typename viennacl::enable_if< viennacl::is_any_sparse_matrix<SparseMatrixType>::value>::type
prod_impl(const SparseMatrixType & sp_mat,
const viennacl::matrix_base<ScalarType> & d_mat,
viennacl::matrix_base<ScalarType> & result)
{
assert( (sp_mat.size1() == result.size1()) && bool("Size check failed for compressed matrix - dense matrix product: size1(sp_mat) != size1(result)"));
assert( (sp_mat.size2() == d_mat.size1()) && bool("Size check failed for compressed matrix - dense matrix product: size2(sp_mat) != size1(d_mat)"));
switch (viennacl::traits::handle(sp_mat).get_active_handle_id())
{
case viennacl::MAIN_MEMORY:
viennacl::linalg::host_based::prod_impl(sp_mat, d_mat, result);
break;
#ifdef VIENNACL_WITH_OPENCL
case viennacl::OPENCL_MEMORY:
viennacl::linalg::opencl::prod_impl(sp_mat, d_mat, result);
break;
#endif
#ifdef VIENNACL_WITH_CUDA
case viennacl::CUDA_MEMORY:
viennacl::linalg::cuda::prod_impl(sp_mat, d_mat, result);
break;
#endif
case viennacl::MEMORY_NOT_INITIALIZED:
throw memory_exception("not initialised!");
default:
throw memory_exception("not implemented");
}
}
typename viennacl::enable_if< viennacl::is_any_sparse_matrix<SparseMatrixType>::value>::type
assign_to_dense(SparseMatrixType const & A,
viennacl::matrix_base<NumericT> & B)
{
assert( (A.size1() == B.size1()) && bool("Size check failed for assignment to dense matrix: size1(A) != size1(B)"));
assert( (A.size2() == B.size1()) && bool("Size check failed for assignment to dense matrix: size2(A) != size2(B)"));
switch (viennacl::traits::handle(A).get_active_handle_id())
{
case viennacl::MAIN_MEMORY:
viennacl::linalg::host_based::amg::assign_to_dense(A, B);
break;
#ifdef VIENNACL_WITH_OPENCL
case viennacl::OPENCL_MEMORY:
viennacl::linalg::opencl::amg::assign_to_dense(A, B);
break;
#endif
#ifdef VIENNACL_WITH_CUDA
case viennacl::CUDA_MEMORY:
viennacl::linalg::cuda::amg::assign_to_dense(A, B);
break;
#endif
case viennacl::MEMORY_NOT_INITIALIZED:
throw memory_exception("not initialised!");
default:
throw memory_exception("not implemented");
}
}
void nmf(viennacl::matrix_base<ScalarType> const & V, viennacl::matrix_base<ScalarType> & W,
viennacl::matrix_base<ScalarType> & H, viennacl::linalg::nmf_config const & conf)
{
assert(V.size1() == W.size1() && V.size2() == H.size2() && bool("Dimensions of W and H don't allow for V = W * H"));
assert(W.size2() == H.size1() && bool("Dimensions of W and H don't match, prod(W, H) impossible"));
switch (viennacl::traits::handle(V).get_active_handle_id())
{
case viennacl::MAIN_MEMORY:
viennacl::linalg::host_based::nmf(V, W, H, conf);
break;
#ifdef VIENNACL_WITH_OPENCL
case viennacl::OPENCL_MEMORY:
viennacl::linalg::opencl::nmf(V,W,H,conf);
break;
#endif
#ifdef VIENNACL_WITH_CUDA
case viennacl::CUDA_MEMORY:
viennacl::linalg::cuda::nmf(V,W,H,conf);
break;
#endif
case viennacl::MEMORY_NOT_INITIALIZED:
throw memory_exception("not initialised!");
default:
throw memory_exception("not implemented");
}
}
void nmf(viennacl::matrix_base<NumericT> const & V,
viennacl::matrix_base<NumericT> & W,
viennacl::matrix_base<NumericT> & H,
viennacl::linalg::nmf_config const & conf)
{
viennacl::hsa::context & ctx = const_cast<viennacl::hsa::context &>(viennacl::traits::hsa_context(V));
const std::string NMF_MUL_DIV_KERNEL = "el_wise_mul_div";
viennacl::linalg::opencl::kernels::nmf<NumericT, viennacl::hsa::context>::init(ctx);
vcl_size_t k = W.size2();
conf.iters_ = 0;
if (viennacl::linalg::norm_frobenius(W) <= 0)
W = viennacl::scalar_matrix<NumericT>(W.size1(), W.size2(), NumericT(1), ctx);
if (viennacl::linalg::norm_frobenius(H) <= 0)
H = viennacl::scalar_matrix<NumericT>(H.size1(), H.size2(), NumericT(1), ctx);
viennacl::matrix_base<NumericT> wn(V.size1(), k, W.row_major(), ctx);
viennacl::matrix_base<NumericT> wd(V.size1(), k, W.row_major(), ctx);
viennacl::matrix_base<NumericT> wtmp(V.size1(), V.size2(), W.row_major(), ctx);
viennacl::matrix_base<NumericT> hn(k, V.size2(), H.row_major(), ctx);
viennacl::matrix_base<NumericT> hd(k, V.size2(), H.row_major(), ctx);
viennacl::matrix_base<NumericT> htmp(k, k, H.row_major(), ctx);
viennacl::matrix_base<NumericT> appr(V.size1(), V.size2(), V.row_major(), ctx);
NumericT last_diff = 0;
NumericT diff_init = 0;
bool stagnation_flag = false;
for (vcl_size_t i = 0; i < conf.max_iterations(); i++)
{
conf.iters_ = i + 1;
{
hn = viennacl::linalg::prod(trans(W), V);
htmp = viennacl::linalg::prod(trans(W), W);
hd = viennacl::linalg::prod(htmp, H);
viennacl::hsa::kernel & mul_div_kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::nmf<NumericT>::program_name(), NMF_MUL_DIV_KERNEL);
viennacl::hsa::enqueue(mul_div_kernel(H, hn, hd, cl_uint(H.internal_size1() * H.internal_size2())));
}
{
wn = viennacl::linalg::prod(V, trans(H));
wtmp = viennacl::linalg::prod(W, H);
wd = viennacl::linalg::prod(wtmp, trans(H));
viennacl::hsa::kernel & mul_div_kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::nmf<NumericT>::program_name(), NMF_MUL_DIV_KERNEL);
viennacl::hsa::enqueue(mul_div_kernel(W, wn, wd, cl_uint(W.internal_size1() * W.internal_size2())));
}
if (i % conf.check_after_steps() == 0) //check for convergence
{
appr = viennacl::linalg::prod(W, H);
appr -= V;
NumericT diff_val = viennacl::linalg::norm_frobenius(appr);
if (i == 0)
diff_init = diff_val;
if (conf.print_relative_error())
std::cout << diff_val / diff_init << std::endl;
// Approximation check
if (diff_val / diff_init < conf.tolerance())
break;
// Stagnation check
if (std::fabs(diff_val - last_diff) / (diff_val * NumericT(conf.check_after_steps())) < conf.stagnation_tolerance()) //avoid situations where convergence stagnates
{
if (stagnation_flag) // iteration stagnates (two iterates with no notable progress)
break;
else
// record stagnation in this iteration
stagnation_flag = true;
} else
// good progress in this iteration, so unset stagnation flag
stagnation_flag = false;
// prepare for next iterate:
last_diff = diff_val;
}
}
}
static vcl_size_t size2(viennacl::matrix_base<T> const & lhs,
ScalarType const & /*rhs*/) { return lhs.size2(); }
static vcl_size_t size1(viennacl::matrix_base<ScalarType> const & lhs,
viennacl::matrix_expression<T2,
T2,
op_trans> const & /*rhs*/) { return lhs.size1(); }
static vcl_size_t size2(viennacl::matrix_expression<T1,
T1,
op_trans> const & /*lhs*/,
viennacl::matrix_base<ScalarType> const & rhs) { return rhs.size2(); }
static std::size_t size1(viennacl::matrix_base<T, F> const & lhs,
ScalarType const & /*rhs*/) { return lhs.size1(); }
static vcl_size_t size2(ScalarType const & /*lhs*/, viennacl::matrix_base<ScalarType> const & rhs) { return rhs.size2(); }