public ICudaResult <double[][]> Multiply(
CUBLAS_OP a_op,
CUBLAS_OP b_op,
double alpha,
double[][] a,
double[][] b,
double beta)
{
// C(m, n) = A(m, k) * B(k, n)
var matrix_a_dimensions = MatrixSizeByOperation(a.Length, a[0].Length, a_op);
var matrix_b_dimensions = MatrixSizeByOperation(b.Length, b[0].Length, b_op);
if (matrix_a_dimensions.Columns != matrix_b_dimensions.Rows)
{
throw new ArgumentOutOfRangeException($"Matrices provided cannot be multipled. Columns in matrix A: {matrix_a_dimensions.Columns} vs rows in matrix B: {matrix_b_dimensions.Rows}");
}
var result = DTM.MatrixMultiplyDouble(
DeviceId,
a_op, b_op,
alpha,
a,
b,
beta);
return(new CudaResult <double[][]>(result.Error, result.Result));
}
internal static (CudaError Error, double[][] Result) MatrixMultiplyDouble(
int device_id,
CUBLAS_OP a_op, CUBLAS_OP b_op,
double alpha,
double[][] a,
double[][] b,
double beta)
{
// .NET does not support marshaling nested arrays between C++ and e.g. C#.
// If you try, you will get the error message, "There is no marshaling support for nested arrays."
// Further, the cuBLAS function cublasSgemm/cublasDgemm does not have pointer-to-pointers as arguments (e.g., float**), so we cannot
// supply a multidimensional array anyway.
// The solution: flatten arrays so that they can passed to CudaSharperLibrary, and then unflatten whatever it passes back.
var d_a = FlattenArray(a.Length, a[0].Length, a);
var d_b = FlattenArray(b.Length, b[0].Length, b);
// C(m,n) = A(m,k) * B(k,n)
// Despite the definition above, this will return the correct size for C. Go figure.
var d_c = new double[a.Length * b.Length];
var transa_op = (int)a_op;
var transb_op = (int)b_op;
var error = SafeNativeMethods.MatrixMultiplyDouble(
device_id,
transa_op, transb_op,
a.Length, b[0].Length, a[0].Length,
alpha,
d_a,
d_b,
beta,
d_c);
return(CudaErrorCodes(error), UnflattenArray(a.Length, b.Length, d_c));
}
private (int Rows, int Columns) MatrixSizeByOperation(int rows, int columns, CUBLAS_OP operation)
{
int matrix_rows = 0;
int matric_columns = 0;
switch (operation)
{
case CUBLAS_OP.DO_NOT_TRANSPOSE:
matrix_rows = rows;
matric_columns = columns;
break;
case CUBLAS_OP.TRANSPOSE:
matrix_rows = columns;
matric_columns = rows;
break;
}
return(matrix_rows, matric_columns);
}