowLen;
int maxIndex = 0;
MakeRandCrsSparseMatrix(Rows, maxRowSize, out AVals, out AIdx, out ARowLen, out maxIndex);
// DisplayCrsMatrix(AVals, AIdx, ARowLen,maxIndex);
MakeRandCrsSparseMatrix(Cols, maxRowSize, out BVals, out BIdx, out BRowLen, out maxIndex);
//DisplayCrsMatrix(BVals, BIdx, BRowLen, maxIndex);
Console.WriteLine("Init takes {0}", t.Elapsed);
t.Start();
CUdeviceptr AValsPtr = cuda.CopyHostToDevice(AVals);
CUdeviceptr AIdxPtr = cuda.CopyHostToDevice(AIdx);
CUdeviceptr ALenghtPtr = cuda.CopyHostToDevice(ARowLen);
int outputSize = Rows * Cols;
float[] output = new float[outputSize];
//allocate memory for output
IntPtr outputPtr2 = cuda.HostAllocate((uint)(outputSize * sizeof(float)), CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
CUdeviceptr dOutput = cuda.GetHostDevicePointer(outputPtr2, 0);
//create dense vector for each column in B matrix
float[] mainVec = new float[maxIndex + 1];
uint memSize = (uint)((maxIndex + 1) * sizeof(float));
CUstream stream0 = cuda.CreateStream();
IntPtr[] mainVecIntPtrs = new IntPtr[2];
//write combined memory allocation
//IntPtr mainVecIPtr = cuda.HostAllocate(memSize,CUDADriver.CU_MEMHOSTALLOC_WRITECOMBINED);
//CUdeviceptr mainVecPtr=cuda.CopyHostToDeviceAsync(mainVecIPtr,memSize,stream0);
//
//mainVecIntPtrs[0] = cuda.HostAllocate(memSize, CUDADriver.CU_MEMHOSTALLOC_WRITECOMBINED);
//mainVecIntPtrs[1] = cuda.HostAllocate(memSize, CUDADriver.CU_MEMHOSTALLOC_WRITECOMBINED);
mainVecIntPtrs[0] = cuda.AllocateHost(memSize);
mainVecIntPtrs[1] = cuda.AllocateHost(memSize);
CUdeviceptr mainVecPtr = cuda.CopyHostToDeviceAsync(mainVecIntPtrs[0], memSize, stream0);
//IntPtr mainVecIPtr = cuda.HostAllocate(memSize,CUDADriver.CU_MEMHOSTALLOC_PORTABLE);
//CUdeviceptr mainVecPtr=cuda.CopyHostToDeviceAsync(mainVecIPtr,memSize,stream0);
//mapped memory allocation
//IntPtr mainVecIPtr = cuda.HostAllocate(memSize, CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
//CUdeviceptr mainVecPtr = cuda.CopyHostToDevice(mainVecIPtr, memSize);
//get texture reference
CUtexref cuTexRef = cuda.GetModuleTexture(module, "vectorTexRef");
cuda.SetTextureFlags(cuTexRef, 0);
cuda.SetTextureAddress(cuTexRef, mainVecPtr, memSize);
Console.WriteLine("copy to device takes {0}", t.Elapsed);
#region set cuda parameters
int Aelements = AVals.Length;
cuda.SetFunctionBlockShape(cuFunc, blockSizeX, blockSizeY, 1);
int offset = 0;
cuda.SetParameter(cuFunc, offset, AValsPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, AIdxPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, ALenghtPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, (uint)Rows);
offset += sizeof(int);
cuda.SetParameter(cuFunc, offset, (uint)Cols);
offset += sizeof(int);
int colIndexParamOffset = offset;
cuda.SetParameter(cuFunc, offset, (uint)0);
offset += sizeof(int);
cuda.SetParameterSize(cuFunc, (uint)offset);
#endregion
Console.WriteLine("start computation");
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
int gridDimX = (int)Math.Ceiling((Rows + 0.0) / (blockSizeX));
int gridDim = (Rows + blockSizeX - 1) / blockSizeX;
Stopwatch timer = Stopwatch.StartNew();
cuda.RecordEvent(start);
for (int rep = 0; rep < repetition; rep++)
{
for (int k = 0; k < Cols; k++)
{
Helpers.InitBuffer(BVals, BIdx, BRowLen, k, mainVecIntPtrs[k % 2]);
cuda.SynchronizeStream(stream0);
cuda.CopyHostToDeviceAsync(mainVecPtr, mainVecIntPtrs[k % 2], memSize, stream0);
cuda.SetParameter(cuFunc, colIndexParamOffset, (uint)k);
cuda.LaunchAsync(cuFunc, gridDimX, 1, stream0);
//cuda.SynchronizeStream(stream0);
////clear host buffer
Helpers.SetBufferIdx(BIdx, BRowLen, k - 1, mainVecIntPtrs[(k + 1) % 2], 0.0f);
//Helpers.InitBuffer(BVals, BIdx, BRowLen, k, mainVecIPtr);
////make asynchronius copy and kernel lauch
//cuda.CopyHostToDeviceAsync(mainVecPtr, mainVecIPtr, memSize, stream0);
//cuda.SetParameter(cuFunc, colIndexParamOffset,(uint) k);
//cuda.LaunchAsync(cuFunc, gridDimX, 1, stream0);
//cuda.SynchronizeStream(stream0);
////clear host buffer
//Helpers.SetBufferIdx(BIdx, BRowLen, k, mainVecIPtr, 0.0f);
}
}
cuda.RecordEvent(end);
cuda.SynchronizeContext();
timer.Stop();
float cudaTime = cuda.ElapsedTime(start, end);
Marshal.Copy(outputPtr2, output, 0, outputSize);
Console.WriteLine("Matrix products with kernel {0}", moduleFunction);
Console.WriteLine(" takes {0} ms stopwatch time {1} ms", cudaTime, timer.Elapsed);
int lenght = displayCount;// Math.Min(displayCount, Rows);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
cuda.Free(AValsPtr);
cuda.Free(AIdxPtr);
cuda.Free(ALenghtPtr);
cuda.Free(dOutput);
cuda.DestroyEvent(start);
cuda.DestroyEvent(end);
cuda.DestroyStream(stream0);
cuda.Free(mainVecPtr);
cuda.DestroyTexture(cuTexRef);
return(output);
}
// http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/online/group__CUEVENT_g349006734f6e7378ea36cb57c239d4c7.html private static extern CUresult nativeEventDestroy(CUevent hEvent);
//http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/online/group__CUEVENT_g93468fbdae4190b79926381a90a94301.html private static extern CUresult nativeEventRecord(CUevent hEvent, CUstream hStream);
static void Main(string[] args)
{
// Create a new instance of CUDA class, select 1st device.
CUDA cuda = new CUDA(0, true);
// Prepare parameters.
int n = 16 * 1024 * 1024;
uint nbytes = (uint)(n * sizeof(int));
int value = 26;
// allocate host memory
int[] a = new int[n];
// allocate device memory
CUdeviceptr d_a = cuda.Allocate <int>(a);
CUDADriver.cuMemsetD8(d_a, 0xff, nbytes);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "asyncAPI.ptx"));
CUfunction func = cuda.GetModuleFunction("increment_kernel");
// set kernel launch configuration
cuda.SetFunctionBlockShape(func, 512, 1, 1);
// create cuda event handles
CUevent start = cuda.CreateEvent();
CUevent stop = cuda.CreateEvent();
// asynchronously issue work to the GPU (all to stream 0)
CUstream stream = new CUstream();
cuda.RecordEvent(start);
cuda.CopyHostToDeviceAsync <int>(d_a, a, stream);
// set parameters for kernel function
cuda.SetParameter(func, 0, (uint)d_a.Pointer);
cuda.SetParameter(func, IntPtr.Size, (uint)value);
cuda.SetParameterSize(func, (uint)(IntPtr.Size + 4));
// actually launch kernel
cuda.LaunchAsync(func, n / 512, 1, stream);
// wait for every thing to finish, then start copy back data
cuda.CopyDeviceToHostAsync <int>(d_a, a, stream);
cuda.RecordEvent(stop);
// print the cpu and gpu times
Console.WriteLine("time spent executing by the GPU: {0} ms", cuda.ElapsedTime(start, stop));
// check the output for correctness
if (CorrectOutput(a, value))
{
Console.WriteLine("Test PASSED");
}
else
{
Console.WriteLine("Test FAILED");
}
// release resources
cuda.DestroyEvent(start);
cuda.DestroyEvent(stop);
cuda.Free(d_a);
}
public void SynchronizeEvent(CUevent e)
{
this.LastError = CUDADriver.cuEventSynchronize(e);
}
// http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/online/group__CUEVENT_g433317083f929b9298f8a88d57aa5017.html private static extern CUresult nativeEventCreate(out CUevent phEvent, CUevent_flags Flags);
// http://developer.download.nvidia.com/compute/cuda/3_1/toolkit/docs/online/group__CUEVENT_ge3ed6a308c602d139373895cb99cb7ab.html private static extern CUresult nativeEventSynchronize(CUevent hEvent);
public static void cuEventRecord(CUevent hEvent)
{
cuEventRecord(hEvent, CUstream.Null);
}
