/// <summary>
/// Factorize the sparse matrix associated to the solver instance.
/// </summary>
public void Factorize()
{
if (_buffer != IntPtr.Zero)
{
throw new Exception("Context already created.");
}
int rows = matrix.RowCount;
int columns = matrix.ColumnCount;
Cuda.Malloc(ref d_x, sizeT * columns);
Cuda.Malloc(ref d_b, sizeT * rows);
// TODO: can the original matrix really be disposed after factorization?
using (var cusparse = new CuSparseContext <T>(stream, matrix, MatrixType.General, transpose))
{
PrepareFactorize();
// Start the timer after the first call to cusolver.
var timer = Stopwatch.StartNew();
Factorize(rows, columns, matrix.NonZerosCount, cusparse);
factorized = true;
timer.Stop();
FactorizationTime = TimeSpan.FromTicks(timer.ElapsedTicks).TotalSeconds;
}
}
/// <summary>
/// Initializes a new instance of the <see cref="CuSparseContext{T}"/> class.
/// </summary>
/// <param name="stream">The <see cref="CudaStream"/>.</param>
/// <param name="A">The sparse matrix.</param>
/// <param name="type">The matrix type.</param>
/// <param name="transpose">A value indicating, whether the storage should be transposed.</param>
public CuSparseContext(CudaStream stream, CompressedColumnStorage <T> A, MatrixType type, bool transpose)
{
Check(NativeMethods.cusparseCreate(ref _p));
Check(NativeMethods.cusparseSetStream(_p, stream.Pointer));
Check(NativeMethods.cusparseCreateMatDescr(ref _matDescr));
Check(NativeMethods.cusparseSetMatType(_matDescr, type));
Check(NativeMethods.cusparseSetMatIndexBase(_matDescr, IndexBase.Zero));
var sizeT = Marshal.SizeOf(typeof(T));
int rows = A.RowCount;
int nnz = A.NonZerosCount;
Cuda.Malloc(ref d_ap, sizeof(int) * (rows + 1));
Cuda.Malloc(ref d_ai, sizeof(int) * nnz);
Cuda.Malloc(ref d_ax, sizeT * nnz);
var handles = new List <GCHandle>();
try
{
// Convert storage to CSR format.
var C = transpose ? A.Transpose(true) : A;
var h_ap = InteropHelper.Pin(C.ColumnPointers, handles);
var h_ai = InteropHelper.Pin(C.RowIndices, handles);
var h_ax = InteropHelper.Pin(C.Values, handles);
Cuda.CopyToDevice(d_ap, h_ap, sizeof(int) * (rows + 1));
Cuda.CopyToDevice(d_ai, h_ai, sizeof(int) * nnz);
Cuda.CopyToDevice(d_ax, h_ax, sizeT * nnz);
}
finally
{
InteropHelper.Free(handles);
}
}
