Skip to content

add new complex vm functions #86

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

Merged
merged 4 commits into from
Jul 14, 2025
Merged

add new complex vm functions #86

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
2db04e6
add new complex vm functions
May 9, 2025
110f02f
address comments
Jul 14, 2025
9bff9c1
use npy_hypot
Jul 14, 2025
4b5aa36
Revert "use npy_hypot"
Jul 14, 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
3 changes: 2 additions & 1 deletion CHANGELOG.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,9 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0

### Added
* Added mkl implementation for floating point data-types of `exp2`, `log2`, `fabs`, `copysign`, `nextafter`, `fmax`, `fmin` and `remainder` functions [gh-81](https://github.com/IntelPython/mkl_umath/pull/81)
* Added mkl implementation for complex data-types of `conjugate` and `absolute` functions [gh-86](https://github.com/IntelPython/mkl_umath/pull/86)

## [0.2.0] (06/DD/2025)
## [0.2.0] (06/03/2025)
This release updates `mkl_umath` to be aligned with both numpy-1.26.x and numpy-2.x.x.

### Added
Expand Down
70 changes: 35 additions & 35 deletions mkl_umath/src/mkl_umath_loops.c.src
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -129,6 +129,7 @@
* when these conditions are not met VML functions may produce incorrect output
*/
#define DISJOINT_OR_SAME(p1, p2, n, s) (((p1) == (p2)) || ((p2) + (n)*(s) < (p1)) || ((p1) + (n)*(s) < (p2)) )
#define DISJOINT_OR_SAME_TWO_DTYPES(p1, p2, n, s1, s2) (((p1) == (p2)) || ((p2) + (n)*(s2) < (p1)) || ((p1) + (n)*(s1) < (p2)) )

/*
* include vectorized functions and dispatchers
Expand Down Expand Up @@ -316,8 +317,7 @@ mkl_umath_@TYPE@_exp(char **args, const npy_intp *dimensions, const npy_intp *st
can_vectorize
,
const @type@ in1 = *(@type@ *)ip1;
const int invalid_cases = npy_isnan(in1) || in1 == NPY_INFINITY || in1 == -NPY_INFINITY;
ignore_fpstatus |= (invalid_cases ? 1 : 0);
ignore_fpstatus = npy_isnan(in1) || in1 == NPY_INFINITY || in1 == -NPY_INFINITY;
*(@type@ *)op1 = @scalarf@(in1);
)
}
Expand Down Expand Up @@ -355,8 +355,7 @@ mkl_umath_@TYPE@_exp2(char **args, const npy_intp *dimensions, const npy_intp *s
can_vectorize
,
const @type@ in1 = *(@type@ *)ip1;
const int invalid_cases = npy_isnan(in1) || in1 == NPY_INFINITY || in1 == -NPY_INFINITY;
ignore_fpstatus |= (invalid_cases ? 1 : 0);
ignore_fpstatus = npy_isnan(in1) || in1 == NPY_INFINITY || in1 == -NPY_INFINITY;
*(@type@ *)op1 = @scalarf@(in1);
)
}
Expand Down Expand Up @@ -493,8 +492,7 @@ mkl_umath_@TYPE@_log2(char **args, const npy_intp *dimensions, const npy_intp *s
can_vectorize
,
const @type@ in1 = *(@type@ *)ip1;
const int invalid_cases = in1 < 0 || in1 == 0 || npy_isnan(in1) || in1 == -NPY_INFINITY;
ignore_fpstatus |= (invalid_cases ? 1 : 0);
ignore_fpstatus = in1 < 0 || in1 == 0 || npy_isnan(in1) || in1 == -NPY_INFINITY;
*(@type@ *)op1 = @scalarf@(in1);
)
}
Expand Down Expand Up @@ -2124,10 +2122,9 @@ mkl_umath_@TYPE@_remainder(char **args, const npy_intp *dimensions, const npy_in
BINARY_LOOP {
const @type@ in1 = *(@type@ *)ip1;
const @type@ in2 = *(@type@ *)ip2;
int invalid_cases = !npy_isnan(in1) && in2 == 0;
invalid_cases |= (in1 == NPY_INFINITY || in1 == -NPY_INFINITY) && !npy_isnan(in2);
invalid_cases |= (in1 != NPY_INFINITY && in1 != -NPY_INFINITY) && (in2 == NPY_INFINITY || in2 == -NPY_INFINITY);
ignore_fpstatus |= (invalid_cases ? 1 : 0);
ignore_fpstatus = !npy_isnan(in1) && in2 == 0;
ignore_fpstatus |= (in1 == NPY_INFINITY || in1 == -NPY_INFINITY) && !npy_isnan(in2);
ignore_fpstatus |= (in1 != NPY_INFINITY && in1 != -NPY_INFINITY) && (in2 == NPY_INFINITY || in2 == -NPY_INFINITY);
divmod@c@(in1, in2, (@type@ *)op1);
}
}
Expand Down Expand Up @@ -2376,10 +2373,10 @@ mkl_umath_@TYPE@_ldexp_long(char **args, const npy_intp *dimensions, const npy_i
* complex types
* #TYPE = CFLOAT, CDOUBLE#
* #ftype = npy_float, npy_double#
* #type = npy_cfloat, npy_cdouble#
* #c = f, #
* #C = F, #
* #s = s, d#
* #SUPPORTED_BY_VML = 1, 1#
* #C = F, #
* #s = c, z#
*/

/* similar to pairwise sum of real floats */
Expand Down Expand Up @@ -2659,44 +2656,47 @@ mkl_umath_@TYPE@__ones_like(char **args, const npy_intp *dimensions, const npy_i
}
}

/* TODO: USE MKL */
void
mkl_umath_@TYPE@_conjugate(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func)) {
UNARY_LOOP {
const @ftype@ in1r = ((@ftype@ *)ip1)[0];
const @ftype@ in1i = ((@ftype@ *)ip1)[1];
((@ftype@ *)op1)[0] = in1r;
((@ftype@ *)op1)[1] = -in1i;
}
const int contig = IS_UNARY_CONT(@type@, @type@);
const int disjoint_or_same = DISJOINT_OR_SAME(args[0], args[1], dimensions[0], sizeof(@type@));
const int can_vectorize = contig && disjoint_or_same;

if(can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
CHUNKED_VML_CALL2(v@s@Conj, dimensions[0], @type@, args[0], args[1]);
/* v@s@Conj(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
} else {
UNARY_LOOP {
const @ftype@ in1r = ((@ftype@ *)ip1)[0];
const @ftype@ in1i = ((@ftype@ *)ip1)[1];
((@ftype@ *)op1)[0] = in1r;
((@ftype@ *)op1)[1] = -in1i;
}
}
}

/* TODO: USE MKL */
void
mkl_umath_@TYPE@_absolute(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func))
{
const int contig = IS_UNARY_CONT(@type@, @ftype@);
const int disjoint_or_same = DISJOINT_OR_SAME_TWO_DTYPES(args[0], args[1], dimensions[0], sizeof(@type@), sizeof(@ftype@));
const int can_vectorize = contig && disjoint_or_same;
int ignore_fpstatus = 0;

// FIXME: abs function VML for complex numbers breaks FFT test_basic.py
//if(steps[0]/2 == sizeof(@ftype@) && steps[1] == sizeof(@ftype@) && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
#if @SUPPORTED_BY_VML@
if(0 == 1) {

if(can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
ignore_fpstatus = 1;
CHUNKED_VML_CALL2(v@s@Abs, dimensions[0], @ftype@, args[0], args[1]);
/* v@s@Abs(dimensions[0], (@ftype@ *) args[0], (@ftype@ *) args[1]); */
} else
#endif
{
CHUNKED_VML_CALL2(v@s@Abs, dimensions[0], @type@, args[0], args[1]);
/* v@s@Abs(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
} else {
UNARY_LOOP {
const @ftype@ in1r = ((@ftype@ *)ip1)[0];
const @ftype@ in1i = ((@ftype@ *)ip1)[1];
if(in1r == 0.0 && in1i == 0.0){
ignore_fpstatus = 1;
}
ignore_fpstatus = npy_isnan(in1r) && npy_isnan(in1i);
*((@ftype@ *)op1) = hypot@c@(in1r, in1i);
}
}
if(ignore_fpstatus) {
feclearexcept(FE_DIVBYZERO | FE_OVERFLOW | FE_UNDERFLOW | FE_INVALID);
feclearexcept(FE_INVALID);
}
}

Expand Down
Loading