summaryrefslogtreecommitdiffstats
path: root/media/libsoundtouch/src/mmx_optimized.cpp
blob: 9062026399f74df61b9797da20d0c61498ca457c (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
////////////////////////////////////////////////////////////////////////////////
///
/// MMX optimized routines. All MMX optimized functions have been gathered into 
/// this single source code file, regardless to their class or original source 
/// code file, in order to ease porting the library to other compiler and 
/// processor platforms.
///
/// The MMX-optimizations are programmed using MMX compiler intrinsics that
/// are supported both by Microsoft Visual C++ and GCC compilers, so this file
/// should compile with both toolsets.
///
/// NOTICE: If using Visual Studio 6.0, you'll need to install the "Visual C++ 
/// 6.0 processor pack" update to support compiler intrinsic syntax. The update
/// is available for download at Microsoft Developers Network, see here:
/// http://msdn.microsoft.com/en-us/vstudio/aa718349.aspx
///
/// Author        : Copyright (c) Olli Parviainen
/// Author e-mail : oparviai 'at' iki.fi
/// SoundTouch WWW: http://www.surina.net/soundtouch
///
////////////////////////////////////////////////////////////////////////////////
//
// Last changed  : $Date: 2015-02-22 15:10:38 +0000 (Sun, 22 Feb 2015) $
// File revision : $Revision: 4 $
//
// $Id: mmx_optimized.cpp 206 2015-02-22 15:10:38Z oparviai $
//
////////////////////////////////////////////////////////////////////////////////
//
// License :
//
//  SoundTouch audio processing library
//  Copyright (c) Olli Parviainen
//
//  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; either
//  version 2.1 of the License, or (at your option) any later version.
//
//  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., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
//
////////////////////////////////////////////////////////////////////////////////

#include "STTypes.h"

#ifdef SOUNDTOUCH_ALLOW_MMX
// MMX routines available only with integer sample type

using namespace soundtouch;

//////////////////////////////////////////////////////////////////////////////
//
// implementation of MMX optimized functions of class 'TDStretchMMX'
//
//////////////////////////////////////////////////////////////////////////////

#include "TDStretch.h"
#include <mmintrin.h>
#include <limits.h>
#include <math.h>


// Calculates cross correlation of two buffers
double TDStretchMMX::calcCrossCorr(const short *pV1, const short *pV2, double &dnorm) const
{
    const __m64 *pVec1, *pVec2;
    __m64 shifter;
    __m64 accu, normaccu;
    long corr, norm;
    int i;
   
    pVec1 = (__m64*)pV1;
    pVec2 = (__m64*)pV2;

    shifter = _m_from_int(overlapDividerBits);
    normaccu = accu = _mm_setzero_si64();

    // Process 4 parallel sets of 2 * stereo samples or 4 * mono samples 
    // during each round for improved CPU-level parallellization.
    for (i = 0; i < channels * overlapLength / 16; i ++)
    {
        __m64 temp, temp2;

        // dictionary of instructions:
        // _m_pmaddwd   : 4*16bit multiply-add, resulting two 32bits = [a0*b0+a1*b1 ; a2*b2+a3*b3]
        // _mm_add_pi32 : 2*32bit add
        // _m_psrad     : 32bit right-shift

        temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]), shifter),
                            _mm_sra_pi32(_mm_madd_pi16(pVec1[1], pVec2[1]), shifter));
        temp2 = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[0], pVec1[0]), shifter),
                            _mm_sra_pi32(_mm_madd_pi16(pVec1[1], pVec1[1]), shifter));
        accu = _mm_add_pi32(accu, temp);
        normaccu = _mm_add_pi32(normaccu, temp2);

        temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]), shifter),
                            _mm_sra_pi32(_mm_madd_pi16(pVec1[3], pVec2[3]), shifter));
        temp2 = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[2], pVec1[2]), shifter),
                            _mm_sra_pi32(_mm_madd_pi16(pVec1[3], pVec1[3]), shifter));
        accu = _mm_add_pi32(accu, temp);
        normaccu = _mm_add_pi32(normaccu, temp2);

        pVec1 += 4;
        pVec2 += 4;
    }

    // copy hi-dword of mm0 to lo-dword of mm1, then sum mmo+mm1
    // and finally store the result into the variable "corr"

    accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
    corr = _m_to_int(accu);

    normaccu = _mm_add_pi32(normaccu, _mm_srli_si64(normaccu, 32));
    norm = _m_to_int(normaccu);

    // Clear MMS state
    _m_empty();

    // Normalize result by dividing by sqrt(norm) - this step is easiest 
    // done using floating point operation
    dnorm = (double)norm;

    return (double)corr / sqrt(dnorm < 1e-9 ? 1.0 : dnorm);
    // Note: Warning about the missing EMMS instruction is harmless
    // as it'll be called elsewhere.
}


/// Update cross-correlation by accumulating "norm" coefficient by previously calculated value
double TDStretchMMX::calcCrossCorrAccumulate(const short *pV1, const short *pV2, double &dnorm) const
{
    const __m64 *pVec1, *pVec2;
    __m64 shifter;
    __m64 accu;
    long corr, lnorm;
    int i;
   
    // cancel first normalizer tap from previous round
    lnorm = 0;
    for (i = 1; i <= channels; i ++)
    {
        lnorm -= (pV1[-i] * pV1[-i]) >> overlapDividerBits;
    }

    pVec1 = (__m64*)pV1;
    pVec2 = (__m64*)pV2;

    shifter = _m_from_int(overlapDividerBits);
    accu = _mm_setzero_si64();

    // Process 4 parallel sets of 2 * stereo samples or 4 * mono samples 
    // during each round for improved CPU-level parallellization.
    for (i = 0; i < channels * overlapLength / 16; i ++)
    {
        __m64 temp;

        // dictionary of instructions:
        // _m_pmaddwd   : 4*16bit multiply-add, resulting two 32bits = [a0*b0+a1*b1 ; a2*b2+a3*b3]
        // _mm_add_pi32 : 2*32bit add
        // _m_psrad     : 32bit right-shift

        temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]), shifter),
                            _mm_sra_pi32(_mm_madd_pi16(pVec1[1], pVec2[1]), shifter));
        accu = _mm_add_pi32(accu, temp);

        temp = _mm_add_pi32(_mm_sra_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]), shifter),
                            _mm_sra_pi32(_mm_madd_pi16(pVec1[3], pVec2[3]), shifter));
        accu = _mm_add_pi32(accu, temp);

        pVec1 += 4;
        pVec2 += 4;
    }

    // copy hi-dword of mm0 to lo-dword of mm1, then sum mmo+mm1
    // and finally store the result into the variable "corr"

    accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
    corr = _m_to_int(accu);

    // Clear MMS state
    _m_empty();

    // update normalizer with last samples of this round
    pV1 = (short *)pVec1;
    for (int j = 1; j <= channels; j ++)
    {
        lnorm += (pV1[-j] * pV1[-j]) >> overlapDividerBits;
    }
    dnorm += (double)lnorm;

    // Normalize result by dividing by sqrt(norm) - this step is easiest 
    // done using floating point operation
    return (double)corr / sqrt((dnorm < 1e-9) ? 1.0 : dnorm);
}


void TDStretchMMX::clearCrossCorrState()
{
    // Clear MMS state
    _m_empty();
    //_asm EMMS;
}



// MMX-optimized version of the function overlapStereo
void TDStretchMMX::overlapStereo(short *output, const short *input) const
{
    const __m64 *pVinput, *pVMidBuf;
    __m64 *pVdest;
    __m64 mix1, mix2, adder, shifter;
    int i;

    pVinput  = (const __m64*)input;
    pVMidBuf = (const __m64*)pMidBuffer;
    pVdest   = (__m64*)output;

    // mix1  = mixer values for 1st stereo sample
    // mix1  = mixer values for 2nd stereo sample
    // adder = adder for updating mixer values after each round
    
    mix1  = _mm_set_pi16(0, overlapLength,   0, overlapLength);
    adder = _mm_set_pi16(1, -1, 1, -1);
    mix2  = _mm_add_pi16(mix1, adder);
    adder = _mm_add_pi16(adder, adder);

    // Overlaplength-division by shifter. "+1" is to account for "-1" deduced in
    // overlapDividerBits calculation earlier.
    shifter = _m_from_int(overlapDividerBits + 1);

    for (i = 0; i < overlapLength / 4; i ++)
    {
        __m64 temp1, temp2;
                
        // load & shuffle data so that input & mixbuffer data samples are paired
        temp1 = _mm_unpacklo_pi16(pVMidBuf[0], pVinput[0]);     // = i0l m0l i0r m0r
        temp2 = _mm_unpackhi_pi16(pVMidBuf[0], pVinput[0]);     // = i1l m1l i1r m1r

        // temp = (temp .* mix) >> shifter
        temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
        temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
        pVdest[0] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit

        // update mix += adder
        mix1 = _mm_add_pi16(mix1, adder);
        mix2 = _mm_add_pi16(mix2, adder);

        // --- second round begins here ---

        // load & shuffle data so that input & mixbuffer data samples are paired
        temp1 = _mm_unpacklo_pi16(pVMidBuf[1], pVinput[1]);       // = i2l m2l i2r m2r
        temp2 = _mm_unpackhi_pi16(pVMidBuf[1], pVinput[1]);       // = i3l m3l i3r m3r

        // temp = (temp .* mix) >> shifter
        temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
        temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
        pVdest[1] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit

        // update mix += adder
        mix1 = _mm_add_pi16(mix1, adder);
        mix2 = _mm_add_pi16(mix2, adder);

        pVinput  += 2;
        pVMidBuf += 2;
        pVdest   += 2;
    }

    _m_empty(); // clear MMS state
}


//////////////////////////////////////////////////////////////////////////////
//
// implementation of MMX optimized functions of class 'FIRFilter'
//
//////////////////////////////////////////////////////////////////////////////

#include "FIRFilter.h"


FIRFilterMMX::FIRFilterMMX() : FIRFilter()
{
    filterCoeffsAlign = NULL;
    filterCoeffsUnalign = NULL;
}


FIRFilterMMX::~FIRFilterMMX()
{
    delete[] filterCoeffsUnalign;
}


// (overloaded) Calculates filter coefficients for MMX routine
void FIRFilterMMX::setCoefficients(const short *coeffs, uint newLength, uint uResultDivFactor)
{
    uint i;
    FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);

    // Ensure that filter coeffs array is aligned to 16-byte boundary
    delete[] filterCoeffsUnalign;
    filterCoeffsUnalign = new short[2 * newLength + 8];
    filterCoeffsAlign = (short *)SOUNDTOUCH_ALIGN_POINTER_16(filterCoeffsUnalign);

    // rearrange the filter coefficients for mmx routines 
    for (i = 0;i < length; i += 4) 
    {
        filterCoeffsAlign[2 * i + 0] = coeffs[i + 0];
        filterCoeffsAlign[2 * i + 1] = coeffs[i + 2];
        filterCoeffsAlign[2 * i + 2] = coeffs[i + 0];
        filterCoeffsAlign[2 * i + 3] = coeffs[i + 2];

        filterCoeffsAlign[2 * i + 4] = coeffs[i + 1];
        filterCoeffsAlign[2 * i + 5] = coeffs[i + 3];
        filterCoeffsAlign[2 * i + 6] = coeffs[i + 1];
        filterCoeffsAlign[2 * i + 7] = coeffs[i + 3];
    }
}



// mmx-optimized version of the filter routine for stereo sound
uint FIRFilterMMX::evaluateFilterStereo(short *dest, const short *src, uint numSamples) const
{
    // Create stack copies of the needed member variables for asm routines :
    uint i, j;
    __m64 *pVdest = (__m64*)dest;

    if (length < 2) return 0;

    for (i = 0; i < (numSamples - length) / 2; i ++)
    {
        __m64 accu1;
        __m64 accu2;
        const __m64 *pVsrc = (const __m64*)src;
        const __m64 *pVfilter = (const __m64*)filterCoeffsAlign;

        accu1 = accu2 = _mm_setzero_si64();
        for (j = 0; j < lengthDiv8 * 2; j ++)
        {
            __m64 temp1, temp2;

            temp1 = _mm_unpacklo_pi16(pVsrc[0], pVsrc[1]);  // = l2 l0 r2 r0
            temp2 = _mm_unpackhi_pi16(pVsrc[0], pVsrc[1]);  // = l3 l1 r3 r1

            accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp1, pVfilter[0]));  // += l2*f2+l0*f0 r2*f2+r0*f0
            accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp2, pVfilter[1]));  // += l3*f3+l1*f1 r3*f3+r1*f1

            temp1 = _mm_unpacklo_pi16(pVsrc[1], pVsrc[2]);  // = l4 l2 r4 r2

            accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp2, pVfilter[0]));  // += l3*f2+l1*f0 r3*f2+r1*f0
            accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp1, pVfilter[1]));  // += l4*f3+l2*f1 r4*f3+r2*f1

            // accu1 += l2*f2+l0*f0 r2*f2+r0*f0
            //       += l3*f3+l1*f1 r3*f3+r1*f1

            // accu2 += l3*f2+l1*f0 r3*f2+r1*f0
            //          l4*f3+l2*f1 r4*f3+r2*f1

            pVfilter += 2;
            pVsrc += 2;
        }
        // accu >>= resultDivFactor
        accu1 = _mm_srai_pi32(accu1, resultDivFactor);
        accu2 = _mm_srai_pi32(accu2, resultDivFactor);

        // pack 2*2*32bits => 4*16 bits
        pVdest[0] = _mm_packs_pi32(accu1, accu2);
        src += 4;
        pVdest ++;
    }

   _m_empty();  // clear emms state

    return (numSamples & 0xfffffffe) - length;
}

#endif  // SOUNDTOUCH_ALLOW_MMX