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);
}
void matrix_print(viennacl::matrix<ScalarType, MatrixLayout>& A)
{
for (unsigned int i = 0; i < A.size1(); i++) {
for (unsigned int j = 0; j < A.size2(); j++)
std::cout << std::fixed << A(i, j) << "\t";
std::cout << "\n";
}
}
void fill_matrix(viennacl::matrix<NumericT> & mat)
{
for (std::size_t i = 0; i < mat.size1(); ++i)
{
for (std::size_t j = 0; j < mat.size2(); ++j)
mat(i, j) = static_cast<NumericT>(-0.5) * random<NumericT>();
mat(i, i) = NumericT(1.0) + NumericT(2.0) * random<NumericT>(); //some extra weight on diagonal for stability
}
}
void matrix_print(viennacl::matrix<ScalarType>& A_orig)
{
ublas::matrix<ScalarType> A(A_orig.size1(), A_orig.size2());
viennacl::copy(A_orig, A);
std::cout << "matrix_print!\n";
for (unsigned int i = 0; i < A.size1(); i++) {
for (unsigned int j = 0; j < A.size2(); j++)
std::cout << A(i, j) << "\t";
std::cout << "\n";
}
}
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;
}
bool test_eigen_val_vec(viennacl::matrix<ScalarType, MatrixLayout> & Q,
std::vector<ScalarType> & eigenvalues,
std::vector<ScalarType> & d,
std::vector<ScalarType> & e)
{
unsigned int Q_size = Q.size2();
ScalarType value = 0;
for(unsigned int j = 0; j < Q_size; j++)
{
// calculate first row
value = (d[0]- eigenvalues[j]) * Q(0, j) + e[1] * Q(1, j);
if(value > EPS)
return false;
// calcuate inner rows
for(unsigned int i = 1; i < Q_size - 1; i++)
{
value = e[i] * Q(i - 1, j) + (d[i]- eigenvalues[j]) * Q(i, j) + e[i + 1] * Q(i + 1, j);
if(value > EPS)
return false;
}
// calculate last row
value = e[Q_size - 1] * Q(Q_size - 2, j) + (d[Q_size - 1] - eigenvalues[j]) * Q(Q_size - 1, j);
if(value > EPS)
return false;
}
return true;
}
bool check_tridiag(viennacl::matrix<ScalarType, MatrixLayout>& A_orig)
{
ublas::matrix<ScalarType> A(A_orig.size1(), A_orig.size2());
viennacl::copy(A_orig, A);
for (unsigned int i = 0; i < A.size1(); i++) {
for (unsigned int j = 0; j < A.size2(); j++) {
if ((std::abs(A(i, j)) > EPS) && ((i - 1) != j) && (i != j) && ((i + 1) != j))
{
// std::cout << "Failed at " << i << " " << j << " " << A(i, j) << "\n";
return false;
}
}
}
return true;
}
bool check_hessenberg(viennacl::matrix<ScalarType, MatrixLayout>& A_orig)
{
ublas::matrix<ScalarType> A(A_orig.size1(), A_orig.size2());
viennacl::copy(A_orig, A);
for (std::size_t i = 0; i < A.size1(); i++) {
for (std::size_t j = 0; j < A.size2(); j++) {
if ((std::abs(A(i, j)) > EPS) && (i > (j + 1)))
{
// std::cout << "Failed at " << i << " " << j << " " << A(i, j) << "\n";
return false;
}
}
}
return true;
}
float matrix_compare(viennacl::matrix<ScalarType>& res,
viennacl::matrix<ScalarType>& ref)
{
std::vector<ScalarType> res_std(res.internal_size());
std::vector<ScalarType> ref_std(ref.internal_size());
viennacl::fast_copy(res, &res_std[0]);
viennacl::fast_copy(ref, &ref_std[0]);
float diff = 0.0;
float mx = 0.0;
for(std::size_t i = 0; i < res_std.size(); i++) {
diff = std::max(diff, std::abs(res_std[i] - ref_std[i]));
mx = std::max(mx, res_std[i]);
}
return diff / mx;
}
ScalarType diff(ublas::matrix<ScalarType> & mat1, viennacl::matrix<ScalarType, F, ALIGNMENT> & mat2)
{
ublas::matrix<ScalarType> mat2_cpu(mat2.size1(), mat2.size2());
copy(mat2, mat2_cpu);
ScalarType ret = 0;
ScalarType act = 0;
for (unsigned int i = 0; i < mat2_cpu.size1(); ++i)
{
for (unsigned int j = 0; j < mat2_cpu.size2(); ++j)
{
act = fabs(mat2_cpu(i,j) - mat1(i,j)) / std::max( fabs(mat2_cpu(i, j)), fabs(mat1(i,j)) );
if (act > ret)
ret = act;
}
}
//std::cout << ret << std::endl;
return ret;
}
void read_matrix_body(std::fstream& f, viennacl::matrix<ScalarType, MatrixLayout>& A)
{
if(!f.is_open())
{
throw std::invalid_argument("File is not opened");
}
boost::numeric::ublas::matrix<ScalarType> h_A(A.size1(), A.size2());
for(std::size_t i = 0; i < h_A.size1(); i++) {
for(std::size_t j = 0; j < h_A.size2(); j++) {
ScalarType val = 0.0;
f >> val;
h_A(i, j) = val;
}
}
viennacl::copy(h_A, A);
}
FeatureNormalize<Scalar>::FeatureNormalize(const viennacl::matrix<Scalar> &X):
blas_matrix_(X.size1(), X.size2()),
normalized_matrix_(X.size1(), X.size2()),
mu_(X.size2()), sigma_(X.size2()) {
viennacl::copy(X, blas_matrix_);
runNormalization();
}
bool check_bidiag(viennacl::matrix<ScalarType>& A)
{
const ScalarType EPS = 0.0001f;
std::vector<ScalarType> aA(A.size1() * A.size2());
viennacl::fast_copy(A, &aA[0]);
for(std::size_t i = 0; i < A.size1(); i++)
{
for(std::size_t j = 0; j < A.size2(); j++)
{
ScalarType val = aA[i * A.size2() + j];
if((fabs(val) > EPS) && (i != j) && ((i + 1) != j))
{
std::cout << "Failed at " << i << " " << j << " " << val << std::endl;
return false;
}
}
}
return true;
}
void svd(viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT> & A,
viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT> & QL,
viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT> & QR)
{
viennacl::ocl::context & ctx = const_cast<viennacl::ocl::context &>(viennacl::traits::opencl_handle(A).context());
viennacl::linalg::opencl::kernels::svd<SCALARTYPE>::init(ctx);
vcl_size_t row_num = A.size1();
vcl_size_t col_num = A.size2();
vcl_size_t to = std::min(row_num, col_num);
//viennacl::vector<SCALARTYPE, ALIGNMENT> d(to);
//viennacl::vector<SCALARTYPE, ALIGNMENT> s(to + 1);
// first stage
detail::bidiag(A, QL, QR);
// second stage
//std::vector<SCALARTYPE> dh(to, 0);
//std::vector<SCALARTYPE> sh(to + 1, 0);
boost::numeric::ublas::vector<SCALARTYPE> dh = boost::numeric::ublas::scalar_vector<SCALARTYPE>(to, 0);
boost::numeric::ublas::vector<SCALARTYPE> sh = boost::numeric::ublas::scalar_vector<SCALARTYPE>(to + 1, 0);
viennacl::linalg::opencl::bidiag_pack_svd(A, dh, sh);
detail::svd_qr_shift( QL, QR, dh, sh);
// Write resulting diagonal matrix with singular values to A:
boost::numeric::ublas::matrix<SCALARTYPE> h_Sigma(row_num, col_num);
h_Sigma.clear();
for (vcl_size_t i = 0; i < to; i++)
h_Sigma(i, i) = dh[i];
copy(h_Sigma, A);
}
void bidiag(viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT> & Ai,
viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT> & QL,
viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT> & QR)
{
vcl_size_t row_num = Ai.size1();
vcl_size_t col_num = Ai.size2();
vcl_size_t to = std::min(row_num, col_num);
vcl_size_t big_to = std::max(row_num, col_num);
//for storing householder vector
viennacl::vector<SCALARTYPE, ALIGNMENT> hh_vector(big_to, viennacl::traits::context(Ai));
QL = viennacl::identity_matrix<SCALARTYPE>(QL.size1(), viennacl::traits::context(QL));
QR = viennacl::identity_matrix<SCALARTYPE>(QR.size1(), viennacl::traits::context(QR));
for (vcl_size_t i = 0; i < to; i++)
{
householder_c(Ai, QL, hh_vector, i, i);
householder_r(Ai, QR, hh_vector, i, i+1);
}
}
bool householder_r(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 (col_start + 1 >= A.size2())
return false;
prepare_householder_vector(A, D, A.size2(), row_start, col_start, col_start, false);
{
viennacl::ocl::kernel& kernel = ctx.get_kernel(viennacl::linalg::opencl::kernels::svd<SCALARTYPE>::program_name(), SVD_HOUSEHOLDER_UPDATE_A_RIGHT_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_QR_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);
}
static size_t size1(viennacl::matrix<ScalarType, F1, A1> const & lhs,
viennacl::matrix_expression<const viennacl::matrix<ScalarType, F2, A2>,
const viennacl::matrix<ScalarType, F2, A2>,
op_trans> const & rhs) { return lhs.size1(); }
void fill_random(viennacl::matrix<ScalarType, MajorT> & A)
{
for (std::size_t i = 0; i < A.size1(); i++)
for (std::size_t j = 0; j < A.size2(); ++j)
A(i, j) = static_cast<ScalarType>(rand()) / ScalarType(RAND_MAX);
}
viennacl::vector_range<viennacl::vector_base<T> > sharedVector(){
viennacl::vector_base<T> tmp(ptr_matrix->handle(), ptr_matrix->internal_size(), 0, 1);
viennacl::vector_range<viennacl::vector_base<T> > v_sub(tmp, r);
return v_sub;
}
viennacl::vector_range<viennacl::vector_base<T> > sharedRow(){
// viennacl::vector_base<T> tmp(ptr_matrix->handle(), ptr_matrix->internal_size(), 0, 1);
viennacl::vector_base<T> tmp(ptr_matrix->handle(), ptr_matrix->size2(), begin, 1);
viennacl::vector_range<viennacl::vector_base<T> > v_sub(tmp, r);
return v_sub;
}
static size_t size(const viennacl::matrix<ScalarType, F, Amat> & lhs,
const viennacl::vector<ScalarType, A> & rhs) { return lhs.size1(); }
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);
}
}
}