Skip to content

FEAT: Adding variable length StringDType to/from QuadDType array casting. #244

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
SwayamInSync wants to merge 10 commits into
base: main
Choose a base branch
from
+336 −14
Open

FEAT: Adding variable length StringDType to/from QuadDType array casting. #244

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
10 commits
Select commit Hold shift + click to select a range
0c6322c
initial stringdtype cast support
SwayamInSync Dec 17, 2025
14dac3d
Merge branch 'main' into strdtype
SwayamInSync Dec 19, 2025
970b62c
adding tests
SwayamInSync Dec 19, 2025
0703921
use the bytes_to_quad helper
SwayamInSync Dec 19, 2025
a3d326a
removed comment, fix given_descrs[0] check order
SwayamInSync Dec 19, 2025
4d14322
consolidated tests, removed some not needed
SwayamInSync Dec 19, 2025
f0e6b1f
StringDtype bypass PyUnicode creation
SwayamInSync Dec 19, 2025
87fd7dc
cover edges
SwayamInSync Dec 19, 2025
407d57a
replace -0 with already present
SwayamInSync Dec 19, 2025
fc4c7fa
remove empty array test
SwayamInSync Dec 19, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
209 changes: 208 additions & 1 deletion quaddtype/numpy_quaddtype/src/casts.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -27,7 +27,7 @@ extern "C" {
#include "dragon4.h"
#include "ops.hpp"

#define NUM_CASTS 38 // 17 to_casts + 17 from_casts + 1 quad_to_quad + 1 void_to_quad
#define NUM_CASTS 40 // 18 to_casts + 18 from_casts + 1 quad_to_quad + 1 void_to_quad
#define QUAD_STR_WIDTH 50 // 42 is enough for scientific notation float128, just keeping some buffer

static NPY_CASTING
Expand Down Expand Up @@ -605,6 +605,175 @@ quad_to_bytes_loop(PyArrayMethod_Context *context, char *const data[],
return 0;
}

// StringDType to QuadDType casting
static NPY_CASTING
stringdtype_to_quad_resolve_descriptors(PyObject *NPY_UNUSED(self), PyArray_DTypeMeta *dtypes[2],
PyArray_Descr *given_descrs[2], PyArray_Descr *loop_descrs[2],
npy_intp *view_offset)
{
Py_INCREF(given_descrs[0]);
loop_descrs[0] = given_descrs[0];
Copy link
Member

@ngoldbaum ngoldbaum Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

If you do this after checking given_descrs[1] then there's no need to decref in the error paths below so it'll be a little clearer


if (given_descrs[1] == NULL) {
loop_descrs[1] = (PyArray_Descr *)new_quaddtype_instance(BACKEND_SLEEF);
if (loop_descrs[1] == nullptr) {
Py_DECREF(loop_descrs[0]);
return (NPY_CASTING)-1;
}
}
else {
Py_INCREF(given_descrs[1]);
loop_descrs[1] = given_descrs[1];
}

// no notion of fix length, so always unsafe
Copy link
Member

@ngoldbaum ngoldbaum Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd just delete this comment, I don't think it's correct. It's unsafe because arbitrary strings aren't generally convertible losslessly to quads.

return NPY_UNSAFE_CASTING;
}

// Note: StringDType elements are always aligned, so Aligned template parameter
// is kept for API consistency but both versions use the same logic
template <bool Aligned>
static int
stringdtype_to_quad_strided_loop(PyArrayMethod_Context *context, char *const data[],
npy_intp const dimensions[], npy_intp const strides[],
void *NPY_UNUSED(auxdata))
{
npy_intp N = dimensions[0];
char *in_ptr = data[0];
char *out_ptr = data[1];
npy_intp in_stride = strides[0];
npy_intp out_stride = strides[1];

PyArray_Descr *const *descrs = context->descriptors;
PyArray_StringDTypeObject *str_descr = (PyArray_StringDTypeObject *)descrs[0];
QuadPrecDTypeObject *descr_out = (QuadPrecDTypeObject *)descrs[1];
QuadBackendType backend = descr_out->backend;

npy_string_allocator *allocator = NpyString_acquire_allocator(str_descr);

while (N--) {
const npy_packed_static_string *ps = (npy_packed_static_string *)in_ptr;
npy_static_string s = {0, NULL};
int is_null = NpyString_load(allocator, ps, &s);

if (is_null == -1) {
NpyString_release_allocator(allocator);
PyErr_SetString(PyExc_MemoryError, "Failed to load string in StringDType to Quad cast");
return -1;
}
else if (is_null) {
// Handle null string - use the default string if available, otherwise error
if (str_descr->has_string_na || str_descr->default_string.buf != NULL) {
s = str_descr->default_string;
}
else {
NpyString_release_allocator(allocator);
PyErr_SetString(PyExc_ValueError, "Cannot convert null string to QuadPrecision");
return -1;
}
}

quad_value out_val;
if (bytes_to_quad_convert(s.buf, s.size, backend, &out_val) < 0) {
Copy link
Member Author

@SwayamInSync SwayamInSync Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

utilising bytes_to_quad_convert instead of unicode_to_quad_convert because the later expects Py_UCS format otherwise both are doing the same thing

NpyString_release_allocator(allocator);
return -1;
}

store_quad<Aligned>(out_ptr, out_val, backend);

in_ptr += in_stride;
out_ptr += out_stride;
}

NpyString_release_allocator(allocator);
return 0;
}

// QuadDType to StringDType casting
static NPY_CASTING
quad_to_stringdtype_resolve_descriptors(PyObject *NPY_UNUSED(self), PyArray_DTypeMeta *dtypes[2],
PyArray_Descr *given_descrs[2], PyArray_Descr *loop_descrs[2],
npy_intp *view_offset)
{
Py_INCREF(given_descrs[0]);
loop_descrs[0] = given_descrs[0];
Copy link
Member

@ngoldbaum ngoldbaum Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

same as https://github.com/numpy/numpy-user-dtypes/pull/244/changes#r2636204832


if (given_descrs[1] == NULL) {
// Default StringDType() already has coerce=True
loop_descrs[1] = (PyArray_Descr *)PyObject_CallNoArgs(
(PyObject *)&PyArray_StringDType);
if (loop_descrs[1] == NULL) {
Py_DECREF(loop_descrs[0]);
return (NPY_CASTING)-1;
}
}
else {
Py_INCREF(given_descrs[1]);
loop_descrs[1] = given_descrs[1];
}

return NPY_SAFE_CASTING;
}

// Note: StringDType elements are always aligned, so Aligned template parameter
// is kept for API consistency but both versions use the same logic
template <bool Aligned>
static int
quad_to_stringdtype_strided_loop(PyArrayMethod_Context *context, char *const data[],
npy_intp const dimensions[], npy_intp const strides[],
void *NPY_UNUSED(auxdata))
{
npy_intp N = dimensions[0];
char *in_ptr = data[0];
char *out_ptr = data[1];
npy_intp in_stride = strides[0];
npy_intp out_stride = strides[1];

PyArray_Descr *const *descrs = context->descriptors;
QuadPrecDTypeObject *descr_in = (QuadPrecDTypeObject *)descrs[0];
PyArray_StringDTypeObject *str_descr = (PyArray_StringDTypeObject *)descrs[1];
QuadBackendType backend = descr_in->backend;

npy_string_allocator *allocator = NpyString_acquire_allocator(str_descr);

while (N--) {
quad_value in_val = load_quad<Aligned>(in_ptr, backend);
Sleef_quad sleef_val = quad_to_sleef_quad(&in_val, backend);

// Get string representation with adaptive notation
// Use a large buffer size to allow for full precision
PyObject *py_str = quad_to_string_adaptive(&sleef_val, QUAD_STR_WIDTH);
Copy link
Member

@ngoldbaum ngoldbaum Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There's no need to create a PyUnicode object here. Just pass the ASCII bytes of the C parsed string to NpyString_Pack.

Copy link
Member Author

@SwayamInSync SwayamInSync Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So actually this function quad_to_string_adaptive uses Dragon4 utilities like Dragon4_Positional_QuadDType and Dragon4_Scientific_QuadDType for conversion and they both returns a PyUnicode_FromString object and from them we extract the cstring from it.

I can modify the dragon4 helper or add one more helper with _cstr suffix that returns cstring and then this would be doable. Let me know if this sounds good?

Copy link
Member

@ngoldbaum ngoldbaum Dec 19, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, there should be a code path that bypasses creating a Python string.

if (py_str == NULL) {
NpyString_release_allocator(allocator);
return -1;
}

Py_ssize_t str_size;
const char *str_buf = PyUnicode_AsUTF8AndSize(py_str, &str_size);
if (str_buf == NULL) {
Py_DECREF(py_str);
NpyString_release_allocator(allocator);
return -1;
}

npy_packed_static_string *out_ps = (npy_packed_static_string *)out_ptr;
if (NpyString_pack(allocator, out_ps, str_buf, (size_t)str_size) < 0) {
Py_DECREF(py_str);
NpyString_release_allocator(allocator);
PyErr_SetString(PyExc_MemoryError, "Failed to pack string in Quad to StringDType cast");
return -1;
}

Py_DECREF(py_str);

in_ptr += in_stride;
out_ptr += out_stride;
}

NpyString_release_allocator(allocator);
return 0;
}

// Tag dispatching to ensure npy_bool/npy_ubyte and npy_half/npy_ushort do not alias in templates
// see e.g. https://stackoverflow.com/q/32522279
struct spec_npy_bool {};
Expand Down Expand Up @@ -1395,6 +1564,44 @@ init_casts_internal(void)
};
add_spec(quad_to_bytes_spec);

// StringDType to QuadPrecision cast
PyArray_DTypeMeta **stringdtype_to_quad_dtypes = new PyArray_DTypeMeta *[2]{&PyArray_StringDType, &QuadPrecDType};
PyType_Slot *stringdtype_to_quad_slots = new PyType_Slot[4]{
{NPY_METH_resolve_descriptors, (void *)&stringdtype_to_quad_resolve_descriptors},
{NPY_METH_strided_loop, (void *)&stringdtype_to_quad_strided_loop<true>},
{NPY_METH_unaligned_strided_loop, (void *)&stringdtype_to_quad_strided_loop<false>},
{0, nullptr}};

PyArrayMethod_Spec *stringdtype_to_quad_spec = new PyArrayMethod_Spec{
.name = "cast_StringDType_to_QuadPrec",
.nin = 1,
.nout = 1,
.casting = NPY_UNSAFE_CASTING,
.flags = static_cast<NPY_ARRAYMETHOD_FLAGS>(NPY_METH_SUPPORTS_UNALIGNED | NPY_METH_REQUIRES_PYAPI),
.dtypes = stringdtype_to_quad_dtypes,
.slots = stringdtype_to_quad_slots,
};
add_spec(stringdtype_to_quad_spec);

// QuadPrecision to StringDType cast
PyArray_DTypeMeta **quad_to_stringdtype_dtypes = new PyArray_DTypeMeta *[2]{&QuadPrecDType, &PyArray_StringDType};
PyType_Slot *quad_to_stringdtype_slots = new PyType_Slot[4]{
{NPY_METH_resolve_descriptors, (void *)&quad_to_stringdtype_resolve_descriptors},
{NPY_METH_strided_loop, (void *)&quad_to_stringdtype_strided_loop<true>},
{NPY_METH_unaligned_strided_loop, (void *)&quad_to_stringdtype_strided_loop<false>},
{0, nullptr}};

PyArrayMethod_Spec *quad_to_stringdtype_spec = new PyArrayMethod_Spec{
.name = "cast_QuadPrec_to_StringDType",
.nin = 1,
.nout = 1,
.casting = NPY_SAFE_CASTING,
.flags = static_cast<NPY_ARRAYMETHOD_FLAGS>(NPY_METH_SUPPORTS_UNALIGNED | NPY_METH_REQUIRES_PYAPI),
.dtypes = quad_to_stringdtype_dtypes,
.slots = quad_to_stringdtype_slots,
};
add_spec(quad_to_stringdtype_spec);

specs[spec_count] = nullptr;
return specs;
}
Expand Down
Loading