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

        // multiply-add mm0*mm6 and mm1*mm7, store results into mm0 and mm1
        // divide mm0 and mm1 by 512 (=right-shift by overlapDividerBits)
        // pack the result into mm0 and store into [edx]
        //
        // Load [eax+8] into mm2 and mm3
        // Load [ebx+8] into mm4
        // unpack words of mm2, mm3 and mm4 into mm2 and mm3
        // multiply-add mm2*mm6 and mm3*mm7, store results into mm2 and mm3
        // divide mm2 and mm3 by 512 (=right-shift by overlapDividerBits)
        // pack the result into mm2 and store into [edx+8]

    
        "movq        (%%eax), %%mm0\n\t"             // mm0 = m1l m1r m0l m0r
        "add         $16, %%edx\n\t"

        "movq        (%%ebx), %%mm3\n\t"             // mm3 = i1l i1r i0l i0r
        "movq        %%mm0, %%mm1\n\t"               // mm1 = m1l m1r m0l m0r

        "movq        8(%%eax), %%mm2\n\t"            // mm2 = m3l m3r m2l m2r
        "punpcklwd   %%mm3, %%mm0\n\t"               // mm0 = i0l m0l i0r m0r

        "movq        8(%%ebx), %%mm4\n\t"            // mm4 = i3l i3r i2l i2r
        "punpckhwd   %%mm3, %%mm1\n\t"               // mm1 = i1l m1l i1r m1r

        "movq        %%mm2, %%mm3\n\t"               // mm3 = m3l m3r m2l m2r
        "punpcklwd   %%mm4, %%mm2\n\t"               // mm2 = i2l m2l i2r m2r

        "pmaddwd     %%mm6, %%mm0\n\t"               // mm0 = i0l*m63+m0l*m62 i0r*m61+m0r*m60
        "punpckhwd   %%mm4, %%mm3\n\t"               // mm3 = i3l m3l i3r m3r

        "movd        %%esi, %%mm4\n\t"               // mm4 = overlapDividerBits

        "pmaddwd     %%mm7, %%mm1\n\t"               // mm1 = i1l*m73+m1l*m72 i1r*m71+m1r*m70
        "paddw       %%mm5, %%mm6\n\t"

        "paddw       %%mm5, %%mm7\n\t"
        "psrad       %%mm4, %%mm0\n\t"               // mmo >>= overlapDividerBits

        "pmaddwd     %%mm6, %%mm2\n\t"               // mm2 = i2l*m63+m2l*m62 i2r*m61+m2r*m60
        "psrad       %%mm4, %%mm1\n\t"               // mm1 >>= overlapDividerBits

        "pmaddwd     %%mm7, %%mm3\n\t"               // mm3 = i3l*m73+m3l*m72 i3r*m71+m3r*m70
        "psrad       %%mm4, %%mm2\n\t"               // mm2 >>= overlapDividerBits

        "packssdw    %%mm1, %%mm0\n\t"               // mm0 = mm1h mm1l mm0h mm0l
        "psrad       %%mm4, %%mm3\n\t"               // mm3 >>= overlapDividerBits

        "add         $16, %%eax\n\t"
        "paddw       %%mm5, %%mm6\n\t"

        "packssdw    %%mm3, %%mm2\n\t"               // mm2 = mm2h mm2l mm3h mm3l
        "paddw       %%mm5, %%mm7\n\t"

        "movq        %%mm0, -16(%%edx)\n\t"
        "add         $16, %%ebx\n\t"

        "movq        %%mm2, -8(%%edx)\n\t"
        "dec         %%ecx\n\t"

        "jnz         2b\n\t"

        "emms\n\t"

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

      : "+rim" (shadow_esi)/* output */
      : "rim" (local_overlapLength),
        "rim" (local_overlapDividerBits),
        "rim" (output),
        "rim" (input),
        "rim" (local_midBuffer)
      /* input */
      : "%edi", "%ecx", "%edx", "%eax", "%ebx" /* regs */
    );
#else
    throw runtime_error("MMX not supported");
#endif
}


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

#include "FIRFilter.h"

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


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


// (overloaded) Calculates filter coefficients for MMX routine
void FIRFilterMMX::setCoefficients(const short *coeffs, uint newLength, uint uResultDivFactor)
{
#ifdef __i386__
    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 *)(((uint)filterCoeffsUnalign + 15) & -16);

    // 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];
    }
#else
    throw runtime_error("MMX not supported");
#endif
}



// mmx-optimized version of the filter routine for stereo sound
uint FIRFilterMMX::evaluateFilterStereo(short *dest, const short *src, const uint numSamples) const
{
#ifdef __i386__
    // Create stack copies of the needed member variables for asm routines :
    uint local_length = length;
    uint local_lengthDiv8 = lengthDiv8;
    uint local_resultDivider = resultDivFactor;
    short *local_filterCoeffs = (short*)filterCoeffsAlign;
    short *local_src = (short *)src;

    asm volatile(
        "\n\t"
        // Load (num_samples-aa_filter_length)/2 to edi as a i
        // Load a pointer to samples to esi
        // Load a pointer to destination to edx

        "movl        %0, %%edi\n\t"
        "subl        %2, %%edi\n\t"
        "movl        %3, %%edx\n\t"
        "sar         $1, %%edi\n"

        // Load filter length/8 to ecx
        // Load pointer to samples from esi to ebx
        // Load counter from edi to ecx
        // Load [ebx] to mm3
        // Load pointer to filter coefficients to eax
        "3:\n\t"
        "movl        %1, %%ebx\n\t"
        "pxor        %%mm0, %%mm0\n\t"

        "movl        %4, %%ecx\n\t"
        "pxor        %%mm7, %%mm7\n\t"

        "movq        (%%ebx), %%mm1\n\t"            // mm1 = l1 r1 l0 r0
        "movl        %5, %%eax\n"
        "4:\n\t"

        "movq        8(%%ebx), %%mm2\n\t"           // mm2 = l3 r3 l2 r2
        "movq        %%mm1, %%mm4\n\t"              // mm4 = l1 r1 l0 r0

        "movq        16(%%ebx), %%mm3\n\t"          // mm3 = l5 r5 l4 r4
        "punpckhwd   %%mm2, %%mm1\n\t"              // mm1 = l3 l1 r3 r1

        "movq        %%mm2, %%mm6\n\t"              // mm6 = l3 r3 l2 r2
        "punpcklwd   %%mm2, %%mm4\n\t"              // mm4 = l2 l0 r2 r0

        "movq        (%%eax), %%mm2\n\t"            // mm2 = f2 f0 f2 f0
        "movq        %%mm1, %%mm5\n\t"              // mm5 = l3 l1 r3 r1

        "punpcklwd   %%mm3, %%mm6\n\t"              // mm6 = l4 l2 r4 r2
        "pmaddwd     %%mm2, %%mm4\n\t"              // mm4 = l2*f2+l0*f0 r2*f2+r0*f0

        "pmaddwd     %%mm2, %%mm5\n\t"              // mm5 = l3*f2+l1*f0 r3*f2+l1*f0
        "movq        8(%%eax), %%mm2\n\t"           // mm2 = f3 f1 f3 f1

        "paddd       %%mm4, %%mm0\n\t"              // mm0 += s02*f02
        "movq        %%mm3, %%mm4\n\t"              // mm4 = l1 r1 l0 r0

        "pmaddwd     %%mm2, %%mm1\n\t"              // mm1 = l3*f3+l1*f1 r3*f3+l1*f1
        "paddd       %%mm5, %%mm7\n\t"              // mm7 += s13*f02

        "pmaddwd     %%mm2, %%mm6\n\t"              // mm6 = l4*f3+l2*f1 r4*f3+f4*f1
        "movq        24(%%ebx), %%mm2\n\t"          // mm2 = l3 r3 l2 r2

        "paddd       %%mm1, %%mm0\n\t"              // mm0 += s31*f31
        "movq        32(%%ebx), %%mm1\n\t"          // mm1 = l5 r5 l4 r4

        "paddd       %%mm6, %%mm7\n\t"              // mm7 += s42*f31
        "punpckhwd   %%mm2, %%mm3\n\t"              // mm3 = l3 l1 r3 r1

        "movq        %%mm2, %%mm6\n\t"              // mm6 = l3 r3 l2 r2
        "punpcklwd   %%mm2, %%mm4\n\t"              // mm4 = l2 l0 r2 r0

        "movq        16(%%eax), %%mm2\n\t"          // mm2 = f2 f0 f2 f0
        "movq        %%mm3, %%mm5\n\t"              // mm5 = l3 l1 r3 r1

        "punpcklwd   %%mm1, %%mm6\n\t"              // mm6 = l4 l2 r4 r2
        "add         $32, %%eax\n\t"

        "pmaddwd     %%mm2, %%mm4\n\t"              // mm4 = l2*f2+l0*f0 r2*f2+r0*f0
        "add         $32, %%ebx\n\t"

        "pmaddwd     %%mm2, %%mm5\n\t"              // mm5 = l3*f2+l1*f0 r3*f2+l1*f0
        "movq        -8(%%eax), %%mm2\n\t"          // mm2 = f3 f1 f3 f1

        "paddd       %%mm4, %%mm0\n\t"              // mm0 += s02*f02
        "pmaddwd     %%mm2, %%mm3\n\t"              // mm3 = l3*f3+l1*f1 r3*f3+l1*f1

        "paddd       %%mm5, %%mm7\n\t"              // mm7 += s13*f02
        "pmaddwd     %%mm2, %%mm6\n\t"              // mm6 = l4*f3+l2*f1 r4*f3+f4*f1

        "paddd       %%mm3, %%mm0\n\t"              // mm0 += s31*f31
        "paddd       %%mm6, %%mm7\n\t"              // mm7 += s42*f31

        "dec         %%ecx\n\t"
        "jnz         4b\n\t"

        // Divide mm0 and mm7 by 8192 (= right-shift by 13),
        // pack and store to [edx]
        "movd        %6, %%mm4\n\t"

        "psrad       %%mm4, %%mm0\n\t"              // divide the result

        "add         $8, %%edx\n\t"
        "psrad       %%mm4, %%mm7\n\t"              // divide the result

        "add         $8, %1\n\t"
        "packssdw    %%mm7, %%mm0\n\t"

        "movq        %%mm0, -8(%%edx)\n\t"
        "dec         %%edi\n\t"

        "jnz         3b\n\t"

        "emms\n\t"
      : /* output */
      : "rim" (numSamples),
        "rim" (local_src),
        "rim" (local_length),
        "rim" (dest),
        "rim" (local_lengthDiv8),
        "rim" (local_filterCoeffs),
        "rim" (local_resultDivider) /* input */
      : "%eax", "%ebx", "%ecx", "%edx", "%edi"/*, "%esi"*/ /* regs */
    );
    return (numSamples & 0xfffffffe) - local_length;
#else
    throw runtime_error("MMX not supported");
    return 0;
#endif
}

#endif // ALLOW_MMX