1// SPDX-License-Identifier: Apache-2.0
2// ----------------------------------------------------------------------------
3// Copyright 2019-2023 Arm Limited
4//
5// Licensed under the Apache License, Version 2.0 (the "License"); you may not
6// use this file except in compliance with the License. You may obtain a copy
7// of the License at:
8//
9//     http://www.apache.org/licenses/LICENSE-2.0
10//
11// Unless required by applicable law or agreed to in writing, software
12// distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
13// WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
14// License for the specific language governing permissions and limitations
15// under the License.
16// ----------------------------------------------------------------------------
17
18/**
19 * @brief 4x32-bit vectors, implemented using plain C++.
20 *
21 * This module implements 4-wide 32-bit float, int, and mask vectors. This
22 * module provides a scalar fallback for VLA code, primarily useful for
23 * debugging VLA algorithms without the complexity of handling SIMD. Only the
24 * baseline level of functionality needed to support VLA is provided.
25 *
26 * Note that the vector conditional operators implemented by this module are
27 * designed to behave like SIMD conditional operators that generate lane masks.
28 * Rather than returning 0/1 booleans like normal C++ code they will return
29 * 0/-1 to give a full lane-width bitmask.
30 *
31 * Note that the documentation for this module still talks about "vectors" to
32 * help developers think about the implied VLA behavior when writing optimized
33 * paths.
34 */
35
36#ifndef ASTC_VECMATHLIB_NONE_4_H_INCLUDED
37#define ASTC_VECMATHLIB_NONE_4_H_INCLUDED
38
39#ifndef ASTCENC_SIMD_INLINE
40	#error "Include astcenc_vecmathlib.h, do not include directly"
41#endif
42
43#include <algorithm>
44#include <cstdio>
45#include <cstring>
46#include <cfenv>
47
48// ============================================================================
49// vfloat4 data type
50// ============================================================================
51
52/**
53 * @brief Data type for 4-wide floats.
54 */
55struct vfloat4
56{
57	/**
58	 * @brief Construct from zero-initialized value.
59	 */
60	ASTCENC_SIMD_INLINE vfloat4() = default;
61
62	/**
63	 * @brief Construct from 4 values loaded from an unaligned address.
64	 *
65	 * Consider using loada() which is better with wider VLA vectors if data is
66	 * aligned to vector length.
67	 */
68	ASTCENC_SIMD_INLINE explicit vfloat4(const float* p)
69	{
70		m[0] = p[0];
71		m[1] = p[1];
72		m[2] = p[2];
73		m[3] = p[3];
74	}
75
76	/**
77	 * @brief Construct from 4 scalar values replicated across all lanes.
78	 *
79	 * Consider using zero() for constexpr zeros.
80	 */
81	ASTCENC_SIMD_INLINE explicit vfloat4(float a)
82	{
83		m[0] = a;
84		m[1] = a;
85		m[2] = a;
86		m[3] = a;
87	}
88
89	/**
90	 * @brief Construct from 4 scalar values.
91	 *
92	 * The value of @c a is stored to lane 0 (LSB) in the SIMD register.
93	 */
94	ASTCENC_SIMD_INLINE explicit vfloat4(float a, float b, float c, float d)
95	{
96		m[0] = a;
97		m[1] = b;
98		m[2] = c;
99		m[3] = d;
100	}
101
102	/**
103	 * @brief Get the scalar value of a single lane.
104	 */
105	template <int l> ASTCENC_SIMD_INLINE float lane() const
106	{
107		return m[l];
108	}
109
110	/**
111	 * @brief Set the scalar value of a single lane.
112	 */
113	template <int l> ASTCENC_SIMD_INLINE void set_lane(float a)
114	{
115		m[l] = a;
116	}
117
118	/**
119	 * @brief Factory that returns a vector of zeros.
120	 */
121	static ASTCENC_SIMD_INLINE vfloat4 zero()
122	{
123		return vfloat4(0.0f);
124	}
125
126	/**
127	 * @brief Factory that returns a replicated scalar loaded from memory.
128	 */
129	static ASTCENC_SIMD_INLINE vfloat4 load1(const float* p)
130	{
131		return vfloat4(*p);
132	}
133
134	/**
135	 * @brief Factory that returns a vector loaded from aligned memory.
136	 */
137	static ASTCENC_SIMD_INLINE vfloat4 loada(const float* p)
138	{
139		return vfloat4(p);
140	}
141
142	/**
143	 * @brief Factory that returns a vector containing the lane IDs.
144	 */
145	static ASTCENC_SIMD_INLINE vfloat4 lane_id()
146	{
147		return vfloat4(0.0f, 1.0f, 2.0f, 3.0f);
148	}
149
150	/**
151	 * @brief Return a swizzled float 2.
152	 */
153	template <int l0, int l1> ASTCENC_SIMD_INLINE vfloat4 swz() const
154	{
155		return  vfloat4(lane<l0>(), lane<l1>(), 0.0f, 0.0f);
156	}
157
158	/**
159	 * @brief Return a swizzled float 3.
160	 */
161	template <int l0, int l1, int l2> ASTCENC_SIMD_INLINE vfloat4 swz() const
162	{
163		return vfloat4(lane<l0>(), lane<l1>(), lane<l2>(), 0.0f);
164	}
165
166	/**
167	 * @brief Return a swizzled float 4.
168	 */
169	template <int l0, int l1, int l2, int l3> ASTCENC_SIMD_INLINE vfloat4 swz() const
170	{
171		return vfloat4(lane<l0>(), lane<l1>(), lane<l2>(), lane<l3>());
172	}
173
174	/**
175	 * @brief The vector ...
176	 */
177	float m[4];
178};
179
180// ============================================================================
181// vint4 data type
182// ============================================================================
183
184/**
185 * @brief Data type for 4-wide ints.
186 */
187struct vint4
188{
189	/**
190	 * @brief Construct from zero-initialized value.
191	 */
192	ASTCENC_SIMD_INLINE vint4() = default;
193
194	/**
195	 * @brief Construct from 4 values loaded from an unaligned address.
196	 *
197	 * Consider using vint4::loada() which is better with wider VLA vectors
198	 * if data is aligned.
199	 */
200	ASTCENC_SIMD_INLINE explicit vint4(const int* p)
201	{
202		m[0] = p[0];
203		m[1] = p[1];
204		m[2] = p[2];
205		m[3] = p[3];
206	}
207
208	/**
209	 * @brief Construct from 4 uint8_t loaded from an unaligned address.
210	 */
211	ASTCENC_SIMD_INLINE explicit vint4(const uint8_t *p)
212	{
213		m[0] = p[0];
214		m[1] = p[1];
215		m[2] = p[2];
216		m[3] = p[3];
217	}
218
219	/**
220	 * @brief Construct from 4 scalar values.
221	 *
222	 * The value of @c a is stored to lane 0 (LSB) in the SIMD register.
223	 */
224	ASTCENC_SIMD_INLINE explicit vint4(int a, int b, int c, int d)
225	{
226		m[0] = a;
227		m[1] = b;
228		m[2] = c;
229		m[3] = d;
230	}
231
232
233	/**
234	 * @brief Construct from 4 scalar values replicated across all lanes.
235	 *
236	 * Consider using vint4::zero() for constexpr zeros.
237	 */
238	ASTCENC_SIMD_INLINE explicit vint4(int a)
239	{
240		m[0] = a;
241		m[1] = a;
242		m[2] = a;
243		m[3] = a;
244	}
245
246	/**
247	 * @brief Get the scalar value of a single lane.
248	 */
249	template <int l> ASTCENC_SIMD_INLINE int lane() const
250	{
251		return m[l];
252	}
253
254	/**
255	 * @brief Set the scalar value of a single lane.
256	 */
257	template <int l> ASTCENC_SIMD_INLINE void set_lane(int a)
258	{
259		m[l] = a;
260	}
261
262	/**
263	 * @brief Factory that returns a vector of zeros.
264	 */
265	static ASTCENC_SIMD_INLINE vint4 zero()
266	{
267		return vint4(0);
268	}
269
270	/**
271	 * @brief Factory that returns a replicated scalar loaded from memory.
272	 */
273	static ASTCENC_SIMD_INLINE vint4 load1(const int* p)
274	{
275		return vint4(*p);
276	}
277
278	/**
279	 * @brief Factory that returns a vector loaded from unaligned memory.
280	 */
281	static ASTCENC_SIMD_INLINE vint4 load(const uint8_t* p)
282	{
283		vint4 data;
284		std::memcpy(&data.m, p, 4 * sizeof(int));
285		return data;
286	}
287
288	/**
289	 * @brief Factory that returns a vector loaded from 16B aligned memory.
290	 */
291	static ASTCENC_SIMD_INLINE vint4 loada(const int* p)
292	{
293		return vint4(p);
294	}
295
296	/**
297	 * @brief Factory that returns a vector containing the lane IDs.
298	 */
299	static ASTCENC_SIMD_INLINE vint4 lane_id()
300	{
301		return vint4(0, 1, 2, 3);
302	}
303
304	/**
305	 * @brief The vector ...
306	 */
307	int m[4];
308};
309
310// ============================================================================
311// vmask4 data type
312// ============================================================================
313
314/**
315 * @brief Data type for 4-wide control plane masks.
316 */
317struct vmask4
318{
319	/**
320	 * @brief Construct from an existing mask value.
321	 */
322	ASTCENC_SIMD_INLINE explicit vmask4(int* p)
323	{
324		m[0] = p[0];
325		m[1] = p[1];
326		m[2] = p[2];
327		m[3] = p[3];
328	}
329
330	/**
331	 * @brief Construct from 1 scalar value.
332	 */
333	ASTCENC_SIMD_INLINE explicit vmask4(bool a)
334	{
335		m[0] = a == false ? 0 : -1;
336		m[1] = a == false ? 0 : -1;
337		m[2] = a == false ? 0 : -1;
338		m[3] = a == false ? 0 : -1;
339	}
340
341	/**
342	 * @brief Construct from 4 scalar values.
343	 *
344	 * The value of @c a is stored to lane 0 (LSB) in the SIMD register.
345	 */
346	ASTCENC_SIMD_INLINE explicit vmask4(bool a, bool b, bool c, bool d)
347	{
348		m[0] = a == false ? 0 : -1;
349		m[1] = b == false ? 0 : -1;
350		m[2] = c == false ? 0 : -1;
351		m[3] = d == false ? 0 : -1;
352	}
353
354	/**
355	 * @brief Get the scalar value of a single lane.
356	 */
357	template <int l> ASTCENC_SIMD_INLINE float lane() const
358	{
359		return m[l] != 0;
360	}
361
362	/**
363	 * @brief The vector ...
364	 */
365	int m[4];
366};
367
368// ============================================================================
369// vmask4 operators and functions
370// ============================================================================
371
372/**
373 * @brief Overload: mask union (or).
374 */
375ASTCENC_SIMD_INLINE vmask4 operator|(vmask4 a, vmask4 b)
376{
377	return vmask4(a.m[0] | b.m[0],
378	              a.m[1] | b.m[1],
379	              a.m[2] | b.m[2],
380	              a.m[3] | b.m[3]);
381}
382
383/**
384 * @brief Overload: mask intersect (and).
385 */
386ASTCENC_SIMD_INLINE vmask4 operator&(vmask4 a, vmask4 b)
387{
388	return vmask4(a.m[0] & b.m[0],
389	              a.m[1] & b.m[1],
390	              a.m[2] & b.m[2],
391	              a.m[3] & b.m[3]);
392}
393
394/**
395 * @brief Overload: mask difference (xor).
396 */
397ASTCENC_SIMD_INLINE vmask4 operator^(vmask4 a, vmask4 b)
398{
399	return vmask4(a.m[0] ^ b.m[0],
400	              a.m[1] ^ b.m[1],
401	              a.m[2] ^ b.m[2],
402	              a.m[3] ^ b.m[3]);
403}
404
405/**
406 * @brief Overload: mask invert (not).
407 */
408ASTCENC_SIMD_INLINE vmask4 operator~(vmask4 a)
409{
410	return vmask4(~a.m[0],
411	              ~a.m[1],
412	              ~a.m[2],
413	              ~a.m[3]);
414}
415
416/**
417 * @brief Return a 1-bit mask code indicating mask status.
418 *
419 * bit0 = lane 0
420 */
421ASTCENC_SIMD_INLINE unsigned int mask(vmask4 a)
422{
423	return ((a.m[0] >> 31) & 0x1) |
424	       ((a.m[1] >> 30) & 0x2) |
425	       ((a.m[2] >> 29) & 0x4) |
426	       ((a.m[3] >> 28) & 0x8);
427}
428
429// ============================================================================
430// vint4 operators and functions
431// ============================================================================
432
433/**
434 * @brief Overload: vector by vector addition.
435 */
436ASTCENC_SIMD_INLINE vint4 operator+(vint4 a, vint4 b)
437{
438	return vint4(a.m[0] + b.m[0],
439	             a.m[1] + b.m[1],
440	             a.m[2] + b.m[2],
441	             a.m[3] + b.m[3]);
442}
443
444/**
445 * @brief Overload: vector by vector subtraction.
446 */
447ASTCENC_SIMD_INLINE vint4 operator-(vint4 a, vint4 b)
448{
449	return vint4(a.m[0] - b.m[0],
450	             a.m[1] - b.m[1],
451	             a.m[2] - b.m[2],
452	             a.m[3] - b.m[3]);
453}
454
455/**
456 * @brief Overload: vector by vector multiplication.
457 */
458ASTCENC_SIMD_INLINE vint4 operator*(vint4 a, vint4 b)
459{
460	return vint4(a.m[0] * b.m[0],
461	             a.m[1] * b.m[1],
462	             a.m[2] * b.m[2],
463	             a.m[3] * b.m[3]);
464}
465
466/**
467 * @brief Overload: vector bit invert.
468 */
469ASTCENC_SIMD_INLINE vint4 operator~(vint4 a)
470{
471	return vint4(~a.m[0],
472	             ~a.m[1],
473	             ~a.m[2],
474	             ~a.m[3]);
475}
476
477/**
478 * @brief Overload: vector by vector bitwise or.
479 */
480ASTCENC_SIMD_INLINE vint4 operator|(vint4 a, vint4 b)
481{
482	return vint4(a.m[0] | b.m[0],
483	             a.m[1] | b.m[1],
484	             a.m[2] | b.m[2],
485	             a.m[3] | b.m[3]);
486}
487
488/**
489 * @brief Overload: vector by vector bitwise and.
490 */
491ASTCENC_SIMD_INLINE vint4 operator&(vint4 a, vint4 b)
492{
493	return vint4(a.m[0] & b.m[0],
494	             a.m[1] & b.m[1],
495	             a.m[2] & b.m[2],
496	             a.m[3] & b.m[3]);
497}
498
499/**
500 * @brief Overload: vector by vector bitwise xor.
501 */
502ASTCENC_SIMD_INLINE vint4 operator^(vint4 a, vint4 b)
503{
504	return vint4(a.m[0] ^ b.m[0],
505	             a.m[1] ^ b.m[1],
506	             a.m[2] ^ b.m[2],
507	             a.m[3] ^ b.m[3]);
508}
509
510/**
511 * @brief Overload: vector by vector equality.
512 */
513ASTCENC_SIMD_INLINE vmask4 operator==(vint4 a, vint4 b)
514{
515	return vmask4(a.m[0] == b.m[0],
516	              a.m[1] == b.m[1],
517	              a.m[2] == b.m[2],
518	              a.m[3] == b.m[3]);
519}
520
521/**
522 * @brief Overload: vector by vector inequality.
523 */
524ASTCENC_SIMD_INLINE vmask4 operator!=(vint4 a, vint4 b)
525{
526	return vmask4(a.m[0] != b.m[0],
527	              a.m[1] != b.m[1],
528	              a.m[2] != b.m[2],
529	              a.m[3] != b.m[3]);
530}
531
532/**
533 * @brief Overload: vector by vector less than.
534 */
535ASTCENC_SIMD_INLINE vmask4 operator<(vint4 a, vint4 b)
536{
537	return vmask4(a.m[0] < b.m[0],
538	              a.m[1] < b.m[1],
539	              a.m[2] < b.m[2],
540	              a.m[3] < b.m[3]);
541}
542
543/**
544 * @brief Overload: vector by vector greater than.
545 */
546ASTCENC_SIMD_INLINE vmask4 operator>(vint4 a, vint4 b)
547{
548	return vmask4(a.m[0] > b.m[0],
549	              a.m[1] > b.m[1],
550	              a.m[2] > b.m[2],
551	              a.m[3] > b.m[3]);
552}
553
554/**
555 * @brief Logical shift left.
556 */
557template <int s> ASTCENC_SIMD_INLINE vint4 lsl(vint4 a)
558{
559	return vint4(a.m[0] << s,
560	             a.m[1] << s,
561	             a.m[2] << s,
562	             a.m[3] << s);
563}
564
565/**
566 * @brief Logical shift right.
567 */
568template <int s> ASTCENC_SIMD_INLINE vint4 lsr(vint4 a)
569{
570	unsigned int as0 = static_cast<unsigned int>(a.m[0]) >> s;
571	unsigned int as1 = static_cast<unsigned int>(a.m[1]) >> s;
572	unsigned int as2 = static_cast<unsigned int>(a.m[2]) >> s;
573	unsigned int as3 = static_cast<unsigned int>(a.m[3]) >> s;
574
575	return vint4(static_cast<int>(as0),
576	             static_cast<int>(as1),
577	             static_cast<int>(as2),
578	             static_cast<int>(as3));
579}
580
581/**
582 * @brief Arithmetic shift right.
583 */
584template <int s> ASTCENC_SIMD_INLINE vint4 asr(vint4 a)
585{
586	return vint4(a.m[0] >> s,
587	             a.m[1] >> s,
588	             a.m[2] >> s,
589	             a.m[3] >> s);
590}
591
592/**
593 * @brief Return the min vector of two vectors.
594 */
595ASTCENC_SIMD_INLINE vint4 min(vint4 a, vint4 b)
596{
597	return vint4(a.m[0] < b.m[0] ? a.m[0] : b.m[0],
598	             a.m[1] < b.m[1] ? a.m[1] : b.m[1],
599	             a.m[2] < b.m[2] ? a.m[2] : b.m[2],
600	             a.m[3] < b.m[3] ? a.m[3] : b.m[3]);
601}
602
603/**
604 * @brief Return the min vector of two vectors.
605 */
606ASTCENC_SIMD_INLINE vint4 max(vint4 a, vint4 b)
607{
608	return vint4(a.m[0] > b.m[0] ? a.m[0] : b.m[0],
609	             a.m[1] > b.m[1] ? a.m[1] : b.m[1],
610	             a.m[2] > b.m[2] ? a.m[2] : b.m[2],
611	             a.m[3] > b.m[3] ? a.m[3] : b.m[3]);
612}
613
614/**
615 * @brief Return the horizontal minimum of a single vector.
616 */
617ASTCENC_SIMD_INLINE vint4 hmin(vint4 a)
618{
619	int b = std::min(a.m[0], a.m[1]);
620	int c = std::min(a.m[2], a.m[3]);
621	return vint4(std::min(b, c));
622}
623
624/**
625 * @brief Return the horizontal maximum of a single vector.
626 */
627ASTCENC_SIMD_INLINE vint4 hmax(vint4 a)
628{
629	int b = std::max(a.m[0], a.m[1]);
630	int c = std::max(a.m[2], a.m[3]);
631	return vint4(std::max(b, c));
632}
633
634/**
635 * @brief Return the horizontal sum of vector lanes as a scalar.
636 */
637ASTCENC_SIMD_INLINE int hadd_s(vint4 a)
638{
639	return a.m[0] + a.m[1] + a.m[2] + a.m[3];
640}
641
642/**
643 * @brief Store a vector to an aligned memory address.
644 */
645ASTCENC_SIMD_INLINE void storea(vint4 a, int* p)
646{
647	p[0] = a.m[0];
648	p[1] = a.m[1];
649	p[2] = a.m[2];
650	p[3] = a.m[3];
651}
652
653/**
654 * @brief Store a vector to an unaligned memory address.
655 */
656ASTCENC_SIMD_INLINE void store(vint4 a, int* p)
657{
658	p[0] = a.m[0];
659	p[1] = a.m[1];
660	p[2] = a.m[2];
661	p[3] = a.m[3];
662}
663
664/**
665 * @brief Store a vector to an unaligned memory address.
666 */
667ASTCENC_SIMD_INLINE void store(vint4 a, uint8_t* p)
668{
669	std::memcpy(p, a.m, sizeof(int) * 4);
670}
671
672/**
673 * @brief Store lowest N (vector width) bytes into an unaligned address.
674 */
675ASTCENC_SIMD_INLINE void store_nbytes(vint4 a, uint8_t* p)
676{
677	std::memcpy(p, a.m, sizeof(uint8_t) * 4);
678}
679
680/**
681 * @brief Gather N (vector width) indices from the array.
682 */
683ASTCENC_SIMD_INLINE vint4 gatheri(const int* base, vint4 indices)
684{
685	return vint4(base[indices.m[0]],
686	             base[indices.m[1]],
687	             base[indices.m[2]],
688	             base[indices.m[3]]);
689}
690
691/**
692 * @brief Pack low 8 bits of N (vector width) lanes into bottom of vector.
693 */
694ASTCENC_SIMD_INLINE vint4 pack_low_bytes(vint4 a)
695{
696	int b0 = a.m[0] & 0xFF;
697	int b1 = a.m[1] & 0xFF;
698	int b2 = a.m[2] & 0xFF;
699	int b3 = a.m[3] & 0xFF;
700
701	int b = b0 | (b1 << 8) | (b2 << 16) | (b3 << 24);
702	return vint4(b, 0, 0, 0);
703}
704
705/**
706 * @brief Return lanes from @c b if MSB of @c cond is set, else @c a.
707 */
708ASTCENC_SIMD_INLINE vint4 select(vint4 a, vint4 b, vmask4 cond)
709{
710	return vint4((cond.m[0] & static_cast<int>(0x80000000)) ? b.m[0] : a.m[0],
711	             (cond.m[1] & static_cast<int>(0x80000000)) ? b.m[1] : a.m[1],
712	             (cond.m[2] & static_cast<int>(0x80000000)) ? b.m[2] : a.m[2],
713	             (cond.m[3] & static_cast<int>(0x80000000)) ? b.m[3] : a.m[3]);
714}
715
716// ============================================================================
717// vfloat4 operators and functions
718// ============================================================================
719
720/**
721 * @brief Overload: vector by vector addition.
722 */
723ASTCENC_SIMD_INLINE vfloat4 operator+(vfloat4 a, vfloat4 b)
724{
725	return vfloat4(a.m[0] + b.m[0],
726	               a.m[1] + b.m[1],
727	               a.m[2] + b.m[2],
728	               a.m[3] + b.m[3]);
729}
730
731/**
732 * @brief Overload: vector by vector subtraction.
733 */
734ASTCENC_SIMD_INLINE vfloat4 operator-(vfloat4 a, vfloat4 b)
735{
736	return vfloat4(a.m[0] - b.m[0],
737	               a.m[1] - b.m[1],
738	               a.m[2] - b.m[2],
739	               a.m[3] - b.m[3]);
740}
741
742/**
743 * @brief Overload: vector by vector multiplication.
744 */
745ASTCENC_SIMD_INLINE vfloat4 operator*(vfloat4 a, vfloat4 b)
746{
747	return vfloat4(a.m[0] * b.m[0],
748	               a.m[1] * b.m[1],
749	               a.m[2] * b.m[2],
750	               a.m[3] * b.m[3]);
751}
752
753/**
754 * @brief Overload: vector by vector division.
755 */
756ASTCENC_SIMD_INLINE vfloat4 operator/(vfloat4 a, vfloat4 b)
757{
758	return vfloat4(a.m[0] / b.m[0],
759	               a.m[1] / b.m[1],
760	               a.m[2] / b.m[2],
761	               a.m[3] / b.m[3]);
762}
763
764/**
765 * @brief Overload: vector by vector equality.
766 */
767ASTCENC_SIMD_INLINE vmask4 operator==(vfloat4 a, vfloat4 b)
768{
769	return vmask4(a.m[0] == b.m[0],
770	              a.m[1] == b.m[1],
771	              a.m[2] == b.m[2],
772	              a.m[3] == b.m[3]);
773}
774
775/**
776 * @brief Overload: vector by vector inequality.
777 */
778ASTCENC_SIMD_INLINE vmask4 operator!=(vfloat4 a, vfloat4 b)
779{
780	return vmask4(a.m[0] != b.m[0],
781	              a.m[1] != b.m[1],
782	              a.m[2] != b.m[2],
783	              a.m[3] != b.m[3]);
784}
785
786/**
787 * @brief Overload: vector by vector less than.
788 */
789ASTCENC_SIMD_INLINE vmask4 operator<(vfloat4 a, vfloat4 b)
790{
791	return vmask4(a.m[0] < b.m[0],
792	              a.m[1] < b.m[1],
793	              a.m[2] < b.m[2],
794	              a.m[3] < b.m[3]);
795}
796
797/**
798 * @brief Overload: vector by vector greater than.
799 */
800ASTCENC_SIMD_INLINE vmask4 operator>(vfloat4 a, vfloat4 b)
801{
802	return vmask4(a.m[0] > b.m[0],
803	              a.m[1] > b.m[1],
804	              a.m[2] > b.m[2],
805	              a.m[3] > b.m[3]);
806}
807
808/**
809 * @brief Overload: vector by vector less than or equal.
810 */
811ASTCENC_SIMD_INLINE vmask4 operator<=(vfloat4 a, vfloat4 b)
812{
813	return vmask4(a.m[0] <= b.m[0],
814	              a.m[1] <= b.m[1],
815	              a.m[2] <= b.m[2],
816	              a.m[3] <= b.m[3]);
817}
818
819/**
820 * @brief Overload: vector by vector greater than or equal.
821 */
822ASTCENC_SIMD_INLINE vmask4 operator>=(vfloat4 a, vfloat4 b)
823{
824	return vmask4(a.m[0] >= b.m[0],
825	              a.m[1] >= b.m[1],
826	              a.m[2] >= b.m[2],
827	              a.m[3] >= b.m[3]);
828}
829
830/**
831 * @brief Return the min vector of two vectors.
832 *
833 * If either lane value is NaN, @c b will be returned for that lane.
834 */
835ASTCENC_SIMD_INLINE vfloat4 min(vfloat4 a, vfloat4 b)
836{
837	return vfloat4(a.m[0] < b.m[0] ? a.m[0] : b.m[0],
838	               a.m[1] < b.m[1] ? a.m[1] : b.m[1],
839	               a.m[2] < b.m[2] ? a.m[2] : b.m[2],
840	               a.m[3] < b.m[3] ? a.m[3] : b.m[3]);
841}
842
843/**
844 * @brief Return the max vector of two vectors.
845 *
846 * If either lane value is NaN, @c b will be returned for that lane.
847 */
848ASTCENC_SIMD_INLINE vfloat4 max(vfloat4 a, vfloat4 b)
849{
850	return vfloat4(a.m[0] > b.m[0] ? a.m[0] : b.m[0],
851	               a.m[1] > b.m[1] ? a.m[1] : b.m[1],
852	               a.m[2] > b.m[2] ? a.m[2] : b.m[2],
853	               a.m[3] > b.m[3] ? a.m[3] : b.m[3]);
854}
855
856/**
857 * @brief Return the absolute value of the float vector.
858 */
859ASTCENC_SIMD_INLINE vfloat4 abs(vfloat4 a)
860{
861	return vfloat4(std::abs(a.m[0]),
862	               std::abs(a.m[1]),
863	               std::abs(a.m[2]),
864	               std::abs(a.m[3]));
865}
866
867/**
868 * @brief Return a float rounded to the nearest integer value.
869 */
870ASTCENC_SIMD_INLINE vfloat4 round(vfloat4 a)
871{
872	assert(std::fegetround() == FE_TONEAREST);
873	return vfloat4(std::nearbyint(a.m[0]),
874	               std::nearbyint(a.m[1]),
875	               std::nearbyint(a.m[2]),
876	               std::nearbyint(a.m[3]));
877}
878
879/**
880 * @brief Return the horizontal minimum of a vector.
881 */
882ASTCENC_SIMD_INLINE vfloat4 hmin(vfloat4 a)
883{
884	float tmp1 = std::min(a.m[0], a.m[1]);
885	float tmp2 = std::min(a.m[2], a.m[3]);
886	return vfloat4(std::min(tmp1, tmp2));
887}
888
889/**
890 * @brief Return the horizontal maximum of a vector.
891 */
892ASTCENC_SIMD_INLINE vfloat4 hmax(vfloat4 a)
893{
894	float tmp1 = std::max(a.m[0], a.m[1]);
895	float tmp2 = std::max(a.m[2], a.m[3]);
896	return vfloat4(std::max(tmp1, tmp2));
897}
898
899/**
900 * @brief Return the horizontal sum of a vector.
901 */
902ASTCENC_SIMD_INLINE float hadd_s(vfloat4 a)
903{
904	// Use halving add, gives invariance with SIMD versions
905	return (a.m[0] + a.m[2]) + (a.m[1] + a.m[3]);
906}
907
908/**
909 * @brief Return the sqrt of the lanes in the vector.
910 */
911ASTCENC_SIMD_INLINE vfloat4 sqrt(vfloat4 a)
912{
913	return vfloat4(std::sqrt(a.m[0]),
914	               std::sqrt(a.m[1]),
915	               std::sqrt(a.m[2]),
916	               std::sqrt(a.m[3]));
917}
918
919/**
920 * @brief Return lanes from @c b if @c cond is set, else @c a.
921 */
922ASTCENC_SIMD_INLINE vfloat4 select(vfloat4 a, vfloat4 b, vmask4 cond)
923{
924	return vfloat4((cond.m[0] & static_cast<int>(0x80000000)) ? b.m[0] : a.m[0],
925	               (cond.m[1] & static_cast<int>(0x80000000)) ? b.m[1] : a.m[1],
926	               (cond.m[2] & static_cast<int>(0x80000000)) ? b.m[2] : a.m[2],
927	               (cond.m[3] & static_cast<int>(0x80000000)) ? b.m[3] : a.m[3]);
928}
929
930/**
931 * @brief Return lanes from @c b if MSB of @c cond is set, else @c a.
932 */
933ASTCENC_SIMD_INLINE vfloat4 select_msb(vfloat4 a, vfloat4 b, vmask4 cond)
934{
935	return vfloat4((cond.m[0] & static_cast<int>(0x80000000)) ? b.m[0] : a.m[0],
936	               (cond.m[1] & static_cast<int>(0x80000000)) ? b.m[1] : a.m[1],
937	               (cond.m[2] & static_cast<int>(0x80000000)) ? b.m[2] : a.m[2],
938	               (cond.m[3] & static_cast<int>(0x80000000)) ? b.m[3] : a.m[3]);
939}
940
941/**
942 * @brief Load a vector of gathered results from an array;
943 */
944ASTCENC_SIMD_INLINE vfloat4 gatherf(const float* base, vint4 indices)
945{
946	return vfloat4(base[indices.m[0]],
947	               base[indices.m[1]],
948	               base[indices.m[2]],
949	               base[indices.m[3]]);
950}
951
952/**
953 * @brief Store a vector to an unaligned memory address.
954 */
955ASTCENC_SIMD_INLINE void store(vfloat4 a, float* ptr)
956{
957	ptr[0] = a.m[0];
958	ptr[1] = a.m[1];
959	ptr[2] = a.m[2];
960	ptr[3] = a.m[3];
961}
962
963/**
964 * @brief Store a vector to an aligned memory address.
965 */
966ASTCENC_SIMD_INLINE void storea(vfloat4 a, float* ptr)
967{
968	ptr[0] = a.m[0];
969	ptr[1] = a.m[1];
970	ptr[2] = a.m[2];
971	ptr[3] = a.m[3];
972}
973
974/**
975 * @brief Return a integer value for a float vector, using truncation.
976 */
977ASTCENC_SIMD_INLINE vint4 float_to_int(vfloat4 a)
978{
979	return vint4(static_cast<int>(a.m[0]),
980	             static_cast<int>(a.m[1]),
981	             static_cast<int>(a.m[2]),
982	             static_cast<int>(a.m[3]));
983}
984
985/**f
986 * @brief Return a integer value for a float vector, using round-to-nearest.
987 */
988ASTCENC_SIMD_INLINE vint4 float_to_int_rtn(vfloat4 a)
989{
990	a = a + vfloat4(0.5f);
991	return vint4(static_cast<int>(a.m[0]),
992	             static_cast<int>(a.m[1]),
993	             static_cast<int>(a.m[2]),
994	             static_cast<int>(a.m[3]));
995}
996
997/**
998 * @brief Return a float value for a integer vector.
999 */
1000ASTCENC_SIMD_INLINE vfloat4 int_to_float(vint4 a)
1001{
1002	return vfloat4(static_cast<float>(a.m[0]),
1003	               static_cast<float>(a.m[1]),
1004	               static_cast<float>(a.m[2]),
1005	               static_cast<float>(a.m[3]));
1006}
1007
1008/**
1009 * @brief Return a float16 value for a float vector, using round-to-nearest.
1010 */
1011ASTCENC_SIMD_INLINE vint4 float_to_float16(vfloat4 a)
1012{
1013	return vint4(
1014		float_to_sf16(a.lane<0>()),
1015		float_to_sf16(a.lane<1>()),
1016		float_to_sf16(a.lane<2>()),
1017		float_to_sf16(a.lane<3>()));
1018}
1019
1020/**
1021 * @brief Return a float16 value for a float scalar, using round-to-nearest.
1022 */
1023static inline uint16_t float_to_float16(float a)
1024{
1025	return float_to_sf16(a);
1026}
1027
1028/**
1029 * @brief Return a float value for a float16 vector.
1030 */
1031ASTCENC_SIMD_INLINE vfloat4 float16_to_float(vint4 a)
1032{
1033	return vfloat4(
1034		sf16_to_float(static_cast<uint16_t>(a.lane<0>())),
1035		sf16_to_float(static_cast<uint16_t>(a.lane<1>())),
1036		sf16_to_float(static_cast<uint16_t>(a.lane<2>())),
1037		sf16_to_float(static_cast<uint16_t>(a.lane<3>())));
1038}
1039
1040/**
1041 * @brief Return a float value for a float16 scalar.
1042 */
1043ASTCENC_SIMD_INLINE float float16_to_float(uint16_t a)
1044{
1045	return sf16_to_float(a);
1046}
1047
1048/**
1049 * @brief Return a float value as an integer bit pattern (i.e. no conversion).
1050 *
1051 * It is a common trick to convert floats into integer bit patterns, perform
1052 * some bit hackery based on knowledge they are IEEE 754 layout, and then
1053 * convert them back again. This is the first half of that flip.
1054 */
1055ASTCENC_SIMD_INLINE vint4 float_as_int(vfloat4 a)
1056{
1057	vint4 r;
1058	std::memcpy(r.m, a.m, 4 * 4);
1059	return r;
1060}
1061
1062/**
1063 * @brief Return a integer value as a float bit pattern (i.e. no conversion).
1064 *
1065 * It is a common trick to convert floats into integer bit patterns, perform
1066 * some bit hackery based on knowledge they are IEEE 754 layout, and then
1067 * convert them back again. This is the second half of that flip.
1068 */
1069ASTCENC_SIMD_INLINE vfloat4 int_as_float(vint4 a)
1070{
1071	vfloat4 r;
1072	std::memcpy(r.m, a.m, 4 * 4);
1073	return r;
1074}
1075
1076/**
1077 * @brief Prepare a vtable lookup table for use with the native SIMD size.
1078 */
1079ASTCENC_SIMD_INLINE void vtable_prepare(vint4 t0, vint4& t0p)
1080{
1081	t0p = t0;
1082}
1083
1084/**
1085 * @brief Prepare a vtable lookup table for use with the native SIMD size.
1086 */
1087ASTCENC_SIMD_INLINE void vtable_prepare(vint4 t0, vint4 t1, vint4& t0p, vint4& t1p)
1088{
1089	t0p = t0;
1090	t1p = t1;
1091}
1092
1093/**
1094 * @brief Prepare a vtable lookup table for use with the native SIMD size.
1095 */
1096ASTCENC_SIMD_INLINE void vtable_prepare(
1097	vint4 t0, vint4 t1, vint4 t2, vint4 t3,
1098	vint4& t0p, vint4& t1p, vint4& t2p, vint4& t3p)
1099{
1100	t0p = t0;
1101	t1p = t1;
1102	t2p = t2;
1103	t3p = t3;
1104}
1105
1106/**
1107 * @brief Perform an 8-bit 16-entry table lookup, with 32-bit indexes.
1108 */
1109ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 idx)
1110{
1111	uint8_t table[16];
1112
1113	std::memcpy(table +  0, t0.m, 4 * sizeof(int));
1114
1115	return vint4(table[idx.lane<0>()],
1116	             table[idx.lane<1>()],
1117	             table[idx.lane<2>()],
1118	             table[idx.lane<3>()]);
1119}
1120
1121
1122/**
1123 * @brief Perform an 8-bit 32-entry table lookup, with 32-bit indexes.
1124 */
1125ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 idx)
1126{
1127	uint8_t table[32];
1128
1129	std::memcpy(table +  0, t0.m, 4 * sizeof(int));
1130	std::memcpy(table + 16, t1.m, 4 * sizeof(int));
1131
1132	return vint4(table[idx.lane<0>()],
1133	             table[idx.lane<1>()],
1134	             table[idx.lane<2>()],
1135	             table[idx.lane<3>()]);
1136}
1137
1138/**
1139 * @brief Perform an 8-bit 64-entry table lookup, with 32-bit indexes.
1140 */
1141ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 t2, vint4 t3, vint4 idx)
1142{
1143	uint8_t table[64];
1144
1145	std::memcpy(table +  0, t0.m, 4 * sizeof(int));
1146	std::memcpy(table + 16, t1.m, 4 * sizeof(int));
1147	std::memcpy(table + 32, t2.m, 4 * sizeof(int));
1148	std::memcpy(table + 48, t3.m, 4 * sizeof(int));
1149
1150	return vint4(table[idx.lane<0>()],
1151	             table[idx.lane<1>()],
1152	             table[idx.lane<2>()],
1153	             table[idx.lane<3>()]);
1154}
1155
1156/**
1157 * @brief Return a vector of interleaved RGBA data.
1158 *
1159 * Input vectors have the value stored in the bottom 8 bits of each lane,
1160 * with high  bits set to zero.
1161 *
1162 * Output vector stores a single RGBA texel packed in each lane.
1163 */
1164ASTCENC_SIMD_INLINE vint4 interleave_rgba8(vint4 r, vint4 g, vint4 b, vint4 a)
1165{
1166	return r + lsl<8>(g) + lsl<16>(b) + lsl<24>(a);
1167}
1168
1169/**
1170 * @brief Store a single vector lane to an unaligned address.
1171 */
1172ASTCENC_SIMD_INLINE void store_lane(uint8_t* base, int data)
1173{
1174	std::memcpy(base, &data, sizeof(int));
1175}
1176
1177/**
1178 * @brief Store a vector, skipping masked lanes.
1179 *
1180 * All masked lanes must be at the end of vector, after all non-masked lanes.
1181 * Input is a byte array of at least 4 bytes per unmasked entry.
1182 */
1183ASTCENC_SIMD_INLINE void store_lanes_masked(uint8_t* base, vint4 data, vmask4 mask)
1184{
1185	if (mask.m[3])
1186	{
1187		store(data, base);
1188	}
1189	else if (mask.m[2])
1190	{
1191		store_lane(base + 0, data.lane<0>());
1192		store_lane(base + 4, data.lane<1>());
1193		store_lane(base + 8, data.lane<2>());
1194	}
1195	else if (mask.m[1])
1196	{
1197		store_lane(base + 0, data.lane<0>());
1198		store_lane(base + 4, data.lane<1>());
1199	}
1200	else if (mask.m[0])
1201	{
1202		store_lane(base + 0, data.lane<0>());
1203	}
1204}
1205
1206#endif // #ifndef ASTC_VECMATHLIB_NONE_4_H_INCLUDED
1207