8000 WIP: ENH: Expand gufunc signature to allow flexible dimension specs by mattip · Pull Request #11132 · numpy/numpy · GitHub
[go: up one dir, main page]
Skip to content

WIP: ENH: Expand gufunc signature to allow flexible dimension specs #11132

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

Closed
wants to merge 8 commits into from
Closed

WIP: ENH: Expand gufunc signature to allow flexible dimension specs #11132

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
e9a5ab5
ENH: Add frozen dimensions to gufunc signatures
mattip Apr 30, 2018
cbbe0f6
convert reserved field to version, set new version to 1
mattip Apr 30, 2018
dbbeb58
add tests for new signatures
mattip Apr 30, 2018
18505d7
test, document, implement flexible signature
mattip Apr 30, 2018
67e752e
improve docs for PyUFunc_FromFuncAndData
mattip May 9, 2018
50d59fd
remove fixed dimension ufuncsignatures
mattip May 26, 2018
d479a5b
belongs on matmul-as-ufunc branch
mattip May 26, 2018
f3d1622
fixes from review
mattip May 26, 2018
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
Next Next commit
ENH: Add frozen dimensions to gufunc signatures
  • Loading branch information
@mattip
mattip committed May 18, 2018
commit e9a5ab5604bab74abf74cd15a740256f39260e77
5 changes: 5 additions & 0 deletions numpy/core/include/numpy/ufuncobject.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -209,6 +209,11 @@ typedef struct _tagPyUFuncObject {
* set by nditer object.
*/
npy_uint32 iter_flags;

/*
* sizes of frozen core dimensions, or -1 if unset
*/
npy_intp *core_dim_szs;
} PyUFuncObject;

#include "arrayobject.h"
Expand Down
44 changes: 44 additions & 0 deletions numpy/core/src/umath/_umath_tests.c.src
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -196,6 +196,40 @@ static void
/**end repeat**/

char *euclidean_pdist_signature = "(n,d)->(p)";
char *cross1d_signature = "(3),(3)->(3)";

/**begin repeat

#TYPE=LONG,DOUBLE#
#typ=npy_long, npy_double#
*/

/*
* This implements the cross product:
* out[n, 0] = in1[n, 1]*in2[n, 2] - in1[n, 2]*in2[n, 1]
* out[n, 1] = in1[n, 2]*in2[n, 0] - in1[n, 0]*in2[n, 2]
* out[n, 2] = in1[n, 0]*in2[n, 1] - in1[n, 1]*in2[n, 0]
*/
static void
@TYPE@_cross1d(char **args, npy_intp *dimensions, npy_intp *steps, void *NPY_UNUSED(func))
{
INIT_OUTER_LOOP_3
npy_intp is1=steps[0], is2=steps[1], os = steps[2];
BEGIN_OUTER_LOOP_3
char *ip1=args[0], *ip2=args[1], *op=args[2];

*(@typ@ *)op = *(@typ@ *)(ip1 + is1) * *(@typ@ *)(ip2 + 2*is2) -
*(@typ@ *)(ip1 + 2*is1) * *(@typ@ *)(ip2 + is2);
op += os;
*(@typ@ *)op = *(@typ@ *)(ip1 + 2*is1) * *(@typ@ *)ip2 -
*(@typ@ *)ip1 * *(@typ@ *)(ip2 + 2*is2);
op += os;
*(@typ@ *)op = *(@typ@ *)ip1 * *(@typ@ *)(ip2 + is2) -
*(@typ@ *)(ip1 + is1) * *(@typ@ *)ip2;
END_OUTER_LOOP
}

/**end repeat**/

/**begin repeat

Expand Down Expand Up @@ -263,6 +297,9 @@ static char innerwt_signatures[] = { NPY_LONG, NPY_LONG, NPY_LONG, NPY_LONG, NPY
static PyUFuncGenericFunction matrix_multiply_functions[] = { LONG_matrix_multiply, FLOAT_matrix_multiply, DOUBLE_matrix_multiply };
static void *matrix_multiply_data[] = { (void *)NULL, (void *)NULL, (void *)NULL };
static char matrix_multiply_signatures[] = { NPY_LONG, NPY_LONG, NPY_LONG, NPY_FLOAT, NPY_FLOAT, NPY_FLOAT, NPY_DOUBLE, NPY_DOUBLE, NPY_DOUBLE };
static PyUFuncGenericFunction cross1d_functions[] = { LONG_cross1d, DOUBLE_cross1d };
static void * cross1d_data[] = { (void *)NULL, (void *)NULL };
static char cross1d_signatures[] = { NPY_LONG, NPY_LONG, NPY_LONG, NPY_DOUBLE, NPY_DOUBLE, NPY_DOUBLE };

static PyUFuncGenericFunction euclidean_pdist_functions[] =
{ FLOAT_euclidean_pdist, DOUBLE_euclidean_pdist };
Expand Down Expand Up @@ -311,6 +348,13 @@ addUfuncs(PyObject *dictionary) {
0, inner1d_signature);
PyDict_SetItemString(dictionary, "inner1d_no_doc", f);
Py_DECREF(f);
f = PyUFunc_FromFuncAndDataAndSignature(cross1d_functions, cross1d_data, cross1d_signatures, 2,
2, 1, PyUFunc_None, "cross1d",
"cross product on the last dimension and broadcast on the rest \n"\
" \"(3),(3)->(3)\" \n",
0, cross1d_signature);
PyDict_SetItemString(dictionary, "cross1d", f);
Py_DECREF(f);
}


Expand Down
65 changes: 59 additions & 6 deletions numpy/core/src/umath/ufunc_object.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -339,6 +339,28 @@ _is_alnum_underscore(char ch)
return _is_alpha_underscore(ch) || (ch >= '0' && ch <= '9');
}

/*
* Convert a string into a number
*/
static npy_intp
_get_size(const char* str)
{
char *stop;
#if defined(_MSC_VER)
#define strtoll _strtoi64
#endif
npy_intp size = (npy_intp)strtoll(str, &stop, 10);
#if defined(_MSC_VER)
#undef strtoll
#endif

if (stop == str || _is_alpha_underscore(*stop)) {
/* not a well formed number */
return -1;
}
return size;
}

/*
* Return the ending position of a variable name
*/
Expand Down Expand Up @@ -406,10 +428,13 @@ _parse_signature(PyUFuncObject *ufunc, const char *signature)
ufunc->core_enabled = 1;
ufunc->core_num_dim_ix = 0;
ufunc->core_num_dims = PyArray_malloc(sizeof(int) * ufunc->nargs);
ufunc->core_dim_ixs = PyArray_malloc(sizeof(int) * len); /* shrink this later */
ufunc->core_offsets = PyArray_malloc(sizeof(int) * ufunc->nargs);
if (ufunc->core_num_dims == NULL || ufunc->core_dim_ixs == NULL
|| ufunc->core_offsets == NULL) {
/* The next two items will be shrunk later */
Copy link
Contributor

Choose a reason for hiding this comment

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

next three

Copy link
Member Author

Choose a reason for hiding this comment

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

removing szs makes it two again, so "fixed"

< 8000 p class="ml-1 mb-2 mt-2" data-show-on-error hidden> Sorry, something went wrong.

ufunc->core_dim_ixs = PyArray_malloc(sizeof(int) * len);
ufunc->core_dim_szs = PyArray_malloc(sizeof(npy_intp) * len);

if (ufunc->core_num_dims == NULL || ufunc->core_dim_ixs == NULL ||
ufunc->core_offsets == NULL || ufunc->core_dim_szs == NULL) {
PyErr_NoMemory();
goto fail;
}
Expand Down Expand Up @@ -438,8 +463,15 @@ _parse_signature(PyUFuncObject *ufunc, const char *signature)
while (signature[i] != ')') {
/* loop over core dimensions */
int j = 0;
if (!_is_alpha_underscore(signature[i])) {
parse_error = "expect dimension name";
npy_intp frozen_size = -1;
if (signature[i] == '\0') {
parse_error = "unexpected end of signature string";
goto fail;
}

if (!_is_alpha_underscore(signature[i]) &&
Copy link
Contributor

Choose a reason for hiding this comment

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

I think here and below it might make more sense to separate out more:

if (_is_alpha...) {
    check on var_names 
}
else {
    frozen_size = _get_...
    if (frozen_size < 0) {
        parse error
    }
    check on same frozen size
}

Copy link
Member Author

Choose a reason for hiding this comment

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

fixed by removing frozen size code

(frozen_size = _get_size(signature + i)) < 0) {
parse_error = "expect dimension name or frozen size";
goto fail;
}
while (j < ufunc->core_num_dim_ix) {
Copy link
Contributor

Choose a reason for hiding this comment

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

This while loop is not necessarily correct for frozen dimensions, since a number does not have to be given exactly the same way (say, a signature of (03)->(3). How about the following:

for (j = 0; j < ufunc->core_num_dim_ix; j++) {
    if (frozen_size >= 0 ? frozen_size == ufunc->core_dim_szs[j]) :
            _is_same_name(signature+i, var_names[j])) {
        break;
    }
}

(Could also be a while, but then at least include the test for (in)equality in the condition!)

Though perhaps it is better to just split the frozen and normal paths altogether (see above)

Copy link
Member Author

Choose a reason for hiding this comment

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

fixed by removing code

Expand All @@ -450,6 +482,7 @@ _parse_signature(PyUFuncObject *ufunc, const char *signature)
}
if (j >= ufunc->core_num_dim_ix) {
var_names[j] = signature+i;
ufunc->core_dim_szs[j] = frozen_size;
ufunc->core_num_dim_ix++;
}
ufunc->core_dim_ixs[cur_core_dim] = j;
Expand Down Expand Up @@ -494,6 +527,9 @@ _parse_signature(PyUFuncObject *ufunc, const char *signature)
}
ufunc->core_dim_ixs = PyArray_realloc(ufunc->core_dim_ixs,
sizeof(int)*cur_core_dim);
// ufunc->core_dim_szs = PyArray_realloc(ufunc->core_dim_szs,
// sizeof(npy_intp)*ufunc->core_num_dim_ix);

/* check for trivial core-signature, e.g. "(),()->()" */
if (cur_core_dim == 0) {
ufunc->core_enabled = 0;
Expand Down Expand Up @@ -2087,10 +2123,14 @@ _get_coredim_sizes(PyUFuncObject *ufunc, PyArrayObject **op,
*/
for (idim = 0; idim < num_dims; ++idim) {
int core_dim_index = ufunc->core_dim_ixs[dim_offset+idim];
npy_intp frozen_size = ufunc->core_dim_szs[core_dim_index];
npy_intp op_dim_size = PyArray_DIM(
op[i], REMAP_AXIS(i, core_start_dim+idim));

if (core_dim_sizes[core_dim_index] == -1) {
if (frozen_size == -1 && core_dim_sizes[core_dim_index] == 1) {
core_dim_sizes[core_dim_index] = op_dim_size;
}
else if (core_dim_sizes[core_dim_index] == -1) {
core_dim_sizes[core_dim_index] = op_dim_size;
}
else if (op_dim_size != core_dim_sizes[core_dim_index]) {
Expand Down Expand Up @@ -2364,6 +2404,17 @@ PyUFunc_GeneralizedFunction(PyUFuncObject *ufunc,
goto fail;
}
}
/*
* Validate the core dimensions of all the operands,
* and collect all of the labeled core dimension sizes
* into the array 'core_dim_sizes'. Initialize them to
* 1, for example in the case where the operand broadcasts
* to a core dimension, it won't be visited.
*/
for (i = 0; i < ufunc->core_num_dim_ix; ++i) {
npy_intp frozen_size = ufunc->core_dim_szs[i];
core_dim_sizes[i] = frozen_size == -1 ? 1 : frozen_size;
}

/* Collect the lengths of the labelled core dimensions */
retval = _get_coredim_sizes(ufunc, op, core_dim_sizes, remap_axis);
Expand Down Expand Up @@ -4671,6 +4722,7 @@ PyUFunc_FromFuncAndDataAndSignature(PyUFuncGenericFunction *func, void **data,
ufunc->core_dim_ixs = NULL;
ufunc->core_offsets = NULL;
ufunc->core_signature = NULL;
ufunc->core_dim_szs = NULL;
Copy link
Contributor

Choose a reason for hiding this comment

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

Should also initialize core_dim_flexible

Copy link
Member Author

Choose a reason for hiding this comment

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

fixed

if (signature != NULL) {
if (_parse_signature(ufunc, signature) != 0) {
Py_DECREF(ufunc);
Expand Down Expand Up @@ -5017,6 +5069,7 @@ ufunc_dealloc(PyUFuncObject *ufunc)
{
PyArray_free(ufunc->core_num_dims);
PyArray_free(ufunc->core_dim_ixs);
PyArray_free(ufunc->core_dim_szs);
Copy link
Contributor

Choose a reason for hiding this comment

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

Should also free ->core_dim_flexible or we'll have a memory leak...

Copy link
Member Author

Choose a reason for hiding this comment

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

fixed

PyArray_free(ufunc->core_offsets);
PyArray_free(ufunc->core_signature);
PyArray_free(ufunc->ptr);
Expand Down
0