public MatrixOutput(CudaProvider cuda, int rows, int columns, int count, bool setToZero)
{
//_cuda = cuda;
_rows = rows;
_columns = columns;
for (var i = 0; i < count; i++)
{
_data.Add(setToZero
? cuda.CreateZeroMatrix(rows, columns)
: cuda.CreateMatrix(rows, columns)
);
}
_ptr = _data.Cast <GpuMatrix>().Select(m => m.Memory.DevicePointer).ToArray();
}
public TensorOutput(CudaProvider cuda, int rows, int columns, int depth, int count, bool setToZero)
{
_cuda = cuda;
_rows = rows;
_columns = columns;
_depth = depth;
_count = count;
// allocate space for the output
_data = setToZero
? cuda.CreateZeroMatrix(rows * columns * depth, count)
: cuda.CreateMatrix(rows * columns * depth, count)
;
// get the pointers
for (var i = 0; i < count; i++)
{
var tensor = _data.Column(i).Split(depth).Cast <GpuVector>().Select(v => v.Memory).ToArray();
_ptr.Add(tensor);
}
}