1 // SPDX-License-Identifier: Apache-2.0
2 // ----------------------------------------------------------------------------
3 // Copyright 2019-2022 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  */
55 struct 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 	 */
vfloat4vfloat468 	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 	 */
vfloat4vfloat481 	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 	 */
vfloat4vfloat494 	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 	 */
lanevfloat4105 	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 	 */
set_lanevfloat4113 	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 	 */
zerovfloat4121 	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 	 */
load1vfloat4129 	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 	 */
loadavfloat4137 	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 	 */
lane_idvfloat4145 	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 	 */
swzvfloat4153 	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 	 */
swzvfloat4161 	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 	 */
swzvfloat4169 	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  */
187 struct 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 	 */
vint4vint4200 	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 	 */
vint4vint4211 	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 	 */
vint4vint4224 	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 	 */
vint4vint4238 	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 	 */
lanevint4249 	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 	 */
set_lanevint4257 	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 	 */
zerovint4265 	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 	 */
load1vint4273 	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 16B aligned memory.
280 	 */
loadavint4281 	static ASTCENC_SIMD_INLINE vint4 loada(const int* p)
282 	{
283 		return vint4(p);
284 	}
285 
286 	/**
287 	 * @brief Factory that returns a vector containing the lane IDs.
288 	 */
lane_idvint4289 	static ASTCENC_SIMD_INLINE vint4 lane_id()
290 	{
291 		return vint4(0, 1, 2, 3);
292 	}
293 
294 	/**
295 	 * @brief The vector ...
296 	 */
297 	int m[4];
298 };
299 
300 // ============================================================================
301 // vmask4 data type
302 // ============================================================================
303 
304 /**
305  * @brief Data type for 4-wide control plane masks.
306  */
307 struct vmask4
308 {
309 	/**
310 	 * @brief Construct from an existing mask value.
311 	 */
vmask4vmask4312 	ASTCENC_SIMD_INLINE explicit vmask4(int* p)
313 	{
314 		m[0] = p[0];
315 		m[1] = p[1];
316 		m[2] = p[2];
317 		m[3] = p[3];
318 	}
319 
320 	/**
321 	 * @brief Construct from 1 scalar value.
322 	 */
vmask4vmask4323 	ASTCENC_SIMD_INLINE explicit vmask4(bool a)
324 	{
325 		m[0] = a == false ? 0 : -1;
326 		m[1] = a == false ? 0 : -1;
327 		m[2] = a == false ? 0 : -1;
328 		m[3] = a == false ? 0 : -1;
329 	}
330 
331 	/**
332 	 * @brief Construct from 4 scalar values.
333 	 *
334 	 * The value of @c a is stored to lane 0 (LSB) in the SIMD register.
335 	 */
vmask4vmask4336 	ASTCENC_SIMD_INLINE explicit vmask4(bool a, bool b, bool c, bool d)
337 	{
338 		m[0] = a == false ? 0 : -1;
339 		m[1] = b == false ? 0 : -1;
340 		m[2] = c == false ? 0 : -1;
341 		m[3] = d == false ? 0 : -1;
342 	}
343 
344 
345 	/**
346 	 * @brief The vector ...
347 	 */
348 	int m[4];
349 };
350 
351 // ============================================================================
352 // vmask4 operators and functions
353 // ============================================================================
354 
355 /**
356  * @brief Overload: mask union (or).
357  */
358 ASTCENC_SIMD_INLINE vmask4 operator|(vmask4 a, vmask4 b)
359 {
360 	return vmask4(a.m[0] | b.m[0],
361 	              a.m[1] | b.m[1],
362 	              a.m[2] | b.m[2],
363 	              a.m[3] | b.m[3]);
364 }
365 
366 /**
367  * @brief Overload: mask intersect (and).
368  */
369 ASTCENC_SIMD_INLINE vmask4 operator&(vmask4 a, vmask4 b)
370 {
371 	return vmask4(a.m[0] & b.m[0],
372 	              a.m[1] & b.m[1],
373 	              a.m[2] & b.m[2],
374 	              a.m[3] & b.m[3]);
375 }
376 
377 /**
378  * @brief Overload: mask difference (xor).
379  */
380 ASTCENC_SIMD_INLINE vmask4 operator^(vmask4 a, vmask4 b)
381 {
382 	return vmask4(a.m[0] ^ b.m[0],
383 	              a.m[1] ^ b.m[1],
384 	              a.m[2] ^ b.m[2],
385 	              a.m[3] ^ b.m[3]);
386 }
387 
388 /**
389  * @brief Overload: mask invert (not).
390  */
391 ASTCENC_SIMD_INLINE vmask4 operator~(vmask4 a)
392 {
393 	return vmask4(~a.m[0],
394 	              ~a.m[1],
395 	              ~a.m[2],
396 	              ~a.m[3]);
397 }
398 
399 /**
400  * @brief Return a 1-bit mask code indicating mask status.
401  *
402  * bit0 = lane 0
403  */
mask(vmask4 a)404 ASTCENC_SIMD_INLINE unsigned int mask(vmask4 a)
405 {
406 	return ((a.m[0] >> 31) & 0x1) |
407 	       ((a.m[1] >> 30) & 0x2) |
408 	       ((a.m[2] >> 29) & 0x4) |
409 	       ((a.m[3] >> 28) & 0x8);
410 }
411 
412 // ============================================================================
413 // vint4 operators and functions
414 // ============================================================================
415 
416 /**
417  * @brief Overload: vector by vector addition.
418  */
419 ASTCENC_SIMD_INLINE vint4 operator+(vint4 a, vint4 b)
420 {
421 	return vint4(a.m[0] + b.m[0],
422 	             a.m[1] + b.m[1],
423 	             a.m[2] + b.m[2],
424 	             a.m[3] + b.m[3]);
425 }
426 
427 /**
428  * @brief Overload: vector by vector subtraction.
429  */
430 ASTCENC_SIMD_INLINE vint4 operator-(vint4 a, vint4 b)
431 {
432 	return vint4(a.m[0] - b.m[0],
433 	             a.m[1] - b.m[1],
434 	             a.m[2] - b.m[2],
435 	             a.m[3] - b.m[3]);
436 }
437 
438 /**
439  * @brief Overload: vector by vector multiplication.
440  */
441 ASTCENC_SIMD_INLINE vint4 operator*(vint4 a, vint4 b)
442 {
443 	return vint4(a.m[0] * b.m[0],
444 	             a.m[1] * b.m[1],
445 	             a.m[2] * b.m[2],
446 	             a.m[3] * b.m[3]);
447 }
448 
449 /**
450  * @brief Overload: vector bit invert.
451  */
452 ASTCENC_SIMD_INLINE vint4 operator~(vint4 a)
453 {
454 	return vint4(~a.m[0],
455 	             ~a.m[1],
456 	             ~a.m[2],
457 	             ~a.m[3]);
458 }
459 
460 /**
461  * @brief Overload: vector by vector bitwise or.
462  */
463 ASTCENC_SIMD_INLINE vint4 operator|(vint4 a, vint4 b)
464 {
465 	return vint4(a.m[0] | b.m[0],
466 	             a.m[1] | b.m[1],
467 	             a.m[2] | b.m[2],
468 	             a.m[3] | b.m[3]);
469 }
470 
471 /**
472  * @brief Overload: vector by vector bitwise and.
473  */
474 ASTCENC_SIMD_INLINE vint4 operator&(vint4 a, vint4 b)
475 {
476 	return vint4(a.m[0] & b.m[0],
477 	             a.m[1] & b.m[1],
478 	             a.m[2] & b.m[2],
479 	             a.m[3] & b.m[3]);
480 }
481 
482 /**
483  * @brief Overload: vector by vector bitwise xor.
484  */
485 ASTCENC_SIMD_INLINE vint4 operator^(vint4 a, vint4 b)
486 {
487 	return vint4(a.m[0] ^ b.m[0],
488 	             a.m[1] ^ b.m[1],
489 	             a.m[2] ^ b.m[2],
490 	             a.m[3] ^ b.m[3]);
491 }
492 
493 /**
494  * @brief Overload: vector by vector equality.
495  */
496 ASTCENC_SIMD_INLINE vmask4 operator==(vint4 a, vint4 b)
497 {
498 	return vmask4(a.m[0] == b.m[0],
499 	              a.m[1] == b.m[1],
500 	              a.m[2] == b.m[2],
501 	              a.m[3] == b.m[3]);
502 }
503 
504 /**
505  * @brief Overload: vector by vector inequality.
506  */
507 ASTCENC_SIMD_INLINE vmask4 operator!=(vint4 a, vint4 b)
508 {
509 	return vmask4(a.m[0] != b.m[0],
510 	              a.m[1] != b.m[1],
511 	              a.m[2] != b.m[2],
512 	              a.m[3] != b.m[3]);
513 }
514 
515 /**
516  * @brief Overload: vector by vector less than.
517  */
518 ASTCENC_SIMD_INLINE vmask4 operator<(vint4 a, vint4 b)
519 {
520 	return vmask4(a.m[0] < b.m[0],
521 	              a.m[1] < b.m[1],
522 	              a.m[2] < b.m[2],
523 	              a.m[3] < b.m[3]);
524 }
525 
526 /**
527  * @brief Overload: vector by vector greater than.
528  */
529 ASTCENC_SIMD_INLINE vmask4 operator>(vint4 a, vint4 b)
530 {
531 	return vmask4(a.m[0] > b.m[0],
532 	              a.m[1] > b.m[1],
533 	              a.m[2] > b.m[2],
534 	              a.m[3] > b.m[3]);
535 }
536 
537 /**
538  * @brief Logical shift left.
539  */
lsl(vint4 a)540 template <int s> ASTCENC_SIMD_INLINE vint4 lsl(vint4 a)
541 {
542 	return vint4(a.m[0] << s,
543 	             a.m[1] << s,
544 	             a.m[2] << s,
545 	             a.m[3] << s);
546 }
547 
548 /**
549  * @brief Logical shift right.
550  */
lsr(vint4 a)551 template <int s> ASTCENC_SIMD_INLINE vint4 lsr(vint4 a)
552 {
553 	unsigned int as0 = static_cast<unsigned int>(a.m[0]) >> s;
554 	unsigned int as1 = static_cast<unsigned int>(a.m[1]) >> s;
555 	unsigned int as2 = static_cast<unsigned int>(a.m[2]) >> s;
556 	unsigned int as3 = static_cast<unsigned int>(a.m[3]) >> s;
557 
558 	return vint4(static_cast<int>(as0),
559 	             static_cast<int>(as1),
560 	             static_cast<int>(as2),
561 	             static_cast<int>(as3));
562 }
563 
564 /**
565  * @brief Arithmetic shift right.
566  */
asr(vint4 a)567 template <int s> ASTCENC_SIMD_INLINE vint4 asr(vint4 a)
568 {
569 	return vint4(a.m[0] >> s,
570 	             a.m[1] >> s,
571 	             a.m[2] >> s,
572 	             a.m[3] >> s);
573 }
574 
575 /**
576  * @brief Return the min vector of two vectors.
577  */
min(vint4 a,vint4 b)578 ASTCENC_SIMD_INLINE vint4 min(vint4 a, vint4 b)
579 {
580 	return vint4(a.m[0] < b.m[0] ? a.m[0] : b.m[0],
581 	             a.m[1] < b.m[1] ? a.m[1] : b.m[1],
582 	             a.m[2] < b.m[2] ? a.m[2] : b.m[2],
583 	             a.m[3] < b.m[3] ? a.m[3] : b.m[3]);
584 }
585 
586 /**
587  * @brief Return the min vector of two vectors.
588  */
max(vint4 a,vint4 b)589 ASTCENC_SIMD_INLINE vint4 max(vint4 a, vint4 b)
590 {
591 	return vint4(a.m[0] > b.m[0] ? a.m[0] : b.m[0],
592 	             a.m[1] > b.m[1] ? a.m[1] : b.m[1],
593 	             a.m[2] > b.m[2] ? a.m[2] : b.m[2],
594 	             a.m[3] > b.m[3] ? a.m[3] : b.m[3]);
595 }
596 
597 /**
598  * @brief Return the horizontal minimum of a single vector.
599  */
hmin(vint4 a)600 ASTCENC_SIMD_INLINE vint4 hmin(vint4 a)
601 {
602 	int b = std::min(a.m[0], a.m[1]);
603 	int c = std::min(a.m[2], a.m[3]);
604 	return vint4(std::min(b, c));
605 }
606 
607 /**
608  * @brief Return the horizontal maximum of a single vector.
609  */
hmax(vint4 a)610 ASTCENC_SIMD_INLINE vint4 hmax(vint4 a)
611 {
612 	int b = std::max(a.m[0], a.m[1]);
613 	int c = std::max(a.m[2], a.m[3]);
614 	return vint4(std::max(b, c));
615 }
616 
617 /**
618  * @brief Return the horizontal sum of vector lanes as a scalar.
619  */
hadd_s(vint4 a)620 ASTCENC_SIMD_INLINE int hadd_s(vint4 a)
621 {
622 	return a.m[0] + a.m[1] + a.m[2] + a.m[3];
623 }
624 
625 /**
626  * @brief Store a vector to an aligned memory address.
627  */
storea(vint4 a,int * p)628 ASTCENC_SIMD_INLINE void storea(vint4 a, int* p)
629 {
630 	p[0] = a.m[0];
631 	p[1] = a.m[1];
632 	p[2] = a.m[2];
633 	p[3] = a.m[3];
634 }
635 
636 /**
637  * @brief Store a vector to an unaligned memory address.
638  */
store(vint4 a,int * p)639 ASTCENC_SIMD_INLINE void store(vint4 a, int* p)
640 {
641 	p[0] = a.m[0];
642 	p[1] = a.m[1];
643 	p[2] = a.m[2];
644 	p[3] = a.m[3];
645 }
646 
647 /**
648  * @brief Store lowest N (vector width) bytes into an unaligned address.
649  */
store_nbytes(vint4 a,uint8_t * p)650 ASTCENC_SIMD_INLINE void store_nbytes(vint4 a, uint8_t* p)
651 {
652 	int* pi = reinterpret_cast<int*>(p);
653 	*pi = a.m[0];
654 }
655 
656 /**
657  * @brief Gather N (vector width) indices from the array.
658  */
gatheri(const int * base,vint4 indices)659 ASTCENC_SIMD_INLINE vint4 gatheri(const int* base, vint4 indices)
660 {
661 	return vint4(base[indices.m[0]],
662 	             base[indices.m[1]],
663 	             base[indices.m[2]],
664 	             base[indices.m[3]]);
665 }
666 
667 /**
668  * @brief Pack low 8 bits of N (vector width) lanes into bottom of vector.
669  */
pack_low_bytes(vint4 a)670 ASTCENC_SIMD_INLINE vint4 pack_low_bytes(vint4 a)
671 {
672 	int b0 = a.m[0] & 0xFF;
673 	int b1 = a.m[1] & 0xFF;
674 	int b2 = a.m[2] & 0xFF;
675 	int b3 = a.m[3] & 0xFF;
676 
677 	int b = b0 | (b1 << 8) | (b2 << 16) | (b3 << 24);
678 	return vint4(b, 0, 0, 0);
679 }
680 
681 /**
682  * @brief Return lanes from @c b if MSB of @c cond is set, else @c a.
683  */
select(vint4 a,vint4 b,vmask4 cond)684 ASTCENC_SIMD_INLINE vint4 select(vint4 a, vint4 b, vmask4 cond)
685 {
686 	return vint4((cond.m[0] & static_cast<int>(0x80000000)) ? b.m[0] : a.m[0],
687 	             (cond.m[1] & static_cast<int>(0x80000000)) ? b.m[1] : a.m[1],
688 	             (cond.m[2] & static_cast<int>(0x80000000)) ? b.m[2] : a.m[2],
689 	             (cond.m[3] & static_cast<int>(0x80000000)) ? b.m[3] : a.m[3]);
690 }
691 
692 // ============================================================================
693 // vfloat4 operators and functions
694 // ============================================================================
695 
696 /**
697  * @brief Overload: vector by vector addition.
698  */
699 ASTCENC_SIMD_INLINE vfloat4 operator+(vfloat4 a, vfloat4 b)
700 {
701 	return vfloat4(a.m[0] + b.m[0],
702 	               a.m[1] + b.m[1],
703 	               a.m[2] + b.m[2],
704 	               a.m[3] + b.m[3]);
705 }
706 
707 /**
708  * @brief Overload: vector by vector subtraction.
709  */
710 ASTCENC_SIMD_INLINE vfloat4 operator-(vfloat4 a, vfloat4 b)
711 {
712 	return vfloat4(a.m[0] - b.m[0],
713 	               a.m[1] - b.m[1],
714 	               a.m[2] - b.m[2],
715 	               a.m[3] - b.m[3]);
716 }
717 
718 /**
719  * @brief Overload: vector by vector multiplication.
720  */
721 ASTCENC_SIMD_INLINE vfloat4 operator*(vfloat4 a, vfloat4 b)
722 {
723 	return vfloat4(a.m[0] * b.m[0],
724 	               a.m[1] * b.m[1],
725 	               a.m[2] * b.m[2],
726 	               a.m[3] * b.m[3]);
727 }
728 
729 /**
730  * @brief Overload: vector by vector division.
731  */
732 ASTCENC_SIMD_INLINE vfloat4 operator/(vfloat4 a, vfloat4 b)
733 {
734 	return vfloat4(a.m[0] / b.m[0],
735 	               a.m[1] / b.m[1],
736 	               a.m[2] / b.m[2],
737 	               a.m[3] / b.m[3]);
738 }
739 
740 /**
741  * @brief Overload: vector by vector equality.
742  */
743 ASTCENC_SIMD_INLINE vmask4 operator==(vfloat4 a, vfloat4 b)
744 {
745 	return vmask4(a.m[0] == b.m[0],
746 	              a.m[1] == b.m[1],
747 	              a.m[2] == b.m[2],
748 	              a.m[3] == b.m[3]);
749 }
750 
751 /**
752  * @brief Overload: vector by vector inequality.
753  */
754 ASTCENC_SIMD_INLINE vmask4 operator!=(vfloat4 a, vfloat4 b)
755 {
756 	return vmask4(a.m[0] != b.m[0],
757 	              a.m[1] != b.m[1],
758 	              a.m[2] != b.m[2],
759 	              a.m[3] != b.m[3]);
760 }
761 
762 /**
763  * @brief Overload: vector by vector less than.
764  */
765 ASTCENC_SIMD_INLINE vmask4 operator<(vfloat4 a, vfloat4 b)
766 {
767 	return vmask4(a.m[0] < b.m[0],
768 	              a.m[1] < b.m[1],
769 	              a.m[2] < b.m[2],
770 	              a.m[3] < b.m[3]);
771 }
772 
773 /**
774  * @brief Overload: vector by vector greater than.
775  */
776 ASTCENC_SIMD_INLINE vmask4 operator>(vfloat4 a, vfloat4 b)
777 {
778 	return vmask4(a.m[0] > b.m[0],
779 	              a.m[1] > b.m[1],
780 	              a.m[2] > b.m[2],
781 	              a.m[3] > b.m[3]);
782 }
783 
784 /**
785  * @brief Overload: vector by vector less than or equal.
786  */
787 ASTCENC_SIMD_INLINE vmask4 operator<=(vfloat4 a, vfloat4 b)
788 {
789 	return vmask4(a.m[0] <= b.m[0],
790 	              a.m[1] <= b.m[1],
791 	              a.m[2] <= b.m[2],
792 	              a.m[3] <= b.m[3]);
793 }
794 
795 /**
796  * @brief Overload: vector by vector greater than or equal.
797  */
798 ASTCENC_SIMD_INLINE vmask4 operator>=(vfloat4 a, vfloat4 b)
799 {
800 	return vmask4(a.m[0] >= b.m[0],
801 	              a.m[1] >= b.m[1],
802 	              a.m[2] >= b.m[2],
803 	              a.m[3] >= b.m[3]);
804 }
805 
806 /**
807  * @brief Return the min vector of two vectors.
808  *
809  * If either lane value is NaN, @c b will be returned for that lane.
810  */
min(vfloat4 a,vfloat4 b)811 ASTCENC_SIMD_INLINE vfloat4 min(vfloat4 a, vfloat4 b)
812 {
813 	return vfloat4(a.m[0] < b.m[0] ? a.m[0] : b.m[0],
814 	               a.m[1] < b.m[1] ? a.m[1] : b.m[1],
815 	               a.m[2] < b.m[2] ? a.m[2] : b.m[2],
816 	               a.m[3] < b.m[3] ? a.m[3] : b.m[3]);
817 }
818 
819 /**
820  * @brief Return the max vector of two vectors.
821  *
822  * If either lane value is NaN, @c b will be returned for that lane.
823  */
max(vfloat4 a,vfloat4 b)824 ASTCENC_SIMD_INLINE vfloat4 max(vfloat4 a, vfloat4 b)
825 {
826 	return vfloat4(a.m[0] > b.m[0] ? a.m[0] : b.m[0],
827 	               a.m[1] > b.m[1] ? a.m[1] : b.m[1],
828 	               a.m[2] > b.m[2] ? a.m[2] : b.m[2],
829 	               a.m[3] > b.m[3] ? a.m[3] : b.m[3]);
830 }
831 
832 /**
833  * @brief Return the absolute value of the float vector.
834  */
abs(vfloat4 a)835 ASTCENC_SIMD_INLINE vfloat4 abs(vfloat4 a)
836 {
837 	return vfloat4(std::abs(a.m[0]),
838 	               std::abs(a.m[1]),
839 	               std::abs(a.m[2]),
840 	               std::abs(a.m[3]));
841 }
842 
843 /**
844  * @brief Return a float rounded to the nearest integer value.
845  */
round(vfloat4 a)846 ASTCENC_SIMD_INLINE vfloat4 round(vfloat4 a)
847 {
848 	assert(std::fegetround() == FE_TONEAREST);
849 	return vfloat4(std::nearbyint(a.m[0]),
850 	               std::nearbyint(a.m[1]),
851 	               std::nearbyint(a.m[2]),
852 	               std::nearbyint(a.m[3]));
853 }
854 
855 /**
856  * @brief Return the horizontal minimum of a vector.
857  */
hmin(vfloat4 a)858 ASTCENC_SIMD_INLINE vfloat4 hmin(vfloat4 a)
859 {
860 	float tmp1 = std::min(a.m[0], a.m[1]);
861 	float tmp2 = std::min(a.m[2], a.m[3]);
862 	return vfloat4(std::min(tmp1, tmp2));
863 }
864 
865 /**
866  * @brief Return the horizontal maximum of a vector.
867  */
hmax(vfloat4 a)868 ASTCENC_SIMD_INLINE vfloat4 hmax(vfloat4 a)
869 {
870 	float tmp1 = std::max(a.m[0], a.m[1]);
871 	float tmp2 = std::max(a.m[2], a.m[3]);
872 	return vfloat4(std::max(tmp1, tmp2));
873 }
874 
875 /**
876  * @brief Return the horizontal sum of a vector.
877  */
hadd_s(vfloat4 a)878 ASTCENC_SIMD_INLINE float hadd_s(vfloat4 a)
879 {
880 	// Use halving add, gives invariance with SIMD versions
881 	return (a.m[0] + a.m[2]) + (a.m[1] + a.m[3]);
882 }
883 
884 /**
885  * @brief Return the sqrt of the lanes in the vector.
886  */
sqrt(vfloat4 a)887 ASTCENC_SIMD_INLINE vfloat4 sqrt(vfloat4 a)
888 {
889 	return vfloat4(std::sqrt(a.m[0]),
890 	               std::sqrt(a.m[1]),
891 	               std::sqrt(a.m[2]),
892 	               std::sqrt(a.m[3]));
893 }
894 
895 /**
896  * @brief Return lanes from @c b if @c cond is set, else @c a.
897  */
select(vfloat4 a,vfloat4 b,vmask4 cond)898 ASTCENC_SIMD_INLINE vfloat4 select(vfloat4 a, vfloat4 b, vmask4 cond)
899 {
900 	return vfloat4((cond.m[0] & static_cast<int>(0x80000000)) ? b.m[0] : a.m[0],
901 	               (cond.m[1] & static_cast<int>(0x80000000)) ? b.m[1] : a.m[1],
902 	               (cond.m[2] & static_cast<int>(0x80000000)) ? b.m[2] : a.m[2],
903 	               (cond.m[3] & static_cast<int>(0x80000000)) ? b.m[3] : a.m[3]);
904 }
905 
906 /**
907  * @brief Return lanes from @c b if MSB of @c cond is set, else @c a.
908  */
select_msb(vfloat4 a,vfloat4 b,vmask4 cond)909 ASTCENC_SIMD_INLINE vfloat4 select_msb(vfloat4 a, vfloat4 b, vmask4 cond)
910 {
911 	return vfloat4((cond.m[0] & static_cast<int>(0x80000000)) ? b.m[0] : a.m[0],
912 	               (cond.m[1] & static_cast<int>(0x80000000)) ? b.m[1] : a.m[1],
913 	               (cond.m[2] & static_cast<int>(0x80000000)) ? b.m[2] : a.m[2],
914 	               (cond.m[3] & static_cast<int>(0x80000000)) ? b.m[3] : a.m[3]);
915 }
916 
917 /**
918  * @brief Load a vector of gathered results from an array;
919  */
gatherf(const float * base,vint4 indices)920 ASTCENC_SIMD_INLINE vfloat4 gatherf(const float* base, vint4 indices)
921 {
922 	return vfloat4(base[indices.m[0]],
923 	               base[indices.m[1]],
924 	               base[indices.m[2]],
925 	               base[indices.m[3]]);
926 }
927 
928 /**
929  * @brief Store a vector to an unaligned memory address.
930  */
store(vfloat4 a,float * ptr)931 ASTCENC_SIMD_INLINE void store(vfloat4 a, float* ptr)
932 {
933 	ptr[0] = a.m[0];
934 	ptr[1] = a.m[1];
935 	ptr[2] = a.m[2];
936 	ptr[3] = a.m[3];
937 }
938 
939 /**
940  * @brief Store a vector to an aligned memory address.
941  */
storea(vfloat4 a,float * ptr)942 ASTCENC_SIMD_INLINE void storea(vfloat4 a, float* ptr)
943 {
944 	ptr[0] = a.m[0];
945 	ptr[1] = a.m[1];
946 	ptr[2] = a.m[2];
947 	ptr[3] = a.m[3];
948 }
949 
950 /**
951  * @brief Return a integer value for a float vector, using truncation.
952  */
float_to_int(vfloat4 a)953 ASTCENC_SIMD_INLINE vint4 float_to_int(vfloat4 a)
954 {
955 	return vint4(static_cast<int>(a.m[0]),
956 	             static_cast<int>(a.m[1]),
957 	             static_cast<int>(a.m[2]),
958 	             static_cast<int>(a.m[3]));
959 }
960 
961 /**f
962  * @brief Return a integer value for a float vector, using round-to-nearest.
963  */
float_to_int_rtn(vfloat4 a)964 ASTCENC_SIMD_INLINE vint4 float_to_int_rtn(vfloat4 a)
965 {
966 	return vint4(static_cast<int>(a.m[0] + 0.5f),
967 	             static_cast<int>(a.m[1] + 0.5f),
968 	             static_cast<int>(a.m[2] + 0.5f),
969 	             static_cast<int>(a.m[3] + 0.5f));
970 }
971 
972 /**
973  * @brief Return a float value for a integer vector.
974  */
int_to_float(vint4 a)975 ASTCENC_SIMD_INLINE vfloat4 int_to_float(vint4 a)
976 {
977 	return vfloat4(static_cast<float>(a.m[0]),
978 	               static_cast<float>(a.m[1]),
979 	               static_cast<float>(a.m[2]),
980 	               static_cast<float>(a.m[3]));
981 }
982 
983 /**
984  * @brief Return a float16 value for a float vector, using round-to-nearest.
985  */
float_to_float16(vfloat4 a)986 ASTCENC_SIMD_INLINE vint4 float_to_float16(vfloat4 a)
987 {
988 	return vint4(
989 		float_to_sf16(a.lane<0>()),
990 		float_to_sf16(a.lane<1>()),
991 		float_to_sf16(a.lane<2>()),
992 		float_to_sf16(a.lane<3>()));
993 }
994 
995 /**
996  * @brief Return a float16 value for a float scalar, using round-to-nearest.
997  */
float_to_float16(float a)998 static inline uint16_t float_to_float16(float a)
999 {
1000 	return float_to_sf16(a);
1001 }
1002 
1003 /**
1004  * @brief Return a float value for a float16 vector.
1005  */
float16_to_float(vint4 a)1006 ASTCENC_SIMD_INLINE vfloat4 float16_to_float(vint4 a)
1007 {
1008 	return vfloat4(
1009 		sf16_to_float(static_cast<uint16_t>(a.lane<0>())),
1010 		sf16_to_float(static_cast<uint16_t>(a.lane<1>())),
1011 		sf16_to_float(static_cast<uint16_t>(a.lane<2>())),
1012 		sf16_to_float(static_cast<uint16_t>(a.lane<3>())));
1013 }
1014 
1015 /**
1016  * @brief Return a float value for a float16 scalar.
1017  */
float16_to_float(uint16_t a)1018 ASTCENC_SIMD_INLINE float float16_to_float(uint16_t a)
1019 {
1020 	return sf16_to_float(a);
1021 }
1022 
1023 /**
1024  * @brief Return a float value as an integer bit pattern (i.e. no conversion).
1025  *
1026  * It is a common trick to convert floats into integer bit patterns, perform
1027  * some bit hackery based on knowledge they are IEEE 754 layout, and then
1028  * convert them back again. This is the first half of that flip.
1029  */
float_as_int(vfloat4 a)1030 ASTCENC_SIMD_INLINE vint4 float_as_int(vfloat4 a)
1031 {
1032 	vint4 r;
1033 	memcpy(r.m, a.m, 4 * 4);
1034 	return r;
1035 }
1036 
1037 /**
1038  * @brief Return a integer value as a float bit pattern (i.e. no conversion).
1039  *
1040  * It is a common trick to convert floats into integer bit patterns, perform
1041  * some bit hackery based on knowledge they are IEEE 754 layout, and then
1042  * convert them back again. This is the second half of that flip.
1043  */
int_as_float(vint4 a)1044 ASTCENC_SIMD_INLINE vfloat4 int_as_float(vint4 a)
1045 {
1046 	vfloat4 r;
1047 	memcpy(r.m, a.m, 4 * 4);
1048 	return r;
1049 }
1050 
1051 /**
1052  * @brief Prepare a vtable lookup table for use with the native SIMD size.
1053  */
vtable_prepare(vint4 t0,vint4 & t0p)1054 ASTCENC_SIMD_INLINE void vtable_prepare(vint4 t0, vint4& t0p)
1055 {
1056 	t0p = t0;
1057 }
1058 
1059 /**
1060  * @brief Prepare a vtable lookup table for use with the native SIMD size.
1061  */
vtable_prepare(vint4 t0,vint4 t1,vint4 & t0p,vint4 & t1p)1062 ASTCENC_SIMD_INLINE void vtable_prepare(vint4 t0, vint4 t1, vint4& t0p, vint4& t1p)
1063 {
1064 	t0p = t0;
1065 	t1p = t1;
1066 }
1067 
1068 /**
1069  * @brief Prepare a vtable lookup table for use with the native SIMD size.
1070  */
vtable_prepare(vint4 t0,vint4 t1,vint4 t2,vint4 t3,vint4 & t0p,vint4 & t1p,vint4 & t2p,vint4 & t3p)1071 ASTCENC_SIMD_INLINE void vtable_prepare(
1072 	vint4 t0, vint4 t1, vint4 t2, vint4 t3,
1073 	vint4& t0p, vint4& t1p, vint4& t2p, vint4& t3p)
1074 {
1075 	t0p = t0;
1076 	t1p = t1;
1077 	t2p = t2;
1078 	t3p = t3;
1079 }
1080 
1081 /**
1082  * @brief Perform an 8-bit 32-entry table lookup, with 32-bit indexes.
1083  */
vtable_8bt_32bi(vint4 t0,vint4 idx)1084 ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 idx)
1085 {
1086 	uint8_t table[16];
1087 	storea(t0, reinterpret_cast<int*>(table +  0));
1088 
1089 	return vint4(table[idx.lane<0>()],
1090 	             table[idx.lane<1>()],
1091 	             table[idx.lane<2>()],
1092 	             table[idx.lane<3>()]);
1093 }
1094 
1095 
1096 /**
1097  * @brief Perform an 8-bit 32-entry table lookup, with 32-bit indexes.
1098  */
vtable_8bt_32bi(vint4 t0,vint4 t1,vint4 idx)1099 ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 idx)
1100 {
1101 	uint8_t table[32];
1102 	storea(t0, reinterpret_cast<int*>(table +  0));
1103 	storea(t1, reinterpret_cast<int*>(table + 16));
1104 
1105 	return vint4(table[idx.lane<0>()],
1106 	             table[idx.lane<1>()],
1107 	             table[idx.lane<2>()],
1108 	             table[idx.lane<3>()]);
1109 }
1110 
1111 /**
1112  * @brief Perform an 8-bit 64-entry table lookup, with 32-bit indexes.
1113  */
vtable_8bt_32bi(vint4 t0,vint4 t1,vint4 t2,vint4 t3,vint4 idx)1114 ASTCENC_SIMD_INLINE vint4 vtable_8bt_32bi(vint4 t0, vint4 t1, vint4 t2, vint4 t3, vint4 idx)
1115 {
1116 	uint8_t table[64];
1117 	storea(t0, reinterpret_cast<int*>(table +  0));
1118 	storea(t1, reinterpret_cast<int*>(table + 16));
1119 	storea(t2, reinterpret_cast<int*>(table + 32));
1120 	storea(t3, reinterpret_cast<int*>(table + 48));
1121 
1122 	return vint4(table[idx.lane<0>()],
1123 	             table[idx.lane<1>()],
1124 	             table[idx.lane<2>()],
1125 	             table[idx.lane<3>()]);
1126 }
1127 
1128 /**
1129  * @brief Return a vector of interleaved RGBA data.
1130  *
1131  * Input vectors have the value stored in the bottom 8 bits of each lane,
1132  * with high  bits set to zero.
1133  *
1134  * Output vector stores a single RGBA texel packed in each lane.
1135  */
interleave_rgba8(vint4 r,vint4 g,vint4 b,vint4 a)1136 ASTCENC_SIMD_INLINE vint4 interleave_rgba8(vint4 r, vint4 g, vint4 b, vint4 a)
1137 {
1138 	return r + lsl<8>(g) + lsl<16>(b) + lsl<24>(a);
1139 }
1140 
1141 /**
1142  * @brief Store a vector, skipping masked lanes.
1143  *
1144  * All masked lanes must be at the end of vector, after all non-masked lanes.
1145  */
store_lanes_masked(int * base,vint4 data,vmask4 mask)1146 ASTCENC_SIMD_INLINE void store_lanes_masked(int* base, vint4 data, vmask4 mask)
1147 {
1148 	if (mask.m[3])
1149 	{
1150 		store(data, base);
1151 	}
1152 	else if (mask.m[2])
1153 	{
1154 		base[0] = data.lane<0>();
1155 		base[1] = data.lane<1>();
1156 		base[2] = data.lane<2>();
1157 	}
1158 	else if (mask.m[1])
1159 	{
1160 		base[0] = data.lane<0>();
1161 		base[1] = data.lane<1>();
1162 	}
1163 	else if (mask.m[0])
1164 	{
1165 		base[0] = data.lane<0>();
1166 	}
1167 }
1168 
1169 #endif // #ifndef ASTC_VECMATHLIB_NONE_4_H_INCLUDED
1170