summaryrefslogtreecommitdiff
path: root/indra/llmath/llvector4a.cpp
blob: b62c17302fae9a976515cc2f4abb18fa46c8fe56 (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
/** 
 * @file llvector4a.cpp
 * @brief SIMD vector implementation
 *
 * $LicenseInfo:firstyear=2010&license=viewergpl$
 * 
 * Copyright (c) 2007-2010, Linden Research, Inc.
 * 
 * Second Life Viewer Source Code
 * The source code in this file ("Source Code") is provided by Linden Lab
 * to you under the terms of the GNU General Public License, version 2.0
 * ("GPL"), unless you have obtained a separate licensing agreement
 * ("Other License"), formally executed by you and Linden Lab.  Terms of
 * the GPL can be found in doc/GPL-license.txt in this distribution, or
 * online at http://secondlifegrid.net/programs/open_source/licensing/gplv2
 * 
 * There are special exceptions to the terms and conditions of the GPL as
 * it is applied to this Source Code. View the full text of the exception
 * in the file doc/FLOSS-exception.txt in this software distribution, or
 * online at
 * http://secondlifegrid.net/programs/open_source/licensing/flossexception
 * 
 * By copying, modifying or distributing this software, you acknowledge
 * that you have read and understood your obligations described above,
 * and agree to abide by those obligations.
 * 
 * ALL LINDEN LAB SOURCE CODE IS PROVIDED "AS IS." LINDEN LAB MAKES NO
 * WARRANTIES, EXPRESS, IMPLIED OR OTHERWISE, REGARDING ITS ACCURACY,
 * COMPLETENESS OR PERFORMANCE.
 * $/LicenseInfo$
 */

#include "llmath.h"
#include "llquantize.h"

extern const LLQuad F_ZERO_4A		= { 0, 0, 0, 0 };
extern const LLQuad F_APPROXIMATELY_ZERO_4A = { 
	F_APPROXIMATELY_ZERO,
	F_APPROXIMATELY_ZERO,
	F_APPROXIMATELY_ZERO,
	F_APPROXIMATELY_ZERO
};

extern const LLVector4a LL_V4A_ZERO = reinterpret_cast<const LLVector4a&> ( F_ZERO_4A );
extern const LLVector4a LL_V4A_EPSILON = reinterpret_cast<const LLVector4a&> ( F_APPROXIMATELY_ZERO_4A );

/*static */void LLVector4a::memcpyNonAliased16(F32* __restrict dst, const F32* __restrict src, size_t bytes)
{
	assert(src != NULL);
	assert(dst != NULL);
	assert(bytes > 0);
	assert((bytes % sizeof(F32))== 0); 
	
	F32* end = dst + (bytes / sizeof(F32) );

	if (bytes > 64)
	{
		F32* begin_64 = LL_NEXT_ALIGNED_ADDRESS_64(dst);
		
		//at least 64 (16*4) bytes before the end of the destination, switch to 16 byte copies
		F32* end_64 = end-16;
		
		_mm_prefetch((char*)begin_64, _MM_HINT_NTA);
		_mm_prefetch((char*)begin_64 + 64, _MM_HINT_NTA);
		_mm_prefetch((char*)begin_64 + 128, _MM_HINT_NTA);
		_mm_prefetch((char*)begin_64 + 192, _MM_HINT_NTA);
		
		while (dst < begin_64)
		{
			copy4a(dst, src);
			dst += 4;
			src += 4;
		}
		
		while (dst < end_64)
		{
			_mm_prefetch((char*)src + 512, _MM_HINT_NTA);
			_mm_prefetch((char*)dst + 512, _MM_HINT_NTA);
			copy4a(dst, src);
			copy4a(dst+4, src+4);
			copy4a(dst+8, src+8);
			copy4a(dst+12, src+12);
			
			dst += 16;
			src += 16;
		}
	}

	while (dst < end)
	{
		copy4a(dst, src);
		dst += 4;
		src += 4;
	}
}

void LLVector4a::setRotated( const LLRotation& rot, const LLVector4a& vec )
{
	const LLVector4a col0 = rot.getColumn(0);
	const LLVector4a col1 = rot.getColumn(1);
	const LLVector4a col2 = rot.getColumn(2);

	LLVector4a result = _mm_load_ss( vec.getF32ptr() );
	result.splat<0>( result );
	result.mul( col0 );

	{
		LLVector4a yyyy = _mm_load_ss( vec.getF32ptr() +  1 );
		yyyy.splat<0>( yyyy );
		yyyy.mul( col1 ); 
		result.add( yyyy );
	}

	{
		LLVector4a zzzz = _mm_load_ss( vec.getF32ptr() +  2 );
		zzzz.splat<0>( zzzz );
		zzzz.mul( col2 );
		result.add( zzzz );
	}

	*this = result;
}

void LLVector4a::setRotated( const LLQuaternion2& quat, const LLVector4a& vec )
{
	const LLVector4a& quatVec = quat.getVector4a();
	LLVector4a temp; temp.setCross3(quatVec, vec);
	temp.add( temp );
	
	const LLVector4a realPart( quatVec.getScalarAt<3>() );
	LLVector4a tempTimesReal; tempTimesReal.setMul( temp, realPart );

	mQ = vec;
	add( tempTimesReal );
	
	LLVector4a imagCrossTemp; imagCrossTemp.setCross3( quatVec, temp );
	add(imagCrossTemp);
}

void LLVector4a::quantize8( const LLVector4a& low, const LLVector4a& high )
{
	LLVector4a val(mQ);
	LLVector4a delta; delta.setSub( high, low );

	{
		val.clamp(low, high);
		val.sub(low);

		// 8-bit quantization means we can do with just 12 bits of reciprocal accuracy
		const LLVector4a oneOverDelta = _mm_rcp_ps(delta.mQ);
// 		{
// 			static LL_ALIGN_16( const F32 F_TWO_4A[4] ) = { 2.f, 2.f, 2.f, 2.f };
// 			LLVector4a two; two.load4a( F_TWO_4A );
// 
// 			// Here we use _mm_rcp_ps plus one round of newton-raphson
// 			// We wish to find 'x' such that x = 1/delta
// 			// As a first approximation, we take x0 = _mm_rcp_ps(delta)
// 			// Then x1 = 2 * x0 - a * x0^2 or x1 = x0 * ( 2 - a * x0 )
// 			// See Intel AP-803 http://ompf.org/!/Intel_application_note_AP-803.pdf
// 			const LLVector4a recipApprox = _mm_rcp_ps(delta.mQ);
// 			oneOverDelta.setMul( delta, recipApprox );
// 			oneOverDelta.setSub( two, oneOverDelta );
// 			oneOverDelta.mul( recipApprox );
// 		}

		val.mul(oneOverDelta);
		val.mul(*reinterpret_cast<const LLVector4a*>(F_U8MAX_4A));
	}

	val = _mm_cvtepi32_ps(_mm_cvtps_epi32( val.mQ ));

	{
		val.mul(*reinterpret_cast<const LLVector4a*>(F_OOU8MAX_4A));
		val.mul(delta);
		val.add(low);
	}

	{
		LLVector4a maxError; maxError.setMul(delta, *reinterpret_cast<const LLVector4a*>(F_OOU8MAX_4A));
		LLVector4a absVal; absVal.setAbs( val );
		setSelectWithMask( absVal.lessThan( maxError ), F_ZERO_4A, val );
	}	
}

void LLVector4a::quantize16( const LLVector4a& low, const LLVector4a& high )
{
	LLVector4a val(mQ);
	LLVector4a delta; delta.setSub( high, low );

	{
		val.clamp(low, high);
		val.sub(low);

		// 16-bit quantization means we need a round of Newton-Raphson
		LLVector4a oneOverDelta;
		{
			static LL_ALIGN_16( const F32 F_TWO_4A[4] ) = { 2.f, 2.f, 2.f, 2.f };
			LLVector4a two; two.load4a( F_TWO_4A );

			// Here we use _mm_rcp_ps plus one round of newton-raphson
			// We wish to find 'x' such that x = 1/delta
			// As a first approximation, we take x0 = _mm_rcp_ps(delta)
			// Then x1 = 2 * x0 - a * x0^2 or x1 = x0 * ( 2 - a * x0 )
			// See Intel AP-803 http://ompf.org/!/Intel_application_note_AP-803.pdf
			const LLVector4a recipApprox = _mm_rcp_ps(delta.mQ);
			oneOverDelta.setMul( delta, recipApprox );
			oneOverDelta.setSub( two, oneOverDelta );
			oneOverDelta.mul( recipApprox );
		}

		val.mul(oneOverDelta);
		val.mul(*reinterpret_cast<const LLVector4a*>(F_U16MAX_4A));
	}

	val = _mm_cvtepi32_ps(_mm_cvtps_epi32( val.mQ ));

	{
		val.mul(*reinterpret_cast<const LLVector4a*>(F_OOU16MAX_4A));
		val.mul(delta);
		val.add(low);
	}

	{
		LLVector4a maxError; maxError.setMul(delta, *reinterpret_cast<const LLVector4a*>(F_OOU16MAX_4A));
		LLVector4a absVal; absVal.setAbs( val );
		setSelectWithMask( absVal.lessThan( maxError ), F_ZERO_4A, val );
	}	
}