Skip to content

Commit acf9f15

Browse files
antonwolfyantonwolfy
authored
Merge pull request #155 from IntelPython/fix-compilation-error
Resolve compilation errors with ICX compiler from 2026.0 release
2 parents bb8d15c + 726c611 commit acf9f15

2 files changed

Lines changed: 45 additions & 30 deletions

File tree

‎CHANGELOG.md‎

Lines changed: 2 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -5,6 +5,8 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
55
and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
66

77
## [dev] - YYYY-MM-DD
8+
### Fixed
9+
* Build with ICX compiler from 2026.0 release [gh-155](https://github.com/IntelPython/mkl_umath/pull/155)
810

911
### Removed
1012
* Dropped support for Python 3.9 [gh-125](https://github.com/IntelPython/mkl_umath/pull/125)

‎mkl_umath/src/mkl_umath_loops.c.src‎

Lines changed: 43 additions & 30 deletions
Original file line numberDiff line numberDiff line change
@@ -73,22 +73,23 @@
7373

7474
#define MKL_INT_MAX ((npy_intp) ((~((MKL_UINT) 0)) >> 1))
7575

76-
#define CHUNKED_VML_CALL2(vml_func, n, type, in1, op1) \
76+
#define CHUNKED_VML_CALL2(vml_func, n, type, mkl_type, mkl_ftype, in1, op1) \
7777
do { \
7878
npy_intp _n_ = (n); \
7979
const npy_intp _chunk_size = MKL_INT_MAX; \
8080
type *in1p = (type *) (in1); \
8181
type *op1p = (type *) (op1); \
8282
while (_n_ > 0) { \
8383
npy_intp _current_chunk = (_n_ > _chunk_size) ? _chunk_size : _n_; \
84-
vml_func((MKL_INT) _current_chunk, in1p, op1p); \
84+
vml_func((MKL_INT) _current_chunk, (mkl_type *)(void *) in1p, \
85+
(mkl_ftype *)(void *) op1p); \
8586
_n_ -= _current_chunk; \
8687
in1p += _current_chunk; \
8788
op1p += _current_chunk; \
8889
} \
8990
} while (0)
9091

91-
#define CHUNKED_VML_CALL3(vml_func, n, type, in1, in2, op1) \
92+
#define CHUNKED_VML_CALL3(vml_func, n, type, mkl_type, in1, in2, op1) \
9293
do { \
9394
npy_intp _n_ = (n); \
9495
const npy_intp _chunk_size = MKL_INT_MAX; \
@@ -97,7 +98,8 @@
9798
type *op1p = (type *) (op1); \
9899
while (_n_ > 0) { \
99100
npy_intp _current_chunk = (_n_ > _chunk_size) ? _chunk_size : _n_; \
100-
vml_func((MKL_INT) _current_chunk, in1p, in2p, op1p); \
101+
vml_func((MKL_INT) _current_chunk, (mkl_type *)(void *) in1p, \
102+
(mkl_type *)(void *) in2p, (mkl_type *)(void *) op1p); \
101103
_n_ -= _current_chunk; \
102104
in1p += _current_chunk; \
103105
in2p += _current_chunk; \
@@ -323,7 +325,7 @@ mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp
323325

324326
if (IS_BINARY_CONT(@type@, @type@)) {
325327
if (dimensions[0] > VML_ASM_THRESHOLD && disjoint_or_same1 && disjoint_or_same2) {
326-
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, args[0], args[1], args[2]);
328+
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, @type@, args[0], args[1], args[2]);
327329
/* v@s@@VML@(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
328330
}
329331
else {
@@ -482,7 +484,7 @@ mkl_umath_@TYPE@_multiply(char **args, const npy_intp *dimensions, const npy_int
482484

483485
if (IS_BINARY_CONT(@type@, @type@)) {
484486
if (dimensions[0] > VML_ASM_THRESHOLD && disjoint_or_same1 && disjoint_or_same2) {
485-
CHUNKED_VML_CALL3(v@s@Mul, dimensions[0], @type@, args[0], args[1], args[2]);
487+
CHUNKED_VML_CALL3(v@s@Mul, dimensions[0], @type@, @type@, args[0], args[1], args[2]);
486488
/* v@s@Mul(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
487489
}
488490
else {
@@ -633,7 +635,7 @@ mkl_umath_@TYPE@_divide(char **args, const npy_intp *dimensions, const npy_intp
633635

634636
if (IS_BINARY_CONT(@type@, @type@)) {
635637
if (dimensions[0] > VML_D_THRESHOLD && disjoint_or_same1 && disjoint_or_same2) {
636-
CHUNKED_VML_CALL3(v@s@Div, dimensions[0], @type@, args[0], args[1], args[2]);
638+
CHUNKED_VML_CALL3(v@s@Div, dimensions[0], @type@, @type@, args[0], args[1], args[2]);
637639
/* v@s@Div(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
638640
}
639641
else {
@@ -784,7 +786,7 @@ mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp
784786
const int can_vectorize = contig && disjoint_or_same1 && disjoint_or_same2;
785787

786788
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
787-
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, args[0], args[1], args[2]);
789+
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, @type@, args[0], args[1], args[2]);
788790
/* v@s@@VML@(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
789791
}
790792
else {
@@ -822,7 +824,7 @@ mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp
822824
const int can_vectorize = contig && disjoint_or_same1 && disjoint_or_same2;
823825

824826
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
825-
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, args[0], args[1], args[2]);
827+
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, @type@, args[0], args[1], args[2]);
826828
/* v@s@@VML@(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
827829
}
828830
else {
@@ -849,7 +851,7 @@ mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp
849851

850852
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD)
851853
{
852-
CHUNKED_VML_CALL2(v@s@@VML@, dimensions[0], @type@, args[0], args[1]);
854+
CHUNKED_VML_CALL2(v@s@@VML@, dimensions[0], @type@, @type@, @type@, args[0], args[1]);
853855
/* v@s@@VML@(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
854856
}
855857
else {
@@ -880,7 +882,7 @@ mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp
880882
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD)
881883
{
882884
ignore_fpstatus = 1;
883-
CHUNKED_VML_CALL2(v@s@@VML@, dimensions[0], @type@, args[0], args[1]);
885+
CHUNKED_VML_CALL2(v@s@@VML@, dimensions[0], @type@, @type@, @type@, args[0], args[1]);
884886
/* v@s@Exp(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
885887
}
886888
else {
@@ -909,7 +911,7 @@ mkl_umath_@TYPE@_absolute(char **args, const npy_intp *dimensions, const npy_int
909911
const int can_vectorize = contig && disjoint_or_same;
910912

911913
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
912-
CHUNKED_VML_CALL2(v@s@Abs, dimensions[0], @type@, args[0], args[1]);
914+
CHUNKED_VML_CALL2(v@s@Abs, dimensions[0], @type@, @type@, @type@, args[0], args[1]);
913915
/* v@s@Abs(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
914916
}
915917
else {
@@ -931,7 +933,7 @@ mkl_umath_@TYPE@_reciprocal(char **args, const npy_intp *dimensions, const npy_i
931933
const int can_vectorize = contig && disjoint_or_same;
932934

933935
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
934-
CHUNKED_VML_CALL2(v@s@Inv, dimensions[0], @type@, args[0], args[1]);
936+
CHUNKED_VML_CALL2(v@s@Inv, dimensions[0], @type@, @type@, @type@, args[0], args[1]);
935937
/* v@s@Inv(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
936938
}
937939
else {
@@ -950,7 +952,7 @@ mkl_umath_@TYPE@_square(char **args, const npy_intp *dimensions, const npy_intp
950952
const int can_vectorize = contig && disjoint_or_same;
951953

952954
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
953-
CHUNKED_VML_CALL2(v@s@Sqr, dimensions[0], @type@, args[0], args[1]);
955+
CHUNKED_VML_CALL2(v@s@Sqr, dimensions[0], @type@, @type@, @type@, args[0], args[1]);
954956
/* v@s@Sqr(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
955957
}
956958
else {
@@ -1187,7 +1189,9 @@ mkl_umath_@TYPE@_ldexp_long(char **args, const npy_intp *dimensions, const npy_i
11871189
* complex types
11881190
* #TYPE = CFLOAT, CDOUBLE#
11891191
* #ftype = npy_float, npy_double#
1190-
* #type = npy_cfloat, npy_cdouble#
1192+
* #type = npy_cfloat, npy_cdouble#
1193+
* #mkl_type = MKL_Complex8, MKL_Complex16#
1194+
* #mkl_ftype = float, double#
11911195
* #c = f, #
11921196
* #C = F, #
11931197
* #s = c, z#
@@ -1281,14 +1285,14 @@ pairwise_sum_@TYPE@(@ftype@ *rr, @ftype@ * ri, char * a, npy_intp n, npy_intp st
12811285
void
12821286
mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func))
12831287
{
1284-
const int contig = IS_BINARY_CONT(@type@, @type@);
1288+
const int contig = IS_BINARY_CONT(@mkl_type@, @mkl_type@);
12851289
const int disjoint_or_same1 = DISJOINT_OR_SAME(args[0], args[2], dimensions[0], sizeof(@type@));
12861290
const int disjoint_or_same2 = DISJOINT_OR_SAME(args[1], args[2], dimensions[0], sizeof(@type@));
12871291
const int can_vectorize = contig && disjoint_or_same1 && disjoint_or_same2;
12881292

12891293
if (can_vectorize && dimensions[0] > VML_ASM_THRESHOLD) {
1290-
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, args[0], args[1], args[2]);
1291-
/* v@s@@VML@(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
1294+
CHUNKED_VML_CALL3(v@s@@VML@, dimensions[0], @type@, @mkl_type@, args[0], args[1], args[2]);
1295+
/* v@s@@VML@(dimensions[0], (@mkl_type@*) args[0], (@mkl_type@*) args[1], (@mkl_type@*) args[2]); */
12921296
}
12931297
else {
12941298
if (IS_BINARY_REDUCE && @PW@) {
@@ -1319,14 +1323,14 @@ mkl_umath_@TYPE@_@kind@(char **args, const npy_intp *dimensions, const npy_intp
13191323
void
13201324
mkl_umath_@TYPE@_multiply(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func))
13211325
{
1322-
const int contig = IS_BINARY_CONT(@type@, @type@);
1326+
const int contig = IS_BINARY_CONT(@mkl_type@, @mkl_type@);
13231327
const int disjoint_or_same1 = DISJOINT_OR_SAME(args[0], args[2], dimensions[0], sizeof(@type@));
13241328
const int disjoint_or_same2 = DISJOINT_OR_SAME(args[1], args[2], dimensions[0], sizeof(@type@));
13251329
const int can_vectorize = contig && disjoint_or_same1 && disjoint_or_same2;
13261330

13271331
if (can_vectorize && dimensions[0] > VML_ASM_THRESHOLD) {
1328-
CHUNKED_VML_CALL3(v@s@Mul, dimensions[0], @type@, args[0], args[1], args[2]);
1329-
/* v@s@Mul(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
1332+
CHUNKED_VML_CALL3(v@s@Mul, dimensions[0], @type@, @mkl_type@, args[0], args[1], args[2]);
1333+
/* v@s@Mul(dimensions[0], (@mkl_type@*) args[0], (@mkl_type@*) args[1], (@mkl_type@*) args[2]); */
13301334
}
13311335
else {
13321336
BINARY_LOOP {
@@ -1343,14 +1347,14 @@ mkl_umath_@TYPE@_multiply(char **args, const npy_intp *dimensions, const npy_int
13431347
void
13441348
mkl_umath_@TYPE@_divide(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func))
13451349
{
1346-
const int contig = IS_BINARY_CONT(@type@, @type@);
1350+
const int contig = IS_BINARY_CONT(@mkl_type@, @mkl_type@);
13471351
const int disjoint_or_same1 = DISJOINT_OR_SAME(args[0], args[2], dimensions[0], sizeof(@type@));
13481352
const int disjoint_or_same2 = DISJOINT_OR_SAME(args[1], args[2], dimensions[0], sizeof(@type@));
13491353
const int can_vectorize = contig && disjoint_or_same1 && disjoint_or_same2;
13501354

13511355
if (can_vectorize && dimensions[0] > VML_D_THRESHOLD) {
1352-
CHUNKED_VML_CALL3(v@s@Div, dimensions[0], @type@, args[0], args[1], args[2]);
1353-
/* v@s@Div(dimensions[0], (@type@*) args[0], (@type@*) args[1], (@type@*) args[2]); */
1356+
CHUNKED_VML_CALL3(v@s@Div, dimensions[0], @type@, @mkl_type@, args[0], args[1], args[2]);
1357+
/* v@s@Div(dimensions[0], (@mkl_type@*) args[0], (@mkl_type@*) args[1], (@mkl_type@*) args[2]); */
13541358
}
13551359
else {
13561360
BINARY_LOOP {
@@ -1493,13 +1497,13 @@ mkl_umath_@TYPE@_reciprocal(char **args, const npy_intp *dimensions, const npy_i
14931497

14941498
void
14951499
mkl_umath_@TYPE@_conjugate(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func)) {
1496-
const int contig = IS_UNARY_CONT(@type@, @type@);
1500+
const int contig = IS_UNARY_CONT(@mkl_type@, @mkl_type@);
14971501
const int disjoint_or_same = DISJOINT_OR_SAME(args[0], args[1], dimensions[0], sizeof(@type@));
14981502
const int can_vectorize = contig && disjoint_or_same;
14991503

15001504
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
1501-
CHUNKED_VML_CALL2(v@s@Conj, dimensions[0], @type@, args[0], args[1]);
1502-
/* v@s@Conj(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
1505+
CHUNKED_VML_CALL2(v@s@Conj, dimensions[0], @type@, @mkl_type@, @mkl_type@, args[0], args[1]);
1506+
/* v@s@Conj(dimensions[0], (@mkl_type@*) args[0], (@mkl_type@*) args[1]); */
15031507
}
15041508
else {
15051509
UNARY_LOOP {
@@ -1511,18 +1515,27 @@ mkl_umath_@TYPE@_conjugate(char **args, const npy_intp *dimensions, const npy_in
15111515
}
15121516
}
15131517

1518+
/**begin repeat
1519+
* complex types
1520+
* #TYPE = CFLOAT, CDOUBLE#
1521+
* #ftype = npy_float, npy_double#
1522+
* #type = npy_cfloat, npy_cdouble#
1523+
* #mkl_ftype = float, double#
1524+
* #mkl_type = MKL_Complex8, MKL_Complex16#
1525+
1526+
*/
15141527
void
15151528
mkl_umath_@TYPE@_absolute(char **args, const npy_intp *dimensions, const npy_intp *steps, void *NPY_UNUSED(func))
15161529
{
1517-
const int contig = IS_UNARY_CONT(@type@, @ftype@);
1530+
const int contig = IS_UNARY_CONT(@mkl_type@, @mkl_ftype@);
15181531
const int disjoint_or_same = DISJOINT_OR_SAME_TWO_DTYPES(args[0], args[1], dimensions[0], sizeof(@type@), sizeof(@ftype@));
15191532
const int can_vectorize = contig && disjoint_or_same;
15201533
int ignore_fpstatus = 0;
15211534

15221535
if (can_vectorize && dimensions[0] > VML_TRANSCEDENTAL_THRESHOLD) {
15231536
ignore_fpstatus = 1;
1524-
CHUNKED_VML_CALL2(v@s@Abs, dimensions[0], @type@, args[0], args[1]);
1525-
/* v@s@Abs(dimensions[0], (@type@*) args[0], (@type@*) args[1]); */
1537+
CHUNKED_VML_CALL2(v@s@Abs, dimensions[0], @type@, @mkl_type@, @mkl_ftype@, args[0], args[1]);
1538+
/* v@s@Abs(dimensions[0], (@mkl_type@*) args[0], (@mkl_ftype@*) args[1]); */
15261539
}
15271540
else {
15281541
UNARY_LOOP {

0 commit comments

Comments
 (0)