double luT(
viennacl::matrix<T> &vclX,
viennacl::vector_base<T> &vclD) {
double logdet=-99.9;
viennacl::linalg::lu_factorize(vclX);
// pointer to the actual diagonal
viennacl::vector_base<T> diagOfVar(
vclX.handle(), vclX.size1(), 0,
vclX.internal_size2() + 1);
// compute log determinant
vclD = viennacl::linalg::element_log(diagOfVar);
logdet = viennacl::linalg::sum(vclD);
// OPERATION_UNARY_LOG_TYPE
//http://viennacl.sourceforge.net/doc/scheduler_8cpp-example.html#a11
// put the diagonals in D, and 1's on the diagonal of L
vclD = diagOfVar;
//diagOfVar = T(1); // problem here
return(logdet);
}
void init_random(viennacl::matrix<T, F> & M)
{
std::vector<T> cM(M.internal_size());
for (std::size_t i = 0; i < M.size1(); ++i)
for (std::size_t j = 0; j < M.size2(); ++j)
cM[F::mem_index(i, j, M.internal_size1(), M.internal_size2())] = T(rand())/T(RAND_MAX);
viennacl::fast_copy(&cM[0],&cM[0] + cM.size(),M);
}
bool householder_c(viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT>& A,
viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT>& Q,
viennacl::vector<SCALARTYPE, ALIGNMENT>& D,
vcl_size_t row_start, vcl_size_t col_start)
{
viennacl::ocl::context & ctx = const_cast<viennacl::ocl::context &>(viennacl::traits::opencl_handle(A).context());
if (row_start + 1 >= A.size1())
return false;
prepare_householder_vector(A, D, A.size1(), row_start, col_start, row_start, true);
{
viennacl::ocl::kernel& kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::svd<SCALARTYPE>::program_name(), SVD_HOUSEHOLDER_UPDATE_A_LEFT_KERNEL);
viennacl::ocl::enqueue(kernel(
A,
D,
static_cast<cl_uint>(row_start),
static_cast<cl_uint>(col_start),
static_cast<cl_uint>(A.size1()),
static_cast<cl_uint>(A.size2()),
static_cast<cl_uint>(A.internal_size2()),
viennacl::ocl::local_mem(static_cast<cl_uint>(128 * sizeof(SCALARTYPE)))
));
}
{
viennacl::ocl::kernel& kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::svd<SCALARTYPE>::program_name(), SVD_HOUSEHOLDER_UPDATE_QL_KERNEL);
viennacl::ocl::enqueue(kernel(
Q,
D,
static_cast<cl_uint>(A.size1()),
// static_cast<cl_uint>(A.size2()),
static_cast<cl_uint>(Q.internal_size2()),
viennacl::ocl::local_mem(static_cast<cl_uint>(128 * sizeof(SCALARTYPE)))
));
}
return true;
}
NumericT diff(std::vector<std::vector<NumericT> > const & A1, viennacl::matrix<NumericT> const & A2)
{
std::vector<NumericT> host_values(A2.internal_size());
for (std::size_t i=0; i<A2.size1(); ++i)
for (std::size_t j=0; j<A2.size2(); ++j)
host_values[i*A2.internal_size2() + j] = A1[i][j];
std::vector<NumericT> device_values(A2.internal_size());
viennacl::fast_copy(A2, &device_values[0]);
viennacl::vector<NumericT> vcl_device_values(A2.internal_size()); // workaround to avoid code duplication
viennacl::copy(device_values, vcl_device_values);
return diff(host_values, vcl_device_values);
}
void nmf(viennacl::matrix<ScalarType> const & v,
viennacl::matrix<ScalarType> & w,
viennacl::matrix<ScalarType> & h,
std::size_t k,
ScalarType eps = 0.000001,
std::size_t max_iter = 10000,
std::size_t check_diff_every_step = 100)
{
viennacl::linalg::kernels::nmf<ScalarType, 1>::init();
w.resize(v.size1(), k);
h.resize(k, v.size2());
std::vector<ScalarType> stl_w(w.internal_size1() * w.internal_size2());
std::vector<ScalarType> stl_h(h.internal_size1() * h.internal_size2());
for (std::size_t j = 0; j < stl_w.size(); j++)
stl_w[j] = static_cast<ScalarType>(rand()) / RAND_MAX;
for (std::size_t j = 0; j < stl_h.size(); j++)
stl_h[j] = static_cast<ScalarType>(rand()) / RAND_MAX;
viennacl::matrix<ScalarType> wn(v.size1(), k);
viennacl::matrix<ScalarType> wd(v.size1(), k);
viennacl::matrix<ScalarType> wtmp(v.size1(), v.size2());
viennacl::matrix<ScalarType> hn(k, v.size2());
viennacl::matrix<ScalarType> hd(k, v.size2());
viennacl::matrix<ScalarType> htmp(k, k);
viennacl::matrix<ScalarType> appr(v.size1(), v.size2());
viennacl::vector<ScalarType> diff(v.size1() * v.size2());
viennacl::fast_copy(&stl_w[0], &stl_w[0] + stl_w.size(), w);
viennacl::fast_copy(&stl_h[0], &stl_h[0] + stl_h.size(), h);
ScalarType last_diff = 0.0f;
for (std::size_t i = 0; i < max_iter; i++)
{
{
hn = viennacl::linalg::prod(trans(w), v);
htmp = viennacl::linalg::prod(trans(w), w);
hd = viennacl::linalg::prod(htmp, h);
viennacl::ocl::kernel & mul_div_kernel = viennacl::ocl::get_kernel(viennacl::linalg::kernels::nmf<ScalarType, 1>::program_name(),
NMF_MUL_DIV_KERNEL);
viennacl::ocl::enqueue(mul_div_kernel(h, hn, hd, cl_uint(stl_h.size())));
}
{
wn = viennacl::linalg::prod(v, trans(h));
wtmp = viennacl::linalg::prod(w, h);
wd = viennacl::linalg::prod(wtmp, trans(h));
viennacl::ocl::kernel & mul_div_kernel = viennacl::ocl::get_kernel(viennacl::linalg::kernels::nmf<ScalarType, 1>::program_name(),
NMF_MUL_DIV_KERNEL);
viennacl::ocl::enqueue(mul_div_kernel(w, wn, wd, cl_uint(stl_w.size())));
}
if (i % check_diff_every_step == 0)
{
appr = viennacl::linalg::prod(w, h);
viennacl::ocl::kernel & sub_kernel = viennacl::ocl::get_kernel(viennacl::linalg::kernels::nmf<ScalarType, 1>::program_name(),
NMF_SUB_KERNEL);
//this is a cheat. i.e save difference of two matrix into vector to get norm_2
viennacl::ocl::enqueue(sub_kernel(appr, v, diff, cl_uint(v.size1() * v.size2())));
ScalarType diff_val = viennacl::linalg::norm_2(diff);
if((diff_val < eps) || (fabs(diff_val - last_diff) < eps))
{
//std::cout << "Breaked at diff - " << diff_val << "\n";
break;
}
last_diff = diff_val;
//printf("Iteration #%lu - %.5f \n", i, diff_val);
}
}
}
viennacl::vector_range<viennacl::vector_base<T> > sharedCol(){
// viennacl::vector_base<T> tmp(ptr_matrix->handle(), ptr_matrix->internal_size(), 0, 1);
// std::cout << "returning column" << std::endl;
viennacl::vector_base<T> tmp(ptr_matrix->handle(), ptr_matrix->size1(), begin, ptr_matrix->internal_size2());
// std::cout << "got column" << std::endl;
viennacl::vector_range<viennacl::vector_base<T> > v_sub(tmp, r);
// std::cout << "got range" << std::endl;
return v_sub;
}