mmx_gcc.cpp

Audacity是一款用於錄音和編輯聲音的、免費的開放源碼軟體。它可以執行於Mac OS X、Microsoft Windows、GNU/Linux和其它作業系統

/*****************************************************************************
 *
 * gcc version of the 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.
 *
 * This file is to be compiled on any platform with the GNU C compiler.
 * Compiler. Please see 'mmx_win.cpp' for the x86 Windows version of this
 * file.
 *
 * Author          : Copyright (c) Olli Parviainen
 * Author e-mail   : oparviai @ iki.fi
 * File created    : 13-Jan-2002
 *
 * Last changed  : $Date: 2004/10/26 19:09:36 $
 * File revision : $Revision: 1.2 $
 *
 * $Id: mmx_gcc.cpp,v 1.2 2004/10/26 19:09:36 vjohnson Exp $
 *
 * Acknowledgements:
 * Adopted for gcc : Stuart Lamble <sjl @ debian.lib.monash.edu.au>
 * Adopted for gcc3: Shachar Raindel <shacharr @ users.sourceforge.net>
 *
 * License :
 * 
 *  SoundTouch sound 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 <stdexcept>
#include <string>
#include "cpu_detect.h"

#ifndef __GNUC__
#error "wrong platform - this source code file is for the GNU C compiler."
#endif

using namespace std;


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

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

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

// these are declared in 'TDStretch.cpp'
extern int scanOffsets[4][24];

// Calculates cross correlation of two buffers
inline int TDStretchMMX::calcCrossCorrStereo(const short *pV1, const short *pV2) const
{
#ifdef __i386__
    int corr;
    uint local_overlapLength = overlapLength;
    uint local_overlapDividerBits = overlapDividerBits;

    asm volatile(
        // Calculate cross-correlation between the tempOffset and tmpbid_buffer.

        // Process 4 parallel batches of 2 * stereo samples each during one
        // round to improve CPU-level parallellization.

        // load address of sloped pV2 buffer to eax
        // load address of mixing point of the sample data buffer to ebx
        // load counter to ecx = overlapLength / 8 - 1
        // empty the mm0

        // prepare to the first round by loading 
        // load mm1 = eax[0]
        // load mm2 = eax[1];

        "\n\tmovl        %1, %%eax"
        "\n\tmovl        %2, %%ebx"

        "\n\tmovq        (%%eax), %%mm1"
        "\n\tmovl        %3, %%ecx"

        "\n\tmovq        8(%%eax), %%mm2"
        "\n\tshr         $3, %%ecx"

        "\n\tpxor        %%mm0, %%mm0"
        "\n\tsub         $1, %%ecx"

        "\n\tmovd        %4, %%mm5"

        "\n1:"
        // multiply-add mm1 = mm1 * ebx[0]
        // multiply-add mm2 = mm2 * ebx[1]
        //
        // add mm2 += mm1
        // mm2 >>= mm5 (=overlapDividerBits)
        // add mm0 += mm2
        //
        // load mm3 = eax[2]
        // multiply-add mm3 = mm3 * ebx[2]
        //
        // load mm4 = eax[3]
        // multiply-add mm4 = mm4 * ebx[3]
        //
        // add mm3 += mm4
        // mm3 >>= mm5 (=overlapDividerBits)
        // add mm0 += mm3
        //
        // add eax += 4
        // add ebx += 4
        // load mm1 = eax[0] (~eax[4])
        // load mm2 = eax[1] (~eax[5])
        //
        // loop

        "\n\tpmaddwd     (%%ebx), %%mm1"   // qword ptr [ebx]
        "\n\tmovq        16(%%eax), %%mm3" // qword ptr [eax+16]

        "\n\tpmaddwd     8(%%ebx), %%mm2"  // qword ptr [ebx+8]
        "\n\tmovq        24(%%eax), %%mm4" // qword ptr [eax+24]

        "\n\tpmaddwd     16(%%ebx), %%mm3" // qword ptr [ebx+16]
        "\n\tpaddd       %%mm1, %%mm2"

        "\n\tpmaddwd     24(%%ebx), %%mm4" // qword ptr [ebx+24]
        "\n\tmovq        32(%%eax), %%mm1" // qword ptr [eax+32]

        "\n\tpsrad       %%mm5, %%mm2"
        "\n\tadd         $32, %%eax"

        "\n\tpaddd       %%mm4, %%mm3"
        "\n\tpaddd       %%mm2, %%mm0"

        "\n\tmovq        8(%%eax), %%mm2"  // qword ptr [eax+8]
        "\n\tpsrad       %%mm5, %%mm3"

        "\n\tadd         $32, %%ebx"
        "\n\tpaddd       %%mm3, %%mm0"

        "\n\tdec         %%ecx"
        "\n\tjnz         1b"

        // Finalize the last partial loop:

        "\n\tmovq        16(%%eax), %%mm3" // qword ptr [eax+16]
        "\n\tpmaddwd     (%%ebx), %%mm1"   // qword ptr [ebx]

        "\n\tmovq        24(%%eax), %%mm4" // qword ptr [eax+24]
        "\n\tpmaddwd     8(%%ebx), %%mm2"  // qword ptr [ebx+8]

        "\n\tpmaddwd     16(%%ebx), %%mm3" // qword ptr [ebx+16]
        "\n\tpaddd       %%mm1, %%mm2"

        "\n\tpmaddwd     24(%%ebx), %%mm4" // qword ptr [ebx+24]
        "\n\tpsrad       %%mm5, %%mm2"

        "\n\tpaddd       %%mm4, %%mm3"
        "\n\tpaddd       %%mm2, %%mm0"

        "\n\tpsrad       %%mm5, %%mm3"
        "\n\tpaddd       %%mm3, %%mm0"

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

        "\n\tmovq        %%mm0, %%mm1"
        "\n\tpsrlq       $32, %%mm1"
        "\n\tpaddd       %%mm1, %%mm0"
        "\n\tmovd        %%mm0, %0"
      : "=rm" (corr)
      : "rim" (pV1), "rim" (pV2), "rim" (local_overlapLength),
        "rim" (local_overlapDividerBits)
      : "%ecx", "%ebx", "%eax"
    );
    return corr;
    
    // Note: Warning about the missing EMMS instruction is harmless
    // as it'll be called elsewhere.
#else
    throw runtime_error("MMX not supported");
#endif
}

void TDStretchMMX::clearCrossCorrState()
{
#ifdef __i386__
    asm volatile("EMMS");
#endif
}


// MMX-optimized version of the function overlapStereo
void TDStretchMMX::overlapStereo(short *output, const short *input) const
{
#ifdef __i386__
    short *local_midBuffer = pMidBuffer;
    uint local_overlapLength = overlapLength;
    uint local_overlapDividerBits = overlapDividerBits;
    unsigned long shadow_esi;

    asm volatile(
        "\n\t"
        // load sliding mixing value counter to mm6 and mm7
        // load counter value to ecx = overlapLength / 4
        // load divider-shifter value to esi
        // load mixing value adder to mm5
        // load address of midBuffer to eax
        // load address of inputBuffer added with ovlOffset to ebx
        // load address of end of the outputBuffer to edx
        //
        // We need to preserve esi, since gcc uses it for the
        // stack frame.

        "movl        %%esi, %0\n\t"

        "movl        %1, %%eax\n\t"               // ecx = 0x0000 OVL_
        "movl        $0x0002fffe, %%edi\n\t"      // ecx = 0x0002 fffe

        "movl        %2, %%edx\n\t"
        "movd        %%eax, %%mm6\n\t"            // mm6 = 0x0000 0000 0000 OVL_

        "movl        %%eax, %%ecx\n\t"
        "sub         $1, %%eax\n\t"

        "punpckldq   %%mm6, %%mm6\n\t"            // mm6 = 0x0000 OVL_ 0000 OVL_

        "or          $0x00010000, %%eax\n\t"      // eax = 0x0001 overlapLength-1
        "movl        %4, %%ebx\n\t"

        "movd        %%edi, %%mm5\n\t"            // mm5 = 0x0000 0000 0002 fffe
        "movd        %%eax, %%mm7\n\t"            // mm7 = 0x0000 0000 0001 01ff

        "movl        %5, %%eax\n\t"               // dword ptr local_midBuffer
        "punpckldq   %%mm5, %%mm5\n\t"            // mm5 = 0x0002 fffe 0002 fffe

        "shr         $2, %%ecx\n\t"               // ecx = overlapLength / 2
        "punpckldq   %%mm7, %%mm7\n\t"            // mm7 = 0x0001 01ff 0001 01ff

        "push        %%edx\n\t"
        "pop         %%esi\n"
        "movl        %3, %%edx\n\t"

        "2:\n\t"
        // Process two parallel batches of 2+2 stereo samples during each round 
        // to improve CPU-level parallellization.
        //
        // Load [eax] into mm0 and mm1
        // Load [ebx] into mm3
        // unpack words of mm0, mm1 and mm3 into mm0 and mm1