private void SetCudaDataForEllpack()
{
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
CudaHelpers.TransformToEllpackRFormat(out vecVals, out vecColIdx, out vecLenght, TrainedModel.SupportElements);
evalBlocks = (sizeSV + evalThreads - 1) / evalThreads;
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
}
private void SetCudaDataForEllpack()
{
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
CudaHelpers.TransformToEllpackRFormat(out vecVals, out vecColIdx, out vecLenght, TrainedModel.SupportElements);
float[] selfSum = TrainedModel.SupportElements.AsParallel().Select(c => c.Values.Sum()).ToArray();
evalBlocks = (sizeSV + evalThreads - 1) / evalThreads;
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
selfSumPtr = cuda.CopyHostToDevice(selfSum);
}
public override void Init()
{
linKernel.ProblemElements = problemElements;
linKernel.Y = Y;
linKernel.Init();
base.Init();
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
int align = preFetch;
CudaHelpers.TransformToEllpackRFormat(out vecVals, out vecColIdx, out vecLenght, problemElements, align);
// CudaHelpers.TransformToEllpackRFormat(out vecVals, out vecColIdx, out vecLenght, problemElements);
selfLinDot = linKernel.DiagonalDotCache;
#region cuda initialization
InitCudaModule();
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
labelsPtr = cuda.CopyHostToDevice(Y);
selfLinDotPtr = cuda.CopyHostToDevice(selfLinDot);
uint memSize = (uint)(2 * problemElements.Length * sizeof(float));
//allocate mapped memory for our results
//CUDARuntime.cudaSetDeviceFlags(CUDARuntime.cudaDeviceMapHost);
// var e= CUDADriver.cuMemHostAlloc(ref outputIntPtr, memSize, 8);
//CUDARuntime.cudaHostAlloc(ref outputIntPtr, memSize, CUDARuntime.cudaHostAllocMapped);
//var errMsg=CUDARuntime.cudaGetErrorString(e);
//cuda.HostRegister(outputIntPtr,memSize, Cuda)
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 dimenson, so many
//indexes will be zero, but cuda computation is faster
VectorI = new float[problemElements[0].Dim + 1];
VectorJ = new float[problemElements[0].Dim + 1];
CudaHelpers.FillDenseVector(problemElements[0], VectorI);
CudaHelpers.FillDenseVector(problemElements[1], VectorJ);
CudaHelpers.SetTextureMemory(cuda, cuModule, ref cuVecI_TexRef, cuVecITexRefName, VectorI, ref VecIPtr);
CudaHelpers.SetTextureMemory(cuda, cuModule, ref cuVecJ_TexRef, cuVecJTexRefName, VectorJ, ref VecJPtr);
}
