x86-64.c

各种工程计算的库函数

void dsp_x86_64_iirf (float *fpVect, long lDataLength, const float *fpCoeff, 
    float *fpX, float *fpY)
{
    long lDataCntr;
    
    X86_64_ASM (
        "movss %0, %%xmm1\n\t" \
        "movss %1, %%xmm2\n\t" \
        "movss %2, %%xmm3\n\t" \
        "movss %3, %%xmm4\n\t" \
        "prefetchnta %4\n\t"
        :
        : "m" (fpX[1]),
          "m" (fpX[2]),
          "m" (fpY[0]),
          "m" (fpY[1]),
          "m" (fpCoeff[0])
        : "xmm1", "xmm2", "xmm3", "xmm4", "memory");
    for (lDataCntr = 0; lDataCntr < lDataLength; lDataCntr++)
    {
        X86_64_ASM (
            "movss %%xmm1, %%xmm0\n\t" \
            "movss %%xmm2, %%xmm1\n\t" \
            "movss %1, %%xmm2\n\t" \
            \
            "movss %2, %%xmm5\n\t" \
            "mulss %%xmm2, %%xmm5\n\t" \
            "movss %3, %%xmm6\n\t" \
            "mulss %%xmm1, %%xmm6\n\t" \
            "addss %%xmm6, %%xmm5\n\t" \
            "movss %4, %%xmm6\n\t" \
            "mulss %%xmm0, %%xmm6\n\t" \
            "addss %%xmm6, %%xmm5\n\t" \
            \
            "movss %5, %%xmm6\n\t" \
            "mulss %%xmm4, %%xmm6\n\t" \
            "movss %6, %%xmm7\n\t" \
            "mulss %%xmm3, %%xmm7\n\t" \
            "addss %%xmm7, %%xmm6\n\t" \
            \
            "addss %%xmm5, %%xmm6\n\t" \
            "movss %%xmm4, %%xmm3\n\t" \
            "movss %%xmm6, %%xmm4\n\t" \
            \
            "movss %%xmm6, %0\n\t"
            : "=m" (fpVect[lDataCntr])
            : "m0" (fpVect[lDataCntr]),
              "m" (fpCoeff[0]),
              "m" (fpCoeff[1]),
              "m" (fpCoeff[2]),
              "m" (fpCoeff[3]),
              "m" (fpCoeff[4])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
              "memory");
    }
    X86_64_ASM (
        "movss %%xmm0, %0\n\t" \
        "movss %%xmm1, %1\n\t" \
        "movss %%xmm2, %2\n\t" \
        "movss %%xmm3, %3\n\t" \
        "movss %%xmm4, %4\n\t"
        : "=m" (fpX[0]),
          "=m" (fpX[1]),
          "=m" (fpX[2]),
          "=m" (fpY[0]),
          "=m" (fpY[1])
        :
        : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "memory");
}


void dsp_x86_64_iir (double *dpVect, long lDataLength, const double *dpCoeff, 
    double *dpX, double *dpY)
{
    long lDataCntr;
    
    X86_64_ASM (
        "movsd %0, %%xmm1\n\t" \
        "movsd %1, %%xmm2\n\t" \
        "movsd %2, %%xmm3\n\t" \
        "movsd %3, %%xmm4\n\t" \
        "prefetchnta %4\n\t" \
        "prefetchnta %5\n\t"
        :
        : "m" (dpX[1]),
          "m" (dpX[2]),
          "m" (dpY[0]),
          "m" (dpY[1]),
          "m" (dpCoeff[0]),
          "m" (dpCoeff[3])
        : "xmm1", "xmm2", "xmm3", "xmm4", "memory");
    for (lDataCntr = 0; lDataCntr < lDataLength; lDataCntr++)
    {
        X86_64_ASM (
            "movsd %%xmm1, %%xmm0\n\t" \
            "movsd %%xmm2, %%xmm1\n\t" \
            "movsd %1, %%xmm2\n\t" \
            \
            "movsd %2, %%xmm5\n\t" \
            "mulsd %%xmm2, %%xmm5\n\t" \
            "movsd %3, %%xmm6\n\t" \
            "mulsd %%xmm1, %%xmm6\n\t" \
            "addsd %%xmm6, %%xmm5\n\t" \
            "movsd %4, %%xmm6\n\t" \
            "mulsd %%xmm0, %%xmm6\n\t" \
            "addsd %%xmm6, %%xmm5\n\t" \
            \
            "movsd %5, %%xmm6\n\t" \
            "mulsd %%xmm4, %%xmm6\n\t" \
            "movsd %6, %%xmm7\n\t" \
            "mulsd %%xmm3, %%xmm7\n\t" \
            "addsd %%xmm7, %%xmm6\n\t" \
            \
            "addsd %%xmm5, %%xmm6\n\t" \
            "movsd %%xmm4, %%xmm3\n\t" \
            "movsd %%xmm6, %%xmm4\n\t" \
            \
            "movsd %%xmm6, %0\n\t"
            : "=m" (dpVect[lDataCntr])
            : "m0" (dpVect[lDataCntr]),
              "m" (dpCoeff[0]),
              "m" (dpCoeff[1]),
              "m" (dpCoeff[2]),
              "m" (dpCoeff[3]),
              "m" (dpCoeff[4])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
              "memory");
    }
    X86_64_ASM (
        "movsd %%xmm0, %0\n\t" \
        "movsd %%xmm1, %1\n\t" \
        "movsd %%xmm2, %2\n\t" \
        "movsd %%xmm3, %3\n\t" \
        "movsd %%xmm4, %4\n\t"
        : "=m" (dpX[0]),
          "=m" (dpX[1]),
          "=m" (dpX[2]),
          "=m" (dpY[0]),
          "=m" (dpY[1])
        :
        : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "memory");
}


void dsp_x86_64_iirf_nip (float *fpDest, const float *fpSrc, long lDataLength, 
    const float *fpCoeff, float *fpX, float *fpY)
{
    long lDataCntr;
    
    X86_64_ASM (
        "movss %0, %%xmm1\n\t" \
        "movss %1, %%xmm2\n\t" \
        "movss %2, %%xmm3\n\t" \
        "movss %3, %%xmm4\n\t" \
        "prefetchnta %4\n\t"
        :
        : "m" (fpX[1]),
          "m" (fpX[2]),
          "m" (fpY[0]),
          "m" (fpY[1]),
          "m" (fpCoeff[0])
        : "xmm1", "xmm2", "xmm3", "xmm4", "memory");
    for (lDataCntr = 0; lDataCntr < lDataLength; lDataCntr++)
    {
        X86_64_ASM (
            "movss %%xmm1, %%xmm0\n\t" \
            "movss %%xmm2, %%xmm1\n\t" \
            "movss %1, %%xmm2\n\t" \
            \
            "movss %2, %%xmm5\n\t" \
            "mulss %%xmm2, %%xmm5\n\t" \
            "movss %3, %%xmm6\n\t" \
            "mulss %%xmm1, %%xmm6\n\t" \
            "addss %%xmm6, %%xmm5\n\t" \
            "movss %4, %%xmm6\n\t" \
            "mulss %%xmm0, %%xmm6\n\t" \
            "addss %%xmm6, %%xmm5\n\t" \
            \
            "movss %5, %%xmm6\n\t" \
            "mulss %%xmm4, %%xmm6\n\t" \
            "movss %6, %%xmm7\n\t" \
            "mulss %%xmm3, %%xmm7\n\t" \
            "addss %%xmm7, %%xmm6\n\t" \
            \
            "addss %%xmm5, %%xmm6\n\t" \
            "movss %%xmm4, %%xmm3\n\t" \
            "movss %%xmm6, %%xmm4\n\t" \
            \
            "movss %%xmm6, %0\n\t"
            : "=m" (fpDest[lDataCntr])
            : "m" (fpSrc[lDataCntr]),
              "m" (fpCoeff[0]),
              "m" (fpCoeff[1]),
              "m" (fpCoeff[2]),
              "m" (fpCoeff[3]),
              "m" (fpCoeff[4])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
              "memory");
    }
    X86_64_ASM (
        "movss %%xmm0, %0\n\t" \
        "movss %%xmm1, %1\n\t" \
        "movss %%xmm2, %2\n\t" \
        "movss %%xmm3, %3\n\t" \
        "movss %%xmm4, %4\n\t"
        : "=m" (fpX[0]),
          "=m" (fpX[1]),
          "=m" (fpX[2]),
          "=m" (fpY[0]),
          "=m" (fpY[1])
        :
        : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "memory");
}


void dsp_x86_64_iir_nip (double *dpDest, const double *dpSrc, long lDataLength, 
    const double *dpCoeff, double *dpX, double *dpY)
{
    long lDataCntr;
    
    X86_64_ASM (
        "movsd %0, %%xmm1\n\t" \
        "movsd %1, %%xmm2\n\t" \
        "movsd %2, %%xmm3\n\t" \
        "movsd %3, %%xmm4\n\t" \
        "prefetchnta %4\n\t" \
        "prefetchnta %5\n\t"
        :
        : "m" (dpX[1]),
          "m" (dpX[2]),
          "m" (dpY[0]),
          "m" (dpY[1]),
          "m" (dpCoeff[0]),
          "m" (dpCoeff[3])
        : "xmm1", "xmm2", "xmm3", "xmm4", "memory");
    for (lDataCntr = 0; lDataCntr < lDataLength; lDataCntr++)
    {
        X86_64_ASM (
            "movsd %%xmm1, %%xmm0\n\t" \
            "movsd %%xmm2, %%xmm1\n\t" \
            "movsd %1, %%xmm2\n\t" \
            \
            "movsd %2, %%xmm5\n\t" \
            "mulsd %%xmm2, %%xmm5\n\t" \
            "movsd %3, %%xmm6\n\t" \
            "mulsd %%xmm1, %%xmm6\n\t" \
            "addsd %%xmm6, %%xmm5\n\t" \
            "movsd %4, %%xmm6\n\t" \
            "mulsd %%xmm0, %%xmm6\n\t" \
            "addsd %%xmm6, %%xmm5\n\t" \
            \
            "movsd %5, %%xmm6\n\t" \
            "mulsd %%xmm4, %%xmm6\n\t" \
            "movsd %6, %%xmm7\n\t" \
            "mulsd %%xmm3, %%xmm7\n\t" \
            "addsd %%xmm7, %%xmm6\n\t" \
            \
            "addsd %%xmm5, %%xmm6\n\t" \
            "movsd %%xmm4, %%xmm3\n\t" \
            "movsd %%xmm6, %%xmm4\n\t" \
            \
            "movsd %%xmm6, %0\n\t"
            : "=m" (dpDest[lDataCntr])
            : "m" (dpSrc[lDataCntr]),
              "m" (dpCoeff[0]),
              "m" (dpCoeff[1]),
              "m" (dpCoeff[2]),
              "m" (dpCoeff[3]),
              "m" (dpCoeff[4])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", "xmm7",
              "memory");
    }
    X86_64_ASM (
        "movsd %%xmm0, %0\n\t" \
        "movsd %%xmm1, %1\n\t" \
        "movsd %%xmm2, %2\n\t" \
        "movsd %%xmm3, %3\n\t" \
        "movsd %%xmm4, %4\n\t"
        : "=m" (dpX[0]),
          "=m" (dpX[1]),
          "=m" (dpX[2]),
          "=m" (dpY[0]),
          "=m" (dpY[1])
        :
        : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "memory");
}


#ifdef __cplusplus
}
#endif

#endif