void lu(Array<T> &lower, Array<T> &upper, Array<int> &pivot, const Array<T> &in)
{
dim4 iDims = in.dims();
int M = iDims[0];
int N = iDims[1];
Array<T> in_copy = copyArray<T>(in);
//////////////////////////////////////////
// LU inplace
int *pivotPtr = pinnedAlloc<int>(min(M, N));
T *inPtr = pinnedAlloc<T> (in_copy.elements());
copyData(inPtr, in);
getrf_func<T>()(AF_LAPACK_COL_MAJOR, M, N,
inPtr, in_copy.strides()[1],
pivotPtr);
convertPivot(&pivotPtr, M, min(M, N));
pivot = createHostDataArray<int>(af::dim4(M), pivotPtr);
//////////////////////////////////////////
// SPLIT into lower and upper
dim4 ldims(M, min(M, N));
dim4 udims(min(M, N), N);
T *lowerPtr = pinnedAlloc<T>(ldims.elements());
T *upperPtr = pinnedAlloc<T>(udims.elements());
dim4 lst(1, ldims[0], ldims[0] * ldims[1], ldims[0] * ldims[1] * ldims[2]);
dim4 ust(1, udims[0], udims[0] * udims[1], udims[0] * udims[1] * udims[2]);
lu_split<T>(lowerPtr, upperPtr, inPtr, ldims, udims, iDims,
lst, ust, in_copy.strides());
lower = createHostDataArray<T>(ldims, lowerPtr);
upper = createHostDataArray<T>(udims, upperPtr);
lower.eval();
upper.eval();
pinnedFree(lowerPtr);
pinnedFree(upperPtr);
pinnedFree(pivotPtr);
pinnedFree(inPtr);
}
void lu(Array<T> &lower, Array<T> &upper, Array<int> &pivot,
const Array<T> &in) {
lower.eval();
upper.eval();
pivot.eval();
in.eval();
dim4 iDims = in.dims();
int M = iDims[0];
int N = iDims[1];
Array<T> in_copy = copyArray<T>(in);
pivot = lu_inplace(in_copy);
// SPLIT into lower and upper
dim4 ldims(M, min(M, N));
dim4 udims(min(M, N), N);
lower = createEmptyArray<T>(ldims);
upper = createEmptyArray<T>(udims);
getQueue().enqueue(kernel::lu_split<T>, lower, upper, in_copy);
}
