void init_random(viennacl::vector<T> & x)
{
std::vector<T> cx(x.internal_size());
for (std::size_t i = 0; i < cx.size(); ++i)
cx[i] = T(rand())/T(RAND_MAX);
viennacl::fast_copy(&cx[0], &cx[0] + cx.size(), x.begin());
}
ScalarType diff_2(ublas::vector<ScalarType> & v1, viennacl::vector<ScalarType> & v2)
{
ublas::vector<ScalarType> v2_cpu(v2.size());
viennacl::copy(v2.begin(), v2.end(), v2_cpu.begin());
return norm_2(v1 - v2_cpu) / norm_2(v1);
}
NumericT diff(std::vector<NumericT> const & v1, viennacl::vector<NumericT> const & v2)
{
std::vector<NumericT> v2_cpu(v2.size());
viennacl::backend::finish();
viennacl::copy(v2.begin(), v2.end(), v2_cpu.begin());
for (std::size_t i=0;i<v1.size(); ++i)
{
if ( std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) ) > 0 )
v2_cpu[i] = std::fabs(v2_cpu[i] - v1[i]) / std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) );
else
v2_cpu[i] = 0.0;
if (v2_cpu[i] > 0.0001)
{
//std::cout << "Neighbor: " << i-1 << ": " << v1[i-1] << " vs. " << v2_cpu[i-1] << std::endl;
std::cout << "Error at entry " << i << ": " << v1[i] << " vs. " << v2[i] << std::endl;
//std::cout << "Neighbor: " << i+1 << ": " << v1[i+1] << " vs. " << v2_cpu[i+1] << std::endl;
exit(EXIT_FAILURE);
}
}
NumericT inf_norm = 0;
for (std::size_t i=0;i<v2_cpu.size(); ++i)
inf_norm = std::max<NumericT>(inf_norm, std::fabs(v2_cpu[i]));
return inf_norm;
}
ScalarType diff(ublas::vector<ScalarType> & v1, viennacl::vector<ScalarType> & v2)
{
ublas::vector<ScalarType> v2_cpu(v2.size());
viennacl::backend::finish();
viennacl::copy(v2.begin(), v2.end(), v2_cpu.begin());
for (unsigned int i=0;i<v1.size(); ++i)
{
if ( std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) ) > 0 )
{
//if (std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) ) < 1e-10 ) //absolute tolerance (avoid round-off issues)
// v2_cpu[i] = 0;
//else
v2_cpu[i] = std::fabs(v2_cpu[i] - v1[i]) / std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) );
}
else
v2_cpu[i] = 0.0;
if (v2_cpu[i] > 0.0001)
{
//std::cout << "Neighbor: " << i-1 << ": " << v1[i-1] << " vs. " << v2_cpu[i-1] << std::endl;
std::cout << "Error at entry " << i << ": " << v1[i] << " vs. " << v2_cpu[i] << std::endl;
//std::cout << "Neighbor: " << i+1 << ": " << v1[i+1] << " vs. " << v2_cpu[i+1] << std::endl;
exit(EXIT_FAILURE);
}
}
return norm_inf(v2_cpu);
}
void apply(viennacl::vector<ScalarType, ALIGNMENT> & vec) const
{
assert(system_matrix.size1() == vec.size());
//run kernel:
viennacl::ocl::kernel & k = viennacl::ocl::get_kernel(viennacl::linalg::kernels::vector<ScalarType, ALIGNMENT>::program_name(),
"diag_precond");
viennacl::ocl::enqueue( k(diag_A_inv, vec, static_cast<cl_uint>(vec.size())) );
}
void prod_impl(const viennacl::hankel_matrix<SCALARTYPE, ALIGNMENT> & mat,
const viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> & vec,
viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> & result)
{
assert(mat.size1() == result.size());
assert(mat.size2() == vec.size());
prod_impl(mat.elements(), vec, result);
viennacl::detail::fft::reverse(result);
}
ScalarType diff ( ublas::vector<ScalarType> & v1, viennacl::vector<ScalarType,Alignment> & v2 ) {
ublas::vector<ScalarType> v2_cpu ( v2.size() );
viennacl::copy( v2.begin(), v2.end(), v2_cpu.begin() );
for ( unsigned int i=0; i<v1.size(); ++i ) {
if ( std::max ( fabs ( v2_cpu[i] ), fabs ( v1[i] ) ) > 0 )
v2_cpu[i] = fabs ( v2_cpu[i] - v1[i] ) / std::max ( fabs ( v2_cpu[i] ), fabs ( v1[i] ) );
else
v2_cpu[i] = 0.0;
}
return norm_inf ( v2_cpu );
}
void prod_impl(const viennacl::compressed_matrix<TYPE, ALIGNMENT> & mat,
const viennacl::vector<TYPE, VECTOR_ALIGNMENT> & vec,
viennacl::vector<TYPE, VECTOR_ALIGNMENT> & result,
size_t NUM_THREADS = 0)
{
assert(mat.size1() == result.size());
assert(mat.size2() == vec.size());
viennacl::ocl::kernel & k = viennacl::ocl::get_kernel(viennacl::linalg::kernels::compressed_matrix<TYPE, ALIGNMENT>::program_name(), "vec_mul");
viennacl::ocl::enqueue(k(mat.handle1(), mat.handle2(), mat.handle(), vec, result, static_cast<cl_uint>(mat.size1())));
}
bool
bisect(const viennacl::vector<NumericT> & diagonal, const viennacl::vector<NumericT> & superdiagonal, viennacl::vector<NumericT> & eigenvalues)
{
assert(diagonal.size() == superdiagonal.size() &&
diagonal.size() == eigenvalues.size() &&
bool("Input vectors do not have the same sizes!"));
bool bResult = false;
// flag if the matrix size is due to explicit user request
// desired precision of eigenvalues
NumericT precision = static_cast<NumericT>(0.00001);
const unsigned int mat_size = static_cast<unsigned int>(diagonal.size());
// set up input
viennacl::linalg::detail::InputData<NumericT> input(diagonal, superdiagonal, mat_size);
NumericT lg = FLT_MAX;
NumericT ug = -FLT_MAX;
// compute Gerschgorin interval
viennacl::linalg::detail::computeGerschgorin(input.std_a, input.std_b, mat_size, lg, ug);
// decide wheter the algorithm for small or for large matrices will be started
if (mat_size <= VIENNACL_BISECT_MAX_SMALL_MATRIX)
{
// initialize memory for result
viennacl::linalg::detail::ResultDataSmall<NumericT> result(mat_size);
// run the kernel
viennacl::linalg::detail::computeEigenvaluesSmallMatrix(input, result, mat_size, lg, ug, precision);
// get the result from the device and do some sanity checks,
viennacl::linalg::detail::processResultSmallMatrix(result, mat_size);
copy(result.std_eigenvalues, eigenvalues);
bResult = true;
}
else
{
// initialize memory for result
viennacl::linalg::detail::ResultDataLarge<NumericT> result(mat_size);
// run the kernel
viennacl::linalg::detail::computeEigenvaluesLargeMatrix(input, result, mat_size, lg, ug, precision);
// get the result from the device and do some sanity checks
bResult = viennacl::linalg::detail::processResultDataLargeMatrix(result, mat_size);
copy(result.std_eigenvalues, eigenvalues);
}
return bResult;
}
static void test_scan_values(viennacl::vector<ScalarType> const & input,
viennacl::vector<ScalarType> & result,
bool is_inclusive_scan)
{
std::vector<ScalarType> host_input(input.size());
std::vector<ScalarType> host_result(result.size());
viennacl::copy(input, host_input);
viennacl::copy(result, host_result);
ScalarType sum = 0;
if (is_inclusive_scan)
{
for(viennacl::vcl_size_t i = 0; i < input.size(); i++)
{
sum += host_input[i];
host_input[i] = sum;
}
}
else
{
for(viennacl::vcl_size_t i = 0; i < input.size(); i++)
{
ScalarType tmp = host_input[i];
host_input[i] = sum;
sum += tmp;
}
}
for(viennacl::vcl_size_t i = 0; i < input.size(); i++)
{
if (host_input[i] != host_result[i])
{
std::cout << "Fail at vector index " << i << std::endl;
std::cout << " result[" << i << "] = " << host_result[i] << std::endl;
std::cout << " Reference = " << host_input[i] << std::endl;
if (i > 0)
{
std::cout << " previous result[" << i-1 << "] = " << host_result[i-1] << std::endl;
std::cout << " previous Reference = " << host_input[i-1] << std::endl;
}
exit(EXIT_FAILURE);
}
}
std::cout << "PASSED!" << std::endl;
}
void split_calc_distance(std::vector<double>& to_sort,viennacl::ocl::context* p_context, int num_splits, naive_knn& knn, viennacl::vector<double>& distances, dense_sliding_window& sliding_window, int num_instances, viennacl::vector<double>& sample)
{
int len = num_instances / num_splits;
auto gpu_begin = distances.begin();
auto gpu_end = gpu_begin + len;
int last = num_instances - len * num_splits;
int current = 0;
knn.calc_distance(distances, sliding_window, current, current+len, sample);
current += len;
for (; current < num_instances; current += len)
{
p_context->get_queue().finish();
viennacl::copy(gpu_begin, gpu_end, to_sort.begin());
knn.calc_distance(distances, sliding_window, current, current+len, sample);
std::sort(to_sort.begin(), to_sort.end());
}
p_context->get_queue().finish();
viennacl::copy(gpu_begin, gpu_end, to_sort.begin());
std::sort(to_sort.begin(), to_sort.end());
if (last > 0)
{
//knn.calc_distance(distances, sliding_window, current -len, current + last, sample);
}
}
static void init_vector(viennacl::vector<ScalarType>& vcl_v)
{
std::vector<ScalarType> v(vcl_v.size());
for (std::size_t i = 0; i < v.size(); ++i)
v[i] = ScalarType(i % 7 + 1);
viennacl::copy(v, vcl_v);
}
ScalarType diff(ublas::vector<ScalarType> & v1, viennacl::vector<ScalarType> & v2)
{
ublas::vector<ScalarType> v2_cpu(v2.size());
viennacl::backend::finish(); //workaround for a bug in APP SDK 2.7 on Trinity APUs (with Catalyst 12.8)
viennacl::copy(v2.begin(), v2.end(), v2_cpu.begin());
for (std::size_t i=0;i<v1.size(); ++i)
{
if ( std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) ) > 0 )
v2_cpu[i] = std::fabs(v2_cpu[i] - v1[i]) / std::max( std::fabs(v2_cpu[i]), std::fabs(v1[i]) );
else
v2_cpu[i] = 0.0;
}
return norm_inf(v2_cpu);
}
void prepare_householder_vector(
viennacl::matrix<SCALARTYPE, row_major, ALIGNMENT>& A,
viennacl::vector<SCALARTYPE, ALIGNMENT>& D,
vcl_size_t size,
vcl_size_t row_start,
vcl_size_t col_start,
vcl_size_t start,
bool is_column
)
{
boost::numeric::ublas::vector<SCALARTYPE> tmp = boost::numeric::ublas::scalar_vector<SCALARTYPE>(size, 0);
copy_vec(A, D, row_start, col_start, is_column);
fast_copy(D.begin(), D.begin() + vcl_ptrdiff_t(size - start), tmp.begin() + start);
//std::cout << "1: " << tmp << "\n";
detail::householder_vector(tmp, start);
fast_copy(tmp, D);
//std::cout << "2: " << D << "\n";
}
void prod_impl(const viennacl::toeplitz_matrix<SCALARTYPE, ALIGNMENT> & mat,
const viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> & vec,
viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> & result)
{
assert(mat.size1() == result.size());
assert(mat.size2() == vec.size());
viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> tmp(vec.size() * 4);
tmp.clear();
viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> tmp2(vec.size() * 4);
viennacl::vector<SCALARTYPE, VECTOR_ALIGNMENT> tep(mat.elements().size() * 2);
viennacl::detail::fft::real_to_complex(mat.elements(), tep, mat.elements().size());
copy(vec, tmp);
viennacl::detail::fft::real_to_complex(tmp, tmp2, vec.size() * 2);
viennacl::linalg::convolve(tep, tmp2, tmp);
viennacl::detail::fft::complex_to_real(tmp, tmp2, vec.size() * 2);
copy(tmp2.begin(), tmp2.begin() + vec.size(), result.begin());
}
void apply(viennacl::vector<NumericT> & vec) const
{
if (vec.handle().get_active_handle_id() != viennacl::MAIN_MEMORY)
{
if (tag_.use_level_scheduling())
{
//std::cout << "Using multifrontal on GPU..." << std::endl;
detail::level_scheduling_substitute(vec,
multifrontal_L_row_index_arrays_,
multifrontal_L_row_buffers_,
multifrontal_L_col_buffers_,
multifrontal_L_element_buffers_,
multifrontal_L_row_elimination_num_list_);
vec = viennacl::linalg::element_div(vec, multifrontal_U_diagonal_);
detail::level_scheduling_substitute(vec,
multifrontal_U_row_index_arrays_,
multifrontal_U_row_buffers_,
multifrontal_U_col_buffers_,
multifrontal_U_element_buffers_,
multifrontal_U_row_elimination_num_list_);
}
else
{
viennacl::context host_context(viennacl::MAIN_MEMORY);
viennacl::context old_context = viennacl::traits::context(vec);
viennacl::switch_memory_context(vec, host_context);
viennacl::linalg::inplace_solve(LU_, vec, unit_lower_tag());
viennacl::linalg::inplace_solve(LU_, vec, upper_tag());
viennacl::switch_memory_context(vec, old_context);
}
}
else //apply ILUT directly:
{
viennacl::linalg::inplace_solve(LU_, vec, unit_lower_tag());
viennacl::linalg::inplace_solve(LU_, vec, upper_tag());
}
}
static void apply(::viennacl::vector<V> &x)
{
x.clear();
}
static bool same_size( const viennacl::vector< T > &x1 , const viennacl::vector< T > &x2 )
{
return x1.size() == x2.size();
}
static void resize( viennacl::vector< T > &x1 , const viennacl::vector< T > &x2 )
{
x1.resize( x2.size() , false );
}
void fill_vector(viennacl::vector<NumericT> & vec)
{
for (std::size_t i = 0; i < vec.size(); ++i)
vec(i) = NumericT(1.0) + NumericT(2.0) * random<NumericT>(); //some extra weight on diagonal for stability
}
void clear(viennacl::vector<ScalarType, AlignmentV> & vec)
{
vec.clear();
}
void clear(viennacl::vector<ScalarType, ALIGNMENT> & vec)
{
vec.clear();
}
viennacl::vector<ScalarType> solve(MatrixType const & A, //MatrixType const & A,
viennacl::vector<ScalarType> const & rhs,
bicgstab_tag const & tag,
viennacl::linalg::no_precond)
{
viennacl::vector<ScalarType> result = viennacl::zero_vector<ScalarType>(rhs.size(), viennacl::traits::context(rhs));
viennacl::vector<ScalarType> residual = rhs;
viennacl::vector<ScalarType> p = rhs;
viennacl::vector<ScalarType> r0star = rhs;
viennacl::vector<ScalarType> Ap = rhs;
viennacl::vector<ScalarType> s = rhs;
viennacl::vector<ScalarType> As = rhs;
// Layout of temporary buffer:
// chunk 0: <residual, r_0^*>
// chunk 1: <As, As>
// chunk 2: <As, s>
// chunk 3: <Ap, r_0^*>
// chunk 4: <As, r_0^*>
// chunk 5: <s, s>
vcl_size_t buffer_size_per_vector = 256;
vcl_size_t num_buffer_chunks = 6;
viennacl::vector<ScalarType> inner_prod_buffer = viennacl::zero_vector<ScalarType>(num_buffer_chunks*buffer_size_per_vector, viennacl::traits::context(rhs)); // temporary buffer
std::vector<ScalarType> host_inner_prod_buffer(inner_prod_buffer.size());
ScalarType norm_rhs_host = viennacl::linalg::norm_2(residual);
ScalarType beta;
ScalarType alpha;
ScalarType omega;
ScalarType residual_norm = norm_rhs_host;
inner_prod_buffer[0] = norm_rhs_host * norm_rhs_host;
ScalarType r_dot_r0 = 0;
ScalarType As_dot_As = 0;
ScalarType As_dot_s = 0;
ScalarType Ap_dot_r0 = 0;
ScalarType As_dot_r0 = 0;
ScalarType s_dot_s = 0;
if (norm_rhs_host <= 0) //solution is zero if RHS norm is zero
return result;
for (vcl_size_t i = 0; i < tag.max_iterations(); ++i)
{
tag.iters(i+1);
// Ap = A*p_j
// Ap_dot_r0 = <Ap, r_0^*>
viennacl::linalg::pipelined_bicgstab_prod(A, p, Ap, r0star,
inner_prod_buffer, buffer_size_per_vector, 3*buffer_size_per_vector);
//////// first (weak) synchronization point ////
///// method 1: compute alpha on host:
//
//// we only need the second chunk of the buffer for computing Ap_dot_r0:
//viennacl::fast_copy(inner_prod_buffer.begin(), inner_prod_buffer.end(), host_inner_prod_buffer.begin());
//Ap_dot_r0 = std::accumulate(host_inner_prod_buffer.begin() + buffer_size_per_vector, host_inner_prod_buffer.begin() + 2 * buffer_size_per_vector, ScalarType(0));
//alpha = residual_dot_r0 / Ap_dot_r0;
//// s_j = r_j - alpha_j q_j
//s = residual - alpha * Ap;
///// method 2: compute alpha on device:
// s = r - alpha * Ap
// <s, s> first stage
// dump alpha at end of inner_prod_buffer
viennacl::linalg::pipelined_bicgstab_update_s(s, residual, Ap,
inner_prod_buffer, buffer_size_per_vector, 5*buffer_size_per_vector);
// As = A*s_j
// As_dot_As = <As, As>
// As_dot_s = <As, s>
// As_dot_r0 = <As, r_0^*>
viennacl::linalg::pipelined_bicgstab_prod(A, s, As, r0star,
inner_prod_buffer, buffer_size_per_vector, 4*buffer_size_per_vector);
//////// second (strong) synchronization point ////
viennacl::fast_copy(inner_prod_buffer.begin(), inner_prod_buffer.end(), host_inner_prod_buffer.begin());
r_dot_r0 = std::accumulate(host_inner_prod_buffer.begin(), host_inner_prod_buffer.begin() + buffer_size_per_vector, ScalarType(0));
As_dot_As = std::accumulate(host_inner_prod_buffer.begin() + buffer_size_per_vector, host_inner_prod_buffer.begin() + 2 * buffer_size_per_vector, ScalarType(0));
As_dot_s = std::accumulate(host_inner_prod_buffer.begin() + 2 * buffer_size_per_vector, host_inner_prod_buffer.begin() + 3 * buffer_size_per_vector, ScalarType(0));
Ap_dot_r0 = std::accumulate(host_inner_prod_buffer.begin() + 3 * buffer_size_per_vector, host_inner_prod_buffer.begin() + 4 * buffer_size_per_vector, ScalarType(0));
As_dot_r0 = std::accumulate(host_inner_prod_buffer.begin() + 4 * buffer_size_per_vector, host_inner_prod_buffer.begin() + 5 * buffer_size_per_vector, ScalarType(0));
s_dot_s = std::accumulate(host_inner_prod_buffer.begin() + 5 * buffer_size_per_vector, host_inner_prod_buffer.begin() + 6 * buffer_size_per_vector, ScalarType(0));
alpha = r_dot_r0 / Ap_dot_r0;
beta = -1.0 * As_dot_r0 / Ap_dot_r0;
omega = As_dot_s / As_dot_As;
residual_norm = std::sqrt(s_dot_s - 2.0 * omega * As_dot_s + omega * omega * As_dot_As);
if (std::fabs(residual_norm / norm_rhs_host) < tag.tolerance())
break;
// x_{j+1} = x_j + alpha * p_j + omega * s_j
// r_{j+1} = s_j - omega * t_j
// p_{j+1} = r_{j+1} + beta * (p_j - omega * q_j)
// and compute first stage of r_dot_r0 = <r_{j+1}, r_o^*> for use in next iteration
viennacl::linalg::pipelined_bicgstab_vector_update(result, alpha, p, omega, s,
residual, As,
beta, Ap,
r0star, inner_prod_buffer, buffer_size_per_vector);
}
//store last error estimate:
tag.error(residual_norm / norm_rhs_host);
return result;
}
static size_t size(const viennacl::vector<ScalarType, A> & lhs,
const RHS & rhs) { return lhs.size(); }
static size_t size2(viennacl::vector<ScalarType, A1> & lhs,
viennacl::vector<ScalarType, A2> & rhs) { return rhs.size2(); }
viennacl::vector<unsigned int> bucket_select(int N, const viennacl::vector<basic_type>& in)
{
viennacl::vector<unsigned int> src(in.size(), viennacl::traits::context(in));
viennacl::vector<unsigned int> dst(in.size() , viennacl::traits::context(in));
// load kernels
static bool init = false;
static int num_groups = 1;
static int wg_size = 128;
if (!init)
{
FILE * tmp = fopen("bucket_select.cl", "rb");
fseek(tmp, 0, SEEK_END);
std::vector<char> binary;
binary.resize(ftell(tmp));
rewind(tmp);
fread(&binary[0], binary.size(), 1, tmp);
fclose(tmp);
binary.push_back(0);
static viennacl::context g_context = viennacl::ocl::current_context();
static bool init = false;
viennacl::ocl::context* ctx = g_context.opencl_pcontext();
std::cout << "Device " << ctx->current_device().name() << std::endl;
ctx->build_options("-cl-std=CL2.0 -D CL_VERSION_2_0");
std::string program_text(&binary[0]);
ctx->add_program(program_text, std::string("test"));
init = true;
}
viennacl::ocl::kernel scan_kernel = viennacl::ocl::current_context().get_kernel("test", "scan_buckets");
viennacl::ocl::kernel scatter_kernel = viennacl::ocl::current_context().get_kernel("test", "scatter_buckets");
viennacl::ocl::kernel init_offsets_kernel = viennacl::ocl::current_context().get_kernel("test", "init_offsets");
scan_kernel.local_work_size(0, wg_size);
scan_kernel.global_work_size(0, wg_size * num_groups);
scatter_kernel.local_work_size(0, wg_size);
scatter_kernel.global_work_size(0, wg_size * num_groups);
init_offsets_kernel.local_work_size(0, wg_size);
init_offsets_kernel.global_work_size(0, wg_size* num_groups);
cl_uint size = src.size();
viennacl::ocl::enqueue(init_offsets_kernel(size, src));
int position = 0;
viennacl::vector<unsigned int> result(N, viennacl::traits::context(in));
int num_buckets = 10;
viennacl::vector<unsigned int> global_histogram((num_buckets + 1) * num_groups, viennacl::traits::context(in)); // -wg_size
viennacl::vector<unsigned int> global_histogram_prefix((num_buckets + 1) * num_groups + 1, viennacl::traits::context(in));
std::vector< unsigned int > global_histogram_cpu((num_buckets + 1) * num_groups + 1);
int scan_start = 0;
int scan_end = in.size();
basic_type pivot;
basic_type base_value;
int split_bucket = 0;
base_value = 0;
pivot = std::numeric_limits<basic_type>::max() / num_buckets;
assert(pivot > 0);
while (position < N)
{
int main = (scan_end / wg_size) * wg_size; // floor to multiple wg size
int loop_end = main == scan_end ? main : main + wg_size; // add wg size if needed
viennacl::ocl::enqueue(scan_kernel(in,
src,
scan_end,
loop_end,
viennacl::ocl::local_mem(sizeof(cl_uint) *wg_size),
base_value,
pivot,
num_buckets,
global_histogram));
viennacl::linalg::exclusive_scan(global_histogram, global_histogram_prefix);
viennacl::copy(global_histogram_prefix, global_histogram_cpu);
global_histogram_cpu[global_histogram_cpu.size() - 1] = global_histogram_cpu[global_histogram_cpu.size() - 2]; // fix last element
for (split_bucket = 1; split_bucket < num_buckets; ++split_bucket)
{
int offset = global_histogram_cpu[num_groups * split_bucket];
if (offset >= N)
break;
}
viennacl::ocl::enqueue(scatter_kernel(
in,
src,
scan_end,
loop_end,
viennacl::ocl::local_mem(sizeof(cl_uint) *wg_size),
(basic_type)base_value,
(basic_type)pivot,
num_buckets,
split_bucket,
global_histogram_prefix,
dst
));
int hist_max = global_histogram_cpu[num_groups * split_bucket];
int hist_min = global_histogram_cpu[num_groups * (split_bucket - 1)];
//#ifdef DEBUG_RADIX_SELECT
std::vector<unsigned int> dst_cpu(in.size());
std::vector<unsigned int> src_cpu(in.size());
viennacl::copy(dst, dst_cpu);
viennacl::copy(src, src_cpu);
//#endif
if (hist_max == N)
break;
if (hist_max> N && hist_min < N)
{
scan_start = global_histogram_cpu[num_groups * (split_bucket - 1)];
scan_end = global_histogram_cpu[num_groups * split_bucket];
if (scan_start > 0)
{
viennacl::copy(dst.begin(), dst.begin() + scan_start, result.begin() + position);
position += scan_start;
}
//#ifdef DEBUG_RADIX_SELECT
std::vector<unsigned int> result_cpu(in.size());
viennacl::copy(result, result_cpu);
//#endif
if (position >= N)
break;
if (scan_end == dst.size() && scan_start == 0)
dst.fast_swap(src);
else
viennacl::copy(dst.begin() + scan_start, dst.begin() + scan_end, src.begin());
scan_end -= scan_start;
}
base_value += pivot * (split_bucket-1);
// update pivot
pivot = pivot / num_buckets;
if (pivot == 0)
break;
}
if (position <N)
viennacl::copy(dst.begin(), dst.begin() + (N - position), result.begin() + position);
return result;
}