Function qr

src/backend/opencl/qr.cpp:30–79

template<typename T>
void qr(Array<T> &q, Array<T> &r, Array<T> &t, const Array<T> &orig) {
    if (OpenCLCPUOffload()) { return cpu::qr(q, r, t, orig); }

    const dim4 NullShape(0, 0, 0, 0);

    dim4 iDims = orig.dims();
    int M      = iDims[0];
    int N      = iDims[1];

    dim4 endPadding(M - iDims[0], max(M, N) - iDims[1], 0, 0);
    Array<T> in =
        (endPadding == NullShape
             ? copyArray(orig)
             : padArrayBorders(orig, NullShape, endPadding, AF_PAD_ZERO));
    in.resetDims(iDims);

    int MN = std::min(M, N);
    int NB = magma_get_geqrf_nb<T>(M);

    int NUM      = (2 * MN + ((N + 31) / 32) * 32) * NB;
    Array<T> tmp = createEmptyArray<T>(dim4(NUM));

    std::vector<T> h_tau(MN);

    int info           = 0;
    cl::Buffer *in_buf = in.get();
    cl::Buffer *dT     = tmp.get();

    magma_geqrf3_gpu<T>(M, N, (*in_buf)(), in.getOffset(), in.strides()[1],
                        &h_tau[0], (*dT)(), tmp.getOffset(), getQueue()(),
                        &info);

    r = createEmptyArray<T>(in.dims());
    kernel::triangle<T>(r, in, true, false);

    cl::Buffer *r_buf = r.get();
    magmablas_swapdblk<T>(MN - 1, NB, (*r_buf)(), r.getOffset(), r.strides()[1],
                          1, (*dT)(), tmp.getOffset() + MN * NB, NB, 0,
                          getQueue()());

    q = in;  // No need to copy
    q.resetDims(dim4(M, M));
    cl::Buffer *q_buf = q.get();

    magma_ungqr_gpu<T>(q.dims()[0], q.dims()[1], std::min(M, N), (*q_buf)(),
                       q.getOffset(), q.strides()[1], &h_tau[0], (*dT)(),
                       tmp.getOffset(), NB, getQueue()(), &info);

    t = createHostDataArray(dim4(MN), &h_tau[0]);
}

template<typename T>
Array<T> qr_inplace(Array<T> &in) {