sse_optimized.cpp 17.5 KB
Newer Older
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
////////////////////////////////////////////////////////////////////////////////
///
/// SSE optimized routines for Pentium-III, Athlon-XP and later CPUs. All SSE 
/// 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 SSE-optimizations are programmed using SSE 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 SSE instruction set. The update is 
/// available for download at Microsoft Developers Network, see here:
/// http://msdn.microsoft.com/en-us/vstudio/aa718349.aspx
///
/// If the above URL is expired or removed, go to "http://msdn.microsoft.com" and 
/// perform a search with keywords "processor pack".
///
/// Author        : Copyright (c) Olli Parviainen
/// Author e-mail : oparviai 'at' iki.fi
/// SoundTouch WWW: http://www.surina.net/soundtouch
///
////////////////////////////////////////////////////////////////////////////////
//
// Last changed  : $Date: 2009-12-28 22:32:57 +0200 (Mon, 28 Dec 2009) $
// File revision : $Revision: 4 $
//
// $Id: sse_optimized.cpp 80 2009-12-28 20:32:57Z 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 "cpu_detect.h"
#include "STTypes.h"

using namespace soundtouch;

#ifdef ALLOW_SSE

// SSE routines available only with float sample type    

//////////////////////////////////////////////////////////////////////////////
//
// implementation of SSE optimized functions of class 'TDStretchSSE'
//
//////////////////////////////////////////////////////////////////////////////

#include "TDStretch.h"
#include <xmmintrin.h>
#include <math.h>

// Calculates cross correlation of two buffers
double TDStretchSSE::calcCrossCorrStereo(const float *pV1, const float *pV2) const
{
    int i;
    const float *pVec1;
    const __m128 *pVec2;
    __m128 vSum, vNorm;

    // Note. It means a major slow-down if the routine needs to tolerate 
    // unaligned __m128 memory accesses. It's way faster if we can skip 
    // unaligned slots and use _mm_load_ps instruction instead of _mm_loadu_ps.
    // This can mean up to ~ 10-fold difference (incl. part of which is
    // due to skipping every second round for stereo sound though).
    //
    // Compile-time define ALLOW_NONEXACT_SIMD_OPTIMIZATION is provided
    // for choosing if this little cheating is allowed.

#ifdef ALLOW_NONEXACT_SIMD_OPTIMIZATION
    // Little cheating allowed, return valid correlation only for 
    // aligned locations, meaning every second round for stereo sound.

    #define _MM_LOAD    _mm_load_ps

    if (((ulong)pV1) & 15) return -1e50;    // skip unaligned locations

#else
    // No cheating allowed, use unaligned load & take the resulting
    // performance hit.
    #define _MM_LOAD    _mm_loadu_ps
#endif 

    // ensure overlapLength is divisible by 8
    assert((overlapLength % 8) == 0);

    // Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
    // Note: pV2 _must_ be aligned to 16-bit boundary, pV1 need not.
    pVec1 = (const float*)pV1;
    pVec2 = (const __m128*)pV2;
    vSum = vNorm = _mm_setzero_ps();

    // Unroll the loop by factor of 4 * 4 operations
    for (i = 0; i < overlapLength / 8; i ++) 
    {
        __m128 vTemp;
        // vSum += pV1[0..3] * pV2[0..3]
        vTemp = _MM_LOAD(pVec1);
        vSum  = _mm_add_ps(vSum,  _mm_mul_ps(vTemp ,pVec2[0]));
        vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));

        // vSum += pV1[4..7] * pV2[4..7]
        vTemp = _MM_LOAD(pVec1 + 4);
        vSum  = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[1]));
        vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));

        // vSum += pV1[8..11] * pV2[8..11]
        vTemp = _MM_LOAD(pVec1 + 8);
        vSum  = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[2]));
        vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));

        // vSum += pV1[12..15] * pV2[12..15]
        vTemp = _MM_LOAD(pVec1 + 12);
        vSum  = _mm_add_ps(vSum, _mm_mul_ps(vTemp, pVec2[3]));
        vNorm = _mm_add_ps(vNorm, _mm_mul_ps(vTemp ,vTemp));

        pVec1 += 16;
        pVec2 += 4;
    }

    // return value = vSum[0] + vSum[1] + vSum[2] + vSum[3]
    float *pvNorm = (float*)&vNorm;
    double norm = sqrt(pvNorm[0] + pvNorm[1] + pvNorm[2] + pvNorm[3]);
    if (norm < 1e-9) norm = 1.0;    // to avoid div by zero

    float *pvSum = (float*)&vSum;
    return (double)(pvSum[0] + pvSum[1] + pvSum[2] + pvSum[3]) / norm;

    /* This is approximately corresponding routine in C-language yet without normalization:
    double corr, norm;
    uint i;

    // Calculates the cross-correlation value between 'pV1' and 'pV2' vectors
    corr = norm = 0.0;
    for (i = 0; i < overlapLength / 8; i ++) 
    {
        corr += pV1[0] * pV2[0] +
                pV1[1] * pV2[1] +
                pV1[2] * pV2[2] +
                pV1[3] * pV2[3] +
                pV1[4] * pV2[4] +
                pV1[5] * pV2[5] +
                pV1[6] * pV2[6] +
                pV1[7] * pV2[7] +
                pV1[8] * pV2[8] +
                pV1[9] * pV2[9] +
                pV1[10] * pV2[10] +
                pV1[11] * pV2[11] +
                pV1[12] * pV2[12] +
                pV1[13] * pV2[13] +
                pV1[14] * pV2[14] +
                pV1[15] * pV2[15];

	for (j = 0; j < 15; j ++) norm += pV1[j] * pV1[j];

        pV1 += 16;
        pV2 += 16;
    }
    return corr / sqrt(norm);
    */

    /* This is a bit outdated, corresponding routine in assembler. This may be teeny-weeny bit
       faster than intrinsic version, but more difficult to maintain & get compiled on multiple
       platforms.

    uint overlapLengthLocal = overlapLength;
    float corr;

    _asm 
    {
        // Very important note: data in 'pV2' _must_ be aligned to 
        // 16-byte boundary!

        // give prefetch hints to CPU of what data are to be needed soonish
        // give more aggressive hints on pV1 as that changes while pV2 stays
        // same between runs
        prefetcht0 [pV1]
        prefetcht0 [pV2]
        prefetcht0 [pV1 + 32]

        mov     eax, dword ptr pV1
        mov     ebx, dword ptr pV2

        xorps   xmm0, xmm0

        mov     ecx, overlapLengthLocal
        shr     ecx, 3  // div by eight

    loop1:
        prefetcht0 [eax + 64]     // give a prefetch hint to CPU what data are to be needed soonish
        prefetcht0 [ebx + 32]     // give a prefetch hint to CPU what data are to be needed soonish
        movups  xmm1, [eax]
        mulps   xmm1, [ebx]
        addps   xmm0, xmm1

        movups  xmm2, [eax + 16]
        mulps   xmm2, [ebx + 16]
        addps   xmm0, xmm2

        prefetcht0 [eax + 96]     // give a prefetch hint to CPU what data are to be needed soonish
        prefetcht0 [ebx + 64]     // give a prefetch hint to CPU what data are to be needed soonish

        movups  xmm3, [eax + 32]
        mulps   xmm3, [ebx + 32]
        addps   xmm0, xmm3

        movups  xmm4, [eax + 48]
        mulps   xmm4, [ebx + 48]
        addps   xmm0, xmm4

        add     eax, 64
        add     ebx, 64

        dec     ecx
        jnz     loop1

        // add the four floats of xmm0 together and return the result. 

        movhlps xmm1, xmm0          // move 3 & 4 of xmm0 to 1 & 2 of xmm1
        addps   xmm1, xmm0
        movaps  xmm2, xmm1
        shufps  xmm2, xmm2, 0x01    // move 2 of xmm2 as 1 of xmm2
        addss   xmm2, xmm1
        movss   corr, xmm2
    }

    return (double)corr;
    */
}


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

#include "FIRFilter.h"

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


FIRFilterSSE::~FIRFilterSSE()
{
    delete[] filterCoeffsUnalign;
    filterCoeffsAlign = NULL;
    filterCoeffsUnalign = NULL;
}


// (overloaded) Calculates filter coefficients for SSE routine
void FIRFilterSSE::setCoefficients(const float *coeffs, uint newLength, uint uResultDivFactor)
{
    uint i;
    float fDivider;

    FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);

    // Scale the filter coefficients so that it won't be necessary to scale the filtering result
    // also rearrange coefficients suitably for 3DNow!
    // Ensure that filter coeffs array is aligned to 16-byte boundary
    delete[] filterCoeffsUnalign;
    filterCoeffsUnalign = new float[2 * newLength + 4];
    filterCoeffsAlign = (float *)(((unsigned long)filterCoeffsUnalign + 15) & (ulong)-16);

    fDivider = (float)resultDivider;

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



// SSE-optimized version of the filter routine for stereo sound
uint FIRFilterSSE::evaluateFilterStereo(float *dest, const float *source, uint numSamples) const
{
    int count = (int)((numSamples - length) & (uint)-2);
    int j;

    assert(count % 2 == 0);

    if (count < 2) return 0;

    assert(source != NULL);
    assert(dest != NULL);
    assert((length % 8) == 0);
    assert(filterCoeffsAlign != NULL);
    assert(((ulong)filterCoeffsAlign) % 16 == 0);

    // filter is evaluated for two stereo samples with each iteration, thus use of 'j += 2'
    for (j = 0; j < count; j += 2)
    {
        const float *pSrc;
        const __m128 *pFil;
        __m128 sum1, sum2;
        uint i;

        pSrc = (const float*)source;              // source audio data
        pFil = (const __m128*)filterCoeffsAlign;  // filter coefficients. NOTE: Assumes coefficients 
                                                  // are aligned to 16-byte boundary
        sum1 = sum2 = _mm_setzero_ps();

        for (i = 0; i < length / 8; i ++) 
        {
            // Unroll loop for efficiency & calculate filter for 2*2 stereo samples 
            // at each pass

            // sum1 is accu for 2*2 filtered stereo sound data at the primary sound data offset
            // sum2 is accu for 2*2 filtered stereo sound data for the next sound sample offset.

            sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc)    , pFil[0]));
            sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 2), pFil[0]));

            sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 4), pFil[1]));
            sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 6), pFil[1]));

            sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 8) ,  pFil[2]));
            sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 10), pFil[2]));

            sum1 = _mm_add_ps(sum1, _mm_mul_ps(_mm_loadu_ps(pSrc + 12), pFil[3]));
            sum2 = _mm_add_ps(sum2, _mm_mul_ps(_mm_loadu_ps(pSrc + 14), pFil[3]));

            pSrc += 16;
            pFil += 4;
        }

        // Now sum1 and sum2 both have a filtered 2-channel sample each, but we still need
        // to sum the two hi- and lo-floats of these registers together.

        // post-shuffle & add the filtered values and store to dest.
        _mm_storeu_ps(dest, _mm_add_ps(
                    _mm_shuffle_ps(sum1, sum2, _MM_SHUFFLE(1,0,3,2)),   // s2_1 s2_0 s1_3 s1_2
                    _mm_shuffle_ps(sum1, sum2, _MM_SHUFFLE(3,2,1,0))    // s2_3 s2_2 s1_1 s1_0
                    ));
        source += 4;
        dest += 4;
    }

    // Ideas for further improvement:
    // 1. If it could be guaranteed that 'source' were always aligned to 16-byte 
    //    boundary, a faster aligned '_mm_load_ps' instruction could be used.
    // 2. If it could be guaranteed that 'dest' were always aligned to 16-byte 
    //    boundary, a faster '_mm_store_ps' instruction could be used.

    return (uint)count;

    /* original routine in C-language. please notice the C-version has differently 
       organized coefficients though.
    double suml1, suml2;
    double sumr1, sumr2;
    uint i, j;

    for (j = 0; j < count; j += 2)
    {
        const float *ptr;
        const float *pFil;

        suml1 = sumr1 = 0.0;
        suml2 = sumr2 = 0.0;
        ptr = src;
        pFil = filterCoeffs;
        for (i = 0; i < lengthLocal; i ++) 
        {
            // unroll loop for efficiency.

            suml1 += ptr[0] * pFil[0] + 
                     ptr[2] * pFil[2] +
                     ptr[4] * pFil[4] +
                     ptr[6] * pFil[6];

            sumr1 += ptr[1] * pFil[1] + 
                     ptr[3] * pFil[3] +
                     ptr[5] * pFil[5] +
                     ptr[7] * pFil[7];

            suml2 += ptr[8] * pFil[0] + 
                     ptr[10] * pFil[2] +
                     ptr[12] * pFil[4] +
                     ptr[14] * pFil[6];

            sumr2 += ptr[9] * pFil[1] + 
                     ptr[11] * pFil[3] +
                     ptr[13] * pFil[5] +
                     ptr[15] * pFil[7];

            ptr += 16;
            pFil += 8;
        }
        dest[0] = (float)suml1;
        dest[1] = (float)sumr1;
        dest[2] = (float)suml2;
        dest[3] = (float)sumr2;

        src += 4;
        dest += 4;
    }
    */


    /* Similar routine in assembly, again obsoleted due to maintainability
    _asm
    {
        // Very important note: data in 'src' _must_ be aligned to 
        // 16-byte boundary!
        mov     edx, count
        mov     ebx, dword ptr src
        mov     eax, dword ptr dest
        shr     edx, 1

    loop1:
        // "outer loop" : during each round 2*2 output samples are calculated

        // give prefetch hints to CPU of what data are to be needed soonish
        prefetcht0 [ebx]
        prefetcht0 [filterCoeffsLocal]

        mov     esi, ebx
        mov     edi, filterCoeffsLocal
        xorps   xmm0, xmm0
        xorps   xmm1, xmm1
        mov     ecx, lengthLocal

    loop2:
        // "inner loop" : during each round eight FIR filter taps are evaluated for 2*2 samples
        prefetcht0 [esi + 32]     // give a prefetch hint to CPU what data are to be needed soonish
        prefetcht0 [edi + 32]     // give a prefetch hint to CPU what data are to be needed soonish

        movups  xmm2, [esi]         // possibly unaligned load
        movups  xmm3, [esi + 8]     // possibly unaligned load
        mulps   xmm2, [edi]
        mulps   xmm3, [edi]
        addps   xmm0, xmm2
        addps   xmm1, xmm3

        movups  xmm4, [esi + 16]    // possibly unaligned load
        movups  xmm5, [esi + 24]    // possibly unaligned load
        mulps   xmm4, [edi + 16]
        mulps   xmm5, [edi + 16]
        addps   xmm0, xmm4
        addps   xmm1, xmm5

        prefetcht0 [esi + 64]     // give a prefetch hint to CPU what data are to be needed soonish
        prefetcht0 [edi + 64]     // give a prefetch hint to CPU what data are to be needed soonish

        movups  xmm6, [esi + 32]    // possibly unaligned load
        movups  xmm7, [esi + 40]    // possibly unaligned load
        mulps   xmm6, [edi + 32]
        mulps   xmm7, [edi + 32]
        addps   xmm0, xmm6
        addps   xmm1, xmm7

        movups  xmm4, [esi + 48]    // possibly unaligned load
        movups  xmm5, [esi + 56]    // possibly unaligned load
        mulps   xmm4, [edi + 48]
        mulps   xmm5, [edi + 48]
        addps   xmm0, xmm4
        addps   xmm1, xmm5

        add     esi, 64
        add     edi, 64
        dec     ecx
        jnz     loop2

        // Now xmm0 and xmm1 both have a filtered 2-channel sample each, but we still need
        // to sum the two hi- and lo-floats of these registers together.

        movhlps xmm2, xmm0          // xmm2 = xmm2_3 xmm2_2 xmm0_3 xmm0_2
        movlhps xmm2, xmm1          // xmm2 = xmm1_1 xmm1_0 xmm0_3 xmm0_2
        shufps  xmm0, xmm1, 0xe4    // xmm0 = xmm1_3 xmm1_2 xmm0_1 xmm0_0
        addps   xmm0, xmm2

        movaps  [eax], xmm0
        add     ebx, 16
        add     eax, 16

        dec     edx
        jnz     loop1
    }
    */
}

#endif  // ALLOW_SSE