summaryrefslogtreecommitdiff
path: root/indra/llmath/llvector4a.inl
blob: 69d3d01efe485d3a4c677d1c6e730f0ea573b1e8 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
/** 
 * @file llvector4a.inl
 * @brief LLVector4a inline function implementations
 *
 * $LicenseInfo:firstyear=2010&license=viewerlgpl$
 * Second Life Viewer Source Code
 * Copyright (C) 2010, Linden Research, Inc.
 * 
 * This library is free software; you can redistribute it and/or
 * modify it under the terms of the GNU Lesser General Public
 * License as published by the Free Software Foundation;
 * version 2.1 of the License only.
 * 
 * This library is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
 * Lesser General Public License for more details.
 * 
 * You should have received a copy of the GNU Lesser General Public
 * License along with this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
 * 
 * Linden Research, Inc., 945 Battery Street, San Francisco, CA  94111  USA
 * $/LicenseInfo$
 */

////////////////////////////////////
// LOAD/STORE
////////////////////////////////////

// Load from 16-byte aligned src array (preferred method of loading)
inline void LLVector4a::load4a(const F32* src)
{
	mQ = _mm_load_ps(src);
}

// Load from unaligned src array (NB: Significantly slower than load4a)
inline void LLVector4a::loadua(const F32* src)
{
	mQ = _mm_loadu_ps(src);
}

// Load only three floats beginning at address 'src'. Slowest method.
inline void LLVector4a::load3(const F32* src)
{
	// mQ = { 0.f, src[2], src[1], src[0] } = { W, Z, Y, X }
	// NB: This differs from the convention of { Z, Y, X, W }
	mQ = _mm_set_ps(0.f, src[2], src[1], src[0]);
}	

// Store to a 16-byte aligned memory address
inline void LLVector4a::store4a(F32* dst) const
{
	_mm_store_ps(dst, mQ);
}

////////////////////////////////////
// BASIC GET/SET 
////////////////////////////////////

// Return a "this" as an F32 pointer. Do not use unless you have a very good reason.  (Not sure? Ask Falcon)
F32* LLVector4a::getF32ptr()
{
	return (F32*) &mQ;
}

// Return a "this" as a const F32 pointer. Do not use unless you have a very good reason.  (Not sure? Ask Falcon)
const F32* const LLVector4a::getF32ptr() const
{
	return (const F32* const) &mQ;
}

// Read-only access a single float in this vector. Do not use in proximity to any function call that manipulates
// the data at the whole vector level or you will incur a substantial penalty. Consider using the splat functions instead
inline F32 LLVector4a::operator[](const S32 idx) const
{
	return ((F32*)&mQ)[idx];
}	

// Prefer this method for read-only access to a single element. Prefer the templated version if the elem is known at compile time.
inline LLSimdScalar LLVector4a::getScalarAt(const S32 idx) const
{
	// Return appropriate LLQuad. It will be cast to LLSimdScalar automatically (should be effectively a nop)
	switch (idx)
	{
		case 0:
			return mQ;
		case 1:
			return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(1, 1, 1, 1));
		case 2:
			return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(2, 2, 2, 2));
		case 3:
		default:
			return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(3, 3, 3, 3));
	}
}

// Prefer this method for read-only access to a single element. Prefer the templated version if the elem is known at compile time.
template <int N> LL_FORCE_INLINE LLSimdScalar LLVector4a::getScalarAt() const
{
	return _mm_shuffle_ps(mQ, mQ, _MM_SHUFFLE(N, N, N, N));
}

template<> LL_FORCE_INLINE LLSimdScalar LLVector4a::getScalarAt<0>() const
{
	return mQ;
}

// Set to an x, y, z and optional w provided
inline void LLVector4a::set(F32 x, F32 y, F32 z, F32 w)
{
	mQ = _mm_set_ps(w, z, y, x);
}

// Set to all zeros
inline void LLVector4a::clear()
{
	mQ = LLVector4a::getZero().mQ;
}

inline void LLVector4a::splat(const F32 x)
{
	mQ = _mm_set1_ps(x);	
}

inline void LLVector4a::splat(const LLSimdScalar& x)
{
	mQ = _mm_shuffle_ps( x.getQuad(), x.getQuad(), _MM_SHUFFLE(0,0,0,0) );
}

// Set all 4 elements to element N of src, with N known at compile time
template <int N> void LLVector4a::splat(const LLVector4a& src)
{
	mQ = _mm_shuffle_ps(src.mQ, src.mQ, _MM_SHUFFLE(N, N, N, N) );
}

// Set all 4 elements to element i of v, with i NOT known at compile time
inline void LLVector4a::splat(const LLVector4a& v, U32 i)
{
	switch (i)
	{
		case 0:
			mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(0, 0, 0, 0));
			break;
		case 1:
			mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(1, 1, 1, 1));
			break;
		case 2:
			mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(2, 2, 2, 2));
			break;
		case 3:
			mQ = _mm_shuffle_ps(v.mQ, v.mQ, _MM_SHUFFLE(3, 3, 3, 3));
			break;
	}
}

// Select bits from sourceIfTrue and sourceIfFalse according to bits in mask
inline void LLVector4a::setSelectWithMask( const LLVector4Logical& mask, const LLVector4a& sourceIfTrue, const LLVector4a& sourceIfFalse )
{
	// ((( sourceIfTrue ^ sourceIfFalse ) & mask) ^ sourceIfFalse )
	// E.g., sourceIfFalse = 1010b, sourceIfTrue = 0101b, mask = 1100b
	// (sourceIfTrue ^ sourceIfFalse) = 1111b --> & mask = 1100b --> ^ sourceIfFalse = 0110b, 
	// as expected (01 from sourceIfTrue, 10 from sourceIfFalse)
	// Courtesy of Mark++, http://markplusplus.wordpress.com/2007/03/14/fast-sse-select-operation/
	mQ = _mm_xor_ps( sourceIfFalse, _mm_and_ps( mask, _mm_xor_ps( sourceIfTrue, sourceIfFalse ) ) );
}

////////////////////////////////////
// ALGEBRAIC
////////////////////////////////////

// Set this to the element-wise (a + b)
inline void LLVector4a::setAdd(const LLVector4a& a, const LLVector4a& b)
{
	mQ = _mm_add_ps(a.mQ, b.mQ);
}

// Set this to element-wise (a - b)
inline void LLVector4a::setSub(const LLVector4a& a, const LLVector4a& b)
{
	mQ = _mm_sub_ps(a.mQ, b.mQ);
}

// Set this to element-wise multiply (a * b)
inline void LLVector4a::setMul(const LLVector4a& a, const LLVector4a& b)
{
	mQ = _mm_mul_ps(a.mQ, b.mQ);
}

// Set this to element-wise quotient (a / b)
inline void LLVector4a::setDiv(const LLVector4a& a, const LLVector4a& b)
{
	mQ = _mm_div_ps( a.mQ, b.mQ );
}

// Set this to the element-wise absolute value of src
inline void LLVector4a::setAbs(const LLVector4a& src)
{
	static const LL_ALIGN_16(U32 F_ABS_MASK_4A[4]) = { 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF, 0x7FFFFFFF };
	mQ = _mm_and_ps(src.mQ, *reinterpret_cast<const LLQuad*>(F_ABS_MASK_4A));
}

// Add to each component in this vector the corresponding component in rhs
inline void LLVector4a::add(const LLVector4a& rhs)
{
	mQ = _mm_add_ps(mQ, rhs.mQ);	
}

// Subtract from each component in this vector the corresponding component in rhs
inline void LLVector4a::sub(const LLVector4a& rhs)
{
	mQ = _mm_sub_ps(mQ, rhs.mQ);
}

// Multiply each component in this vector by the corresponding component in rhs
inline void LLVector4a::mul(const LLVector4a& rhs)
{
	mQ = _mm_mul_ps(mQ, rhs.mQ);	
}

// Divide each component in this vector by the corresponding component in rhs
inline void LLVector4a::div(const LLVector4a& rhs)
{
	// TODO: Check accuracy, maybe add divFast
	mQ = _mm_div_ps(mQ, rhs.mQ);
}

// Multiply this vector by x in a scalar fashion
inline void LLVector4a::mul(const F32 x) 
{
	LLVector4a t;
	t.splat(x);
	
	mQ = _mm_mul_ps(mQ, t.mQ);
}

// Set this to (a x b) (geometric cross-product)
inline void LLVector4a::setCross3(const LLVector4a& a, const LLVector4a& b)
{
	// Vectors are stored in memory in w, z, y, x order from high to low
	// Set vector1 = { a[W], a[X], a[Z], a[Y] }
	const LLQuad vector1 = _mm_shuffle_ps( a.mQ, a.mQ, _MM_SHUFFLE( 3, 0, 2, 1 ));
	// Set vector2 = { b[W], b[Y], b[X], b[Z] }
	const LLQuad vector2 = _mm_shuffle_ps( b.mQ, b.mQ, _MM_SHUFFLE( 3, 1, 0, 2 ));
	// mQ = { a[W]*b[W], a[X]*b[Y], a[Z]*b[X], a[Y]*b[Z] }
	mQ = _mm_mul_ps( vector1, vector2 );
	// vector3 = { a[W], a[Y], a[X], a[Z] }
	const LLQuad vector3 = _mm_shuffle_ps( a.mQ, a.mQ, _MM_SHUFFLE( 3, 1, 0, 2 ));
	// vector4 = { b[W], b[X], b[Z], b[Y] }
	const LLQuad vector4 = _mm_shuffle_ps( b.mQ, b.mQ, _MM_SHUFFLE( 3, 0, 2, 1 ));
	// mQ = { 0, a[X]*b[Y] - a[Y]*b[X], a[Z]*b[X] - a[X]*b[Z], a[Y]*b[Z] - a[Z]*b[Y] }
	mQ = _mm_sub_ps( mQ, _mm_mul_ps( vector3, vector4 ));
}

/* This function works, but may be slightly slower than the one below on older machines
 inline void LLVector4a::setAllDot3(const LLVector4a& a, const LLVector4a& b)
 {
 // ab = { a[W]*b[W], a[Z]*b[Z], a[Y]*b[Y], a[X]*b[X] }
 const LLQuad ab = _mm_mul_ps( a.mQ, b.mQ );
 // yzxw = { a[W]*b[W], a[Z]*b[Z], a[X]*b[X], a[Y]*b[Y] }
 const LLQuad wzxy = _mm_shuffle_ps( ab, ab, _MM_SHUFFLE(3, 2, 0, 1 ));
 // xPlusY = { 2*a[W]*b[W], 2 * a[Z] * b[Z], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] }
 const LLQuad xPlusY = _mm_add_ps(ab, wzxy);
 // xPlusYSplat = { a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } 
 const LLQuad xPlusYSplat = _mm_movelh_ps(xPlusY, xPlusY);
 // zSplat = { a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z] }
 const LLQuad zSplat = _mm_shuffle_ps( ab, ab, _MM_SHUFFLE( 2, 2, 2, 2 ));
 // mQ = { a[Z] * b[Z] + a[Y] * b[Y] + a[X] * b[X], same, same, same }
 mQ = _mm_add_ps(zSplat, xPlusYSplat);
 }*/

// Set all elements to the dot product of the x, y, and z elements in a and b
inline void LLVector4a::setAllDot3(const LLVector4a& a, const LLVector4a& b)
{
	// ab = { a[W]*b[W], a[Z]*b[Z], a[Y]*b[Y], a[X]*b[X] }
	const LLQuad ab = _mm_mul_ps( a.mQ, b.mQ );
	// yzxw = { a[W]*b[W], a[Z]*b[Z], a[X]*b[X], a[Y]*b[Y] }
	const __m128i wzxy = _mm_shuffle_epi32(_mm_castps_si128(ab), _MM_SHUFFLE(3, 2, 0, 1 ));
	// xPlusY = { 2*a[W]*b[W], 2 * a[Z] * b[Z], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] }
	const LLQuad xPlusY = _mm_add_ps(ab, _mm_castsi128_ps(wzxy));
	// xPlusYSplat = { a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } 
	const LLQuad xPlusYSplat = _mm_movelh_ps(xPlusY, xPlusY);
	// zSplat = { a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z], a[Z]*b[Z] }
	const __m128i zSplat = _mm_shuffle_epi32(_mm_castps_si128(ab), _MM_SHUFFLE( 2, 2, 2, 2 ));
	// mQ = { a[Z] * b[Z] + a[Y] * b[Y] + a[X] * b[X], same, same, same }
	mQ = _mm_add_ps(_mm_castsi128_ps(zSplat), xPlusYSplat);
}

// Set all elements to the dot product of the x, y, z, and w elements in a and b
inline void LLVector4a::setAllDot4(const LLVector4a& a, const LLVector4a& b)
{
	// ab = { a[W]*b[W], a[Z]*b[Z], a[Y]*b[Y], a[X]*b[X] }
	const LLQuad ab = _mm_mul_ps( a.mQ, b.mQ );
	// yzxw = { a[W]*b[W], a[Z]*b[Z], a[X]*b[X], a[Y]*b[Y] }
	const __m128i zwxy = _mm_shuffle_epi32(_mm_castps_si128(ab), _MM_SHUFFLE(2, 3, 0, 1 ));
	// zPlusWandXplusY = { a[W]*b[W] + a[Z]*b[Z], a[Z] * b[Z] + a[W]*b[W], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] }
	const LLQuad zPlusWandXplusY = _mm_add_ps(ab, _mm_castsi128_ps(zwxy));
	// xPlusYSplat = { a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y], a[Y]*b[Y] + a[X] * b[X], a[X] * b[X] + a[Y] * b[Y] } 
	const LLQuad xPlusYSplat = _mm_movelh_ps(zPlusWandXplusY, zPlusWandXplusY);
	const LLQuad zPlusWSplat = _mm_movehl_ps(zPlusWandXplusY, zPlusWandXplusY);

	// mQ = { a[W]*b[W] + a[Z] * b[Z] + a[Y] * b[Y] + a[X] * b[X], same, same, same }
	mQ = _mm_add_ps(xPlusYSplat, zPlusWSplat);
}

// Return the 3D dot product of this vector and b
inline LLSimdScalar LLVector4a::dot3(const LLVector4a& b) const
{
	const LLQuad ab = _mm_mul_ps( mQ, b.mQ );
	const LLQuad splatY = _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128(ab), _MM_SHUFFLE(1, 1, 1, 1) ) );
	const LLQuad splatZ = _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128(ab), _MM_SHUFFLE(2, 2, 2, 2) ) );
	const LLQuad xPlusY = _mm_add_ps( ab, splatY );
	return _mm_add_ps( xPlusY, splatZ );	
}

// Return the 4D dot product of this vector and b
inline LLSimdScalar LLVector4a::dot4(const LLVector4a& b) const
{
	// ab = { w, z, y, x }
 	const LLQuad ab = _mm_mul_ps( mQ, b.mQ );
 	// upperProdsInLowerElems = { y, x, y, x }
	const LLQuad upperProdsInLowerElems = _mm_movehl_ps( ab, ab );
	// sumOfPairs = { w+y, z+x, 2y, 2x }
 	const LLQuad sumOfPairs = _mm_add_ps( upperProdsInLowerElems, ab );
	// shuffled = { z+x, z+x, z+x, z+x }
	const LLQuad shuffled = _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( sumOfPairs ), _MM_SHUFFLE(1, 1, 1, 1) ) );
	return _mm_add_ss( sumOfPairs, shuffled );
}

// Normalize this vector with respect to the x, y, and z components only. Accurate to 22 bites of precision. W component is destroyed
// Note that this does not consider zero length vectors!
inline void LLVector4a::normalize3()
{
	// lenSqrd = a dot a
	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );
	// rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 }
	const LLQuad rsqrt = _mm_rsqrt_ps(lenSqrd.mQ);
	static const LLQuad half = { 0.5f, 0.5f, 0.5f, 0.5f };
	static const LLQuad three = {3.f, 3.f, 3.f, 3.f };
	// Now we do one round of Newton-Raphson approximation to get full accuracy
	// According to the Newton-Raphson method, given a first 'w' for the root of f(x) = 1/x^2 - a (i.e., x = 1/sqrt(a))
	// the next better approximation w[i+1] = w - f(w)/f'(w) = w - (1/w^2 - a)/(-2*w^(-3))
	// w[i+1] = w + 0.5 * (1/w^2 - a) * w^3 = w + 0.5 * (w - a*w^3) = 1.5 * w - 0.5 * a * w^3
	// = 0.5 * w * (3 - a*w^2)
	// Our first approx is w = rsqrt. We need out = a * w[i+1] (this is the input vector 'a', not the 'a' from the above formula
	// which is actually lenSqrd). So out = a * [0.5*rsqrt * (3 - lenSqrd*rsqrt*rsqrt)]
	const LLQuad AtimesRsqrt = _mm_mul_ps( lenSqrd.mQ, rsqrt );
	const LLQuad AtimesRsqrtTimesRsqrt = _mm_mul_ps( AtimesRsqrt, rsqrt );
	const LLQuad threeMinusAtimesRsqrtTimesRsqrt = _mm_sub_ps(three, AtimesRsqrtTimesRsqrt );
	const LLQuad nrApprox = _mm_mul_ps(half, _mm_mul_ps(rsqrt, threeMinusAtimesRsqrtTimesRsqrt));
	mQ = _mm_mul_ps( mQ, nrApprox );
}

// Normalize this vector with respect to all components. Accurate to 22 bites of precision.
// Note that this does not consider zero length vectors!
inline void LLVector4a::normalize4()
{
	// lenSqrd = a dot a
	LLVector4a lenSqrd; lenSqrd.setAllDot4( *this, *this );
	// rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 }
	const LLQuad rsqrt = _mm_rsqrt_ps(lenSqrd.mQ);
	static const LLQuad half = { 0.5f, 0.5f, 0.5f, 0.5f };
	static const LLQuad three = {3.f, 3.f, 3.f, 3.f };
	// Now we do one round of Newton-Raphson approximation to get full accuracy
	// According to the Newton-Raphson method, given a first 'w' for the root of f(x) = 1/x^2 - a (i.e., x = 1/sqrt(a))
	// the next better approximation w[i+1] = w - f(w)/f'(w) = w - (1/w^2 - a)/(-2*w^(-3))
	// w[i+1] = w + 0.5 * (1/w^2 - a) * w^3 = w + 0.5 * (w - a*w^3) = 1.5 * w - 0.5 * a * w^3
	// = 0.5 * w * (3 - a*w^2)
	// Our first approx is w = rsqrt. We need out = a * w[i+1] (this is the input vector 'a', not the 'a' from the above formula
	// which is actually lenSqrd). So out = a * [0.5*rsqrt * (3 - lenSqrd*rsqrt*rsqrt)]
	const LLQuad AtimesRsqrt = _mm_mul_ps( lenSqrd.mQ, rsqrt );
	const LLQuad AtimesRsqrtTimesRsqrt = _mm_mul_ps( AtimesRsqrt, rsqrt );
	const LLQuad threeMinusAtimesRsqrtTimesRsqrt = _mm_sub_ps(three, AtimesRsqrtTimesRsqrt );
	const LLQuad nrApprox = _mm_mul_ps(half, _mm_mul_ps(rsqrt, threeMinusAtimesRsqrtTimesRsqrt));
	mQ = _mm_mul_ps( mQ, nrApprox );
}

// Normalize this vector with respect to the x, y, and z components only. Accurate to 22 bites of precision. W component is destroyed
// Note that this does not consider zero length vectors!
inline LLSimdScalar LLVector4a::normalize3withLength()
{
	// lenSqrd = a dot a
	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );
	// rsqrt = approximate reciprocal square (i.e., { ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2, ~1/len(a)^2 }
	const LLQuad rsqrt = _mm_rsqrt_ps(lenSqrd.mQ);
	static const LLQuad half = { 0.5f, 0.5f, 0.5f, 0.5f };
	static const LLQuad three = {3.f, 3.f, 3.f, 3.f };
	// Now we do one round of Newton-Raphson approximation to get full accuracy
	// According to the Newton-Raphson method, given a first 'w' for the root of f(x) = 1/x^2 - a (i.e., x = 1/sqrt(a))
	// the next better approximation w[i+1] = w - f(w)/f'(w) = w - (1/w^2 - a)/(-2*w^(-3))
	// w[i+1] = w + 0.5 * (1/w^2 - a) * w^3 = w + 0.5 * (w - a*w^3) = 1.5 * w - 0.5 * a * w^3
	// = 0.5 * w * (3 - a*w^2)
	// Our first approx is w = rsqrt. We need out = a * w[i+1] (this is the input vector 'a', not the 'a' from the above formula
	// which is actually lenSqrd). So out = a * [0.5*rsqrt * (3 - lenSqrd*rsqrt*rsqrt)]
	const LLQuad AtimesRsqrt = _mm_mul_ps( lenSqrd.mQ, rsqrt );
	const LLQuad AtimesRsqrtTimesRsqrt = _mm_mul_ps( AtimesRsqrt, rsqrt );
	const LLQuad threeMinusAtimesRsqrtTimesRsqrt = _mm_sub_ps(three, AtimesRsqrtTimesRsqrt );
	const LLQuad nrApprox = _mm_mul_ps(half, _mm_mul_ps(rsqrt, threeMinusAtimesRsqrtTimesRsqrt));
	mQ = _mm_mul_ps( mQ, nrApprox );
	return _mm_sqrt_ss(lenSqrd);
}

// Normalize this vector with respect to the x, y, and z components only. Accurate only to 10-12 bits of precision. W component is destroyed
// Note that this does not consider zero length vectors!
inline void LLVector4a::normalize3fast()
{
	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );
	const LLQuad approxRsqrt = _mm_rsqrt_ps(lenSqrd.mQ);
	mQ = _mm_mul_ps( mQ, approxRsqrt );
}

inline void LLVector4a::normalize3fast_checked(LLVector4a* d)
{
	if (!isFinite3())
	{
		*this = d ? *d : LLVector4a(0,1,0,1);
		return;
	}

	LLVector4a lenSqrd; lenSqrd.setAllDot3( *this, *this );

	if (lenSqrd.getF32ptr()[0] <= FLT_EPSILON)
	{
		*this = d ? *d : LLVector4a(0,1,0,1);
		return;
	}

	const LLQuad approxRsqrt = _mm_rsqrt_ps(lenSqrd.mQ);
	mQ = _mm_mul_ps( mQ, approxRsqrt );
}

// Return true if this vector is normalized with respect to x,y,z up to tolerance
inline LLBool32 LLVector4a::isNormalized3( F32 tolerance ) const
{
	static LL_ALIGN_16(const U32 ones[4]) = { 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 };
	LLSimdScalar tol = _mm_load_ss( &tolerance );
	tol = _mm_mul_ss( tol, tol );
	LLVector4a lenSquared; lenSquared.setAllDot3( *this, *this );
	lenSquared.sub( *reinterpret_cast<const LLVector4a*>(ones) );
	lenSquared.setAbs(lenSquared);
	return _mm_comile_ss( lenSquared, tol );		
}

// Return true if this vector is normalized with respect to all components up to tolerance
inline LLBool32 LLVector4a::isNormalized4( F32 tolerance ) const
{
	static LL_ALIGN_16(const U32 ones[4]) = { 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 };
	LLSimdScalar tol = _mm_load_ss( &tolerance );
	tol = _mm_mul_ss( tol, tol );
	LLVector4a lenSquared; lenSquared.setAllDot4( *this, *this );
	lenSquared.sub( *reinterpret_cast<const LLVector4a*>(ones) );
	lenSquared.setAbs(lenSquared);
	return _mm_comile_ss( lenSquared, tol );		
}

// Set all elements to the length of vector 'v' 
inline void LLVector4a::setAllLength3( const LLVector4a& v )
{
	LLVector4a lenSqrd;
	lenSqrd.setAllDot3(v, v);
	
	mQ = _mm_sqrt_ps(lenSqrd.mQ);
}

// Get this vector's length
inline LLSimdScalar LLVector4a::getLength3() const
{
	return _mm_sqrt_ss( dot3( (const LLVector4a)mQ ) );
}

// Set the components of this vector to the minimum of the corresponding components of lhs and rhs
inline void LLVector4a::setMin(const LLVector4a& lhs, const LLVector4a& rhs)
{
	mQ = _mm_min_ps(lhs.mQ, rhs.mQ);
}

// Set the components of this vector to the maximum of the corresponding components of lhs and rhs
inline void LLVector4a::setMax(const LLVector4a& lhs, const LLVector4a& rhs)
{
	mQ = _mm_max_ps(lhs.mQ, rhs.mQ);
}

// Set this to  lhs + (rhs-lhs)*c
inline void LLVector4a::setLerp(const LLVector4a& lhs, const LLVector4a& rhs, F32 c)
{
	LLVector4a t;
	t.setSub(rhs,lhs);
	t.mul(c);
	setAdd(lhs, t);
}

inline LLBool32 LLVector4a::isFinite3() const
{
	static LL_ALIGN_16(const U32 nanOrInfMask[4]) = { 0x7f800000, 0x7f800000, 0x7f800000, 0x7f800000 };
	ll_assert_aligned(nanOrInfMask,16);
	const __m128i nanOrInfMaskV = *reinterpret_cast<const __m128i*> (nanOrInfMask);
	const __m128i maskResult = _mm_and_si128( _mm_castps_si128(mQ), nanOrInfMaskV );
	const LLVector4Logical equalityCheck = _mm_castsi128_ps(_mm_cmpeq_epi32( maskResult, nanOrInfMaskV ));
	return !equalityCheck.areAnySet( LLVector4Logical::MASK_XYZ );
}
	
inline LLBool32 LLVector4a::isFinite4() const
{
	static LL_ALIGN_16(const U32 nanOrInfMask[4]) = { 0x7f800000, 0x7f800000, 0x7f800000, 0x7f800000 };
	const __m128i nanOrInfMaskV = *reinterpret_cast<const __m128i*> (nanOrInfMask);
	const __m128i maskResult = _mm_and_si128( _mm_castps_si128(mQ), nanOrInfMaskV );
	const LLVector4Logical equalityCheck = _mm_castsi128_ps(_mm_cmpeq_epi32( maskResult, nanOrInfMaskV ));
	return !equalityCheck.areAnySet( LLVector4Logical::MASK_XYZW );
}

inline void LLVector4a::setRotatedInv( const LLRotation& rot, const LLVector4a& vec )
{
	LLRotation inv; inv.setTranspose( rot );
	setRotated( inv, vec );
}

inline void LLVector4a::setRotatedInv( const LLQuaternion2& quat, const LLVector4a& vec )
{
	LLQuaternion2 invRot; invRot.setConjugate( quat );
	setRotated(invRot, vec);
}

inline void LLVector4a::clamp( const LLVector4a& low, const LLVector4a& high )
{
	const LLVector4Logical highMask = greaterThan( high );
	const LLVector4Logical lowMask = lessThan( low );

	setSelectWithMask( highMask, high, *this );
	setSelectWithMask( lowMask, low, *this );
}


////////////////////////////////////
// LOGICAL
////////////////////////////////////	
// The functions in this section will compare the elements in this vector
// to those in rhs and return an LLVector4Logical with all bits set in elements
// where the comparison was true and all bits unset in elements where the comparison
// was false. See llvector4logica.h
////////////////////////////////////
// WARNING: Other than equals3 and equals4, these functions do NOT account
// for floating point tolerance. You should include the appropriate tolerance
// in the inputs.
////////////////////////////////////

inline LLVector4Logical LLVector4a::greaterThan(const LLVector4a& rhs) const
{	
	return _mm_cmpgt_ps(mQ, rhs.mQ);
}

inline LLVector4Logical LLVector4a::lessThan(const LLVector4a& rhs) const
{
	return _mm_cmplt_ps(mQ, rhs.mQ);
}

inline LLVector4Logical LLVector4a::greaterEqual(const LLVector4a& rhs) const
{
	return _mm_cmpge_ps(mQ, rhs.mQ);
}

inline LLVector4Logical LLVector4a::lessEqual(const LLVector4a& rhs) const
{
	return _mm_cmple_ps(mQ, rhs.mQ);
}

inline LLVector4Logical LLVector4a::equal(const LLVector4a& rhs) const
{
	return _mm_cmpeq_ps(mQ, rhs.mQ);
}

// Returns true if this and rhs are componentwise equal up to the specified absolute tolerance
inline bool LLVector4a::equals4(const LLVector4a& rhs, F32 tolerance ) const
{
	LLVector4a diff; diff.setSub( *this, rhs );
	diff.setAbs( diff );
	const LLQuad tol = _mm_set1_ps( tolerance );
	const LLQuad cmp = _mm_cmplt_ps( diff, tol );
	return (_mm_movemask_ps( cmp ) & LLVector4Logical::MASK_XYZW) == LLVector4Logical::MASK_XYZW;
}

inline bool LLVector4a::equals3(const LLVector4a& rhs, F32 tolerance ) const
{
	LLVector4a diff; diff.setSub( *this, rhs );
	diff.setAbs( diff );
	const LLQuad tol = _mm_set1_ps( tolerance );
	const LLQuad t = _mm_cmplt_ps( diff, tol ); 
	return (_mm_movemask_ps( t ) & LLVector4Logical::MASK_XYZ) == LLVector4Logical::MASK_XYZ;
	
}

////////////////////////////////////
// OPERATORS
////////////////////////////////////	

// Do NOT add aditional operators without consulting someone with SSE experience
inline const LLVector4a& LLVector4a::operator= ( const LLVector4a& rhs )
{
	mQ = rhs.mQ;
	return *this;
}

inline const LLVector4a& LLVector4a::operator= ( const LLQuad& rhs )
{
	mQ = rhs;
	return *this;
}

inline LLVector4a::operator LLQuad() const
{
	return mQ;
}