1 ////////////////////////////////////////////////////////////////////////////////
2 ///
3 /// MMX optimized routines. All MMX optimized functions have been gathered into
4 /// this single source code file, regardless to their class or original source
5 /// code file, in order to ease porting the library to other compiler and
6 /// processor platforms.
7 ///
8 /// The MMX-optimizations are programmed using MMX compiler intrinsics that
9 /// are supported both by Microsoft Visual C++ and GCC compilers, so this file
10 /// should compile with both toolsets.
11 ///
12 /// NOTICE: If using Visual Studio 6.0, you'll need to install the "Visual C++
13 /// 6.0 processor pack" update to support compiler intrinsic syntax. The update
14 /// is available for download at Microsoft Developers Network, see here:
15 /// http://msdn.microsoft.com/en-us/vstudio/aa718349.aspx
16 ///
17 /// Author : Copyright (c) Olli Parviainen
18 /// Author e-mail : oparviai 'at' iki.fi
19 /// SoundTouch WWW: http://www.surina.net/soundtouch
20 ///
21 ////////////////////////////////////////////////////////////////////////////////
22 //
23 // Last changed : $Date: 2012-11-08 20:53:01 +0200 (Thu, 08 Nov 2012) $
24 // File revision : $Revision: 4 $
25 //
26 // $Id: mmx_optimized.cpp 160 2012-11-08 18:53:01Z oparviai $
27 //
28 ////////////////////////////////////////////////////////////////////////////////
29 //
30 // License :
31 //
32 // SoundTouch audio processing library
33 // Copyright (c) Olli Parviainen
34 //
35 // This library is free software; you can redistribute it and/or
36 // modify it under the terms of the GNU Lesser General Public
37 // License as published by the Free Software Foundation; either
38 // version 2.1 of the License, or (at your option) any later version.
39 //
40 // This library is distributed in the hope that it will be useful,
41 // but WITHOUT ANY WARRANTY; without even the implied warranty of
42 // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
43 // Lesser General Public License for more details.
44 //
45 // You should have received a copy of the GNU Lesser General Public
46 // License along with this library; if not, write to the Free Software
47 // Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
48 //
49 ////////////////////////////////////////////////////////////////////////////////
50
51 #include "STTypes.h"
52
53 #ifdef SOUNDTOUCH_ALLOW_MMX
54 // MMX routines available only with integer sample type
55
56 using namespace soundtouch;
57
58 //////////////////////////////////////////////////////////////////////////////
59 //
60 // implementation of MMX optimized functions of class 'TDStretchMMX'
61 //
62 //////////////////////////////////////////////////////////////////////////////
63
64 #include "TDStretch.h"
65 #include <mmintrin.h>
66 #include <limits.h>
67 #include <math.h>
68
69
70 // Calculates cross correlation of two buffers
calcCrossCorr(const short * pV1,const short * pV2) const71 double TDStretchMMX::calcCrossCorr(const short *pV1, const short *pV2) const
72 {
73 const __m64 *pVec1, *pVec2;
74 __m64 shifter;
75 __m64 accu, normaccu;
76 long corr, norm;
77 int i;
78
79 pVec1 = (__m64*)pV1;
80 pVec2 = (__m64*)pV2;
81
82 shifter = _m_from_int(overlapDividerBits);
83 normaccu = accu = _mm_setzero_si64();
84
85 // Process 4 parallel sets of 2 * stereo samples or 4 * mono samples
86 // during each round for improved CPU-level parallellization.
87 for (i = 0; i < channels * overlapLength / 16; i ++)
88 {
89 __m64 temp, temp2;
90
91 // dictionary of instructions:
92 // _m_pmaddwd : 4*16bit multiply-add, resulting two 32bits = [a0*b0+a1*b1 ; a2*b2+a3*b3]
93 // _mm_add_pi32 : 2*32bit add
94 // _m_psrad : 32bit right-shift
95
96 temp = _mm_add_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]),
97 _mm_madd_pi16(pVec1[1], pVec2[1]));
98 temp2 = _mm_add_pi32(_mm_madd_pi16(pVec1[0], pVec1[0]),
99 _mm_madd_pi16(pVec1[1], pVec1[1]));
100 accu = _mm_add_pi32(accu, _mm_sra_pi32(temp, shifter));
101 normaccu = _mm_add_pi32(normaccu, _mm_sra_pi32(temp2, shifter));
102
103 temp = _mm_add_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]),
104 _mm_madd_pi16(pVec1[3], pVec2[3]));
105 temp2 = _mm_add_pi32(_mm_madd_pi16(pVec1[2], pVec1[2]),
106 _mm_madd_pi16(pVec1[3], pVec1[3]));
107 accu = _mm_add_pi32(accu, _mm_sra_pi32(temp, shifter));
108 normaccu = _mm_add_pi32(normaccu, _mm_sra_pi32(temp2, shifter));
109
110 pVec1 += 4;
111 pVec2 += 4;
112 }
113
114 // copy hi-dword of mm0 to lo-dword of mm1, then sum mmo+mm1
115 // and finally store the result into the variable "corr"
116
117 accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
118 corr = _m_to_int(accu);
119
120 normaccu = _mm_add_pi32(normaccu, _mm_srli_si64(normaccu, 32));
121 norm = _m_to_int(normaccu);
122
123 // Clear MMS state
124 _m_empty();
125
126 // Normalize result by dividing by sqrt(norm) - this step is easiest
127 // done using floating point operation
128 if (norm == 0) norm = 1; // to avoid div by zero
129
130 return (double)corr / sqrt((double)norm);
131 // Note: Warning about the missing EMMS instruction is harmless
132 // as it'll be called elsewhere.
133 }
134
135
136
clearCrossCorrState()137 void TDStretchMMX::clearCrossCorrState()
138 {
139 // Clear MMS state
140 _m_empty();
141 //_asm EMMS;
142 }
143
144
145
146 // MMX-optimized version of the function overlapStereo
overlapStereo(short * output,const short * input) const147 void TDStretchMMX::overlapStereo(short *output, const short *input) const
148 {
149 const __m64 *pVinput, *pVMidBuf;
150 __m64 *pVdest;
151 __m64 mix1, mix2, adder, shifter;
152 int i;
153
154 pVinput = (const __m64*)input;
155 pVMidBuf = (const __m64*)pMidBuffer;
156 pVdest = (__m64*)output;
157
158 // mix1 = mixer values for 1st stereo sample
159 // mix1 = mixer values for 2nd stereo sample
160 // adder = adder for updating mixer values after each round
161
162 mix1 = _mm_set_pi16(0, overlapLength, 0, overlapLength);
163 adder = _mm_set_pi16(1, -1, 1, -1);
164 mix2 = _mm_add_pi16(mix1, adder);
165 adder = _mm_add_pi16(adder, adder);
166
167 // Overlaplength-division by shifter. "+1" is to account for "-1" deduced in
168 // overlapDividerBits calculation earlier.
169 shifter = _m_from_int(overlapDividerBits + 1);
170
171 for (i = 0; i < overlapLength / 4; i ++)
172 {
173 __m64 temp1, temp2;
174
175 // load & shuffle data so that input & mixbuffer data samples are paired
176 temp1 = _mm_unpacklo_pi16(pVMidBuf[0], pVinput[0]); // = i0l m0l i0r m0r
177 temp2 = _mm_unpackhi_pi16(pVMidBuf[0], pVinput[0]); // = i1l m1l i1r m1r
178
179 // temp = (temp .* mix) >> shifter
180 temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
181 temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
182 pVdest[0] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
183
184 // update mix += adder
185 mix1 = _mm_add_pi16(mix1, adder);
186 mix2 = _mm_add_pi16(mix2, adder);
187
188 // --- second round begins here ---
189
190 // load & shuffle data so that input & mixbuffer data samples are paired
191 temp1 = _mm_unpacklo_pi16(pVMidBuf[1], pVinput[1]); // = i2l m2l i2r m2r
192 temp2 = _mm_unpackhi_pi16(pVMidBuf[1], pVinput[1]); // = i3l m3l i3r m3r
193
194 // temp = (temp .* mix) >> shifter
195 temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
196 temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
197 pVdest[1] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
198
199 // update mix += adder
200 mix1 = _mm_add_pi16(mix1, adder);
201 mix2 = _mm_add_pi16(mix2, adder);
202
203 pVinput += 2;
204 pVMidBuf += 2;
205 pVdest += 2;
206 }
207
208 _m_empty(); // clear MMS state
209 }
210
211
212 //////////////////////////////////////////////////////////////////////////////
213 //
214 // implementation of MMX optimized functions of class 'FIRFilter'
215 //
216 //////////////////////////////////////////////////////////////////////////////
217
218 #include "FIRFilter.h"
219
220
FIRFilterMMX()221 FIRFilterMMX::FIRFilterMMX() : FIRFilter()
222 {
223 filterCoeffsUnalign = NULL;
224 }
225
226
~FIRFilterMMX()227 FIRFilterMMX::~FIRFilterMMX()
228 {
229 delete[] filterCoeffsUnalign;
230 }
231
232
233 // (overloaded) Calculates filter coefficients for MMX routine
setCoefficients(const short * coeffs,uint newLength,uint uResultDivFactor)234 void FIRFilterMMX::setCoefficients(const short *coeffs, uint newLength, uint uResultDivFactor)
235 {
236 uint i;
237 FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);
238
239 // Ensure that filter coeffs array is aligned to 16-byte boundary
240 delete[] filterCoeffsUnalign;
241 filterCoeffsUnalign = new short[2 * newLength + 8];
242 filterCoeffsAlign = (short *)SOUNDTOUCH_ALIGN_POINTER_16(filterCoeffsUnalign);
243
244 // rearrange the filter coefficients for mmx routines
245 for (i = 0;i < length; i += 4)
246 {
247 filterCoeffsAlign[2 * i + 0] = coeffs[i + 0];
248 filterCoeffsAlign[2 * i + 1] = coeffs[i + 2];
249 filterCoeffsAlign[2 * i + 2] = coeffs[i + 0];
250 filterCoeffsAlign[2 * i + 3] = coeffs[i + 2];
251
252 filterCoeffsAlign[2 * i + 4] = coeffs[i + 1];
253 filterCoeffsAlign[2 * i + 5] = coeffs[i + 3];
254 filterCoeffsAlign[2 * i + 6] = coeffs[i + 1];
255 filterCoeffsAlign[2 * i + 7] = coeffs[i + 3];
256 }
257 }
258
259
260
261 // mmx-optimized version of the filter routine for stereo sound
evaluateFilterStereo(short * dest,const short * src,uint numSamples) const262 uint FIRFilterMMX::evaluateFilterStereo(short *dest, const short *src, uint numSamples) const
263 {
264 // Create stack copies of the needed member variables for asm routines :
265 uint i, j;
266 __m64 *pVdest = (__m64*)dest;
267
268 if (length < 2) return 0;
269
270 for (i = 0; i < (numSamples - length) / 2; i ++)
271 {
272 __m64 accu1;
273 __m64 accu2;
274 const __m64 *pVsrc = (const __m64*)src;
275 const __m64 *pVfilter = (const __m64*)filterCoeffsAlign;
276
277 accu1 = accu2 = _mm_setzero_si64();
278 for (j = 0; j < lengthDiv8 * 2; j ++)
279 {
280 __m64 temp1, temp2;
281
282 temp1 = _mm_unpacklo_pi16(pVsrc[0], pVsrc[1]); // = l2 l0 r2 r0
283 temp2 = _mm_unpackhi_pi16(pVsrc[0], pVsrc[1]); // = l3 l1 r3 r1
284
285 accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp1, pVfilter[0])); // += l2*f2+l0*f0 r2*f2+r0*f0
286 accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp2, pVfilter[1])); // += l3*f3+l1*f1 r3*f3+r1*f1
287
288 temp1 = _mm_unpacklo_pi16(pVsrc[1], pVsrc[2]); // = l4 l2 r4 r2
289
290 accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp2, pVfilter[0])); // += l3*f2+l1*f0 r3*f2+r1*f0
291 accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp1, pVfilter[1])); // += l4*f3+l2*f1 r4*f3+r2*f1
292
293 // accu1 += l2*f2+l0*f0 r2*f2+r0*f0
294 // += l3*f3+l1*f1 r3*f3+r1*f1
295
296 // accu2 += l3*f2+l1*f0 r3*f2+r1*f0
297 // l4*f3+l2*f1 r4*f3+r2*f1
298
299 pVfilter += 2;
300 pVsrc += 2;
301 }
302 // accu >>= resultDivFactor
303 accu1 = _mm_srai_pi32(accu1, resultDivFactor);
304 accu2 = _mm_srai_pi32(accu2, resultDivFactor);
305
306 // pack 2*2*32bits => 4*16 bits
307 pVdest[0] = _mm_packs_pi32(accu1, accu2);
308 src += 4;
309 pVdest ++;
310 }
311
312 _m_empty(); // clear emms state
313
314 return (numSamples & 0xfffffffe) - length;
315 }
316
317 #endif // SOUNDTOUCH_ALLOW_MMX
318