def idd_lssolve(queue, m, n, a, krank):
for j in range(n - krank):
blas.trsv(queue, a[:krank,:krank], a[:krank,krank+j],lower=False)
ctx = queue.get_info(cl.command_queue_info.CONTEXT)
prg = cl.Program(ctx, util.get_source('id_kerns.cl')).build()
prg.moveup(queue, [krank, n-krank], None, a.data, np.int32(krank), np.int32(n))
def iddr_id(queue, m, n, a, krank, lst, rnorms):
ctx = queue.get_info(cl.command_queue_info.CONTEXT)
iddr_qrpiv(queue, m, n, a, krank, lst, rnorms)
id_prg = cl.Program(ctx, util.get_source('id_kerns.cl')).build()
id_prg.rnorm(queue, [krank, 1], None, a.data, rnorms.data, np.int32(n))
idd_lssolve(queue, m, n, a, krank)
def iddr_qrpiv(queue, m, n, a, krank, ind, ss):
r = cl_array.Array(queue, a.shape, a.dtype)
r.set(np.zeros(a.shape, a.dtype))
ind.set(np.arange(ind.shape[0], dtype=ind.dtype))
ctx = queue.get_info(cl.command_queue_info.CONTEXT)
qr_prg = cl.Program(ctx, util.get_source('qr_kerns.cl')).build()
# begin debug
# from scipy import linalg
# _,R,P = linalg.qr(a.get(),pivoting=True)
# print(R)
# print(P)
# end debug
ss_l = cl.LocalMemory(m*np.dtype('float64').itemsize)
qr_prg.ss(queue, [m,n], [m,1], a.data, ss.data, ss_l,
np.int32(0),np.int32(n))
kpiv = util.argmax(queue, ss)
qk = cl.LocalMemory(m*np.dtype('float64').itemsize)
aj_qk = cl.LocalMemory(m*np.dtype('float64').itemsize)
nloops = np.min([m,n,krank])
for k in range(nloops):
qr_prg.swap_col(queue, [m, 1], [m, 1], a.data, ss.data, r.data, ind.data,
np.int32(k), np.int32(kpiv), np.int32(n))
qr_prg.proj_rm(queue, [m, n-(k+1)], [m, 1],
a.data, r.data, ss.data,
qk, ss_l, aj_qk, np.int32(k), np.int32(n))
qr_prg.ss(queue, [m,n-(k+1)], [m,1], a.data, ss.data, ss_l, np.int32(k+1),
np.int32(n))
kpiv = util.argmax(queue, ss[k+1:]) + k + 1
# qr_prg.norm(queue, [m,1], [m, 1],
# r.data, ss.data,
# np.int32(nloops-1), np.int32(n))
# print(r.get())
# print(ind)
# copy r into a for output
a[:krank,:] = r[:krank,:]