1
////////////////////////////////////////////////////////////////////////////////
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.
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.
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
17
/// Author : Copyright (c) Olli Parviainen
18
/// Author e-mail : oparviai 'at' iki.fi
19
/// SoundTouch WWW: http://www.surina.net/soundtouch
21
////////////////////////////////////////////////////////////////////////////////
23
// Last changed : $Date: 2014-01-07 20:25:40 +0200 (Tue, 07 Jan 2014) $
24
// File revision : $Revision: 4 $
26
// $Id: mmx_optimized.cpp 184 2014-01-07 18:25:40Z oparviai $
28
////////////////////////////////////////////////////////////////////////////////
32
// SoundTouch audio processing library
33
// Copyright (c) Olli Parviainen
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.
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.
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
49
////////////////////////////////////////////////////////////////////////////////
53
#ifdef SOUNDTOUCH_ALLOW_MMX
54
// MMX routines available only with integer sample type
56
using namespace soundtouch;
58
//////////////////////////////////////////////////////////////////////////////
60
// implementation of MMX optimized functions of class 'TDStretchMMX'
62
//////////////////////////////////////////////////////////////////////////////
64
#include "TDStretch.h"
70
// Calculates cross correlation of two buffers
71
double TDStretchMMX::calcCrossCorr(const short *pV1, const short *pV2, double &dnorm) const
73
const __m64 *pVec1, *pVec2;
82
shifter = _m_from_int(overlapDividerBits);
83
normaccu = accu = _mm_setzero_si64();
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 ++)
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
96
temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]), shifter),
97
_mm_sra_pi32(_mm_madd_pi16(pVec1[1], pVec2[1]), shifter));
98
temp2 = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[0], pVec1[0]), shifter),
99
_mm_sra_pi32(_mm_madd_pi16(pVec1[1], pVec1[1]), shifter));
100
accu = _mm_add_pi32(accu, temp);
101
normaccu = _mm_add_pi32(normaccu, temp2);
103
temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]), shifter),
104
_mm_sra_pi32(_mm_madd_pi16(pVec1[3], pVec2[3]), shifter));
105
temp2 = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[2], pVec1[2]), shifter),
106
_mm_sra_pi32(_mm_madd_pi16(pVec1[3], pVec1[3]), shifter));
107
accu = _mm_add_pi32(accu, temp);
108
normaccu = _mm_add_pi32(normaccu, temp2);
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"
117
accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
118
corr = _m_to_int(accu);
120
normaccu = _mm_add_pi32(normaccu, _mm_srli_si64(normaccu, 32));
121
norm = _m_to_int(normaccu);
126
// Normalize result by dividing by sqrt(norm) - this step is easiest
127
// done using floating point operation
128
dnorm = (double)norm;
130
return (double)corr / sqrt(dnorm < 1e-9 ? 1.0 : dnorm);
131
// Note: Warning about the missing EMMS instruction is harmless
132
// as it'll be called elsewhere.
136
/// Update cross-correlation by accumulating "norm" coefficient by previously calculated value
137
double TDStretchMMX::calcCrossCorrAccumulate(const short *pV1, const short *pV2, double &dnorm) const
139
const __m64 *pVec1, *pVec2;
145
// cancel first normalizer tap from previous round
147
for (i = 1; i <= channels; i ++)
149
lnorm -= (pV1[-i] * pV1[-i]) >> overlapDividerBits;
155
shifter = _m_from_int(overlapDividerBits);
156
accu = _mm_setzero_si64();
158
// Process 4 parallel sets of 2 * stereo samples or 4 * mono samples
159
// during each round for improved CPU-level parallellization.
160
for (i = 0; i < channels * overlapLength / 16; i ++)
164
// dictionary of instructions:
165
// _m_pmaddwd : 4*16bit multiply-add, resulting two 32bits = [a0*b0+a1*b1 ; a2*b2+a3*b3]
166
// _mm_add_pi32 : 2*32bit add
167
// _m_psrad : 32bit right-shift
169
temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]), shifter),
170
_mm_sra_pi32(_mm_madd_pi16(pVec1[1], pVec2[1]), shifter));
171
accu = _mm_add_pi32(accu, temp);
173
temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]), shifter),
174
_mm_sra_pi32(_mm_madd_pi16(pVec1[3], pVec2[3]), shifter));
175
accu = _mm_add_pi32(accu, temp);
181
// copy hi-dword of mm0 to lo-dword of mm1, then sum mmo+mm1
182
// and finally store the result into the variable "corr"
184
accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
185
corr = _m_to_int(accu);
190
// update normalizer with last samples of this round
191
pV1 = (short *)pVec1;
192
for (int j = 1; j <= channels; j ++)
194
lnorm += (pV1[-j] * pV1[-j]) >> overlapDividerBits;
196
dnorm += (double)lnorm;
198
// Normalize result by dividing by sqrt(norm) - this step is easiest
199
// done using floating point operation
200
return (double)corr / sqrt((dnorm < 1e-9) ? 1.0 : dnorm);
204
void TDStretchMMX::clearCrossCorrState()
213
// MMX-optimized version of the function overlapStereo
214
void TDStretchMMX::overlapStereo(short *output, const short *input) const
216
const __m64 *pVinput, *pVMidBuf;
218
__m64 mix1, mix2, adder, shifter;
221
pVinput = (const __m64*)input;
222
pVMidBuf = (const __m64*)pMidBuffer;
223
pVdest = (__m64*)output;
225
// mix1 = mixer values for 1st stereo sample
226
// mix1 = mixer values for 2nd stereo sample
227
// adder = adder for updating mixer values after each round
229
mix1 = _mm_set_pi16(0, overlapLength, 0, overlapLength);
230
adder = _mm_set_pi16(1, -1, 1, -1);
231
mix2 = _mm_add_pi16(mix1, adder);
232
adder = _mm_add_pi16(adder, adder);
234
// Overlaplength-division by shifter. "+1" is to account for "-1" deduced in
235
// overlapDividerBits calculation earlier.
236
shifter = _m_from_int(overlapDividerBits + 1);
238
for (i = 0; i < overlapLength / 4; i ++)
242
// load & shuffle data so that input & mixbuffer data samples are paired
243
temp1 = _mm_unpacklo_pi16(pVMidBuf[0], pVinput[0]); // = i0l m0l i0r m0r
244
temp2 = _mm_unpackhi_pi16(pVMidBuf[0], pVinput[0]); // = i1l m1l i1r m1r
246
// temp = (temp .* mix) >> shifter
247
temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
248
temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
249
pVdest[0] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
251
// update mix += adder
252
mix1 = _mm_add_pi16(mix1, adder);
253
mix2 = _mm_add_pi16(mix2, adder);
255
// --- second round begins here ---
257
// load & shuffle data so that input & mixbuffer data samples are paired
258
temp1 = _mm_unpacklo_pi16(pVMidBuf[1], pVinput[1]); // = i2l m2l i2r m2r
259
temp2 = _mm_unpackhi_pi16(pVMidBuf[1], pVinput[1]); // = i3l m3l i3r m3r
261
// temp = (temp .* mix) >> shifter
262
temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
263
temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
264
pVdest[1] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
266
// update mix += adder
267
mix1 = _mm_add_pi16(mix1, adder);
268
mix2 = _mm_add_pi16(mix2, adder);
275
_m_empty(); // clear MMS state
279
//////////////////////////////////////////////////////////////////////////////
281
// implementation of MMX optimized functions of class 'FIRFilter'
283
//////////////////////////////////////////////////////////////////////////////
285
#include "FIRFilter.h"
288
FIRFilterMMX::FIRFilterMMX() : FIRFilter()
290
filterCoeffsUnalign = NULL;
294
FIRFilterMMX::~FIRFilterMMX()
296
delete[] filterCoeffsUnalign;
300
// (overloaded) Calculates filter coefficients for MMX routine
301
void FIRFilterMMX::setCoefficients(const short *coeffs, uint newLength, uint uResultDivFactor)
304
FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);
306
// Ensure that filter coeffs array is aligned to 16-byte boundary
307
delete[] filterCoeffsUnalign;
308
filterCoeffsUnalign = new short[2 * newLength + 8];
309
filterCoeffsAlign = (short *)SOUNDTOUCH_ALIGN_POINTER_16(filterCoeffsUnalign);
311
// rearrange the filter coefficients for mmx routines
312
for (i = 0;i < length; i += 4)
314
filterCoeffsAlign[2 * i + 0] = coeffs[i + 0];
315
filterCoeffsAlign[2 * i + 1] = coeffs[i + 2];
316
filterCoeffsAlign[2 * i + 2] = coeffs[i + 0];
317
filterCoeffsAlign[2 * i + 3] = coeffs[i + 2];
319
filterCoeffsAlign[2 * i + 4] = coeffs[i + 1];
320
filterCoeffsAlign[2 * i + 5] = coeffs[i + 3];
321
filterCoeffsAlign[2 * i + 6] = coeffs[i + 1];
322
filterCoeffsAlign[2 * i + 7] = coeffs[i + 3];
328
// mmx-optimized version of the filter routine for stereo sound
329
uint FIRFilterMMX::evaluateFilterStereo(short *dest, const short *src, uint numSamples) const
331
// Create stack copies of the needed member variables for asm routines :
333
__m64 *pVdest = (__m64*)dest;
335
if (length < 2) return 0;
337
for (i = 0; i < (numSamples - length) / 2; i ++)
341
const __m64 *pVsrc = (const __m64*)src;
342
const __m64 *pVfilter = (const __m64*)filterCoeffsAlign;
344
accu1 = accu2 = _mm_setzero_si64();
345
for (j = 0; j < lengthDiv8 * 2; j ++)
349
temp1 = _mm_unpacklo_pi16(pVsrc[0], pVsrc[1]); // = l2 l0 r2 r0
350
temp2 = _mm_unpackhi_pi16(pVsrc[0], pVsrc[1]); // = l3 l1 r3 r1
352
accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp1, pVfilter[0])); // += l2*f2+l0*f0 r2*f2+r0*f0
353
accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp2, pVfilter[1])); // += l3*f3+l1*f1 r3*f3+r1*f1
355
temp1 = _mm_unpacklo_pi16(pVsrc[1], pVsrc[2]); // = l4 l2 r4 r2
357
accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp2, pVfilter[0])); // += l3*f2+l1*f0 r3*f2+r1*f0
358
accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp1, pVfilter[1])); // += l4*f3+l2*f1 r4*f3+r2*f1
360
// accu1 += l2*f2+l0*f0 r2*f2+r0*f0
361
// += l3*f3+l1*f1 r3*f3+r1*f1
363
// accu2 += l3*f2+l1*f0 r3*f2+r1*f0
364
// l4*f3+l2*f1 r4*f3+r2*f1
369
// accu >>= resultDivFactor
370
accu1 = _mm_srai_pi32(accu1, resultDivFactor);
371
accu2 = _mm_srai_pi32(accu2, resultDivFactor);
373
// pack 2*2*32bits => 4*16 bits
374
pVdest[0] = _mm_packs_pi32(accu1, accu2);
379
_m_empty(); // clear emms state
381
return (numSamples & 0xfffffffe) - length;
384
#endif // SOUNDTOUCH_ALLOW_MMX