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 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;
}
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);
}
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);
}
