private static int[] CalculateElementsFrequencies <T>(IList <T> elements, BitmapWrapper bitmapWrapper)
{
int frequencesSize = ProjectConstants.BlockSize * ProjectConstants.GridSize;
var elementsFrequencies = new int[frequencesSize];
using (var cuda = new CUDA(0, true))
{
var path = Path.Combine(Path.GetDirectoryName(Assembly.GetExecutingAssembly().Location),
"apriori_count1.cubin");
cuda.LoadModule(path);
var inputData = cuda.CopyHostToDevice(bitmapWrapper.RgbValues);
var inputSetData = cuda.CopyHostToDevice(elements.ToArray());
var answer = cuda.Allocate(new int[frequencesSize]);
CallTheFrequencyCount(cuda, inputData, inputSetData, answer, bitmapWrapper, elements.Count);
cuda.CopyDeviceToHost(answer, elementsFrequencies);
cuda.Free(inputData);
cuda.Free(answer);
cuda.UnloadModule();
}
return(elementsFrequencies);
}
public override void Init()
{
cuda = new CUDA(0, true);
var cuCtx = cuda.CreateContext(0, CUCtxFlags.MapHost);
cuda.SetCurrentContext(cuCtx);
cuModule = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, cudaModuleName));
cuFunc = cuda.GetModuleFunction(cudaEvaluatorKernelName);
cuFuncSign = cuda.GetModuleFunction(cudaSignKernelName);
//reserved memory based on dimension of support vector
//svVector = new float[TrainedModel.SupportElements[0].Count];
stream = cuda.CreateStream();
//memSvSize = (uint)(TrainedModel.SupportElements[0].Count * sizeof(float));
memSvSize = (uint)(TrainedModel.SupportElements[0].Dim * sizeof(float));
//allocates memory for buffers
svVecIntPtrs[0] = cuda.AllocateHost(memSvSize);
svVecIntPtrs[1] = cuda.AllocateHost(memSvSize);
mainVecPtr = cuda.CopyHostToDeviceAsync(svVecIntPtrs[0], memSvSize, stream);
cuSVTexRef = cuda.GetModuleTexture(cuModule, "svTexRef");
cuda.SetTextureFlags(cuSVTexRef, 0);
cuda.SetTextureAddress(cuSVTexRef, mainVecPtr, memSvSize);
//todo: copy labels and alphas
float[] svLabels = new float[TrainedModel.SupportElements.Length];
float[] svAlphas = new float[TrainedModel.SupportElements.Length];
Parallel.For(0, TrainedModel.SupportElementsIndexes.Length,
i => {
int idx = TrainedModel.SupportElementsIndexes[i];
svLabels[i] = TrainedModel.Y[i];
//svLabels[i] = TrainningProblem.Labels[idx];
svAlphas[i] = TrainedModel.Alpha[idx];
});
//for (int i = 0; i < TrainedModel.SupportElementsIndexes.Length; i++)
//{
// int idx = TrainedModel.SupportElementsIndexes[i];
// svLabels[i]= TrainningProblem.Labels[idx];
// svAlphas[i] = TrainedModel.Alpha[idx];
//}
labelsPtr = cuda.CopyHostToDevice(svLabels);
alphasPtr = cuda.CopyHostToDevice(svAlphas);
IsInitialized = true;
}
public virtual void SetMemoryForDenseVector(int i, int j)
{
if (IVectorIdx != i)
{
SparseVec vecI = problemElements[i];
CudaHelpers.FillDenseVector(vecI, VectorI);
cuda.CopyHostToDevice(VecIPtr, VectorI);
}
if (JVectorIdx != j)
{
SparseVec vecJ = problemElements[j];
CudaHelpers.FillDenseVector(vecJ, VectorJ);
cuda.CopyHostToDevice(VecJPtr, VectorJ);
}
}
public virtual void SetMemoryForDenseVector(int mainIndex)
{
SparseVec mainVec = problemElements[mainIndex];
if (mainVectorIdx != mainIndex)
{
CudaHelpers.FillDenseVector(mainVec, mainVector);
cuda.CopyHostToDevice(mainVecPtr, mainVector);
}
}
// It's interesting to change the number of blocks and the number of threads to
// understand how to keep the hardware busy.
//
// Here are some numbers I get on my G80:
// blocks - clocks
// 1 - 3096
// 8 - 3232
// 16 - 3364
// 32 - 4615
// 64 - 9981
//
// With less than 16 blocks some of the multiprocessors of the device are idle. With
// more than 16 you are using all the multiprocessors, but there's only one block per
// multiprocessor and that doesn't allow you to hide the latency of the memory. With
// more than 32 the speed scales linearly.
static void Main(string[] args)
{
// Init CUDA, select 1st device.
CUDA cuda = new CUDA(0, true);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "clock_kernel.cubin"));
CUfunction func = cuda.GetModuleFunction("timedReduction");
int[] timer = new int[NUM_BLOCKS * 2];
float[] input = new float[NUM_THREADS * 2];
for (int i = 0; i < NUM_THREADS * 2; i++)
{
input[i] = (float)i;
}
CUdeviceptr dinput = cuda.CopyHostToDevice <float>(input);
CUdeviceptr doutput = cuda.Allocate((uint)(sizeof(float) * NUM_BLOCKS));
CUdeviceptr dtimer = cuda.Allocate <int>(timer);
cuda.SetParameter(func, 0, (uint)dinput.Pointer);
cuda.SetParameter(func, IntPtr.Size, (uint)doutput.Pointer);
cuda.SetParameter(func, IntPtr.Size * 2, (uint)dtimer.Pointer);
cuda.SetParameterSize(func, (uint)(IntPtr.Size * 3));
//timedReduction<<<NUM_BLOCKS, NUM_THREADS, sizeof(float) * 2 * NUM_THREADS>>>(dinput, doutput, dtimer);
cuda.SetFunctionBlockShape(func, NUM_THREADS, 1, 1);
cuda.SetFunctionSharedSize(func, (uint)(sizeof(float) * 2 * NUM_THREADS));
cuda.Launch(func, NUM_BLOCKS, 1);
cuda.CopyDeviceToHost <int>(dtimer, timer);
cuda.Free(dinput);
cuda.Free(doutput);
cuda.Free(dtimer);
foreach (int i in timer)
{
Console.WriteLine(i);
}
Console.WriteLine("Test PASSED");
int minStart = timer[0];
int maxEnd = timer[NUM_BLOCKS];
for (int i = 1; i < NUM_BLOCKS; i++)
{
minStart = timer[i] < minStart ? timer[i] : minStart;
maxEnd = timer[NUM_BLOCKS + i] > maxEnd ? timer[NUM_BLOCKS + i] : maxEnd;
}
Console.WriteLine("time = {0}", maxEnd - minStart);
}
private static int[] CalculateCandidatesFrequencies <T>(IList <List <int> > candidates, BitmapWrapper bitmapWrapper, double borderValue)
{
using (var cuda = new CUDA(0, true))
{
var elementsFrequencies = new int[candidates.Count];
var inputSets = new int[candidates.Count * candidates[0].Count];
var ind = 0;
foreach (var candidate in candidates)
{
foreach (var i in candidate)
{
inputSets[ind] = i;
ind++;
}
}
var path = Path.Combine(Path.GetDirectoryName(Assembly.GetExecutingAssembly().Location),
Names.CudaAprioriCountModule);
cuda.LoadModule(path);
var inputData = cuda.CopyHostToDevice(bitmapWrapper.RgbValues);
var inputSetData = cuda.CopyHostToDevice(inputSets);
var frequenciesOnHost = cuda.Allocate(new int[candidates.Count]);
var sw = new Stopwatch();
sw.Start();
CallTheSetsFrequenciesCount(cuda, inputData, inputSetData, frequenciesOnHost, bitmapWrapper, candidates[0].Count, candidates.Count, borderValue);
sw.Stop();
//Console.WriteLine("CalculateCandidatesFrequencies: {0} ms", sw.ElapsedMilliseconds);
cuda.CopyDeviceToHost(frequenciesOnHost, elementsFrequencies);
cuda.Free(inputData);
cuda.Free(inputSetData);
cuda.Free(frequenciesOnHost);
cuda.UnloadModule();
return(elementsFrequencies);
}
}
//
// A sorting network is a sorting algorith, where the sequence of comparisons
// is not data-dependent. That makes them suitable for parallel implementations.
//
// Bitonic sort is one of the fastest sorting networks, consisting of o(n log^2 n)
// comparators. It has a simple implemention and it's very efficient when sorting
// a small number of elements:
//
// http://citeseer.ist.psu.edu/blelloch98experimental.html
//
// This implementation is based on:
//
// http://www.tools-of-computing.com/tc/CS/Sorts/bitonic_sort.htm
//
static void Main(string[] args)
{
const int NUM = 256;
// Init CUDA, select 1st device.
CUDA cuda = new CUDA(0, true);
// create values
int[] values = new int[NUM];
Random rand = new Random();
for (int i = 0; i < NUM; i++)
{
values[i] = rand.Next();
}
// allocate memory and copy to device
CUdeviceptr dvalues = cuda.CopyHostToDevice <int>(values);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "bitonic.cubin"));
CUfunction func = cuda.GetModuleFunction("bitonicSort");
cuda.SetParameter(func, 0, (uint)dvalues.Pointer);
cuda.SetParameterSize(func, (uint)IntPtr.Size);
//bitonicSort<<<1, NUM, sizeof(int) * NUM>>>(dvalues);
cuda.SetFunctionBlockShape(func, NUM, 1, 1);
cuda.SetFunctionSharedSize(func, sizeof(int) * NUM);
cuda.Launch(func, 1, 1);
cuda.CopyDeviceToHost <int>(dvalues, values);
cuda.Free(dvalues);
bool passed = true;
for (int i = 1; i < NUM; i++)
{
if (values[i - 1] > values[i])
{
passed = false;
break;
}
}
Console.WriteLine("Test {0}", passed ? "PASSED" : "FAILED");
}
//
// A sorting network is a sorting algorith, where the sequence of comparisons
// is not data-dependent. That makes them suitable for parallel implementations.
//
// Bitonic sort is one of the fastest sorting networks, consisting of o(n log^2 n)
// comparators. It has a simple implemention and it's very efficient when sorting
// a small number of elements:
//
// http://citeseer.ist.psu.edu/blelloch98experimental.html
//
// This implementation is based on:
//
// http://www.tools-of-computing.com/tc/CS/Sorts/bitonic_sort.htm
//
static void Main(string[] args)
{
const int NUM = 256;
// Init CUDA, select 1st device.
CUDA cuda = new CUDA(0, true);
// create values
int[] values = new int[NUM];
Random rand = new Random();
for (int i = 0; i < NUM; i++)
{
values[i] = rand.Next();
}
// allocate memory and copy to device
CUdeviceptr dvalues = cuda.CopyHostToDevice<int>(values);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "bitonic.cubin"));
CUfunction func = cuda.GetModuleFunction("bitonicSort");
cuda.SetParameter(func, 0, (uint)dvalues.Pointer);
cuda.SetParameterSize(func, (uint)IntPtr.Size);
//bitonicSort<<<1, NUM, sizeof(int) * NUM>>>(dvalues);
cuda.SetFunctionBlockShape(func, NUM, 1, 1);
cuda.SetFunctionSharedSize(func, sizeof(int) * NUM);
cuda.Launch(func, 1, 1);
cuda.CopyDeviceToHost<int>(dvalues, values);
cuda.Free(dvalues);
bool passed = true;
for (int i = 1; i < NUM; i++)
{
if (values[i - 1] > values[i])
{
passed = false;
break;
}
}
Console.WriteLine("Test {0}", passed ? "PASSED" : "FAILED");
}
private int[] CalculateElementsFrequencies(BitmapWrapper bitmapWrapper)
{
var elementsFrequencies = new int[bitmapWrapper.Width];
using (var cuda = new CUDA(0, true))
{
var path = Path.Combine(Path.GetDirectoryName(Assembly.GetExecutingAssembly().Location),
Names.CudaAprioriCountModule);
cuda.LoadModule(path);
var inputData = cuda.CopyHostToDevice(bitmapWrapper.RgbValues);
var answer = cuda.Allocate(new int[bitmapWrapper.Width]);
CallTheFrequencyCount(cuda, inputData, answer, bitmapWrapper);
cuda.CopyDeviceToHost(answer, elementsFrequencies);
cuda.Free(inputData);
cuda.Free(answer);
cuda.UnloadModule();
}
return(elementsFrequencies);
}
public override void AllProducts(int element1, float[] results)
{
//base.AllProducts(element1, results);
var mainVec = problemElements[element1];
if (mainVecIdx != element1)
{
CudaHelpers.FillDenseVector(mainVec, mainVector);
cuGPU.CopyHostToDevice(mainVectorPtr, mainVector);
}
mainVecIdx = element1;
//float elapsed;
//gpu.StartTimer();
gpu.Launch(blockPerGrid, blockSize, cudaFunctionName, valsPtr, idxPtr, vecLenghtPtr, sliceStartPtr, selfLinDotPtr, labelsPtr, outputPtr, mainVecIdx, problemElements.Length, Gamma, align);
//elapsed = gpu.StopTimer();
gpu.Synchronize();
//elapsed = gpu.StopTimer();
Marshal.Copy(outputIntPtr, results, 0, results.Length);
}
//private static void InitMainVector(float[] vecVals, int[] vecIdx, int[] vecLenght, float[] mainVec)
//{
// for (int j = vecLenght[mainIndex]; j < vecLenght[mainIndex + 1]; j++)
// {
// int idx = vecIdx[j];
// float val = vecVals[j];
// mainVec[idx] = val;
// }
//}
private static float[] CuDotProdCSRwriteCombined(int repetition)
{
//always the same values
Random rnd = new Random(1);
CUDA cuda = new CUDA(0, true);
// load module
CUmodule module = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "structKernel.cubin"));
CUfunction cuFunc = cuda.GetModuleFunction("spmv_csr_vector_kernel_wc");
int maxRowSize = avgElements + stdElements - 1;
Console.WriteLine("init arrays");
Stopwatch t = Stopwatch.StartNew();
//temp lists for values, indices and vecotr lenght
List<float> vecValsL = new List<float>(N * maxRowSize / 2);
List<int> vecIdxL = new List<int>(N * maxRowSize / 2);
List<int> vecLenghtL = new List<int>(N);
float[] vecVals;
int[] vecIdx;
int[] vecLenght;
maxIndex = 0;
int vecStartIdx = 0;
for (int i = 0; i < N; i++)
{
int vecSize = avgElements + i % stdElements;
float[] vals = Helpers.InitValues(i, vecSize, maxVal);
vecValsL.AddRange(vals);
int[] index = Helpers.InitIndices(i, vecSize, ref maxIndex);
vecIdxL.AddRange(index);
vecLenghtL.Add(vecStartIdx);
vecStartIdx += vecSize;
}
//for last index
vecLenghtL.Add(vecStartIdx);
vecVals = vecValsL.ToArray();
vecIdx = vecIdxL.ToArray();
vecLenght = vecLenghtL.ToArray();
Console.WriteLine("Init takes {0}", t.Elapsed);
t.Start();
CUdeviceptr valsPtr = cuda.CopyHostToDevice(vecVals);
CUdeviceptr idxPtr = cuda.CopyHostToDevice(vecIdx);
CUdeviceptr vecLenghtPtr = cuda.CopyHostToDevice(vecLenght);
float[] output = new float[N];
//CUdeviceptr dOutput = cuda.Allocate(output);
IntPtr outputPtr2 = cuda.HostAllocate((uint)(N * sizeof(float)), CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
CUdeviceptr dOutput = cuda.GetHostDevicePointer(outputPtr2, 0);
uint memSize = (uint)((maxIndex + 1) * sizeof(float));
uint flags = CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP | CUDADriver.CU_MEMHOSTALLOC_WRITECOMBINED;
uint tt = (uint)CUMemHostAllocFlags.WriteCombined;
uint s = (uint)CUMemHostAllocFlags.DeviceMap;
IntPtr mainVecIntPtr = cuda.HostAllocate(memSize, flags);
CUdeviceptr mainVecPtr = cuda.GetHostDevicePointer(mainVecIntPtr, 0);
Console.WriteLine("copy to device takes {0}", t.Elapsed);
#region set cuda parameters
cuda.SetFunctionBlockShape(cuFunc, threadsPerBlock, 1, 1);
int offset = 0;
cuda.SetParameter(cuFunc, offset, valsPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, idxPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, vecLenghtPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, mainVecPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(cuFunc, offset, (uint)N);
offset += sizeof(int);
cuda.SetParameter(cuFunc, offset, (uint)vecStartIdx);
offset += sizeof(int);
cuda.SetParameterSize(cuFunc, (uint)offset);
#endregion
Console.WriteLine("start computation");
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
mainIndex = StartingIndex;
Stopwatch timer = Stopwatch.StartNew();
cuda.RecordEvent(start);
for (int k = 0; k < repetition; k++)
{
//float[] tempFloatarr = new float[memSize];
Helpers.InitBuffer(vecVals, vecIdx, vecLenght,mainIndex, mainVecIntPtr);
//Marshal.Copy(mainVecIntPtr, tempFloatarr, 0, tempFloatarr.Length);
cuda.Launch(cuFunc, blocksPerGrid, 1);
cuda.SynchronizeContext();
//cuda.CopyDeviceToHost(dOutput, output);
Marshal.Copy(outputPtr2, output, 0, N);
//mainVec = new float[maxIndex + 1];
//Array.Clear(mainVec, 0, mainVec.Length);
//clear previous vector values
Helpers.SetBufferIdx(vecIdx, vecLenght,mainIndex, mainVecIntPtr,0.0f);
mainIndex++;
}
cuda.RecordEvent(end);
cuda.SynchronizeContext();
//cuda.SynchronizeEvent(end);
// cuda.CopyDeviceToHost(dOutput, output);
timer.Stop();
float naiveTime = cuda.ElapsedTime(start, end);
Console.Write("csr vector Dot products with mainIndex {0} and {1}-vectors takes {2} ms stopwatch time {3} ms", mainIndex, N, naiveTime, timer.Elapsed);
int lenght = Math.Min(displayCount, N);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
cuda.Free(valsPtr);
cuda.Free(idxPtr);
cuda.Free(dOutput);
cuda.Free(vecLenghtPtr);
//cuda.DestroyArray(cuArr);
cuda.Free(mainVecPtr);
//cuda.DestroyTexture(cuTexRef);
// cuda.Free(mainVecPtr);
cuda.DestroyEvent(start);
cuda.DestroyEvent(end);
return output;
}
private static unsafe void CuStructPass()
{
int N = 4;
int sparseVecSize = sizeof(SparseVecPtr);
uint size = (uint)(N * sizeof(SparseVecPtr));
CUDA cuda = new CUDA(0, true);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "structKernel.cubin"));
CUfunction structPassFunc = cuda.GetModuleFunction("StructPass");
SparseVecPtr[] vectors = new SparseVecPtr[N];
for (int i = 0; i < N; i++)
{
vectors[i] = new SparseVecPtr();
vectors[i].size = 2;
float[] vals = new float[2] { (float)i + 1 % 5, (float)i + 2 % 7 };
//GCHandle valHandle = GCHandle.Alloc(vals, GCHandleType.Pinned);
//vectors[i].values = valHandle.AddrOfPinnedObject();
int[] index = new int[2] { i % 5, i % 7 };
//GCHandle idxHandle = GCHandle.Alloc(index, GCHandleType.Pinned);
//vectors[i].indices = idxHandle.AddrOfPinnedObject();
//valHandle.Free();
//idxHandle.Free();
CUdeviceptr valsPtr = cuda.CopyHostToDevice(vals);
CUdeviceptr idxPtr = cuda.CopyHostToDevice(index);
vectors[i].indices = new IntPtr(idxPtr.Pointer);
vectors[i].values = (IntPtr)valsPtr.Pointer;
}
GCHandle handle = GCHandle.Alloc(vectors, GCHandleType.Pinned);
IntPtr ptr = handle.AddrOfPinnedObject();
float[] output = new float[N];
//CUdeviceptr dVectors = cuda.CopyHostToDevice(vectors);
CUdeviceptr dVectors = cuda.CopyHostToDevice(ptr, size);
CUdeviceptr dOutput = cuda.Allocate(output);
int threadsPerBlock = 256;
int blocksPerGrid = (N + threadsPerBlock - 1) / threadsPerBlock;
//error = cuFuncSetBlockShape(vecAdd, threadsPerBlock, 1, 1);
cuda.SetFunctionBlockShape(structPassFunc, threadsPerBlock, 1, 1);
int offset = 0;
cuda.SetParameter(structPassFunc, offset, (uint)dVectors.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, (uint)dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, (uint)N);
offset += sizeof(int);
cuda.SetParameterSize(structPassFunc, (uint)offset);
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
cuda.RecordEvent(start);
cuda.Launch(structPassFunc, blocksPerGrid, 1);
cuda.RecordEvent(end);
cuda.SynchronizeContext();
//cuda.SynchronizeEvent(end);
float naiveTime = cuda.ElapsedTime(start, end);
Console.Write("passing struct takes {0}ms", naiveTime);
cuda.CopyDeviceToHost(dOutput, output);
int lenght = Math.Min(10, N);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
}
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)
{
// 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();
}
private void UpdateAlpha(int i, int j)
{
float Qii = QD[i];
float Qjj = QD[j];
float yi = y[i];
float yj = y[j];
float Qij = 0;
//todo: remove it only for testing
//float[] a = new float[alpha.Length];
//cuda.CopyDeviceToHost(kiPtr, a);
//copy Qij form device to host
float[] temp = new float[] { 0.0f };
cuda.CopyDeviceToHost(kiPtr + sizeof(float) * j, temp);
Qij = temp[0];
//copy Grad[i] from device to host
float[] gradI = new float[] { 0.0f };
cuda.CopyDeviceToHost(gradPtr + sizeof(float) * i, gradI);
G[i] = gradI[0];
//copy Grad[j] from device to host
float[] gradJ = new float[] { 0.0f };
cuda.CopyDeviceToHost(gradPtr + sizeof(float) * j, gradJ);
G[j] = gradJ[0];
cuda.SynchronizeContext();
////copy alpha[i]
//float[] aI = new float[] { 0.0f };
//cuda.CopyDeviceToHost(alphaPtr + sizeof(float) * i, aI);
//alpha[i] = aI[0];
//float[] aJ = new float[] { 0.0f };
//cuda.CopyDeviceToHost(alphaPtr + sizeof(float) * j, aJ);
//alpha[j] = aJ[0];
float quad_coef = Qii + Qjj - 2 * yi * yj * Qij;
if (quad_coef <= 0)
{
quad_coef = 1e-12f;
}
float delta = 0;
float diff = 0;
float sum = 0;
if (y[i] != y[j])
{
delta = (-G[i] - G[j]) / quad_coef;
diff = alpha[i] - alpha[j];
alpha[i] += delta;
alpha[j] += delta;
if (diff > 0)
{
if (alpha[j] < 0)
{
alpha[j] = 0;
alpha[i] = diff;
}
}
else
{
if (alpha[i] < 0)
{
alpha[i] = 0;
alpha[j] = -diff;
}
}
if (diff > 0)
{
if (alpha[i] > C)
{
alpha[i] = C;
alpha[j] = C - diff;
}
}
else
{
if (alpha[j] > C)
{
alpha[j] = C;
alpha[i] = C + diff;
}
}
}
else
{
delta = (G[i] - G[j]) / quad_coef;
sum = alpha[i] + alpha[j];
alpha[i] -= delta;
alpha[j] += delta;
if (sum > C)
{
if (alpha[i] > C)
{
alpha[i] = C;
alpha[j] = sum - C;
}
}
else
{
if (alpha[j] < 0)
{
alpha[j] = 0;
alpha[i] = sum;
}
}
if (sum > C)
{
if (alpha[j] > C)
{
alpha[j] = C;
alpha[i] = sum - C;
}
}
else
{
if (alpha[i] < 0)
{
alpha[i] = 0;
alpha[j] = sum;
}
}
}
//set alpha on device
cuda.CopyHostToDevice(alphaPtr + sizeof(float) * i, new float[] { alpha[i] });
cuda.CopyHostToDevice(alphaPtr + sizeof(float) * j, new float[] { alpha[j] });
//todo: remove it
//cuda.SynchronizeContext();
//todo: remove it, only for debuging
//cuda.CopyDeviceToHost(alphaPtr, a);
}
public override float[] Predict(SparseVec[] elements)
{
float[] prediction = new float[elements.Length];
uint reduceSize = (uint)reductionBlocks * sizeof(float);
int loop = (elements.Length + NUM_STREAMS - 1) / NUM_STREAMS;
for (int i = 0; i < loop; i++)
{
for (int s = 0; s < NUM_STREAMS; s++)
{
int idx = i * NUM_STREAMS + s;
if (idx < elements.Length)
{
var vec = elements[idx];
//remove
//float[] svDots = TrainedModel.SupportElements.Select(sv => sv.DotProduct(vec)).ToArray();
//set nonzero values to dense vector accessible through vecIntPtr
CudaHelpers.InitBuffer(vec, mainVecIntPtrs[s]);
#region sync version
cuda.CopyHostToDevice(mainVecCuPtr[s], mainVecIntPtrs[s], vectorsDimMemSize);
cuda.SetParameter(cuFuncEval, kernelResultParamOffset, evalOutputCuPtr[s]);
//cuda.SetParameter(cuFuncEval, vectorSelfDotParamOffset, vec.DotProduct());
SetCudaEvalFuncParamsForVector(vec);
cuda.SetParameter(cuFuncEval, texSelParamOffset, s + 1);
cuda.Launch(cuFuncEval, evalBlocks, 1);
float[] t = new float[sizeSV];
cuda.CopyDeviceToHost(evalOutputCuPtr[s], t);
cuda.SetParameter(cuFuncReduce, offsetMemToReduce, evalOutputCuPtr[s]);
cuda.SetParameter(cuFuncReduce, offsetOutMemReduce, reduceCuPtr[s]);
cuda.Launch(cuFuncReduce, reductionBlocks, 1);
cuda.CopyDeviceToHost(reduceCuPtr[s], reduceIntPtrs[s], reduceSize);
float[] r = new float[reductionBlocks];
cuda.CopyDeviceToHost(reduceCuPtr[s], r);
#endregion
//cuda.CopyHostToDeviceAsync(mainVecCuPtr[s], mainVecIntPtrs[s], vectorsDimMemSize, stream[s]);
////cuFunc user different textures
//cuda.SetParameter(cuFuncEval, kernelResultParamOffset, evalOutputCuPtr[s]);
//cuda.SetParameter(cuFuncEval, vectorSelfDotParamOffset, vec.DotProduct());
//cuda.SetParameter(cuFuncEval, texSelParamOffset, s + 1);
//cuda.LaunchAsync(cuFuncEval, evalBlocks, 1, stream[s]);
//cuda.SetParameter(cuFuncReduce, offsetMemToReduce, evalOutputCuPtr[s]);
//cuda.SetParameter(cuFuncReduce, offsetOutMemReduce, reduceCuPtr[s]);
//cuda.LaunchAsync(cuFuncReduce, reductionBlocks, 1, stream[s]);
//cuda.CopyDeviceToHostAsync(reduceCuPtr[s], reduceIntPtrs[s], reduceSize, stream[s]);
}
}
//wait for all streams
cuda.SynchronizeContext();
for (int s = 0; s < NUM_STREAMS; s++)
{
int idx = i * NUM_STREAMS + s;
if (idx < elements.Length)
{
var vec = elements[idx];
//clear the buffer
//set nonzero values to dense vector accessible thought vecIntPtr
CudaHelpers.SetBufferIdx(vec, mainVecIntPtrs[s], 0.0f);
float evalValue = ReduceOnHost(reduceIntPtrs[s], reductionBlocks);
prediction[idx] = evalValue;
}
}
}
return(prediction);
}
private static void CuAddVec()
{
int N = 50000;
uint size = (uint)N * sizeof(float);
CUDA cuda = new CUDA(0, true);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "structKernel.cubin"));
CUfunction vecAddFunc = cuda.GetModuleFunction("VecAdd");
float[] A = new float[N];
float[] B = new float[N];
float[] C = new float[N];
for (int i = 0; i < A.Length; i++)
{
A[i] = (float)i;
B[i] = (float)i + 0.1f;
}
CUdeviceptr dA = cuda.CopyHostToDevice(A);
CUdeviceptr dB = cuda.CopyHostToDevice(B);
CUdeviceptr dC = cuda.Allocate(A);
int threadsPerBlock = 256;
int blocksPerGrid = (N + threadsPerBlock - 1) / threadsPerBlock;
//error = cuFuncSetBlockShape(vecAdd, threadsPerBlock, 1, 1);
cuda.SetFunctionBlockShape(vecAddFunc, threadsPerBlock, 1, 1);
int offset = 0;
cuda.SetParameter(vecAddFunc, offset, (uint)dA.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(vecAddFunc, offset, (uint)dB.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(vecAddFunc, offset, (uint)dC.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(vecAddFunc, offset, (uint)N);
offset += sizeof(int);
cuda.SetParameterSize(vecAddFunc, (uint)offset);
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
cuda.RecordEvent(start);
cuda.Launch(vecAddFunc, blocksPerGrid, 1);
cuda.RecordEvent(end);
cuda.SynchronizeContext();
//cuda.SynchronizeEvent(end);
float naiveTime = cuda.ElapsedTime(start, end);
Console.Write("adding takes {0}ms", naiveTime);
cuda.CopyDeviceToHost(dC, C);
for (int i = 0; i < 10; i++)
{
Console.WriteLine("{0}-{1}", i, C[i]);
}
}
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>
/// fill data which can be changed like
/// Y - labels
/// diag -
/// </summary>
/// <param name="sub_prob"></param>
/// <param name="w"></param>
/// <param name="Cn"></param>
/// <param name="Cp"></param>
private void FillDataOnCuda(Problem <SparseVec> sub_prob, float[] w, double Cn, double Cp)
{
cuda.CopyHostToDevice(labelsPtr, sub_prob.Y);
diag = new float[] { (float)(0.5 / Cn), 0, (float)(0.5 / Cp) };
cuda.CopyHostToDevice(diagPtr, diag);
cuda.CopyHostToDevice(mainVecPtr, w);
}
private static float[] CuDotProdEllPackTexCached()
{
//always the same values
Random rnd = new Random(1);
CUDA cuda = new CUDA(0, true);
// load module
CUmodule module = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "structKernel.cubin"));
CUfunction structPassFunc = cuda.GetModuleFunction("DotProdEllPackCached");
int maxRowSize = avgElements + stdElements - 1;
Console.WriteLine("init arrays");
Stopwatch t = Stopwatch.StartNew();
float[] vecVals = new float[N * maxRowSize];
int[] vecIdx = new int[N * maxRowSize];
maxIndex = 0;
for (int i = 0; i < N; i++)
{
int vecSize = avgElements + i % stdElements;
float[] vals = Helpers.InitValues(i, vecSize, maxVal);
//values are column-major aligment
for (int z = 0; z < vals.Length; z++)
{
int m = z * N + i;
vecVals[m] = vals[z];
}
//Array.Copy(vals,0,vecVals,i*maxRowSize,vals.Length);
int[] index = Helpers.InitIndices(i, vecSize, ref maxIndex);
//Array.Copy(index, 0, vecIdx, i * maxRowSize, index.Length);
for (int z = 0; z < index.Length; z++)
{
int m = z * N + i;
vecIdx[m] = index[z];
}
}
float[] mainVec = new float[maxIndex + 1];
for (int j = 0; j < maxRowSize; j++)
{
int idx = vecIdx[mainIndex + N * j];
float val = vecVals[mainIndex + N * j];
mainVec[idx] = val;
}
Console.WriteLine("Init takes {0}", t.Elapsed);
t.Start();
CUdeviceptr valsPtr = cuda.CopyHostToDevice(vecVals);
CUdeviceptr idxPtr = cuda.CopyHostToDevice(vecIdx);
CUarray cuArr = cuda.CreateArray(mainVec);
cuda.CopyHostToArray(cuArr, mainVec, 0);
//CUDAArrayDescriptor cuDesc = new CUDAArrayDescriptor();
//cuDesc.Format = CUArrayFormat.Float;
//cuDesc.NumChannels = 1;
//cuDesc.Width = maxIndex+1;
CUtexref cuTexRef = cuda.GetModuleTexture(module, "texRef");
cuda.SetTextureFlags(cuTexRef, 0);
cuda.SetTextureArray(cuTexRef, cuArr);
float[] output = new float[N];
CUdeviceptr dOutput = cuda.Allocate(output);
Console.WriteLine("copy to device takes {0}", t.Elapsed);
cuda.SetFunctionBlockShape(structPassFunc, threadsPerBlock, 1, 1);
int offset = 0;
cuda.SetParameter(structPassFunc, offset, valsPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, idxPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, (uint)maxRowSize);
offset += sizeof(int);
cuda.SetParameter(structPassFunc, offset, (uint)N);
offset += sizeof(int);
cuda.SetParameterSize(structPassFunc, (uint)offset);
Console.WriteLine("start computation");
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
Stopwatch timer = Stopwatch.StartNew();
cuda.RecordEvent(start);
cuda.Launch(structPassFunc, blocksPerGrid, 1);
cuda.RecordEvent(end);
cuda.SynchronizeContext();
//cuda.SynchronizeEvent(end);
timer.Stop();
float naiveTime = cuda.ElapsedTime(start, end);
Console.Write("EllPack Cached Dot products with mainIndex {0} and {1}-vectors takes {2} ms stopwatch time {3} ms", mainIndex, N, naiveTime, timer.Elapsed);
cuda.CopyDeviceToHost(dOutput, output);
int lenght = Math.Min(displayCount, N);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
cuda.Free(valsPtr);
cuda.Free(idxPtr);
cuda.Free(dOutput);
cuda.DestroyArray(cuArr);
cuda.DestroyTexture(cuTexRef);
return output;
}
static void Main(string[] args)
{
// Init and select 1st device.
CUDA cuda = new CUDA(0, true);
// load module
//cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "simpleCUFFT.ptx"));
CUfunction func = new CUfunction();// cuda.GetModuleFunction("ComplexPointwiseMulAndScale");
// The filter size is assumed to be a number smaller than the signal size
const int SIGNAL_SIZE = 50;
const int FILTER_KERNEL_SIZE = 11;
// Allocate host memory for the signal
Float2[] h_signal = new Float2[SIGNAL_SIZE];
// Initalize the memory for the signal
Random r = new Random();
for (int i = 0; i < SIGNAL_SIZE; ++i)
{
h_signal[i].x = r.Next() / (float)int.MaxValue;
h_signal[i].y = 0;
}
// Allocate host memory for the filter
Float2[] h_filter_kernel = new Float2[FILTER_KERNEL_SIZE];
// Initalize the memory for the filter
for (int i = 0; i < FILTER_KERNEL_SIZE; ++i)
{
h_filter_kernel[i].x = r.Next() / (float)int.MaxValue;
h_filter_kernel[i].y = 0;
}
// Pad signal and filter kernel
Float2[] h_padded_signal;
Float2[] h_padded_filter_kernel;
int new_size = PadData(h_signal, out h_padded_signal, SIGNAL_SIZE,
h_filter_kernel, out h_padded_filter_kernel, FILTER_KERNEL_SIZE);
// Allocate device memory for signal
// Copy host memory to device
CUdeviceptr d_signal = cuda.CopyHostToDevice <Float2>(h_padded_signal);
// Allocate device memory for filter kernel
// Copy host memory to device
CUdeviceptr d_filter_kernel = cuda.CopyHostToDevice <Float2>(h_padded_filter_kernel);
// CUFFT plan
CUFFT fft = new CUFFT(cuda);
cufftHandle handle = new cufftHandle();
CUFFTResult fftres = CUFFTDriver.cufftPlan1d(ref handle, new_size, CUFFTType.C2C, 1);
//fft.Plan1D(new_size, CUFFTType.C2C, 1);
return;
// Transform signal and kernel
fft.ExecuteComplexToComplex(d_signal, d_signal, CUFFTDirection.Forward);
fft.ExecuteComplexToComplex(d_filter_kernel, d_filter_kernel, CUFFTDirection.Forward);
// Multiply the coefficients together and normalize the result
// ComplexPointwiseMulAndScale<<<32, 256>>>(d_signal, d_filter_kernel, new_size, 1.0f / new_size);
cuda.SetFunctionBlockShape(func, 256, 1, 1);
cuda.SetParameter(func, 0, (uint)d_signal.Pointer);
cuda.SetParameter(func, IntPtr.Size, (uint)d_filter_kernel.Pointer);
cuda.SetParameter(func, IntPtr.Size * 2, (uint)new_size);
cuda.SetParameter(func, IntPtr.Size * 2 + 4, 1.0f / new_size);
cuda.SetParameterSize(func, (uint)(IntPtr.Size * 2 + 8));
cuda.Launch(func, 32, 1);
// Transform signal back
fft.ExecuteComplexToComplex(d_signal, d_signal, CUFFTDirection.Inverse);
// Copy device memory to host
Float2[] h_convolved_signal = h_padded_signal;
cuda.CopyDeviceToHost <Float2>(d_signal, h_convolved_signal);
// Allocate host memory for the convolution result
Float2[] h_convolved_signal_ref = new Float2[SIGNAL_SIZE];
// Convolve on the host
Convolve(h_signal, SIGNAL_SIZE,
h_filter_kernel, FILTER_KERNEL_SIZE,
h_convolved_signal_ref);
// check result
bool res = cutCompareL2fe(h_convolved_signal_ref, h_convolved_signal, 2 * SIGNAL_SIZE, 1e-5f);
Console.WriteLine("Test {0}", (true == res) ? "PASSED" : "FAILED");
//Destroy CUFFT context
fft.Destroy();
// cleanup memory
cuda.Free(d_signal);
cuda.Free(d_filter_kernel);
}
private static unsafe float[] CuDotProdSparseVecStruct()
{
int sparseVecSize = sizeof(SparseVecPtr);
uint size = (uint)(N * sizeof(SparseVecPtr));
//always the same values
Random rnd = new Random(1);
CUDA cuda = new CUDA(0, true);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "structKernel.cubin"));
//CUfunction structPassFunc = cuda.GetModuleFunction("DotProd");
CUfunction structPassFunc = cuda.GetModuleFunction("DotProd2");
SparseVecPtr[] vectors = new SparseVecPtr[N];
Console.WriteLine("init and copy data");
Stopwatch t = Stopwatch.StartNew();
mainIndex = StartingIndex;
for (int i = 0; i < N; i++)
{
vectors[i] = new SparseVecPtr();
int vecSize = avgElements + i % stdElements;
vectors[i].size = vecSize;
float[] vals = Helpers.InitValues(i, vecSize, maxVal);
int[] index = Helpers.InitIndices(i, vecSize, ref maxIndex);
CUdeviceptr valsPtr = cuda.CopyHostToDevice(vals);
CUdeviceptr idxPtr = cuda.CopyHostToDevice(index);
vectors[i].indices = new IntPtr(idxPtr.Pointer);
vectors[i].values = (IntPtr)valsPtr.Pointer;
}
GCHandle handle = GCHandle.Alloc(vectors, GCHandleType.Pinned);
IntPtr ptr = handle.AddrOfPinnedObject();
float[] output = new float[N];
//CUdeviceptr dVectors = cuda.CopyHostToDevice(vectors);
CUdeviceptr dVectors = cuda.CopyHostToDevice(ptr, size);
CUdeviceptr dOutput = cuda.Allocate(output);
Console.WriteLine("copy and init takes {0}", t.Elapsed);
#region set cuda parameters
cuda.SetFunctionBlockShape(structPassFunc, threadsPerBlock, 1, 1);
int offset = 0;
cuda.SetParameter(structPassFunc, offset, dVectors.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, (uint)mainIndex);
offset += sizeof(int);
cuda.SetParameter(structPassFunc, offset, (uint)N);
offset += sizeof(int);
cuda.SetParameterSize(structPassFunc, (uint)offset);
#endregion
Console.WriteLine("start computation");
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
Stopwatch timer = Stopwatch.StartNew();
cuda.RecordEvent(start);
cuda.Launch(structPassFunc, blocksPerGrid, 1);
cuda.RecordEvent(end);
cuda.SynchronizeContext();
//cuda.SynchronizeEvent(end);
timer.Stop();
float naiveTime = cuda.ElapsedTime(start, end);
Console.Write("Dot products with mainIndex {0} and {1}-vectors takes {2} ms stopwatch time {3} ms", mainIndex, N, naiveTime, timer.Elapsed);
cuda.CopyDeviceToHost(dOutput, output);
int lenght = Math.Min(displayCount, N);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
cuda.Free(dVectors);
cuda.Free(dOutput);
return output;
}
/// <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>
/// Do BB step, updates alpha and "w" vector
/// </summary>
/// <remarks>
/// This method has many side effects:
/// 1. copies data on device from alphaPtr to alphaTmpPtr
/// 2. sets values in deltas array, which are differences between new alpha and old alphas
/// 3. alphaTmpPtr stores new updated alphas
/// </remarks>
/// <param name="step"></param>
/// <param name="sub_prob"></param>
private void DoBBstep(float step, Problem <SparseVec> sub_prob)
{
int blocks = (sub_prob.Elements.Length + threadsPerBlock - 1) / threadsPerBlock;
/*
* Update alpha
*
* 1. copy alpha to alphaTmp
* 2. copy step into device constant
* 3. set parameters
*
*/
cuda.CopyDeviceToDevice(alphaPtr, alphaTmpPtr, alphaMemSize);
float[] stepData = new float[] { step };
cuda.CopyHostToDevice(stepBBPtr, stepData);
cuda.SetParameter(cuFuncUpdateAlpha, gradParamOffsetInUpdateAlpha, gradPtr.Pointer);
cuda.SetParameter(cuFuncUpdateAlpha, alphaParamOffsetInUpdateAlpha, alphaTmpPtr.Pointer);
cuda.Launch(cuFuncUpdateAlpha, blocks, 1);
//float[] updatedAlpha = new float[sub_prob.ElementsCount];
//cuda.CopyDeviceToHost(alphaTmpPtr, updatedAlpha);
//float[] updatedDeltas = new float[sub_prob.ElementsCount];
//cuda.CopyDeviceToHost(deltasPtr, updatedDeltas);
/*
* Update w - based on aplha deltas
*
*/
int bpgUpdateW = (sub_prob.Elements[0].Dim + threadsPerBlock - 1) / threadsPerBlock;
cuda.CopyDeviceToDevice(wVecPtr, wTempVecPtr, wVecMemSize);
cuda.SetParameter(cuFuncUpdateW, wVecParamOffsetInUpdateW, wTempVecPtr.Pointer);
cuda.Launch(cuFuncUpdateW, bpgUpdateW, 1);
//float[] wTest = new float[sub_prob.FeaturesCount];
//cuda.CopyDeviceToHost(wTempVecPtr, wTest);
}
internal static void SetTextureMemory(CUDA cuda, ref CUtexref texture, string texName, float[] data, ref CUdeviceptr memPtr)
{
texture = cuda.GetModuleTexture(texName);
memPtr = cuda.CopyHostToDevice(data);
cuda.SetTextureAddress(texture, memPtr, (uint)(sizeof(float) * data.Length));
}
private void Worker()
{
#if CUDA
int nN = _nMergingDeviceNumber;
string sS = "CUDA-" + nN + "-" + _nIndex;
try
{
Command cCmd;
CUDA cCUDA;
#region CUDA Init
try
{
cCUDA = new CUDA(true);
cCUDA.CreateContext(nN % 1000); // number of cuda in prefs (still alwais 0)
}
catch (Exception ex)
{
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
throw new Exception("CreateContext(" + nN % 1000 + ") error. Try to change CUDA's card number in prefs", ex);
}
(new Logger(sS)).WriteDebug("CreateContext(" + nN % 1000 + ") is ok!");
uint nMemoryReservedForMerge = 2 * 1024 * 1024; //PREFERENCES типа <memory reserved="2097152" />
uint nMemoryStarvationThreshold = cCUDA.TotalMemory / 2; //PREFERENCES через проценты... типа <memory starvation="50%" />
uint nMemoryFree;
string sModule = "CUDAFunctions_" + Preferences.nCUDAVersion + "_x" + (IntPtr.Size * 8);
if (Logger.bDebug)
{
(new Logger(sS)).WriteDebug(sModule + " Current CUDA = [name=" + cCUDA.CurrentDevice.Name + "][compute_capability=" + cCUDA.CurrentDevice.ComputeCapability + "]");
}
cCUDA.LoadModule((byte[])Properties.Resource.ResourceManager.GetObject(sModule)); // $(ProjectDir)Resources\CUDAFunctions.cubin
CUfunction cCUDAFuncMerge = cCUDA.GetModuleFunction("CUDAFrameMerge");
int nThreadsPerBlock = 16; //32 //256 //пришлось уменьшить с 512 до 256 сридов на блок, потому что при добавлении "движения" и операций с float, ловил ошибку: Too Many Resources Requested for Launch (This error means that the number of registers available on the multiprocessor is being exceeded. Reduce the number of threads per block to solve the problem)
cCUDA.SetFunctionBlockShape(cCUDAFuncMerge, nThreadsPerBlock, nThreadsPerBlock, 1);
CUDADriver.cuParamSetSize(cCUDAFuncMerge, 8);
Dictionary <long, CUdeviceptr> ahPMIDs_DevicePointers = new Dictionary <long, CUdeviceptr>();
CUdeviceptr cPMs;
CUdeviceptr cInfos;
CUdeviceptr cAlphaMap;
CUdeviceptr cAlphaMap_info3d;
CUdeviceptr cAlphaMap_info2d;
if (true)
{
//IntPtr[] aPointersByAlpha = new IntPtr[254]; //те самые поинтеры-альфы. Ссылаются на массивы поинтеров B, т.е. BackGrounds
//IntPtr[] aPointersByBackground = new IntPtr[256]; // те самые массивы поинтеров B, т.е. BackGrounds
byte[] aAlphaMap = new byte[(byte.MaxValue - 1) * (byte.MaxValue + 1) * (byte.MaxValue + 1)];
int[] aAlphaMap_info3d = new int[254]; // начала 2d слоёв
ushort[] aAlphaMap_info2d = new ushort[256]; // начала строк в одном 2d
int nResult, nIndx = 0, nIndxInfo = 0, nIndx2d = 0;
for (byte nAlpha = 1; 255 > nAlpha; nAlpha++)
{
aAlphaMap_info3d[nIndxInfo++] = nIndx;
for (ushort nBackground = 0; 256 > nBackground; nBackground++)
{
if (nAlpha == 1)
{
aAlphaMap_info2d[nIndx2d++] = (ushort)nIndx;
}
for (ushort nForeground = 0; 256 > nForeground; nForeground++)
{
if (255 < (nResult = (int)((float)(nAlpha * (nForeground - nBackground)) / 255 + nBackground + 0.5)))
{
nResult = 255;
}
aAlphaMap[nIndx++] = (byte)nResult;
}
//aPointersByBackground[nBackground] = (IntPtr)cCUDA.CopyHostToDevice<byte>(aResults).Pointer;
}
//aPointersByAlpha[nAlpha - 1] = (IntPtr)cCUDA.CopyHostToDevice<IntPtr>(aPointersByBackground).Pointer;
}
cAlphaMap_info3d = cCUDA.CopyHostToDevice <int>(aAlphaMap_info3d);
cAlphaMap = cCUDA.CopyHostToDevice <byte>(aAlphaMap);
cAlphaMap_info2d = cCUDA.CopyHostToDevice <ushort>(aAlphaMap_info2d);
}
CUdeviceptr cAlphaMap2;
CUdeviceptr cAlphaMap2_info2d;
{
byte[] aAlphaMap2 = new byte[(byte.MaxValue - 1) * (byte.MaxValue - 1)];
ushort[] aAlphaMap2_info2d = new ushort[254];
int nIndx = 0, nIndx2d = 0;
for (byte nFGColorAlpha = 1; 255 > nFGColorAlpha; nFGColorAlpha++) // можно использовать симметрию умножения, но х с ней пока
{
aAlphaMap2_info2d[nIndx2d++] = (ushort)nIndx;
for (byte nPixelAlpha = 1; 255 > nPixelAlpha; nPixelAlpha++)
{
aAlphaMap2[nIndx++] = (byte)((float)nFGColorAlpha * nPixelAlpha / 255 + 0.5);
}
}
cAlphaMap2 = cCUDA.CopyHostToDevice <byte>(aAlphaMap2);
cAlphaMap2_info2d = cCUDA.CopyHostToDevice <ushort>(aAlphaMap2_info2d);
}
CUdeviceptr cAlphaMap3;
CUdeviceptr cAlphaMap3_info2d;
{
byte[] aAlphaMap3 = new byte[byte.MaxValue * (byte.MaxValue - 1)];
ushort[] aAlphaMap3_info2d = new ushort[255];
int nIndx = 0, nIndx2d = 0;
for (ushort nFGColorAlpha = 1; 256 > nFGColorAlpha; nFGColorAlpha++)
{
aAlphaMap3_info2d[nIndx2d++] = (ushort)nIndx;
for (byte nMask = 1; 255 > nMask; nMask++)
{
aAlphaMap3[nIndx++] = (byte)(nFGColorAlpha * ((255 - nMask) / 255f) + 0.5);
}
}
cAlphaMap3 = cCUDA.CopyHostToDevice <byte>(aAlphaMap3);
cAlphaMap3_info2d = cCUDA.CopyHostToDevice <ushort>(aAlphaMap3_info2d);
}
#endregion CUDA Init
#if DEBUG
Dictionary <long, DateTime> ahDebug = new Dictionary <long, DateTime>();
Dictionary <long, Area> ahDebugAr = new Dictionary <long, Area>();
#endif
DateTime dtNextTime = DateTime.MinValue, dtNow;
bool bSet;
List <IntPtr> aDPs;
List <PixelsMap> aPMs;
while (true)
{
if (1 > aqQueue.CountGet() && (dtNow = DateTime.Now) > dtNextTime)
{
dtNextTime = dtNow.AddMinutes(20);
#if DEBUG
dtNow = dtNow.Subtract(TimeSpan.FromHours(2));
string sMessage = "";
foreach (long nID in ahDebug.OrderBy(o => o.Value).Select(o => o.Key))
{
if (dtNow > ahDebug[nID])
{
sMessage += "<br>[" + nID + " - " + ahDebug[nID].ToString("HH:mm:ss") + "]" + ahDebugAr[nID].ToString();
}
}
#endif
(new Logger(sS)).WriteDebug("CUDA free memory:" + cCUDA.FreeMemory
#if DEBUG
+ "; possibly timeworn allocations:" + (1 > sMessage.Length ? "no" : sMessage)
#endif
);
}
cCmd = aqQueue.Dequeue(); //если нечего отдать - заснёт
switch (cCmd.eID)
{
case Command.ID.Allocate:
#region
try
{
cCmd.cPM._cException = null;
if (1 > cCmd.cPM._nID)
{
if (0 < cCmd.cPM._nBytesQty)
{
nMemoryFree = cCUDA.FreeMemory;
if (nMemoryReservedForMerge < nMemoryFree - cCmd.cPM._nBytesQty)
{
bMemoryStarvation = (nMemoryFree < nMemoryStarvationThreshold);
(new Logger(sS)).WriteDebug3("pixelmap allocateCUDA [current_id=" + _nCurrentID + "]");
cCmd.cPM._nID = System.Threading.Interlocked.Increment(ref _nCurrentID);
ahPMIDs_DevicePointers.Add(cCmd.cPM._nID, cCUDA.Allocate(cCmd.cPM._nBytesQty));
#if DEBUG
ahDebug.Add(cCmd.cPM._nID, DateTime.Now);
ahDebugAr.Add(cCmd.cPM._nID, cCmd.cPM.stArea);
#endif
}
else
{
bMemoryStarvation = true;
throw new Exception("out of memory in CUDA device during Allocate. Only 2 MBytes reserved for the Merge");
}
}
else
{
throw new Exception("bytes quantity in PixelsMap have to be greater than zero for Allocate [_bDisposed = " + cCmd.cPM._bDisposed + "][_bProcessing = " + cCmd.cPM._bProcessing + "][_stPosition.X = " + cCmd.cPM._stPosition.X + "][_stPosition.Y = " + cCmd.cPM._stPosition.Y + "][_bTemp = " + cCmd.cPM._bTemp + "][_dt = " + cCmd.cPM._dtCreate + "][_nBytesQty = " + cCmd.cPM._nBytesQty + "][_nID = " + cCmd.cPM._nID + "][_nShiftTotalX = " + cCmd.cPM._nShiftTotalX + "][_stArea.nHeight = " + cCmd.cPM._stArea.nHeight + "][_stArea.nWidth = " + cCmd.cPM._stArea.nWidth + "][bKeepAlive = " + cCmd.cPM.bKeepAlive + "][eAlpha = " + cCmd.cPM.eAlpha + "][bCUDA = " + cCmd.cPM.stMergingMethod + "][nAlphaConstant = " + cCmd.cPM.nAlphaConstant + "][nID = " + cCmd.cPM.nID + "][nLength = " + cCmd.cPM.nLength + "][stArea.nHeight = " + cCmd.cPM.stArea.nHeight + "][stArea.nWidth = " + cCmd.cPM.stArea.nWidth + "]");
}
}
else
{
throw new Exception("PixelsMap ID have to be zero for Allocate");
}
}
catch (Exception ex)
{
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
(new Logger(sS)).WriteError(ex);
(new Logger(sS)).WriteDebug("bytes qty:" + cCmd.cPM._nBytesQty);
cCmd.cPM._cException = ex;
}
cCmd.cMRE.Set();
break;
#endregion
case Command.ID.CopyIn:
#region
try
{
cCmd.cPM._cException = null;
if (1 > cCmd.cPM._nID)
{
if (cCUDA.FreeMemory - cCmd.cPM._nBytesQty > nMemoryReservedForMerge)
{
(new Logger(sS)).WriteDebug3("pixelmap copyinCUDA not allocated [pm_id=" + _nCurrentID + "]");
cCmd.cPM._nID = System.Threading.Interlocked.Increment(ref _nCurrentID);
if (cCmd.ahParameters.ContainsKey(typeof(IntPtr)))
{
ahPMIDs_DevicePointers.Add(cCmd.cPM._nID, cCUDA.CopyHostToDevice((IntPtr)cCmd.ahParameters[typeof(IntPtr)], cCmd.cPM._nBytesQty));
}
else if (cCmd.ahParameters.ContainsKey(typeof(byte[])))
{
ahPMIDs_DevicePointers.Add(cCmd.cPM._nID, cCUDA.CopyHostToDevice((byte[])cCmd.ahParameters[typeof(byte[])]));
}
else
{
throw new Exception("unknown parameter type");
}
#if DEBUG
ahDebug.Add(cCmd.cPM._nID, DateTime.Now);
ahDebugAr.Add(cCmd.cPM._nID, cCmd.cPM.stArea);
#endif
}
else
{
throw new Exception("out of memory in CUDA device during CopyIn. Only 2 MBytes reserved for the Merge.");
}
}
else
{
(new Logger(sS)).WriteDebug4("pixelmap copyinCUDA allocated [pm_id=" + _nCurrentID + "]");
if (cCmd.ahParameters.ContainsKey(typeof(IntPtr)))
{
cCUDA.CopyHostToDevice(ahPMIDs_DevicePointers[cCmd.cPM._nID], (IntPtr)cCmd.ahParameters[typeof(IntPtr)], cCmd.cPM._nBytesQty);
}
else if (cCmd.ahParameters.ContainsKey(typeof(byte[])))
{
cCUDA.CopyHostToDevice(ahPMIDs_DevicePointers[cCmd.cPM._nID], (byte[])cCmd.ahParameters[typeof(byte[])]);
}
else
{
throw new Exception("unknown parameter type");
}
}
}
catch (Exception ex)
{
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
(new Logger(sS)).WriteError(ex);
cCmd.cPM._cException = ex;
}
cCmd.cMRE.Set();
#endregion
break;
case Command.ID.CopyOut:
#region
try
{
if (0 < cCmd.cPM._nID)
{
if (!cCmd.ahParameters.ContainsKey(typeof(IntPtr)))
{
if (cCmd.ahParameters.ContainsKey(typeof(byte[])))
{
byte[] aB = (byte[])cCmd.ahParameters[typeof(byte[])];
cCmd.cPM._aBytes = null;
if (cCmd.cPM._nBytesQty != aB.Length)
{
(new Logger(sS)).WriteWarning("wrong array size for copyout [got:" + aB.Length + "][expected:" + cCmd.cPM._nBytesQty + "]");
}
cCUDA.CopyDeviceToHost <byte>(ahPMIDs_DevicePointers[cCmd.cPM._nID], aB);
}
else // не юзается (см. copyout())
{
cCmd.cPM._aBytes = _cBinM.BytesGet((int)cCmd.cPM._nBytesQty, 3);
cCUDA.CopyDeviceToHost <byte>(ahPMIDs_DevicePointers[cCmd.cPM._nID], cCmd.cPM._aBytes.aBytes);
}
}
else
{
cCUDA.CopyDeviceToHost(ahPMIDs_DevicePointers[cCmd.cPM._nID], (IntPtr)cCmd.ahParameters[typeof(IntPtr)], cCmd.cPM._nBytesQty);
}
(new Logger(sS)).WriteDebug5("copy out [id:" + cCmd.cPM._nID + "][ptr:" + ahPMIDs_DevicePointers[cCmd.cPM._nID].Pointer + "]");
}
else
{
throw new Exception("PixelsMap have to be allocated for CopyOut");
}
}
catch (Exception ex)
{
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
(new Logger(sS)).WriteError(ex);
cCmd.cPM._cException = ex;
}
cCmd.cMRE.Set();
#endregion
break;
case Command.ID.Merge:
#region
bSet = false;
try
{
aPMs = (List <PixelsMap>)cCmd.ahParameters[typeof(List <PixelsMap>)];
DisCom.MergeInfo cMergeInfo = (DisCom.MergeInfo)cCmd.ahParameters[typeof(DisCom.MergeInfo)];
aDPs = new List <IntPtr>();
if (1 > cCmd.cPM._nID)
{
throw new Exception("background PixelsMap have to be allocated for Merge");
}
aDPs.Add((IntPtr)ahPMIDs_DevicePointers[cCmd.cPM._nID].Pointer);
for (int nIndx = 0; nIndx < aPMs.Count; nIndx++)
{
if (!ahPMIDs_DevicePointers.ContainsKey(aPMs[nIndx]._nID))
{
throw new Exception("there is a corrupted ID in layers for merge [id:" + aPMs[nIndx]._nID + "]");
}
if (1 > ahPMIDs_DevicePointers[aPMs[nIndx]._nID].Pointer)
{
throw new Exception("there is an empty pointer in layers for merge [id:" + aPMs[nIndx]._nID + "]");
}
aDPs.Add((IntPtr)ahPMIDs_DevicePointers[aPMs[nIndx]._nID].Pointer);
}
cPMs = cCUDA.CopyHostToDevice <IntPtr>(aDPs.ToArray());
cInfos = cCUDA.CopyHostToDevice(cMergeInfo, cMergeInfo.SizeGet()); // operator intptr in DisCom.MergeInfo
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, 0, (IntPtr)cPMs.Pointer);
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size, (IntPtr)cInfos.Pointer);
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 2, (IntPtr)cAlphaMap.Pointer); //
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 3, (IntPtr)cAlphaMap_info3d.Pointer); //
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 4, (IntPtr)cAlphaMap_info2d.Pointer); //
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 5, (IntPtr)cAlphaMap2.Pointer);
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 6, (IntPtr)cAlphaMap2_info2d.Pointer); //
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 7, (IntPtr)cAlphaMap3.Pointer);
cCUDA.SetParameter <IntPtr>(cCUDAFuncMerge, IntPtr.Size * 8, (IntPtr)cAlphaMap3_info2d.Pointer); //
cCUDA.SetParameterSize(cCUDAFuncMerge, (uint)(IntPtr.Size * 9));
int nIterationsX = (0 == cMergeInfo.nBackgroundWidth % nThreadsPerBlock ? cMergeInfo.nBackgroundWidth / nThreadsPerBlock : cMergeInfo.nBackgroundWidth / nThreadsPerBlock + 1);
int nIterationsY = (0 == cMergeInfo.nBackgroundHight % nThreadsPerBlock ? cMergeInfo.nBackgroundHight / nThreadsPerBlock : cMergeInfo.nBackgroundHight / nThreadsPerBlock + 1);
//int nIterationsX = (0 == cMergeInfo.nBackgroundHight % nThreadsPerBlock ? cMergeInfo.nBackgroundHight / nThreadsPerBlock : cMergeInfo.nBackgroundHight / nThreadsPerBlock + 1);
cCUDA.Launch(cCUDAFuncMerge, nIterationsX, nIterationsY);
cCUDA.Free(cPMs);
cCUDA.Free(cInfos);
cCmd.cMRE.Set();
bSet = true;
cMergeInfo.Dispose();
for (int nIndx = 0; nIndx < aPMs.Count; nIndx++)
{
lock (aPMs[nIndx]._cSyncRoot)
aPMs[nIndx]._bProcessing = false;
aPMs[nIndx].Dispose();
}
}
catch (Exception ex)
{
cCmd.cPM._cException = ex;
if (!bSet)
{
cCmd.cMRE.Set();
}
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
(new Logger(sS)).WriteError(ex);
}
#endregion
break;
case Command.ID.Dispose:
#region
(new Logger(sS)).Write(Logger.Level.debug4, "dispose: in");
try
{
if (ahPMIDs_DevicePointers.ContainsKey(cCmd.cPM._nID))
{
if (0 < cCmd.cPM._nID && 0 < ahPMIDs_DevicePointers[cCmd.cPM._nID].Pointer)
{
cCUDA.Free(ahPMIDs_DevicePointers[cCmd.cPM._nID]);
//cCUDA.SynchronizeContext();
bMemoryStarvation = (cCUDA.FreeMemory < nMemoryStarvationThreshold);
(new Logger(sS)).WriteDebug3("dispose [id:" + cCmd.cPM._nID + "][ptr:" + ahPMIDs_DevicePointers[cCmd.cPM._nID].Pointer + "]");
}
ahPMIDs_DevicePointers.Remove(cCmd.cPM._nID);
#if DEBUG
ahDebug.Remove(cCmd.cPM._nID);
ahDebugAr.Remove(cCmd.cPM._nID);
#endif
cCmd.cPM._nID = 0;
}
}
catch (Exception ex)
{
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
(new Logger(sS)).WriteError(ex);
cCmd.cPM._cException = ex;
}
(new Logger(sS)).Write(Logger.Level.debug4, "dispose: out");
#endregion
break;
}
}
}
catch (Exception ex)
{
if (ex is CUDAException)
{
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
}
(new Logger(sS)).WriteError("CUDA STOPPED!!!! [id = " + _nIndex + "]", ex);
}
#endif
}
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);
}
// It's interesting to change the number of blocks and the number of threads to
// understand how to keep the hardware busy.
//
// Here are some numbers I get on my G80:
// blocks - clocks
// 1 - 3096
// 8 - 3232
// 16 - 3364
// 32 - 4615
// 64 - 9981
//
// With less than 16 blocks some of the multiprocessors of the device are idle. With
// more than 16 you are using all the multiprocessors, but there's only one block per
// multiprocessor and that doesn't allow you to hide the latency of the memory. With
// more than 32 the speed scales linearly.
static void Main(string[] args)
{
// Init CUDA, select 1st device.
CUDA cuda = new CUDA(0, true);
// load module
cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "clock_kernel.cubin"));
CUfunction func = cuda.GetModuleFunction("timedReduction");
int[] timer = new int[NUM_BLOCKS * 2];
float[] input = new float[NUM_THREADS * 2];
for (int i = 0; i < NUM_THREADS * 2; i++)
{
input[i] = (float)i;
}
CUdeviceptr dinput = cuda.CopyHostToDevice<float>(input);
CUdeviceptr doutput = cuda.Allocate((uint)(sizeof(float) * NUM_BLOCKS));
CUdeviceptr dtimer = cuda.Allocate<int>(timer);
cuda.SetParameter(func, 0, (uint)dinput.Pointer);
cuda.SetParameter(func, IntPtr.Size, (uint)doutput.Pointer);
cuda.SetParameter(func, IntPtr.Size*2, (uint)dtimer.Pointer);
cuda.SetParameterSize(func, (uint)(IntPtr.Size*3));
//timedReduction<<<NUM_BLOCKS, NUM_THREADS, sizeof(float) * 2 * NUM_THREADS>>>(dinput, doutput, dtimer);
cuda.SetFunctionBlockShape(func, NUM_THREADS, 1, 1);
cuda.SetFunctionSharedSize(func, (uint)(sizeof(float) * 2 * NUM_THREADS));
cuda.Launch(func, NUM_BLOCKS, 1);
cuda.CopyDeviceToHost<int>(dtimer, timer);
cuda.Free(dinput);
cuda.Free(doutput);
cuda.Free(dtimer);
foreach (int i in timer)
{
Console.WriteLine(i);
}
Console.WriteLine("Test PASSED");
int minStart = timer[0];
int maxEnd = timer[NUM_BLOCKS];
for (int i = 1; i < NUM_BLOCKS; i++)
{
minStart = timer[i] < minStart ? timer[i] : minStart;
maxEnd = timer[NUM_BLOCKS + i] > maxEnd ? timer[NUM_BLOCKS + i] : maxEnd;
}
Console.WriteLine("time = {0}", maxEnd - minStart);
}
/// <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);
}
internal static void SetTextureMemory(CUDA cuda, ref CUtexref texture, string texName, float[] data, ref CUdeviceptr memPtr)
{
texture = cuda.GetModuleTexture(texName);
memPtr = cuda.CopyHostToDevice(data);
cuda.SetTextureAddress(texture, memPtr, (uint)(sizeof(float) * data.Length));
}
static private void Worker(object cState)
{
try
{
Command cCmd;
CUDA cCUDA = new CUDA(true);
for (int i = 0; i < 10; i++)
{
try
{
cCUDA.CreateContext(i);
(new Logger()).WriteDebug2(i + ": success");
break;
}
catch (Exception ex)
{
(new Logger()).WriteDebug2(i + ": failed");
if (Logger.bDebug && Logger.Level.debug3 > Logger.eLevelMinimum)
(new Logger()).WriteError(ex);
}
}
uint nMemoryReservedForMerge = 2 * 1024 * 1024; //PREFERENCES типа <memory reserved="2097152" />
uint nMemoryStarvationThreshold = cCUDA.TotalMemory / 2; //PREFERENCES через проценты... типа <memory starvation="50%" />
uint nMemoryFree;
string sModule = "CUDAFunctions_" + Preferences.nCUDAVersion + "_x" + (IntPtr.Size * 8);
if (Logger.bDebug)
(new Logger()).WriteDebug3(sModule);
cCUDA.LoadModule((byte[])Properties.Resource.ResourceManager.GetObject(sModule)); // $(ProjectDir)Resources\CUDAFunctions.cubin
//cCUDA.LoadModule(@"c:\projects\!helpers\video\PixelsMap\Resources\CUDAFunctions.cubin");
CUfunction cCUDAFuncMerge = cCUDA.GetModuleFunction("CUDAFrameMerge");
int nThreadsPerBlock = 256; //пришлось уменьшить с 512 до 256 сридов на блок, потому что при добавлении "движения" и операций с float, ловил ошибку: Too Many Resources Requested for Launch (This error means that the number of registers available on the multiprocessor is being exceeded. Reduce the number of threads per block to solve the problem)
cCUDA.SetFunctionBlockShape(cCUDAFuncMerge, nThreadsPerBlock, 1, 1);
CUDADriver.cuParamSetSize(cCUDAFuncMerge, 8);
Dictionary<ulong, CUdeviceptr> ahDevicePointers = new Dictionary<ulong, CUdeviceptr>();
CUdeviceptr cPMs;
CUdeviceptr cInfos;
CUdeviceptr cAlphaMap;
{
//IntPtr[] aPointersByAlpha = new IntPtr[254]; //те самые поинтеры-альфы. Ссылаются на массивы поинтеров B, т.е. BackGrounds
//IntPtr[] aPointersByBackground = new IntPtr[256]; // те самые массивы поинтеров B, т.е. BackGrounds
byte[] aAlphaMap = new byte[16646144];
int nResult, nIndx = 0;
for (byte nAlpha = 1; 255 > nAlpha; nAlpha++)
{
for (ushort nBackground = 0; 256 > nBackground; nBackground++)
{
for (ushort nForeground = 0; 256 > nForeground; nForeground++)
{
if (255 < (nResult = (int)((float)(nAlpha * (nForeground - nBackground)) / 255 + nBackground + 0.5)))
nResult = 255;
aAlphaMap[nIndx++] = (byte)nResult;
}
//aPointersByBackground[nBackground] = (IntPtr)cCUDA.CopyHostToDevice<byte>(aResults).Pointer;
}
//aPointersByAlpha[nAlpha - 1] = (IntPtr)cCUDA.CopyHostToDevice<IntPtr>(aPointersByBackground).Pointer;
}
cAlphaMap = cCUDA.CopyHostToDevice<byte>(aAlphaMap);
}
//{
// IntPtr[] aPointersByAlpha = new IntPtr[254]; //те самые поинтеры-альфы. Ссылаются на массивы поинтеров B, т.е. BackGrounds
// IntPtr[] aPointersByBackground = new IntPtr[256]; // те самые массивы поинтеров B, т.е. BackGrounds
// byte[] aResults = new byte[256];
// int nResult;
// for (byte nAlpha = 1; 255 > nAlpha; nAlpha++)
// {
// for (ushort nBackground = 0; 256 > nBackground; nBackground++)
// {
// for (ushort nForeground = 0; 256 > nForeground; nForeground++)
// {
// if (255 < (nResult = (int)((float)(nAlpha * (nForeground - nBackground)) / 255 + nBackground + 0.5)))
// nResult = 255;
// aResults[nForeground] = (byte)nResult;
// }
// aPointersByBackground[nBackground] = (IntPtr)cCUDA.CopyHostToDevice<byte>(aResults).Pointer;
// }
// aPointersByAlpha[nAlpha - 1] = (IntPtr)cCUDA.CopyHostToDevice<IntPtr>(aPointersByBackground).Pointer;
// }
// cAlphaMap = cCUDA.CopyHostToDevice<IntPtr>(aPointersByAlpha);
//}
#if DEBUG
Dictionary<ulong, DateTime> ahDebug = new Dictionary<ulong,DateTime>();
#endif
DateTime dtNextTime = DateTime.MinValue, dtNow;
long nStartTick; // logging
while (true)
{
if (1 > _aqCommands.CountGet() && (dtNow = DateTime.Now) > dtNextTime)
{
dtNextTime = dtNow.AddSeconds(60);
#if DEBUG
dtNow = dtNow.Subtract(TimeSpan.FromHours(2));
string sMessage = "";
foreach (ulong nID in ahDebug.Keys)
if (dtNow > ahDebug[nID])
sMessage += "<br>[" + nID + ":" + ahDebug[nID].ToString("HH:mm:ss") + "]";
#endif
(new Logger()).WriteDebug("CUDA free memory:" + cCUDA.FreeMemory
#if DEBUG
+ "; possibly timeworn allocations:" + (1 > sMessage.Length ? "no" : sMessage)
#endif
);
}
while (true)
{
try
{
cCmd = _aqCommands.Dequeue(); //если нечего отдать - заснёт
break;
}
catch (Exception ex)
{
(new Logger()).WriteError(ex);
}
}
_CommandsCount = _aqCommands.nCount;
switch (cCmd.eID)
{
case Command.ID.Allocate:
#region
try
{
cCmd.cPM._cException = null;
if (1 > cCmd.cPM._nID)
{
if (0 < cCmd.cPM._nBytesQty)
{
nMemoryFree = cCUDA.FreeMemory;
if (nMemoryReservedForMerge < nMemoryFree - cCmd.cPM._nBytesQty)
{
bMemoryStarvation = (nMemoryFree < nMemoryStarvationThreshold);
cCmd.cPM._nID = _nCurrentID++;
ahDevicePointers.Add(cCmd.cPM._nID, cCUDA.Allocate(cCmd.cPM._nBytesQty));
#if DEBUG
ahDebug.Add(cCmd.cPM._nID, DateTime.Now);
#endif
}
else
{
bMemoryStarvation = true;
throw new Exception("out of memory in CUDA device during Allocate. Only 2 MBytes reserved for the Merge");
}
}
else
throw new Exception("bytes quantity in PixelsMap have to be greater than zero for Allocate [_bDisposed = " + cCmd.cPM._bDisposed + "][_bProcessing = " + cCmd.cPM._bProcessing + "][_bShiftVertical = " + cCmd.cPM._bShiftVertical + "][_bTemp = " + cCmd.cPM._bTemp + "][_dt = " + cCmd.cPM._dt + "][_nBytesQty = " + cCmd.cPM._nBytesQty + "][_nID = " + cCmd.cPM._nID + "][_nShiftPosition = " + cCmd.cPM._nShiftPosition + "][_stArea.nHeight = " + cCmd.cPM._stArea.nHeight + "][_stArea.nWidth = " + cCmd.cPM._stArea.nWidth + "][bKeepAlive = " + cCmd.cPM.bKeepAlive + "][bBackgroundClear = " + cCmd.cPM.bBackgroundClear + "][eAlpha = " + cCmd.cPM.eAlpha + "][bCUDA = " + cCmd.cPM.bCUDA + "][nAlphaConstant = " + cCmd.cPM.nAlphaConstant + "][nID = " + cCmd.cPM.nID + "][nLength = " + cCmd.cPM.nLength + "][stArea.nHeight = " + cCmd.cPM.stArea.nHeight + "][stArea.nWidth = " + cCmd.cPM.stArea.nWidth + "]");
}
else
throw new Exception("PixelsMap ID have to be zero for Allocate");
}
catch (Exception ex)
{
if (ex is CUDAException)
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
(new Logger()).WriteError(ex);
(new Logger()).WriteDebug("bytes qty:" + cCmd.cPM._nBytesQty);
cCmd.cPM._cException = ex;
}
cCmd.cMRE.Set();
break;
#endregion
case Command.ID.CopyIn:
#region
nStartTick = DateTime.Now.Ticks; // logging
try
{
cCmd.cPM._cException = null;
if (1 > cCmd.cPM._nID)
{
if (cCUDA.FreeMemory - cCmd.cPM._nBytesQty > nMemoryReservedForMerge)
{
cCmd.cPM._nID = _nCurrentID++;
if (cCmd.ahParameters.ContainsKey(typeof(IntPtr)))
ahDevicePointers.Add(cCmd.cPM._nID, cCUDA.CopyHostToDevice((IntPtr)cCmd.ahParameters[typeof(IntPtr)], cCmd.cPM._nBytesQty));
else if (cCmd.ahParameters.ContainsKey(typeof(byte[])))
ahDevicePointers.Add(cCmd.cPM._nID, cCUDA.CopyHostToDevice((byte[])cCmd.ahParameters[typeof(byte[])]));
else
throw new Exception("unknown parameter type");
#if DEBUG
ahDebug.Add(cCmd.cPM._nID, DateTime.Now);
#endif
}
else
throw new Exception("out of memory in CUDA device during CopyIn. Only 2 MBytes reserved for the Merge.");
}
else
{
if (cCmd.ahParameters.ContainsKey(typeof(IntPtr)))
cCUDA.CopyHostToDevice(ahDevicePointers[cCmd.cPM._nID], (IntPtr)cCmd.ahParameters[typeof(IntPtr)], cCmd.cPM._nBytesQty);
else if (cCmd.ahParameters.ContainsKey(typeof(byte[])))
cCUDA.CopyHostToDevice(ahDevicePointers[cCmd.cPM._nID], (byte[])cCmd.ahParameters[typeof(byte[])]);
else
throw new Exception("unknown parameter type");
}
if (ahDevicePointers.ContainsKey(cCmd.cPM._nID))
(new Logger()).WriteDebug5("copy in [id:" + cCmd.cPM._nID + "][ptr:" + ahDevicePointers[cCmd.cPM._nID].Pointer + "]");
else
(new Logger()).WriteDebug5("copy in [id:" + cCmd.cPM._nID + "][ptr: not in dictionary]");
}
catch (Exception ex)
{
if (ex is CUDAException)
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
(new Logger()).WriteError(ex);
cCmd.cPM._cException = ex;
}
if (new TimeSpan(DateTime.Now.Ticks - nStartTick).TotalMilliseconds >= 20) // logging
(new Logger()).WriteNotice("PixelMap: Command.ID.CopyIn: execution time > 20ms: " + new TimeSpan(DateTime.Now.Ticks - nStartTick).TotalMilliseconds +"ms"); // logging
cCmd.cMRE.Set();
break;
#endregion
case Command.ID.CopyOut:
#region
nStartTick = DateTime.Now.Ticks; // logging
try
{
if (0 < cCmd.cPM._nID)
{
if(!cCmd.ahParameters.ContainsKey(typeof(IntPtr)))
{
if(cCmd.ahParameters.ContainsKey(typeof(byte[])))
{
cCmd.cPM._aBytes = (byte[])cCmd.ahParameters[typeof(byte[])];
if(cCmd.cPM._nBytesQty != cCmd.cPM._aBytes.Length)
(new Logger()).WriteWarning("wrong array size for copyout [got:" + cCmd.cPM._aBytes.Length + "][expected:" + cCmd.cPM._nBytesQty + "]");
}
else
cCmd.cPM._aBytes = new byte[cCmd.cPM._nBytesQty];
cCUDA.CopyDeviceToHost<byte>(ahDevicePointers[cCmd.cPM._nID], cCmd.cPM._aBytes);
}
else
cCUDA.CopyDeviceToHost(ahDevicePointers[cCmd.cPM._nID], (IntPtr)cCmd.ahParameters[typeof(IntPtr)], cCmd.cPM._nBytesQty);
(new Logger()).WriteDebug5("copy out [id:" + cCmd.cPM._nID + "][ptr:" + ahDevicePointers[cCmd.cPM._nID].Pointer + "]");
}
else
throw new Exception("PixelsMap have to be allocated for CopyOut");
}
catch (Exception ex)
{
if (ex is CUDAException)
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
(new Logger()).WriteError(ex);
cCmd.cPM._cException = ex;
}
if (new TimeSpan(DateTime.Now.Ticks - nStartTick).TotalMilliseconds >= 20) // logging
(new Logger()).WriteNotice("PixelMap: Command.ID.CopyOut: execution time > 20ms: " + new TimeSpan(DateTime.Now.Ticks - nStartTick).TotalMilliseconds +"ms"); // logging
cCmd.cMRE.Set();
break;
#endregion
case Command.ID.Merge:
#region
try
{
List<PixelsMap> aPMs = (List<PixelsMap>)cCmd.ahParameters[typeof(List<PixelsMap>)];
DisCom.MergeInfo cMergeInfo = (DisCom.MergeInfo)cCmd.ahParameters[typeof(DisCom.MergeInfo)];
List<IntPtr> aDPs = new List<IntPtr>();
if (1 > cCmd.cPM._nID)
throw new Exception("background PixelsMap have to be allocated for Merge");
aDPs.Add((IntPtr)ahDevicePointers[cCmd.cPM._nID].Pointer);
for (int nIndx = 0; nIndx < aPMs.Count; nIndx++)
{
if (!ahDevicePointers.ContainsKey(aPMs[nIndx]._nID))
throw new Exception("there is a corrupted ID in layers for merge [id:" + aPMs[nIndx]._nID + "]");
if (1 > ahDevicePointers[aPMs[nIndx]._nID].Pointer)
throw new Exception("there is an empty pointer in layers for merge [id:" + aPMs[nIndx]._nID + "]");
aDPs.Add((IntPtr)ahDevicePointers[aPMs[nIndx]._nID].Pointer);
}
cPMs = cCUDA.CopyHostToDevice<IntPtr>(aDPs.ToArray());
cInfos = cCUDA.CopyHostToDevice(cMergeInfo, cMergeInfo.SizeGet());
cCUDA.SetParameter<IntPtr>(cCUDAFuncMerge, 0, (IntPtr)cPMs.Pointer);
cCUDA.SetParameter<IntPtr>(cCUDAFuncMerge, IntPtr.Size, (IntPtr)cInfos.Pointer);
cCUDA.SetParameter<IntPtr>(cCUDAFuncMerge, IntPtr.Size * 2, (IntPtr)cAlphaMap.Pointer);
cCUDA.SetParameterSize(cCUDAFuncMerge, (uint)(IntPtr.Size * 3));
int nIterations = (0 == cMergeInfo.nBackgroundSize % nThreadsPerBlock ? cMergeInfo.nBackgroundSize / nThreadsPerBlock : cMergeInfo.nBackgroundSize / nThreadsPerBlock + 1);
cCUDA.Launch(cCUDAFuncMerge, nIterations, 1);
cCmd.cMRE.Set();
cMergeInfo.Dispose();
cCUDA.Free(cPMs);
cCUDA.Free(cInfos);
for (int nIndx = 0; nIndx < aPMs.Count; nIndx++)
{
lock (aPMs[nIndx]._cSyncRoot)
aPMs[nIndx]._bProcessing = false;
aPMs[nIndx].Dispose();
}
}
catch (Exception ex)
{
cCmd.cMRE.Set();
if (ex is CUDAException)
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
(new Logger()).WriteError(ex);
cCmd.cPM._cException = ex;
}
break;
#endregion
case Command.ID.Dispose:
#region
nStartTick = DateTime.Now.Ticks; // logging
(new Logger()).Write(Logger.Level.debug2, "dispose: in");
try
{
if (ahDevicePointers.ContainsKey(cCmd.cPM._nID))
{
if (0 < cCmd.cPM._nID && 0 < ahDevicePointers[cCmd.cPM._nID].Pointer)
{
cCUDA.Free(ahDevicePointers[cCmd.cPM._nID]);
//cCUDA.SynchronizeContext();
bMemoryStarvation = (cCUDA.FreeMemory < nMemoryStarvationThreshold);
(new Logger()).WriteDebug3("dispose [id:" + cCmd.cPM._nID + "][ptr:" + ahDevicePointers[cCmd.cPM._nID].Pointer + "]");
}
ahDevicePointers.Remove(cCmd.cPM._nID);
#if DEBUG
ahDebug.Remove(cCmd.cPM._nID);
#endif
cCmd.cPM._nID = 0;
}
}
catch (Exception ex)
{
if (ex is CUDAException)
ex = new Exception("CUDA Error:" + ((CUDAException)ex).CUDAError.ToString(), ex);
(new Logger()).WriteError(ex);
cCmd.cPM._cException = ex;
}
(new Logger()).Write(Logger.Level.debug2, "dispose: out");
if (new TimeSpan(DateTime.Now.Ticks - nStartTick).TotalMilliseconds >= 20) // logging
(new Logger()).WriteNotice("PixelMap: Command.ID.Dispose: execution time > 20ms: " + new TimeSpan(DateTime.Now.Ticks - nStartTick).TotalMilliseconds +"ms"); // logging
break;
#endregion
}
}
}
catch (Exception ex)
{
(new Logger()).WriteError(ex);
}
}
static void Main(string[] args)
{
// Init and select 1st device.
CUDA cuda = new CUDA(0, true);
// load module
//cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "simpleCUFFT.ptx"));
CUfunction func = new CUfunction();// cuda.GetModuleFunction("ComplexPointwiseMulAndScale");
// The filter size is assumed to be a number smaller than the signal size
const int SIGNAL_SIZE = 50;
const int FILTER_KERNEL_SIZE = 11;
// Allocate host memory for the signal
Float2[] h_signal = new Float2[SIGNAL_SIZE];
// Initalize the memory for the signal
Random r = new Random();
for (int i = 0; i < SIGNAL_SIZE; ++i)
{
h_signal[i].x = r.Next() / (float)int.MaxValue;
h_signal[i].y = 0;
}
// Allocate host memory for the filter
Float2[] h_filter_kernel = new Float2[FILTER_KERNEL_SIZE];
// Initalize the memory for the filter
for (int i = 0; i < FILTER_KERNEL_SIZE; ++i)
{
h_filter_kernel[i].x = r.Next() / (float)int.MaxValue;
h_filter_kernel[i].y = 0;
}
// Pad signal and filter kernel
Float2[] h_padded_signal;
Float2[] h_padded_filter_kernel;
int new_size = PadData(h_signal, out h_padded_signal, SIGNAL_SIZE,
h_filter_kernel, out h_padded_filter_kernel, FILTER_KERNEL_SIZE);
// Allocate device memory for signal
// Copy host memory to device
CUdeviceptr d_signal = cuda.CopyHostToDevice<Float2>(h_padded_signal);
// Allocate device memory for filter kernel
// Copy host memory to device
CUdeviceptr d_filter_kernel = cuda.CopyHostToDevice<Float2>(h_padded_filter_kernel);
// CUFFT plan
CUFFT fft = new CUFFT(cuda);
cufftHandle handle = new cufftHandle();
CUFFTResult fftres = CUFFTDriver.cufftPlan1d(ref handle, new_size, CUFFTType.C2C, 1);
//fft.Plan1D(new_size, CUFFTType.C2C, 1);
return;
// Transform signal and kernel
fft.ExecuteComplexToComplex(d_signal, d_signal, CUFFTDirection.Forward);
fft.ExecuteComplexToComplex(d_filter_kernel, d_filter_kernel, CUFFTDirection.Forward);
// Multiply the coefficients together and normalize the result
// ComplexPointwiseMulAndScale<<<32, 256>>>(d_signal, d_filter_kernel, new_size, 1.0f / new_size);
cuda.SetFunctionBlockShape(func, 256, 1, 1);
cuda.SetParameter(func, 0, (uint)d_signal.Pointer);
cuda.SetParameter(func, IntPtr.Size, (uint)d_filter_kernel.Pointer);
cuda.SetParameter(func, IntPtr.Size * 2, (uint)new_size);
cuda.SetParameter(func, IntPtr.Size * 2 + 4, 1.0f / new_size);
cuda.SetParameterSize(func, (uint)(IntPtr.Size * 2 + 8));
cuda.Launch(func, 32, 1);
// Transform signal back
fft.ExecuteComplexToComplex(d_signal, d_signal, CUFFTDirection.Inverse);
// Copy device memory to host
Float2[] h_convolved_signal = h_padded_signal;
cuda.CopyDeviceToHost<Float2>(d_signal, h_convolved_signal);
// Allocate host memory for the convolution result
Float2[] h_convolved_signal_ref = new Float2[SIGNAL_SIZE];
// Convolve on the host
Convolve(h_signal, SIGNAL_SIZE,
h_filter_kernel, FILTER_KERNEL_SIZE,
h_convolved_signal_ref);
// check result
bool res = cutCompareL2fe(h_convolved_signal_ref, h_convolved_signal, 2 * SIGNAL_SIZE, 1e-5f);
Console.WriteLine("Test {0}", (true == res) ? "PASSED" : "FAILED");
//Destroy CUFFT context
fft.Destroy();
// cleanup memory
cuda.Free(d_signal);
cuda.Free(d_filter_kernel);
}
private static float[] CuRBFCSRCached()
{
//always the same values
Random rnd = new Random(1);
CUDA cuda = new CUDA(0, true);
// load module
CUmodule module = cuda.LoadModule(Path.Combine(Environment.CurrentDirectory, "structKernel.cubin"));
CUfunction structPassFunc = cuda.GetModuleFunction("RBFspmv_csr_vector");
int maxRowSize = avgElements + stdElements - 1;
Console.WriteLine("init arrays");
Stopwatch t = Stopwatch.StartNew();
List<float> vecValsL = new List<float>(N * maxRowSize / 2);
List<int> vecIdxL = new List<int>(N * maxRowSize / 2);
List<int> vecLenghtL = new List<int>(N);
float[] vecVals;
int[] vecIdx;
int[] vecLenght;
float[] selfDot = new float[N];
maxIndex = 0;
int vecStartIdx = 0;
for (int i = 0; i < N; i++)
{
int vecSize = avgElements + i % stdElements;
float[] vals = Helpers.InitValues(i, vecSize, maxVal);
vecValsL.AddRange(vals);
for (int z = 0; z < vals.Length; z++)
{
selfDot[i] += vals[z] * vals[z];
}
int[] index = Helpers.InitIndices(i, vecSize, ref maxIndex);
vecIdxL.AddRange(index);
vecLenghtL.Add(vecStartIdx);
vecStartIdx += vecSize;
}
//for last index
vecLenghtL.Add(vecStartIdx);
vecVals = vecValsL.ToArray();
vecIdx = vecIdxL.ToArray();
vecLenght = vecLenghtL.ToArray();
float[] mainVec = new float[maxIndex + 1];
for (int j = vecLenght[mainIndex]; j < vecLenght[mainIndex + 1]; j++)
{
int idx = vecIdx[j];
float val = vecVals[j];
mainVec[idx] = val;
}
Console.WriteLine("Init takes {0}", t.Elapsed);
t.Start();
CUdeviceptr valsPtr = cuda.CopyHostToDevice(vecVals);
CUdeviceptr idxPtr = cuda.CopyHostToDevice(vecIdx);
CUdeviceptr vecLenghtPtr = cuda.CopyHostToDevice(vecLenght);
CUdeviceptr selfDotPtr = cuda.CopyHostToDevice(selfDot);
//copy to texture
CUarray cuArr = cuda.CreateArray(mainVec);
cuda.CopyHostToArray(cuArr, mainVec, 0);
CUtexref cuTexRef = cuda.GetModuleTexture(module, "texRef");
cuda.SetTextureFlags(cuTexRef, 0);
cuda.SetTextureArray(cuTexRef, cuArr);
float[] output = new float[N];
CUdeviceptr dOutput = cuda.Allocate(output);
Console.WriteLine("copy to device takes {0}", t.Elapsed);
cuda.SetFunctionBlockShape(structPassFunc, threadsPerBlock, 1, 1);
int offset = 0;
cuda.SetParameter(structPassFunc, offset, valsPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, idxPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, vecLenghtPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, selfDotPtr.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, dOutput.Pointer);
offset += IntPtr.Size;
cuda.SetParameter(structPassFunc, offset, (uint)N);
offset += sizeof(int);
cuda.SetParameter(structPassFunc, offset, (uint)mainIndex);
offset += sizeof(int);
cuda.SetParameter(structPassFunc, offset, Gamma);
offset += sizeof(float);
cuda.SetParameter(structPassFunc, offset, (uint)vecStartIdx);
offset += sizeof(int);
cuda.SetParameterSize(structPassFunc, (uint)offset);
Console.WriteLine("start computation");
CUevent start = cuda.CreateEvent();
CUevent end = cuda.CreateEvent();
Stopwatch timer = Stopwatch.StartNew();
cuda.RecordEvent(start);
cuda.Launch(structPassFunc, blocksPerGrid, 1);
cuda.RecordEvent(end);
cuda.SynchronizeContext();
//cuda.SynchronizeEvent(end);
timer.Stop();
float naiveTime = cuda.ElapsedTime(start, end);
Console.Write("csr vector Dot products with mainIndex {0} and {1}-vectors takes {2} ms stopwatch time {3} ms", mainIndex, N, naiveTime, timer.Elapsed);
cuda.CopyDeviceToHost(dOutput, output);
int lenght = Math.Min(displayCount, N);
Console.WriteLine();
for (int i = 0; i < lenght; i++)
{
Console.WriteLine("{0}-{1}", i, output[i]);
}
cuda.Free(valsPtr);
cuda.Free(idxPtr);
cuda.Free(dOutput);
cuda.Free(selfDotPtr);
cuda.Free(vecLenghtPtr);
cuda.DestroyArray(cuArr);
cuda.DestroyTexture(cuTexRef);
cuda.DestroyEvent(start);
cuda.DestroyEvent(end);
return output;
}
