private void SetCudaDataForFormat()
{
evalThreads = threadsPerRow * sliceSize;
int N = sizeSV;
evalBlocks = (int)Math.Ceiling(1.0 * N * threadsPerRow / evalThreads);
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, TrainedModel.SupportElements, threadsPerRow, sliceSize);
float[] selfSum = TrainedModel.SupportElements.AsParallel().Select(c => c.Values.Sum()).ToArray();
//copy data to device, set cuda function parameters
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
sliceStartPtr = cuda.CopyHostToDevice(sliceStart);
selfSumPtr = cuda.CopyHostToDevice(selfSum);
}
private void SetCudaData()
{
float[] vecVals;
int[] vecIdx;
int[] vecLenght;
CudaHelpers.TransformToCSRFormat(out vecVals, out vecIdx, out vecLenght, TrainedModel.SupportElements);
evalBlocks = (sizeSV + evalThreads - 1) / evalThreads;
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecIdx);
vecPointerPtr = cuda.CopyHostToDevice(vecLenght);
}
private void SetCudaDataForERTILP()
{
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
evalBlocks = (int)Math.Ceiling((ThreadsPerRow * sizeSV + 0.0) / evalThreads);
int align = ThreadsPerRow * Prefetch;
CudaHelpers.TransformToERTILPFormat(out vecVals, out vecColIdx, out vecLenght, TrainedModel.SupportElements, align, ThreadsPerRow);
//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[] selfLinDot = TrainedModel.SupportElements.Select(c => c.DotProduct()).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);
selfDotPtr = cuda.CopyHostToDevice(selfLinDot);
}
private void SetCudaDataForFormat()
{
evalThreads = threadsPerRow * sliceSize;
int N = sizeSV;
evalBlocks = (int)Math.Ceiling(1.0 * N * threadsPerRow / evalThreads);
align = (int)Math.Ceiling(1.0 * sliceSize * threadsPerRow / 64) * 64;
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
int[] sliceStart;
CudaHelpers.TransformToSERTILP(out vecVals, out vecColIdx, out sliceStart, out vecLenght, TrainedModel.SupportElements, threadsPerRow, sliceSize, preFechSize);
//copy data to device, set cuda function parameters
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
sliceStartPtr = cuda.CopyHostToDevice(sliceStart);
}
private void SetCudaData()
{
float[] vecVals;
int[] vecIdx;
int[] vecLenght;
CudaHelpers.TransformToCSRFormat(out vecVals, out vecIdx, out vecLenght, TrainedModel.SupportElements);
evalBlocks = (sizeSV+evalThreads-1) / evalThreads;
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecIdx);
vecPointerPtr = cuda.CopyHostToDevice(vecLenght);
}
private void SetCudaDataForEllpack()
{
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
int align = 2;
CudaHelpers.TransformToEllpackRFormat(out vecVals, out vecColIdx, out vecLenght, TrainedModel.SupportElements, align);
float[] selfLinDot = TrainedModel.SupportElements.Select(c => c.DotProduct()).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);
selfDotPtr = cuda.CopyHostToDevice(selfLinDot);
}
private void SetCudaDataForFormat()
{
evalThreads = threadsPerRow * sliceSize;
int N = sizeSV;
evalBlocks = (int)Math.Ceiling(1.0 * N * threadsPerRow / evalThreads);
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, TrainedModel.SupportElements, threadsPerRow, sliceSize);
float[] selfSum = TrainedModel.SupportElements.AsParallel().Select(c => c.Values.Sum()).ToArray();
//copy data to device, set cuda function parameters
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
sliceStartPtr = cuda.CopyHostToDevice(sliceStart);
selfSumPtr = cuda.CopyHostToDevice(selfSum);
}
private void SetCudaDataForFormat()
{
evalThreads = threadsPerRow * sliceSize;
int N = sizeSV;
evalBlocks = (int)Math.Ceiling(1.0 * N * threadsPerRow / evalThreads);
align = (int)Math.Ceiling(1.0 * sliceSize * threadsPerRow / 64) * 64;
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
int[] sliceStart;
CudaHelpers.TransformToSERTILP(out vecVals, out vecColIdx, out sliceStart, out vecLenght, TrainedModel.SupportElements, threadsPerRow, sliceSize, preFechSize);
//copy data to device, set cuda function parameters
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
sliceStartPtr = cuda.CopyHostToDevice(sliceStart);
}
public override T Execute <T>(params object[] args)
{
if (string.IsNullOrEmpty(m_entryMethod))
{
throw new Exception("No method is loaded");
}
// Create cuda context
GASS.CUDA.CUDA cuda = new GASS.CUDA.CUDA(true);
// Load module (PTX) and get function
GASS.CUDA.Types.CUmodule module = cuda.LoadModule(m_ptx);
GASS.CUDA.Types.CUfunction func = cuda.GetModuleFunction(module, m_entryMethod);
// Create data
float[] A = new float[N];
float[] B = new float[N];
float[] C = new float[N];
Random rand = new Random();
for (int i = 0; i < N; i++)
{
A[i] = rand.Next(1000);
B[i] = rand.Next(1000);
C[i] = -1;
}
// Allocate data on device
GASS.CUDA.Types.CUdeviceptr dA = cuda.Allocate <float>(A);
GASS.CUDA.Types.CUdeviceptr dB = cuda.Allocate <float>(B);
GASS.CUDA.Types.CUdeviceptr dC = cuda.Allocate <float>(C);
// Copy data to device
cuda.CopyHostToDevice <float>(dA, A);
cuda.CopyHostToDevice <float>(dB, B);
cuda.CopyHostToDevice <float>(dC, C);
// Load parameters
int offset = 0;
cuda.SetParameter(func, offset, dA);
offset += Marshal.SizeOf(typeof(IntPtr));
cuda.SetParameter(func, offset, dB);
offset += Marshal.SizeOf(typeof(IntPtr));
cuda.SetParameter(func, offset, dC);
offset += Marshal.SizeOf(typeof(IntPtr));
cuda.SetParameterSize(func, (uint)offset);
// Setup execution
int blocksPerGrid = (N + threadsPerBlock - 1) / threadsPerBlock;
cuda.SetFunctionBlockShape(func, threadsPerBlock, 1, 1);
// Launch execution
cuda.Launch(func, blocksPerGrid, 1);
// Wait for and retrive result
cuda.SynchronizeContext();
cuda.CopyDeviceToHost <float>(dC, C);
// Clean device memory
cuda.Free(dA);
cuda.Free(dB);
cuda.Free(dC);
return(default(T));
}
private void SetCudaDataForERTILP()
{
float[] vecVals;
int[] vecColIdx;
int[] vecLenght;
evalBlocks = (int)Math.Ceiling((ThreadsPerRow * sizeSV + 0.0) / evalThreads);
int align = ThreadsPerRow * Prefetch;
CudaHelpers.TransformToERTILPFormat(out vecVals, out vecColIdx, out vecLenght, TrainedModel.SupportElements, align, ThreadsPerRow);
//copy data to device, set cuda function parameters
valsPtr = cuda.CopyHostToDevice(vecVals);
idxPtr = cuda.CopyHostToDevice(vecColIdx);
vecLengthPtr = cuda.CopyHostToDevice(vecLenght);
}
