/external/pysoundtouch14/libsoundtouch/mmx_optimized.cpp

http://echo-nest-remix.googlecode.com/ · C++ · 305 lines · 138 code · 68 blank · 99 comment · 6 complexity · 19b2dd27bb322136dc97dea54d3c85b3 MD5 · raw file

  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: 2009-01-25 16:13:39 +0200 (Sun, 25 Jan 2009) $
  24. // File revision : $Revision: 4 $
  25. //
  26. // $Id: mmx_optimized.cpp 51 2009-01-25 14:13:39Z 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. #include "STTypes.h"
  51. #ifdef ALLOW_MMX
  52. // MMX routines available only with integer sample type
  53. //#if !(WIN32 || __i386__ || __x86_64__)
  54. //#error "wrong platform - this source code file is exclusively for x86 platforms"
  55. //#endif
  56. using namespace soundtouch;
  57. //////////////////////////////////////////////////////////////////////////////
  58. //
  59. // implementation of MMX optimized functions of class 'TDStretchMMX'
  60. //
  61. //////////////////////////////////////////////////////////////////////////////
  62. #include "TDStretch.h"
  63. #include <mmintrin.h>
  64. #include <limits.h>
  65. // Calculates cross correlation of two buffers
  66. long TDStretchMMX::calcCrossCorrStereo(const short *pV1, const short *pV2) const
  67. {
  68. const __m64 *pVec1, *pVec2;
  69. __m64 shifter;
  70. __m64 accu;
  71. long corr;
  72. int i;
  73. pVec1 = (__m64*)pV1;
  74. pVec2 = (__m64*)pV2;
  75. shifter = _m_from_int(overlapDividerBits);
  76. accu = _mm_setzero_si64();
  77. // Process 4 parallel sets of 2 * stereo samples each during each
  78. // round to improve CPU-level parallellization.
  79. for (i = 0; i < overlapLength / 8; i ++)
  80. {
  81. __m64 temp;
  82. // dictionary of instructions:
  83. // _m_pmaddwd : 4*16bit multiply-add, resulting two 32bits = [a0*b0+a1*b1 ; a2*b2+a3*b3]
  84. // _mm_add_pi32 : 2*32bit add
  85. // _m_psrad : 32bit right-shift
  86. temp = _mm_add_pi32(_mm_madd_pi16(pVec1[0], pVec2[0]),
  87. _mm_madd_pi16(pVec1[1], pVec2[1]));
  88. accu = _mm_add_pi32(accu, _mm_sra_pi32(temp, shifter));
  89. temp = _mm_add_pi32(_mm_madd_pi16(pVec1[2], pVec2[2]),
  90. _mm_madd_pi16(pVec1[3], pVec2[3]));
  91. accu = _mm_add_pi32(accu, _mm_sra_pi32(temp, shifter));
  92. pVec1 += 4;
  93. pVec2 += 4;
  94. }
  95. // copy hi-dword of mm0 to lo-dword of mm1, then sum mmo+mm1
  96. // and finally store the result into the variable "corr"
  97. accu = _mm_add_pi32(accu, _mm_srli_si64(accu, 32));
  98. corr = _m_to_int(accu);
  99. // Clear MMS state
  100. _m_empty();
  101. return corr;
  102. // Note: Warning about the missing EMMS instruction is harmless
  103. // as it'll be called elsewhere.
  104. }
  105. void TDStretchMMX::clearCrossCorrState()
  106. {
  107. // Clear MMS state
  108. _m_empty();
  109. //_asm EMMS;
  110. }
  111. // MMX-optimized version of the function overlapStereo
  112. void TDStretchMMX::overlapStereo(short *output, const short *input) const
  113. {
  114. const __m64 *pVinput, *pVMidBuf;
  115. __m64 *pVdest;
  116. __m64 mix1, mix2, adder, shifter;
  117. int i;
  118. pVinput = (const __m64*)input;
  119. pVMidBuf = (const __m64*)pMidBuffer;
  120. pVdest = (__m64*)output;
  121. // mix1 = mixer values for 1st stereo sample
  122. // mix1 = mixer values for 2nd stereo sample
  123. // adder = adder for updating mixer values after each round
  124. mix1 = _mm_set_pi16(0, overlapLength, 0, overlapLength);
  125. adder = _mm_set_pi16(1, -1, 1, -1);
  126. mix2 = _mm_add_pi16(mix1, adder);
  127. adder = _mm_add_pi16(adder, adder);
  128. shifter = _m_from_int(overlapDividerBits);
  129. for (i = 0; i < overlapLength / 4; i ++)
  130. {
  131. __m64 temp1, temp2;
  132. // load & shuffle data so that input & mixbuffer data samples are paired
  133. temp1 = _mm_unpacklo_pi16(pVMidBuf[0], pVinput[0]); // = i0l m0l i0r m0r
  134. temp2 = _mm_unpackhi_pi16(pVMidBuf[0], pVinput[0]); // = i1l m1l i1r m1r
  135. // temp = (temp .* mix) >> shifter
  136. temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
  137. temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
  138. pVdest[0] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
  139. // update mix += adder
  140. mix1 = _mm_add_pi16(mix1, adder);
  141. mix2 = _mm_add_pi16(mix2, adder);
  142. // --- second round begins here ---
  143. // load & shuffle data so that input & mixbuffer data samples are paired
  144. temp1 = _mm_unpacklo_pi16(pVMidBuf[1], pVinput[1]); // = i2l m2l i2r m2r
  145. temp2 = _mm_unpackhi_pi16(pVMidBuf[1], pVinput[1]); // = i3l m3l i3r m3r
  146. // temp = (temp .* mix) >> shifter
  147. temp1 = _mm_sra_pi32(_mm_madd_pi16(temp1, mix1), shifter);
  148. temp2 = _mm_sra_pi32(_mm_madd_pi16(temp2, mix2), shifter);
  149. pVdest[1] = _mm_packs_pi32(temp1, temp2); // pack 2*2*32bit => 4*16bit
  150. // update mix += adder
  151. mix1 = _mm_add_pi16(mix1, adder);
  152. mix2 = _mm_add_pi16(mix2, adder);
  153. pVinput += 2;
  154. pVMidBuf += 2;
  155. pVdest += 2;
  156. }
  157. _m_empty(); // clear MMS state
  158. }
  159. //////////////////////////////////////////////////////////////////////////////
  160. //
  161. // implementation of MMX optimized functions of class 'FIRFilter'
  162. //
  163. //////////////////////////////////////////////////////////////////////////////
  164. #include "FIRFilter.h"
  165. FIRFilterMMX::FIRFilterMMX() : FIRFilter()
  166. {
  167. filterCoeffsUnalign = NULL;
  168. }
  169. FIRFilterMMX::~FIRFilterMMX()
  170. {
  171. delete[] filterCoeffsUnalign;
  172. }
  173. // (overloaded) Calculates filter coefficients for MMX routine
  174. void FIRFilterMMX::setCoefficients(const short *coeffs, uint newLength, uint uResultDivFactor)
  175. {
  176. uint i;
  177. FIRFilter::setCoefficients(coeffs, newLength, uResultDivFactor);
  178. // Ensure that filter coeffs array is aligned to 16-byte boundary
  179. delete[] filterCoeffsUnalign;
  180. filterCoeffsUnalign = new short[2 * newLength + 8];
  181. filterCoeffsAlign = (short *)(((ulong)filterCoeffsUnalign + 15) & -16);
  182. // rearrange the filter coefficients for mmx routines
  183. for (i = 0;i < length; i += 4)
  184. {
  185. filterCoeffsAlign[2 * i + 0] = coeffs[i + 0];
  186. filterCoeffsAlign[2 * i + 1] = coeffs[i + 2];
  187. filterCoeffsAlign[2 * i + 2] = coeffs[i + 0];
  188. filterCoeffsAlign[2 * i + 3] = coeffs[i + 2];
  189. filterCoeffsAlign[2 * i + 4] = coeffs[i + 1];
  190. filterCoeffsAlign[2 * i + 5] = coeffs[i + 3];
  191. filterCoeffsAlign[2 * i + 6] = coeffs[i + 1];
  192. filterCoeffsAlign[2 * i + 7] = coeffs[i + 3];
  193. }
  194. }
  195. // mmx-optimized version of the filter routine for stereo sound
  196. uint FIRFilterMMX::evaluateFilterStereo(short *dest, const short *src, uint numSamples) const
  197. {
  198. // Create stack copies of the needed member variables for asm routines :
  199. uint i, j;
  200. __m64 *pVdest = (__m64*)dest;
  201. if (length < 2) return 0;
  202. for (i = 0; i < numSamples / 2; i ++)
  203. {
  204. __m64 accu1;
  205. __m64 accu2;
  206. const __m64 *pVsrc = (const __m64*)src;
  207. const __m64 *pVfilter = (const __m64*)filterCoeffsAlign;
  208. accu1 = accu2 = _mm_setzero_si64();
  209. for (j = 0; j < lengthDiv8 * 2; j ++)
  210. {
  211. __m64 temp1, temp2;
  212. temp1 = _mm_unpacklo_pi16(pVsrc[0], pVsrc[1]); // = l2 l0 r2 r0
  213. temp2 = _mm_unpackhi_pi16(pVsrc[0], pVsrc[1]); // = l3 l1 r3 r1
  214. accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp1, pVfilter[0])); // += l2*f2+l0*f0 r2*f2+r0*f0
  215. accu1 = _mm_add_pi32(accu1, _mm_madd_pi16(temp2, pVfilter[1])); // += l3*f3+l1*f1 r3*f3+r1*f1
  216. temp1 = _mm_unpacklo_pi16(pVsrc[1], pVsrc[2]); // = l4 l2 r4 r2
  217. accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp2, pVfilter[0])); // += l3*f2+l1*f0 r3*f2+r1*f0
  218. accu2 = _mm_add_pi32(accu2, _mm_madd_pi16(temp1, pVfilter[1])); // += l4*f3+l2*f1 r4*f3+r2*f1
  219. // accu1 += l2*f2+l0*f0 r2*f2+r0*f0
  220. // += l3*f3+l1*f1 r3*f3+r1*f1
  221. // accu2 += l3*f2+l1*f0 r3*f2+r1*f0
  222. // l4*f3+l2*f1 r4*f3+r2*f1
  223. pVfilter += 2;
  224. pVsrc += 2;
  225. }
  226. // accu >>= resultDivFactor
  227. accu1 = _mm_srai_pi32(accu1, resultDivFactor);
  228. accu2 = _mm_srai_pi32(accu2, resultDivFactor);
  229. // pack 2*2*32bits => 4*16 bits
  230. pVdest[0] = _mm_packs_pi32(accu1, accu2);
  231. src += 4;
  232. pVdest ++;
  233. }
  234. _m_empty(); // clear emms state
  235. return (numSamples & 0xfffffffe) - length;
  236. }
  237. #endif // ALLOW_MMX