/**
* This function promotes the requested `axis` from
* a strided dim to a var dim. It modifies `shape`, `coord`,
* `elem`, and `arr` to point to a new array, and
* copies the data over.
*/
static void promote_nd_arr_dim(std::vector<intptr_t> &shape,
std::vector<afpd_coordentry> &coord,
afpd_dtype &elem, nd::array &arr, intptr_t axis,
bool copy_final_coord)
{
vector<afpd_coordentry> newcoord;
afpd_dtype newelem;
newelem.dtp = elem.dtp;
// Convert the axis into a var dim
shape[axis] = -1;
// Create the new array
nd::array newarr = allocate_nd_arr(shape, newcoord, newelem, axis);
// Copy the data up to, but not including, the current `coord`
// from the old `arr` to the new one. The recursion stops
// at `axis`, where all subsequent dimensions are handled by the
// created kernel.
ckernel_builder<kernel_request_host> k;
if (elem.dtp.get_type_id() != uninitialized_type_id) {
make_assignment_kernel(&k, 0, newcoord[axis].tp, newcoord[axis].arrmeta_ptr,
coord[axis].tp, coord[axis].arrmeta_ptr,
kernel_request_strided, &eval::default_eval_context);
}
copy_to_promoted_nd_arr(shape, newarr.get_readwrite_originptr(), newcoord,
newelem, arr.get_readonly_originptr(), coord, elem, k,
0, axis, copy_final_coord, true);
arr.swap(newarr);
coord.swap(newcoord);
elem.swap(newelem);
}
nd::array dynd::format_json(const nd::array& n)
{
// Create a UTF-8 string
nd::array result = nd::empty(ndt::make_string());
// Initialize the output with some memory
output_data out;
out.blockref = reinterpret_cast<const string_type_metadata *>(result.get_ndo_meta())->blockref;
out.api = get_memory_block_pod_allocator_api(out.blockref);
out.api->allocate(out.blockref, 1024, 1, &out.out_begin, &out.out_capacity_end);
out.out_end = out.out_begin;
if (!n.get_type().is_expression()) {
::format_json(out, n.get_type(), n.get_ndo_meta(), n.get_readonly_originptr());
} else {
nd::array tmp = n.eval();
::format_json(out, tmp.get_type(), tmp.get_ndo_meta(), tmp.get_readonly_originptr());
}
// Shrink the memory to fit, and set the pointers in the output
string_type_data *d = reinterpret_cast<string_type_data *>(result.get_readwrite_originptr());
d->begin = out.out_begin;
d->end = out.out_capacity_end;
out.api->resize(out.blockref, out.out_end - out.out_begin, &d->begin, &d->end);
// Finalize processing and mark the result as immutable
result.get_type().extended()->metadata_finalize_buffers(result.get_ndo_meta());
result.flag_as_immutable();
return result;
}
nd::array dynd::format_json(const nd::array &n, bool struct_as_list)
{
// Create a UTF-8 string
nd::array result = nd::empty(ndt::string_type::make());
// Initialize the output with some memory
output_data out;
out.out_string.resize(1024);
out.out_begin = out.out_string.begin();
out.out_capacity_end = out.out_string.end();
out.out_end = out.out_begin;
out.struct_as_list = struct_as_list;
if (!n.get_type().is_expression()) {
::format_json(out, n.get_type(), n.get_arrmeta(), n.get_readonly_originptr());
} else {
nd::array tmp = n.eval();
::format_json(out, tmp.get_type(), tmp.get_arrmeta(), tmp.get_readonly_originptr());
}
// Shrink the memory to fit, and set the pointers in the output
string *d = reinterpret_cast<string *>(result.get_readwrite_originptr());
d->assign(out.out_string.data(), out.out_end - out.out_begin);
// Finalize processing and mark the result as immutable
result.get_type().extended()->arrmeta_finalize_buffers(result.get_arrmeta());
result.flag_as_immutable();
return result;
}
static void json_as_buffer(const nd::array &json, nd::array &out_tmp_ref,
const char *&begin, const char *&end)
{
// Check the type of 'json', and get pointers to the begin/end of a UTF-8
// buffer
ndt::type json_type = json.get_type().value_type();
switch (json_type.get_kind()) {
case string_kind: {
const ndt::base_string_type *sdt =
json_type.extended<ndt::base_string_type>();
switch (sdt->get_encoding()) {
case string_encoding_ascii:
case string_encoding_utf_8:
out_tmp_ref = json.eval();
// The data is already UTF-8, so use the buffer directly
sdt->get_string_range(&begin, &end, out_tmp_ref.get_arrmeta(),
out_tmp_ref.get_readonly_originptr());
break;
default: {
// The data needs to be converted to UTF-8 before parsing
ndt::type utf8_tp = ndt::string_type::make(string_encoding_utf_8);
out_tmp_ref = json.ucast(utf8_tp).eval();
sdt = static_cast<const ndt::base_string_type *>(utf8_tp.extended());
sdt->get_string_range(&begin, &end, out_tmp_ref.get_arrmeta(),
out_tmp_ref.get_readonly_originptr());
break;
}
}
break;
}
case bytes_kind: {
out_tmp_ref = json.eval();
const ndt::base_bytes_type *bdt =
json_type.extended<ndt::base_bytes_type>();
bdt->get_bytes_range(&begin, &end, out_tmp_ref.get_arrmeta(),
out_tmp_ref.get_readonly_originptr());
break;
}
default: {
stringstream ss;
ss << "Input for JSON parsing must be either bytes (interpreted as UTF-8) "
"or a string, not \"" << json_type << "\"";
throw runtime_error(ss.str());
break;
}
}
}
void dynd::typed_data_assign(const ndt::type &dst_tp, const char *dst_arrmeta,
char *dst_data, const nd::array &src_arr,
const eval::eval_context *ectx)
{
typed_data_assign(dst_tp, dst_arrmeta, dst_data, src_arr.get_type(),
src_arr.get_arrmeta(), src_arr.get_readonly_originptr(),
ectx);
}
const char *get_category_data_from_value(size_t value) const {
if (value >= get_category_count()) {
throw std::runtime_error("category value is out of bounds");
}
return m_categories.get_readonly_originptr() +
m_value_to_category_index[value] *
reinterpret_cast<const strided_dim_type_arrmeta *>(
m_categories.get_arrmeta())->stride;
}
const char *get_category_data_from_value(uint32_t value) const
{
if (value >= get_category_count()) {
throw std::runtime_error("category value is out of bounds");
}
return m_categories.get_readonly_originptr() +
unchecked_fixed_dim_get<intptr_t>(m_value_to_category_index,
value) *
reinterpret_cast<const fixed_dim_type_arrmeta *>(
m_categories.get_arrmeta())->stride;
}
uint32_t categorical_type::get_value_from_category(const nd::array& category) const
{
if (category.get_type() == m_category_tp) {
// If the type is right, get the category value directly
return get_value_from_category(category.get_arrmeta(), category.get_readonly_originptr());
} else {
// Otherwise convert to the correct type, then get the category value
nd::array c = nd::empty(m_category_tp);
c.val_assign(category);
return get_value_from_category(c.get_arrmeta(), c.get_readonly_originptr());
}
}
static nd::array view_concrete(const nd::array &arr, const ndt::type &tp)
{
// Allocate a result array to attempt the view in it
nd::array result(make_array_memory_block(tp.get_arrmeta_size()));
// Copy the fields
result.get_ndo()->data.ptr = arr.get_ndo()->data.ptr;
if (arr.get_ndo()->data.ref == NULL) {
// Embedded data, need reference to the array
result.get_ndo()->data.ref = arr.get_memblock().release();
} else {
// Use the same data reference, avoid producing a chain
result.get_ndo()->data.ref = arr.get_data_memblock().release();
}
result.get_ndo()->m_type = ndt::type(tp).release();
result.get_ndo()->m_flags = arr.get_ndo()->m_flags;
// First handle a special case of viewing outermost "var" as "fixed[#]"
if (arr.get_type().get_type_id() == var_dim_type_id && tp.get_type_id() == fixed_dim_type_id) {
const var_dim_type_arrmeta *in_am = reinterpret_cast<const var_dim_type_arrmeta *>(arr.get_arrmeta());
const var_dim_type_data *in_dat = reinterpret_cast<const var_dim_type_data *>(arr.get_readonly_originptr());
fixed_dim_type_arrmeta *out_am = reinterpret_cast<fixed_dim_type_arrmeta *>(result.get_arrmeta());
out_am->dim_size = tp.extended<ndt::fixed_dim_type>()->get_fixed_dim_size();
out_am->stride = in_am->stride;
if ((intptr_t)in_dat->size == out_am->dim_size) {
// Use the more specific data reference from the var arrmeta if possible
if (in_am->blockref != NULL) {
memory_block_decref(result.get_ndo()->data.ref);
memory_block_incref(in_am->blockref);
result.get_ndo()->data.ref = in_am->blockref;
}
result.get_ndo()->data.ptr = in_dat->begin + in_am->offset;
// Try to copy the rest of the arrmeta as a view
if (try_view(arr.get_type().extended<ndt::base_dim_type>()->get_element_type(),
arr.get_arrmeta() + sizeof(var_dim_type_arrmeta),
tp.extended<ndt::base_dim_type>()->get_element_type(),
result.get_arrmeta() + sizeof(fixed_dim_type_arrmeta), arr.get_memblock().get())) {
return result;
}
}
}
// Otherwise try to copy the arrmeta as a view
else if (try_view(arr.get_type(), arr.get_arrmeta(), tp, result.get_arrmeta(), arr.get_memblock().get())) {
// If it succeeded, return it
return result;
}
stringstream ss;
ss << "Unable to view nd::array of type " << arr.get_type();
ss << " as type " << tp;
throw type_error(ss.str());
}
/**
* Substitutes the field types for contiguous array of types
*/
static nd::array
substitute_type_array(const nd::array &type_array,
const std::map<std::string, ndt::type> &typevars,
bool concrete)
{
intptr_t field_count = type_array.get_dim_size();
const ndt::type *field_types =
reinterpret_cast<const ndt::type *>(type_array.get_readonly_originptr());
nd::array tmp_field_types(nd::empty(field_count, ndt::make_type()));
ndt::type *ftraw =
reinterpret_cast<ndt::type *>(tmp_field_types.get_readwrite_originptr());
for (intptr_t i = 0; i < field_count; ++i) {
ftraw[i] = ndt::substitute(field_types[i], typevars, concrete);
}
return tmp_field_types;
}
/**
* This function promotes the dtype the array currently has with
* `tp`, allocates a new one, then copies all the data up to the
* current index in `coord`. This modifies coord and elem in place.
*/
static void promote_nd_arr_dtype(const std::vector<intptr_t> &shape,
std::vector<afpd_coordentry> &coord,
afpd_dtype &elem, nd::array &arr,
const ndt::type &tp)
{
intptr_t ndim = shape.size();
vector<afpd_coordentry> newcoord;
afpd_dtype newelem;
if (elem.dtp.get_type_id() == uninitialized_type_id) {
// If the `elem` dtype is uninitialized, it means a dummy
// array was created to capture dimensional structure until
// the first value is encountered
newelem.dtp = tp;
} else {
newelem.dtp = promote_types_arithmetic(elem.dtp, tp);
}
// Create the new array
nd::array newarr = allocate_nd_arr(shape, newcoord, newelem, ndim);
// Copy the data up to, but not including, the current `coord`
// from the old `arr` to the new one
ckernel_builder<kernel_request_host> k;
if (elem.dtp.get_type_id() != uninitialized_type_id) {
make_assignment_kernel(&k, 0, newelem.dtp, newelem.arrmeta_ptr, elem.dtp,
elem.arrmeta_ptr, kernel_request_strided,
&eval::default_eval_context);
} else {
// An assignment kernel which copies one byte - will only
// be called with count==0 when dtp is uninitialized
make_assignment_kernel(&k, 0, ndt::type::make<char>(), NULL,
ndt::type::make<char>(), NULL,
kernel_request_strided, &eval::default_eval_context);
}
copy_to_promoted_nd_arr(shape, newarr.get_readwrite_originptr(), newcoord,
newelem, arr.get_readonly_originptr(), coord, elem, k,
0, ndim, false, true);
arr.swap(newarr);
coord.swap(newcoord);
elem.swap(newelem);
}
nd::callable::callable(const nd::array &rhs)
{
if (!rhs.is_null()) {
if (rhs.get_type().get_type_id() == callable_type_id) {
const callable_type_data *af =
reinterpret_cast<const callable_type_data *>(
rhs.get_readonly_originptr());
if (af->instantiate != NULL) {
// It's valid: callable type, contains instantiate function.
m_value = rhs;
} else {
throw invalid_argument("Require a non-empty callable, "
"provided callable has NULL "
"instantiate function");
}
} else {
stringstream ss;
ss << "Cannot implicitly convert nd::array of type "
<< rhs.get_type().value_type() << " to callable";
throw type_error(ss.str());
}
}
}
inline const uintptr_t *get_data_offsets_raw() const {
return reinterpret_cast<const uintptr_t *>(
m_data_offsets.get_readonly_originptr());
}
const type *get_field_types_raw() const
{
return reinterpret_cast<const type *>(
m_field_types.get_readonly_originptr());
}
const uintptr_t *get_data_offsets(const char *DYND_UNUSED(arrmeta)) const {
return reinterpret_cast<const uintptr_t *>(
m_data_offsets.get_readonly_originptr());
}
nd::array dynd::struct_concat(nd::array lhs, nd::array rhs)
{
nd::array res;
if (lhs.is_null()) {
res = rhs;
return res;
}
if (rhs.is_null()) {
res = lhs;
return res;
}
const ndt::type &lhs_tp = lhs.get_type(), &rhs_tp = rhs.get_type();
if (lhs_tp.get_kind() != struct_kind) {
stringstream ss;
ss << "Cannot concatenate array with type " << lhs_tp << " as a struct";
throw invalid_argument(ss.str());
}
if (rhs_tp.get_kind() != struct_kind) {
stringstream ss;
ss << "Cannot concatenate array with type " << rhs_tp << " as a struct";
throw invalid_argument(ss.str());
}
// Make an empty shell struct by concatenating the fields together
intptr_t lhs_n = lhs_tp.extended<ndt::base_struct_type>()->get_field_count();
intptr_t rhs_n = rhs_tp.extended<ndt::base_struct_type>()->get_field_count();
intptr_t res_n = lhs_n + rhs_n;
nd::array res_field_names = nd::empty(res_n, ndt::string_type::make());
nd::array res_field_types = nd::empty(res_n, ndt::make_type());
res_field_names(irange(0, lhs_n)).vals() = lhs_tp.extended<ndt::base_struct_type>()->get_field_names();
res_field_names(irange(lhs_n, res_n)).vals() = rhs_tp.extended<ndt::base_struct_type>()->get_field_names();
res_field_types(irange(0, lhs_n)).vals() = lhs_tp.extended<ndt::base_struct_type>()->get_field_types();
res_field_types(irange(lhs_n, res_n)).vals() = rhs_tp.extended<ndt::base_struct_type>()->get_field_types();
ndt::type res_tp = ndt::struct_type::make(res_field_names, res_field_types);
const ndt::type *res_field_tps = res_tp.extended<ndt::base_struct_type>()->get_field_types_raw();
res = nd::empty_shell(res_tp);
// Initialize the default data offsets for the struct arrmeta
ndt::struct_type::fill_default_data_offsets(res_n, res_tp.extended<ndt::base_struct_type>()->get_field_types_raw(),
reinterpret_cast<uintptr_t *>(res.get_arrmeta()));
// Get information about the arrmeta layout of the input and res
const uintptr_t *lhs_arrmeta_offsets = lhs_tp.extended<ndt::base_struct_type>()->get_arrmeta_offsets_raw();
const uintptr_t *rhs_arrmeta_offsets = rhs_tp.extended<ndt::base_struct_type>()->get_arrmeta_offsets_raw();
const uintptr_t *res_arrmeta_offsets = res_tp.extended<ndt::base_struct_type>()->get_arrmeta_offsets_raw();
const char *lhs_arrmeta = lhs.get_arrmeta();
const char *rhs_arrmeta = rhs.get_arrmeta();
char *res_arrmeta = res.get_arrmeta();
// Copy the arrmeta from the input arrays
for (intptr_t i = 0; i < lhs_n; ++i) {
const ndt::type &tp = res_field_tps[i];
if (!tp.is_builtin()) {
tp.extended()->arrmeta_copy_construct(res_arrmeta + res_arrmeta_offsets[i], lhs_arrmeta + lhs_arrmeta_offsets[i],
lhs.get_data_memblock().get());
}
}
for (intptr_t i = 0; i < rhs_n; ++i) {
const ndt::type &tp = res_field_tps[i + lhs_n];
if (!tp.is_builtin()) {
tp.extended()->arrmeta_copy_construct(res_arrmeta + res_arrmeta_offsets[i + lhs_n],
rhs_arrmeta + rhs_arrmeta_offsets[i], rhs.get_data_memblock().get());
}
}
// Get information about the data layout of the input and res
const uintptr_t *lhs_data_offsets = lhs_tp.extended<ndt::base_struct_type>()->get_data_offsets(lhs.get_arrmeta());
const uintptr_t *rhs_data_offsets = rhs_tp.extended<ndt::base_struct_type>()->get_data_offsets(rhs.get_arrmeta());
const uintptr_t *res_data_offsets = res_tp.extended<ndt::base_struct_type>()->get_data_offsets(res.get_arrmeta());
const char *lhs_data = lhs.get_readonly_originptr();
const char *rhs_data = rhs.get_readonly_originptr();
char *res_data = res.get_readwrite_originptr();
// Copy the data from the input arrays
for (intptr_t i = 0; i < lhs_n; ++i) {
const ndt::type &tp = res_field_tps[i];
typed_data_copy(tp, res_arrmeta + res_arrmeta_offsets[i], res_data + res_data_offsets[i],
lhs_arrmeta + lhs_arrmeta_offsets[i], lhs_data + lhs_data_offsets[i]);
}
for (intptr_t i = 0; i < rhs_n; ++i) {
const ndt::type &tp = res_field_tps[i + lhs_n];
typed_data_copy(tp, res_arrmeta + res_arrmeta_offsets[i + lhs_n], res_data + res_data_offsets[i + lhs_n],
rhs_arrmeta + rhs_arrmeta_offsets[i], rhs_data + rhs_data_offsets[i]);
}
return res;
}
/**
* Adds a ckernel layer for processing one dimension of the reduction.
* This is for a strided dimension which is being broadcast, and is
* the final dimension before the accumulation operation.
*/
static size_t make_strided_inner_broadcast_dimension_kernel(
const callable_type_data *elwise_reduction_const,
const ndt::callable_type *elwise_reduction_tp,
const callable_type_data *dst_initialization_const,
const ndt::callable_type *dst_initialization_tp, void *ckb,
intptr_t ckb_offset, intptr_t dst_stride, intptr_t src_stride,
intptr_t src_size, const ndt::type &dst_tp, const char *dst_arrmeta,
const ndt::type &src_tp, const char *src_arrmeta, bool right_associative,
const nd::array &reduction_identity, kernel_request_t kernreq,
const eval::eval_context *ectx)
{
callable_type_data *elwise_reduction =
const_cast<callable_type_data *>(elwise_reduction_const);
callable_type_data *dst_initialization =
const_cast<callable_type_data *>(dst_initialization_const);
intptr_t root_ckb_offset = ckb_offset;
strided_inner_broadcast_kernel_extra *e =
reinterpret_cast<ckernel_builder<kernel_request_host> *>(ckb)
->alloc_ck<strided_inner_broadcast_kernel_extra>(ckb_offset);
e->destructor = &strided_inner_broadcast_kernel_extra::destruct;
// Cannot have both a dst_initialization kernel and a reduction identity
if (dst_initialization != NULL && !reduction_identity.is_null()) {
throw invalid_argument(
"make_lifted_reduction_ckernel: cannot specify"
" both a dst_initialization kernel and a reduction_identity");
}
if (reduction_identity.is_null()) {
// Get the function pointer for the first_call, for the case with
// no reduction identity
if (kernreq == kernel_request_single) {
e->set_first_call_function(
&strided_inner_broadcast_kernel_extra::single_first);
} else if (kernreq == kernel_request_strided) {
e->set_first_call_function(
&strided_inner_broadcast_kernel_extra::strided_first);
} else {
stringstream ss;
ss << "make_lifted_reduction_ckernel: unrecognized request "
<< (int)kernreq;
throw runtime_error(ss.str());
}
} else {
// Get the function pointer for the first_call, for the case with
// a reduction identity
if (kernreq == kernel_request_single) {
e->set_first_call_function(
&strided_inner_broadcast_kernel_extra::single_first_with_ident);
} else if (kernreq == kernel_request_strided) {
e->set_first_call_function(
&strided_inner_broadcast_kernel_extra::strided_first_with_ident);
} else {
stringstream ss;
ss << "make_lifted_reduction_ckernel: unrecognized request "
<< (int)kernreq;
throw runtime_error(ss.str());
}
if (reduction_identity.get_type() != dst_tp) {
stringstream ss;
ss << "make_lifted_reduction_ckernel: reduction identity type ";
ss << reduction_identity.get_type() << " does not match dst type ";
ss << dst_tp;
throw runtime_error(ss.str());
}
e->ident_data = reduction_identity.get_readonly_originptr();
e->ident_ref = reduction_identity.get_memblock().release();
}
// The function pointer for followup accumulation calls
e->set_followup_call_function(
&strided_inner_broadcast_kernel_extra::strided_followup);
// The striding parameters
e->dst_stride = dst_stride;
e->src_stride = src_stride;
e->size = src_size;
// Validate that the provided callables are unary operations,
// and have the correct types
if (elwise_reduction_tp->get_npos() != 1 &&
elwise_reduction_tp->get_npos() != 2) {
stringstream ss;
ss << "make_lifted_reduction_ckernel: elwise reduction ckernel ";
ss << "funcproto must be unary or a binary expr with all equal types";
throw runtime_error(ss.str());
}
if (elwise_reduction_tp->get_return_type() != dst_tp) {
stringstream ss;
ss << "make_lifted_reduction_ckernel: elwise reduction ckernel ";
ss << "dst type is " << elwise_reduction_tp->get_return_type();
ss << ", expected " << dst_tp;
throw type_error(ss.str());
}
if (elwise_reduction_tp->get_pos_type(0) != src_tp) {
stringstream ss;
ss << "make_lifted_reduction_ckernel: elwise reduction ckernel ";
ss << "src type is " << elwise_reduction_tp->get_return_type();
ss << ", expected " << src_tp;
throw type_error(ss.str());
}
if (dst_initialization != NULL) {
check_dst_initialization(dst_initialization_tp, dst_tp, src_tp);
}
if (elwise_reduction_tp->get_npos() == 2) {
ckb_offset = kernels::wrap_binary_as_unary_reduction_ckernel(
ckb, ckb_offset, right_associative, kernel_request_strided);
ndt::type src_tp_doubled[2] = {src_tp, src_tp};
const char *src_arrmeta_doubled[2] = {src_arrmeta, src_arrmeta};
ckb_offset = elwise_reduction->instantiate(
elwise_reduction->static_data, 0, NULL, ckb, ckb_offset, dst_tp,
dst_arrmeta, elwise_reduction_tp->get_npos(), src_tp_doubled,
src_arrmeta_doubled, kernel_request_strided, ectx, nd::array(),
std::map<nd::string, ndt::type>());
} else {
ckb_offset = elwise_reduction->instantiate(
elwise_reduction->static_data, 0, NULL, ckb, ckb_offset, dst_tp,
dst_arrmeta, elwise_reduction_tp->get_npos(), &src_tp, &src_arrmeta,
kernel_request_strided, ectx, nd::array(),
std::map<nd::string, ndt::type>());
}
// Make sure there's capacity for the next ckernel
reinterpret_cast<ckernel_builder<kernel_request_host> *>(ckb)
->reserve(ckb_offset + sizeof(ckernel_prefix));
// Need to retrieve 'e' again because it may have moved
e = reinterpret_cast<ckernel_builder<kernel_request_host> *>(ckb)
->get_at<strided_inner_broadcast_kernel_extra>(root_ckb_offset);
e->dst_init_kernel_offset = ckb_offset - root_ckb_offset;
if (dst_initialization != NULL) {
ckb_offset = dst_initialization->instantiate(
dst_initialization->static_data, 0, NULL, ckb, ckb_offset, dst_tp,
dst_arrmeta, elwise_reduction_tp->get_npos(), &src_tp, &src_arrmeta,
kernel_request_strided, ectx, nd::array(),
std::map<nd::string, ndt::type>());
} else if (reduction_identity.is_null()) {
ckb_offset =
make_assignment_kernel(ckb, ckb_offset, dst_tp, dst_arrmeta, src_tp,
src_arrmeta, kernel_request_strided, ectx);
} else {
ckb_offset = make_assignment_kernel(
ckb, ckb_offset, dst_tp, dst_arrmeta, reduction_identity.get_type(),
reduction_identity.get_arrmeta(), kernel_request_strided, ectx);
}
return ckb_offset;
}
categorical_type::categorical_type(const nd::array& categories, bool presorted)
: base_type(categorical_type_id, custom_kind, 4, 4, type_flag_scalar, 0, 0, 0)
{
intptr_t category_count;
if (presorted) {
// This is construction shortcut, for the case when the categories are already
// sorted. No validation of this is done, the caller should have ensured it
// was correct already, typically by construction.
m_categories = categories.eval_immutable();
m_category_tp = m_categories.get_type().at(0);
category_count = categories.get_dim_size();
m_value_to_category_index.resize(category_count);
m_category_index_to_value.resize(category_count);
for (size_t i = 0; i != (size_t)category_count; ++i) {
m_value_to_category_index[i] = i;
m_category_index_to_value[i] = i;
}
} else {
// Process the categories array to make sure it's valid
const ndt::type& cdt = categories.get_type();
if (cdt.get_type_id() != strided_dim_type_id) {
throw dynd::type_error("categorical_type only supports construction from a strided array of categories");
}
m_category_tp = categories.get_type().at(0);
if (!m_category_tp.is_scalar()) {
throw dynd::type_error("categorical_type only supports construction from a 1-dimensional strided array of categories");
}
category_count = categories.get_dim_size();
intptr_t categories_stride = reinterpret_cast<const strided_dim_type_arrmeta *>(categories.get_arrmeta())->stride;
const char *categories_element_arrmeta = categories.get_arrmeta() + sizeof(strided_dim_type_arrmeta);
comparison_ckernel_builder k;
::make_comparison_kernel(&k, 0,
m_category_tp, categories_element_arrmeta,
m_category_tp, categories_element_arrmeta,
comparison_type_sorting_less, &eval::default_eval_context);
cmp less(k.get_function(), k.get());
set<const char *, cmp> uniques(less);
m_value_to_category_index.resize(category_count);
m_category_index_to_value.resize(category_count);
// create the mapping from indices of (to be lexicographically sorted) categories to values
for (size_t i = 0; i != (size_t)category_count; ++i) {
m_category_index_to_value[i] = i;
const char *category_value = categories.get_readonly_originptr() +
i * categories_stride;
if (uniques.find(category_value) == uniques.end()) {
uniques.insert(category_value);
} else {
stringstream ss;
ss << "categories must be unique: category value ";
m_category_tp.print_data(ss, categories_element_arrmeta, category_value);
ss << " appears more than once";
throw std::runtime_error(ss.str());
}
}
// TODO: Putting everything in a set already caused a sort operation to occur,
// there's no reason we should need a second sort.
std::sort(m_category_index_to_value.begin(), m_category_index_to_value.end(),
sorter(categories.get_readonly_originptr(), categories_stride,
k.get_function(), k.get()));
// invert the m_category_index_to_value permutation
for (uint32_t i = 0; i < m_category_index_to_value.size(); ++i) {
m_value_to_category_index[m_category_index_to_value[i]] = i;
}
m_categories = make_sorted_categories(uniques, m_category_tp,
categories_element_arrmeta);
}
// Use the number of categories to set which underlying integer storage to use
if (category_count <= 256) {
m_storage_type = ndt::make_type<uint8_t>();
} else if (category_count <= 65536) {
m_storage_type = ndt::make_type<uint16_t>();
} else {
m_storage_type = ndt::make_type<uint32_t>();
}
m_members.data_size = m_storage_type.get_data_size();
m_members.data_alignment = (uint8_t)m_storage_type.get_data_alignment();
}
const arrfunc_type_data *get_is_avail_arrfunc() const
{
return reinterpret_cast<const arrfunc_type_data *>(
m_nafunc.get_readonly_originptr());
}
const arrfunc_type_data *get_assign_na_arrfunc() const
{
return reinterpret_cast<const arrfunc_type_data *>(
m_nafunc.get_readonly_originptr()) +
1;
}
inline const arrfunc_type_data *get() const
{
return !m_value.is_null() ? reinterpret_cast<const arrfunc_type_data *>(
m_value.get_readonly_originptr())
: NULL;
}
array_iter(const nd::array& op0, const nd::array& op1) {
init(op0.get_type(), op0.get_ndo_meta(), op0.get_readwrite_originptr(),
op1.get_type(), op1.get_ndo_meta(), op1.get_readonly_originptr());
}