Skip to content

Commit eb7d6d2

Browse files
JohanMabilleJohanMabille
committed
Fixed fmin, fmax and fabs for integer batches (ARM)
1 parent 8e13e57 commit eb7d6d2

10 files changed

+84
-160
lines changed

‎include/xsimd/types/xsimd_neon_int16.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515

1616
#include "xsimd_base.hpp"
1717
#include "xsimd_neon_bool.hpp"
18+
#include "xsimd_neon_int_base.hpp"
1819
#include "xsimd_neon_utils.hpp"
1920

2021
namespace xsimd
@@ -224,6 +225,7 @@ namespace xsimd
224225

225226
template <>
226227
struct batch_kernel<int16_t, 8>
228+
: neon_int_kernel_base<batch<int16_t, 8>>
227229
{
228230
using batch_type = batch<int16_t, 8>;
229231
using value_type = int16_t;
@@ -323,26 +325,6 @@ namespace xsimd
323325
return vabsq_s16(rhs);
324326
}
325327

326-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
327-
{
328-
return x * y + z;
329-
}
330-
331-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
332-
{
333-
return x * y - z;
334-
}
335-
336-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
337-
{
338-
return -x * y + z;
339-
}
340-
341-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
342-
{
343-
return -x * y - z;
344-
}
345-
346328
static value_type hadd(const batch_type& rhs)
347329
{
348330
#if XSIMD_ARM_INSTR_SET >= XSIMD_ARM8_64_NEON_VERSION

‎include/xsimd/types/xsimd_neon_int32.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515

1616
#include "xsimd_base.hpp"
1717
#include "xsimd_neon_bool.hpp"
18+
#include "xsimd_neon_int_base.hpp"
1819
#include "xsimd_neon_utils.hpp"
1920

2021
namespace xsimd
@@ -304,6 +305,7 @@ namespace xsimd
304305
{
305306
template <>
306307
struct batch_kernel<int32_t, 4>
308+
: neon_int_kernel_base<batch<int32_t, 4>>
307309
{
308310
using batch_type = batch<int32_t, 4>;
309311
using value_type = int32_t;
@@ -407,26 +409,6 @@ namespace xsimd
407409
return vabsq_s32(rhs);
408410
}
409411

410-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
411-
{
412-
return x * y + z;
413-
}
414-
415-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
416-
{
417-
return x * y - z;
418-
}
419-
420-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
421-
{
422-
return -x * y + z;
423-
}
424-
425-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
426-
{
427-
return -x * y - z;
428-
}
429-
430412
static value_type hadd(const batch_type& rhs)
431413
{
432414
#if XSIMD_ARM_INSTR_SET >= XSIMD_ARM8_64_NEON_VERSION

‎include/xsimd/types/xsimd_neon_int64.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -12,6 +12,7 @@
1212
#define XSIMD_NEON_INT64_HPP
1313

1414
#include "xsimd_base.hpp"
15+
#include "xsimd_neon_int_base.hpp"
1516

1617
namespace xsimd
1718
{
@@ -303,6 +304,7 @@ namespace xsimd
303304
{
304305
template <>
305306
struct batch_kernel<int64_t, 2>
307+
: neon_int_kernel_base<batch<int64_t, 2>>
306308
{
307309
using batch_type = batch<int64_t, 2>;
308310
using value_type = int64_t;
@@ -424,26 +426,6 @@ namespace xsimd
424426
#endif
425427
}
426428

427-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
428-
{
429-
return x * y + z;
430-
}
431-
432-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
433-
{
434-
return x * y - z;
435-
}
436-
437-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
438-
{
439-
return -x * y + z;
440-
}
441-
442-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
443-
{
444-
return -x * y - z;
445-
}
446-
447429
static value_type hadd(const batch_type& rhs)
448430
{
449431
#if XSIMD_ARM_INSTR_SET >= XSIMD_ARM8_64_NEON_VERSION

‎include/xsimd/types/xsimd_neon_int8.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515

1616
#include "xsimd_base.hpp"
1717
#include "xsimd_neon_bool.hpp"
18+
#include "xsimd_neon_int_base.hpp"
1819
#include "xsimd_neon_utils.hpp"
1920

2021
namespace xsimd
@@ -224,6 +225,7 @@ namespace xsimd
224225

225226
template <>
226227
struct batch_kernel<int8_t, 16>
228+
: neon_int_kernel_base<batch<int8_t, 16>>
227229
{
228230
using batch_type = batch<int8_t, 16>;
229231
using value_type = int8_t;
@@ -323,26 +325,6 @@ namespace xsimd
323325
return vabsq_s8(rhs);
324326
}
325327

326-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
327-
{
328-
return x * y + z;
329-
}
330-
331-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
332-
{
333-
return x * y - z;
334-
}
335-
336-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
337-
{
338-
return -x * y + z;
339-
}
340-
341-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
342-
{
343-
return -x * y - z;
344-
}
345-
346328
// Not implemented yet
347329
static value_type hadd(const batch_type& rhs)
348330
{

‎include/xsimd/types/xsimd_neon_int_base.hpp‎

Lines changed: 61 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,61 @@
1+
/***************************************************************************
2+
* Copyright (c) Johan Mabille, Sylvain Corlay, Wolf Vollprecht and *
3+
* Martin Renou *
4+
* Copyright (c) QuantStack *
5+
* *
6+
* Distributed under the terms of the BSD 3-Clause License. *
7+
* *
8+
* The full license is in the file LICENSE, distributed with this software. *
9+
****************************************************************************/
10+
11+
#ifndef XSIMD_NEON_INT_BASE_HPP
12+
#define XSIMD_NEON_INT_BASE_HPP
13+
14+
namespace xsimd
15+
{
16+
namespace detail
17+
{
18+
template <class B>
19+
struct neon_int_kernel_base
20+
{
21+
using batch_type = B;
22+
23+
static batch_type fmin(const batch_type& lhs, const batch_type& rhs)
24+
{
25+
return min(lhs, rhs);
26+
}
27+
28+
static batch_type fmax(const batch_type& lhs, const batch_type& rhs)
29+
{
30+
return max(lhs, rhs);
31+
}
32+
33+
static batch_type fabs(const batch_type& rhs)
34+
{
35+
return abs(rhs);
36+
}
37+
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
38+
{
39+
return x * y + z;
40+
}
41+
42+
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
43+
{
44+
return x * y - z;
45+
}
46+
47+
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
48+
{
49+
return -x * y + z;
50+
}
51+
52+
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
53+
{
54+
return -x * y - z;
55+
}
56+
};
57+
}
58+
}
59+
60+
#endif
61+

‎include/xsimd/types/xsimd_neon_uint16.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515

1616
#include "xsimd_base.hpp"
1717
#include "xsimd_neon_bool.hpp"
18+
#include "xsimd_neon_int_base.hpp"
1819
#include "xsimd_neon_utils.hpp"
1920

2021
namespace xsimd
@@ -197,6 +198,7 @@ namespace xsimd
197198
{
198199
template <>
199200
struct batch_kernel<uint16_t, 8>
201+
: neon_int_kernel_base<batch<uint16_t, 8>>
200202
{
201203
using batch_type = batch<uint16_t, 8>;
202204
using value_type = uint16_t;
@@ -296,26 +298,6 @@ namespace xsimd
296298
return rhs;
297299
}
298300

299-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
300-
{
301-
return x * y + z;
302-
}
303-
304-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
305-
{
306-
return x * y - z;
307-
}
308-
309-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
310-
{
311-
return -x * y + z;
312-
}
313-
314-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
315-
{
316-
return -x * y - z;
317-
}
318-
319301
// Not implemented yet
320302
static value_type hadd(const batch_type& rhs)
321303
{

‎include/xsimd/types/xsimd_neon_uint32.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -15,6 +15,7 @@
1515

1616
#include "xsimd_base.hpp"
1717
#include "xsimd_neon_bool.hpp"
18+
#include "xsimd_neon_int_base.hpp"
1819
#include "xsimd_neon_utils.hpp"
1920

2021
namespace xsimd
@@ -295,6 +296,7 @@ namespace xsimd
295296
{
296297
template <>
297298
struct batch_kernel<uint32_t, 4>
299+
: neon_int_kernel_base<batch<uint32_t, 4>>
298300
{
299301
using batch_type = batch<uint32_t, 4>;
300302
using value_type = uint32_t;
@@ -398,26 +400,6 @@ namespace xsimd
398400
return rhs;
399401
}
400402

401-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
402-
{
403-
return x * y + z;
404-
}
405-
406-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
407-
{
408-
return x * y - z;
409-
}
410-
411-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
412-
{
413-
return -x * y + z;
414-
}
415-
416-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
417-
{
418-
return -x * y - z;
419-
}
420-
421403
static value_type hadd(const batch_type& rhs)
422404
{
423405
#if XSIMD_ARM_INSTR_SET >= XSIMD_ARM8_64_NEON_VERSION

‎include/xsimd/types/xsimd_neon_uint64.hpp‎

Lines changed: 2 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -12,6 +12,7 @@
1212
#define XSIMD_NEON_UINT64_HPP
1313

1414
#include "xsimd_base.hpp"
15+
#include "xsimd_neon_int_base.hpp"
1516

1617
namespace xsimd
1718
{
@@ -299,6 +300,7 @@ namespace xsimd
299300
{
300301
template <>
301302
struct batch_kernel<uint64_t, 2>
303+
: neon_int_kernel_base<batch<uint64_t, 2>>
302304
{
303305
using batch_type = batch<uint64_t, 2>;
304306
using value_type = uint64_t;
@@ -452,26 +454,6 @@ namespace xsimd
452454
return rhs;
453455
}
454456

455-
static batch_type fma(const batch_type& x, const batch_type& y, const batch_type& z)
456-
{
457-
return x * y + z;
458-
}
459-
460-
static batch_type fms(const batch_type& x, const batch_type& y, const batch_type& z)
461-
{
462-
return x * y - z;
463-
}
464-
465-
static batch_type fnma(const batch_type& x, const batch_type& y, const batch_type& z)
466-
{
467-
return -x * y + z;
468-
}
469-
470-
static batch_type fnms(const batch_type& x, const batch_type& y, const batch_type& z)
471-
{
472-
return -x * y - z;
473-
}
474-
475457
static value_type hadd(const batch_type& rhs)
476458
{
477459
#if XSIMD_ARM_INSTR_SET >= XSIMD_ARM8_64_NEON_VERSION

0 commit comments

Comments
 (0)