public override void Init()
{
base.Init();
blockSize = threadsPerRow * sliceSize;
int N = problemElements.Length;
blocksPerGrid = (int)Math.Ceiling(1.0 * N * threadsPerRow / blockSize);
align = (int)Math.Ceiling(1.0 * sliceSize * threadsPerRow / 64) * 64;
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
int[] sliceStart;
CudaHelpers.TransformToSlicedEllpack(out vecVals, out vecColIdx, out sliceStart, out vecLenght, problemElements, threadsPerRow, sliceSize);
selfSum = problemElements.AsParallel().Select(x => x.Values.Sum()).ToArray();
#region cuda initialization
InitCudaModule();
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
sliceStartPtr = cuda.CopyHostToDevice(sliceStart);
labelsPtr = cuda.CopyHostToDevice(Y);
//!!!!!
selfSumPtr = cuda.CopyHostToDevice(selfSum);
uint memSize = (uint)(problemElements.Length * sizeof(float));
outputIntPtr = cuda.HostAllocate(memSize, CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
outputPtr = cuda.GetHostDevicePointer(outputIntPtr, 0);
//normal memory allocation
//outputPtr = cuda.Allocate((uint)(sizeof(float) * problemElements.Length));
#endregion
SetCudaFunctionParameters();
//allocate memory for main vector, size of this vector is the same as dimension, so many
//indexes will be zero, but cuda computation is faster
mainVector = new float[problemElements[0].Dim + 1];
CudaHelpers.FillDenseVector(problemElements[0], mainVector);
CudaHelpers.SetTextureMemory(cuda, cuModule, ref cuMainVecTexRef, cudaMainVecTexRefName, mainVector, ref mainVecPtr);
// CudaHelpers.SetTextureMemory(cuda,cuModule,ref cuLabelsTexRef, cudaLabelsTexRefName, Y, ref labelsPtr);
}
public override void Init()
{
linKernel.ProblemElements = problemElements;
linKernel.Y = Y;
linKernel.Init();
base.Init();
blockSize = threadsPerRow * sliceSize;
int N = problemElements.Length;
blockPerGrid = (int)Math.Ceiling(1.0 * N * threadsPerRow / blockSize);
align = (int)Math.Ceiling(1.0 * sliceSize * threadsPerRow / 64) * 64;
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
int[] sliceStart;
CudaHelpers.TransformToSlicedEllpack(out vecVals, out vecColIdx, out sliceStart, out vecLenght, problemElements, threadsPerRow, sliceSize);
selfLinDot = linKernel.DiagonalDotCache;
#region cudafy initialization
InitCudaModule();
//copy data to device, set cuda function parameters
valsPtr = gpu.CopyToDevice(vecVals);
idxPtr = gpu.CopyToDevice(vecColIdx);
vecLenghtPtr = gpu.CopyToDevice(vecLenght);
sliceStartPtr = gpu.CopyToDevice(sliceStart);
//!!!!!
selfLinDotPtr = gpu.CopyToDevice(selfLinDot);
labelsPtr = gpu.CopyToDevice(Y);
//gpu.CopyToConstantMemory(new float[] { Gamma }, GammaDev);
//float[] GammaDev =new float[] { Gamma };
//float[] GammaDevPtr = gpu.Allocate<float>(1);
//gpu.CopyToConstantMemory<float>(GammaDev,GammaDevPtr);
//float[] Gammas = new float[] { Gamma };
//float[] GammaDev = gpu.Allocate<float>(1);
//gpu.CopyToConstantMemory<float>(Gammas, GammaDev);
int memSize = (problemElements.Length * sizeof(float));
//allocate mapped memory for our results
//outputIntPtr = gpu.HostAllocate<float>(problemElements.Length); // .HostAllocate(memSize, CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
//outputPtr = gpu.GetDeviceMemoryFromIntPtr(outputIntPtr);// cuda.GetHostDevicePointer(outputIntPtr, 0);
outputIntPtr = cuGPU.HostAllocate((uint)memSize, CUDADriver.CU_MEMHOSTALLOC_DEVICEMAP);
outputPtr = cuGPU.GetHostDevicePointer(outputIntPtr, 0);
#endregion
//allocate memory for main vector, size of this vector is the same as dimenson, so many
//indexes will be zero, but cuda computation is faster
mainVector = new float[problemElements[0].Dim + 1];
CudaHelpers.FillDenseVector(problemElements[0], mainVector);
CudaHelpers.SetTextureMemory(cuGPU, ref cuMainVecTexRef, cudaMainVecTexRefName, mainVector, ref mainVectorPtr);
//CudaHelpers.SetTextureMemory(cuGPU, ref cuLabelsTexRef, cudaLabelsTexRefName, Y, ref labelsPtr);
}
