private void SetCudaData(Problem <SparseVec> sub_prob)
{
int vecDim = sub_prob.FeaturesCount;//.Elements[0].Dim;
/*
* copy vectors to CUDA device
*/
#region copy trainning examples to GPU
float[] vecVals;
int[] vecIdx;
int[] vecLenght;
CudaHelpers.TransformToCSRFormat(out vecVals, out vecIdx, out vecLenght, sub_prob.Elements);
valsCSRPtr = cuda.CopyHostToDevice(vecVals);
idxCSRPtr = cuda.CopyHostToDevice(vecIdx);
vecLenghtCSRPtr = cuda.CopyHostToDevice(vecLenght);
Stopwatch timer = Stopwatch.StartNew();
CudaHelpers.TransformToCSCFormat2(out vecVals, out vecIdx, out vecLenght, sub_prob.Elements);
timer.Stop();
valsCSCPtr = cuda.CopyHostToDevice(vecVals);
idxCSCPtr = cuda.CopyHostToDevice(vecIdx);
vecLenghtCSCPtr = cuda.CopyHostToDevice(vecLenght);
// float[] vecVals2;
// int[] vecIdx2;
// int[] vecLenght2;
// Stopwatch timer2 = Stopwatch.StartNew();
// CudaHelpers.TransformToCSCFormat2(out vecVals2, out vecIdx2, out vecLenght2, sub_prob.Elements);
// timer2.Stop();
//var a= vecIdx.SequenceEqual(vecIdx2);
//var b= vecVals.SequenceEqual(vecVals2);
//var c= vecLenght.SequenceEqual(vecLenght2);
#endregion
/*
* allocate memory for gradient
*/
alphaMemSize = (uint)(sub_prob.ElementsCount * sizeof(float));
gradPtr = cuda.Allocate(alphaMemSize);
gradOldPtr = cuda.Allocate(alphaMemSize);
alphaPtr = cuda.Allocate(alphaMemSize);
alphaOldPtr = cuda.Allocate(alphaMemSize);
alphaTmpPtr = cuda.Allocate(alphaMemSize);
/*
* reduction blocks for computing Obj
*/
GetNumThreadsAndBlocks(vecDim, 64, threadsPerBlock, ref threadsForReduceObjW, ref bpgReduceW);
reduceObjW = new float[bpgReduceW];
uint reduceWBytes = (uint)bpgReduceW * sizeof(float);
reduceObjWPtr = cuda.Allocate(reduceWBytes);
/*
* reduction size for kernels which operate on alpha
*/
int reductionSize = problem.ElementsCount;
threadsForReduceObjAlpha = 0;
GetNumThreadsAndBlocks(problem.ElementsCount, 64, threadsPerBlock, ref threadsForReduceObjAlpha, ref bpgReduceAlpha);
uint alphaReductionBytes = (uint)bpgReduceAlpha * sizeof(float);
/*
* reduction array for computing objective function value
*/
reduceObjAlpha = new float[bpgReduceAlpha];
reduceObjAlphaPtr = cuda.Allocate(alphaReductionBytes);
/*
* reduction array for computing gradient max norm
*/
reduceGradMaxNorm = new float[bpgReduceAlpha];
reduceGradMaxNormPtr = cuda.Allocate(alphaReductionBytes);
/*
* reduction arrays for computing BB step
*/
alphaPartReduce = new float[bpgReduceAlpha];
gradPartReduce = new float[bpgReduceAlpha];
alphaGradPartReduce = new float[bpgReduceAlpha];
reduceBBAlphaGradPtr = cuda.Allocate(alphaReductionBytes);
reduceBBAlphaPtr = cuda.Allocate(alphaReductionBytes);
reduceBBGradPtr = cuda.Allocate(alphaReductionBytes);
//float[] wVec = new float[vecDim];
wVecMemSize = (uint)vecDim * sizeof(float);
wTempVecPtr = cuda.Allocate(wVecMemSize);
//move W wector
SetTextureMemory(ref cuWVecTexRef, cudaWVecTexRefName, ref wVecPtr, wVecMemSize);
//set texture memory for labels
SetTextureMemory(ref cuLabelsTexRef, cudaLabelsTexRefName, sub_prob.Y, ref labelsPtr);
SetTextureMemory(ref cuDeltasTexRef, "deltasTexRef", ref deltasPtr, alphaMemSize);
diagPtr = cuda.GetModuleGlobal(cuModule, "diag_shift");
stepBBPtr = cuda.GetModuleGlobal(cuModule, "stepBB");
float[] stepData = new float[] { 0.1f };
cuda.CopyHostToDevice(stepBBPtr, stepData);
SetCudaParameters(sub_prob);
}