public override void Init()
{
base.Init();
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
//change the blocksPerGrid, because we launch many threads per row
blocksPerGrid = (int)Math.Ceiling((ThreadsPerRow * problemElements.Length + 0.0) / threadsPerBlock);
int align = ThreadsPerRow * Prefetch;
CudaHelpers.TransformToERTILPFormat(out vecVals, out vecColIdx, out vecLenght, problemElements, align, ThreadsPerRow);
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);
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);
if (MakeDenseVectorOnGPU)
{
vecBuilder = new EllpackDenseVectorBuilder(cuda, mainVecPtr, valsPtr, idxPtr, vecLengthPtr, problemElements.Length, problemElements[0].Dim);
vecBuilder.Init();
}
}
public override void Init()
{
linKernel.ProblemElements = problemElements;
linKernel.Y = Y;
linKernel.Init();
base.Init();
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
//change the blocksPerGrid, because we launch many threads per row
blocksPerGrid = (int)Math.Ceiling((ThreadsPerRow * problemElements.Length + 0.0) / threadsPerBlock);
int align = ThreadsPerRow * Prefetch;
CudaHelpers.TransformToERTILPFormat(out vecVals, out vecColIdx, out vecLenght, problemElements, align, ThreadsPerRow);
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);
selfLinDotPtr = cuda.CopyHostToDevice(selfLinDot);
uint memSize = (uint)(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 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);
if (MakeDenseVectorOnGPU)
{
vecBuilder = new EllpackDenseVectorBuilder(cuda, mainVecPtr, valsPtr, idxPtr, vecLengthPtr, problemElements.Length, problemElements[0].Dim);
vecBuilder.Init();
}
}
