x86.c

各种工程计算的库函数

    int iDataCntr;
    
    for (iDataCntr = 0; iDataCntr < iDataLength; iDataCntr++)
    {
        X86_ASM (
            "movsd %1, %%xmm0\n\t" \
            "mulsd %2, %%xmm0\n\t" \
            "movsd %%xmm0, %0\n\t"
            : "=m" (dpDest[iDataCntr])
            : "m" (dpSrc1[iDataCntr]),
              "m" (dpSrc2[iDataCntr])
            : "xmm0", "memory");
    }
}


void dsp_x86_3dnow_cmulf (float *fpDest, const float *fpSrc, int iDataLength)
{
    int iDataCntr;
    pv2sf m64pDest = (pv2sf) fpDest;
    
    X86_ASM (
        "movq %0, %%mm3\n\t"
        :
        : "m" (fpSrc[0])
        : "mm3", "memory");
    for (iDataCntr = 0; iDataCntr < iDataLength; iDataCntr++)
    {
        X86_ASM (
            "movq %1, %%mm0\n\t" \
            "movq %%mm3, %%mm1\n\t" \
            "pswapd %%mm1, %%mm2\n\t" \
            "pfmul %%mm0, %%mm1\n\t" \
            "pfmul %%mm0, %%mm2\n\t" \
            "pfpnacc %%mm2, %%mm1\n\t" \
            "movntq %%mm1, %0\n\t"
            : "=m" (m64pDest[iDataCntr])
            : "m0" (m64pDest[iDataCntr])
            : "mm0", "mm1", "mm2", "mm3", "memory");
    }
    X86_ASM (
        "femms\n\t" \
        "sfence\n\t");
}


void dsp_x86_sse_cmulf (float *fpDest, const float *fpSrc, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    
    X86_ASM (
        "movss %0, %%xmm2\n\t" \
        "movss %1, %%xmm3\n\t"
        :
        : "m" (fpSrc[0]),
          "m" (fpSrc[1])
        : "xmm2", "xmm3", "memory");
    iDataCount = (iDataLength << 1);
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr += 2)
    {
        X86_ASM (
            "movss %2, %%xmm0\n\t" \
            "movss %%xmm0, %%xmm1\n\t" \
            "movss %3, %%xmm4\n\t" \
            \
            "mulss %%xmm2, %%xmm0\n\t" \
            "movss %%xmm4, %%xmm5\n\t" \
            "mulss %%xmm3, %%xmm5\n\t" \
            "subss %%xmm5, %%xmm0\n\t" \
            \
            "mulss %%xmm3, %%xmm1\n\t" \
            "movss %%xmm4, %%xmm5\n\t" \
            "mulss %%xmm2, %%xmm5\n\t" \
            "addss %%xmm5, %%xmm1\n\t" \
            \
            "movss %%xmm0, %0\n\t" \
            "movss %%xmm1, %1\n\t"
            : "=m" (fpDest[iDataCntr]),
              "=m" (fpDest[iDataCntr + 1])
            : "m0" (fpDest[iDataCntr]),
              "m1" (fpDest[iDataCntr + 1])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "memory");
    }
}


void dsp_x86_sse_cmul (double *dpDest, const double *dpSrc, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    
    X86_ASM (
        "movsd %0, %%xmm2\n\t" \
        "movsd %1, %%xmm3\n\t"
        :
        : "m" (dpSrc[0]),
          "m" (dpSrc[1])
        : "xmm2", "xmm3", "memory");
    iDataCount = (iDataLength << 1);
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr += 2)
    {
        X86_ASM (
            "movsd %2, %%xmm0\n\t" \
            "movsd %%xmm0, %%xmm1\n\t" \
            "movsd %3, %%xmm4\n\t" \
            \
            "mulsd %%xmm2, %%xmm0\n\t" \
            "movsd %%xmm4, %%xmm5\n\t" \
            "mulsd %%xmm3, %%xmm5\n\t" \
            "subsd %%xmm5, %%xmm0\n\t" \
            \
            "mulsd %%xmm3, %%xmm1\n\t" \
            "movsd %%xmm4, %%xmm5\n\t" \
            "mulsd %%xmm2, %%xmm5\n\t" \
            "addsd %%xmm5, %%xmm1\n\t" \
            \
            "movsd %%xmm0, %0\n\t" \
            "movsd %%xmm1, %1\n\t"
            : "=m" (dpDest[iDataCntr]),
              "=m" (dpDest[iDataCntr + 1])
            : "m0" (dpDest[iDataCntr]),
              "m1" (dpDest[iDataCntr + 1])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "memory");
    }
}


void dsp_x86_3dnow_cmul2f (float *fpDest, const float *fpSrc, int iDataLength)
{
    int iDataCntr;
    pv2sf m64pDest = (pv2sf) fpDest;
    pv2sf m64pSrc = (pv2sf) fpSrc;
    
    for (iDataCntr = 0; iDataCntr < iDataLength; iDataCntr++)
    {
        X86_ASM (
            "movq %1, %%mm0\n\t" \
            "movq %2, %%mm1\n\t" \
            "pswapd %%mm1, %%mm2\n\t" \
            "pfmul %%mm0, %%mm1\n\t" \
            "pfmul %%mm0, %%mm2\n\t" \
            "pfpnacc %%mm2, %%mm1\n\t" \
            "movntq %%mm1, %0\n\t"
            : "=m" (m64pDest[iDataCntr])
            : "m0" (m64pDest[iDataCntr]),
              "m" (m64pSrc[iDataCntr])
            : "mm0", "mm1", "mm2", "memory");
    }
    X86_ASM (
        "femms\n\t" \
        "sfence\n\t");
}


void dsp_x86_sse_cmul2f (float *fpDest, const float *fpSrc, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    
    iDataCount = (iDataLength << 1);
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr += 2)
    {
        X86_ASM (
            "movss %4, %%xmm2\n\t" \
            "movss %5, %%xmm3\n\t" \
            \
            "movss %2, %%xmm0\n\t" \
            "movss %%xmm0, %%xmm1\n\t" \
            "movss %3, %%xmm4\n\t" \
            \
            "mulss %%xmm2, %%xmm0\n\t" \
            "movss %%xmm4, %%xmm5\n\t" \
            "mulss %%xmm3, %%xmm5\n\t" \
            "subss %%xmm5, %%xmm0\n\t" \
            \
            "mulss %%xmm3, %%xmm1\n\t" \
            "movss %%xmm4, %%xmm5\n\t" \
            "mulss %%xmm2, %%xmm5\n\t" \
            "addss %%xmm5, %%xmm1\n\t" \
            \
            "movss %%xmm0, %0\n\t" \
            "movss %%xmm1, %1\n\t"
            : "=m" (fpDest[iDataCntr]),
              "=m" (fpDest[iDataCntr + 1])
            : "m0" (fpDest[iDataCntr]),
              "m1" (fpDest[iDataCntr + 1]),
              "m" (fpSrc[iDataCntr]),
              "m" (fpSrc[iDataCntr + 1])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "memory");
    }
}


void dsp_x86_sse_cmul2 (double *dpDest, const double *dpSrc, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    
    iDataCount = (iDataLength << 1);
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr += 2)
    {
        X86_ASM (
            "movsd %4, %%xmm2\n\t" \
            "movsd %5, %%xmm3\n\t" \
            \
            "movsd %2, %%xmm0\n\t" \
            "movsd %%xmm0, %%xmm1\n\t" \
            "movsd %3, %%xmm4\n\t" \
            \
            "mulsd %%xmm2, %%xmm0\n\t" \
            "movsd %%xmm4, %%xmm5\n\t" \
            "mulsd %%xmm3, %%xmm5\n\t" \
            "subsd %%xmm5, %%xmm0\n\t" \
            \
            "mulsd %%xmm3, %%xmm1\n\t" \
            "movsd %%xmm4, %%xmm5\n\t" \
            "mulsd %%xmm2, %%xmm5\n\t" \
            "addsd %%xmm5, %%xmm1\n\t" \
            \
            "movsd %%xmm0, %0\n\t" \
            "movsd %%xmm1, %1\n\t"
            : "=m" (dpDest[iDataCntr]),
              "=m" (dpDest[iDataCntr + 1])
            : "m0" (dpDest[iDataCntr]),
              "m1" (dpDest[iDataCntr + 1]),
              "m" (dpSrc[iDataCntr]),
              "m" (dpSrc[iDataCntr + 1])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "memory");
    }
}


void dsp_x86_3dnow_cmul3f (float *fpDest, const float *fpSrc1, 
    const float *fpSrc2, int iDataLength)
{
    int iDataCntr;
    pv2sf m64pDest = (pv2sf) fpDest;
    pv2sf m64pSrc1 = (pv2sf) fpSrc1;
    pv2sf m64pSrc2 = (pv2sf) fpSrc2;
    
    for (iDataCntr = 0; iDataCntr < iDataLength; iDataCntr++)
    {
        X86_ASM (
            "movq %1, %%mm0\n\t" \
            "movq %2, %%mm1\n\t" \
            "pswapd %%mm1, %%mm2\n\t" \
            "pfmul %%mm0, %%mm1\n\t" \
            "pfmul %%mm0, %%mm2\n\t" \
            "pfpnacc %%mm2, %%mm1\n\t" \
            "movntq %%mm1, %0\n\t"
            : "=m" (m64pDest[iDataCntr])
            : "m" (m64pSrc1[iDataCntr]),
              "m" (m64pSrc2[iDataCntr])
            : "mm0", "mm1", "mm2", "memory");
    }
    X86_ASM (
        "femms\n\t" \
        "sfence\n\t");
}


void dsp_x86_sse_cmul3f (float *fpDest, const float *fpSrc1, 
    const float *fpSrc2, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    
    iDataCount = (iDataLength << 1);
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr += 2)
    {
        X86_ASM (
            "movss %4, %%xmm2\n\t" \
            "movss %5, %%xmm3\n\t" \
            \
            "movss %2, %%xmm0\n\t" \
            "movss %%xmm0, %%xmm1\n\t" \
            "movss %3, %%xmm4\n\t" \
            \
            "mulss %%xmm2, %%xmm0\n\t" \
            "movss %%xmm4, %%xmm5\n\t" \
            "mulss %%xmm3, %%xmm5\n\t" \
            "subss %%xmm5, %%xmm0\n\t" \
            \
            "mulss %%xmm3, %%xmm1\n\t" \
            "movss %%xmm4, %%xmm5\n\t" \
            "mulss %%xmm2, %%xmm5\n\t" \
            "addss %%xmm5, %%xmm1\n\t" \
            \
            "movss %%xmm0, %0\n\t" \
            "movss %%xmm1, %1\n\t"
            : "=m" (fpDest[iDataCntr]),
              "=m" (fpDest[iDataCntr + 1])
            : "m" (fpSrc1[iDataCntr]),
              "m" (fpSrc1[iDataCntr + 1]),
              "m" (fpSrc2[iDataCntr]),
              "m" (fpSrc2[iDataCntr + 1])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "memory");
    }
}


void dsp_x86_sse_cmul3 (double *dpDest, const double *dpSrc1, 
    const double *dpSrc2, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    
    iDataCount = (iDataLength << 1);
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr += 2)
    {
        X86_ASM (
            "movsd %4, %%xmm2\n\t" \
            "movsd %5, %%xmm3\n\t" \
            \
            "movsd %2, %%xmm0\n\t" \
            "movsd %%xmm0, %%xmm1\n\t" \
            "movsd %3, %%xmm4\n\t" \
            \
            "mulsd %%xmm2, %%xmm0\n\t" \
            "movsd %%xmm4, %%xmm5\n\t" \
            "mulsd %%xmm3, %%xmm5\n\t" \
            "subsd %%xmm5, %%xmm0\n\t" \
            \
            "mulsd %%xmm3, %%xmm1\n\t" \
            "movsd %%xmm4, %%xmm5\n\t" \
            "mulsd %%xmm2, %%xmm5\n\t" \
            "addsd %%xmm5, %%xmm1\n\t" \
            \
            "movsd %%xmm0, %0\n\t" \
            "movsd %%xmm1, %1\n\t"
            : "=m" (dpDest[iDataCntr]),
              "=m" (dpDest[iDataCntr + 1])
            : "m" (dpSrc1[iDataCntr]),
              "m" (dpSrc1[iDataCntr + 1]),
              "m" (dpSrc2[iDataCntr]),
              "m" (dpSrc2[iDataCntr + 1])
            : "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "memory");
    }
}


void dsp_x86_3dnow_maf (float *fpVect, float fMul, float fAdd, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    pv2sf m64pVect = (pv2sf) fpVect;
    stm64 m64Mul;
    stm64 m64Add;

    m64Mul.f[0] = m64Mul.f[1] = fMul;
    m64Add.f[0] = m64Add.f[1] = fAdd;
    iDataCount = (iDataLength >> 1);
    X86_ASM (
        "movq %0, %%mm1\n\t" \
        "movq %1, %%mm2\n\t"
        :
        : "m" (m64Mul),
          "m" (m64Add)
        : "mm1", "mm2", "memory");
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr++)
    {
        X86_ASM (
            "movq %1, %%mm0\n\t" \
            "pfmul %%mm1, %%mm0\n\t" \
            "pfadd %%mm2, %%mm0\n\t" \
            "movntq %%mm0, %0\n\t"
            : "=m" (m64pVect[iDataCntr])
            : "m0" (m64pVect[iDataCntr])
            : "mm0", "mm1", "mm2", "memory");
    }
    if (iDataLength & 0x1)
    {
        X86_ASM (
            "movd %1, %%mm0\n\t" \
            "pfmul %%mm1, %%mm0\n\t" \
            "pfadd %%mm2, %%mm0\n\t" \
            "movd %%mm0, %0\n\t"
            : "=m" (fpVect[iDataLength - 1])
            : "m0" (fpVect[iDataLength - 1])
            : "mm0", "mm1", "mm2", "memory");
    }
    X86_ASM (
        "femms\n\t" \
        "sfence\n\t");
}


void dsp_x86_sse_maf (float *fpVect, float fMul, float fAdd, int iDataLength)
{
    int iDataCntr;
    
    X86_ASM (
        "movss %0, %%xmm1\n\t" \
        "movss %1, %%xmm2\n\t"
        :
        : "m" (fMul),
          "m" (fAdd)
        : "xmm1", "xmm2", "memory");
    for (iDataCntr = 0; iDataCntr < iDataLength; iDataCntr++)
    {
        X86_ASM (
            "movss %1, %%xmm0\n\t" \
            "mulss %%xmm1, %%xmm0\n\t" \
            "addss %%xmm2, %%xmm0\n\t" \
            "movss %%xmm0, %0\n\t"
            : "=m" (fpVect[iDataCntr])
            : "m0" (fpVect[iDataCntr])
            : "xmm0", "xmm1", "xmm2", "memory");
    }
}


void dsp_x86_sse_ma (double *dpVect, double dMul, double dAdd, int iDataLength)
{
    int iDataCntr;
    
    X86_ASM (
        "movsd %0, %%xmm1\n\t" \
        "movsd %1, %%xmm2\n\t"
        :
        : "m" (dMul),
          "m" (dAdd)
        : "xmm1", "xmm2", "memory");
    for (iDataCntr = 0; iDataCntr < iDataLength; iDataCntr++)
    {
        X86_ASM (
            "movsd %1, %%xmm0\n\t" \
            "mulsd %%xmm1, %%xmm0\n\t" \
            "addsd %%xmm2, %%xmm0\n\t" \
            "movsd %%xmm0, %0\n\t"
            : "=m" (dpVect[iDataCntr])
            : "m0" (dpVect[iDataCntr])
            : "xmm0", "xmm1", "xmm2", "memory");
    }
}


void dsp_x86_3dnow_ma2f (float *fpDest, const float *fpSrc,
    float fMul, float fAdd, int iDataLength)
{
    int iDataCntr;
    int iDataCount;
    pv2sf m64pDest = (pv2sf) fpDest;
    pv2sf m64pSrc = (pv2sf) fpSrc;
    stm64 m64Mul;
    stm64 m64Add;

    m64Mul.f[0] = m64Mul.f[1] = fMul;
    m64Add.f[0] = m64Add.f[1] = fAdd;
    iDataCount = (iDataLength >> 1);
    X86_ASM (
        "movq %0, %%mm1\n\t" \
        "movq %1, %%mm2\n\t"
        :
        : "m" (m64Mul),
          "m" (m64Add)
        : "mm1", "mm2", "memory");
    for (iDataCntr = 0; iDataCntr < iDataCount; iDataCntr++)
    {
        X86_ASM (
            "movq %1, %%mm0\n\t" \
            "pfmul %%mm1, %%mm0\n\t" \
            "pfadd %%mm2, %%mm0\n\t" \
            "movntq %%mm0, %0\n\t"
            : "=m" (m64pDest[iDataCntr])
            : "m" (m64pSrc[iDataCntr])
            : "mm0", "mm1", "mm2", "memory");
    }
    if (iDataLength & 0x1)
    {
        X86_ASM (
            "movd %1, %%mm0\n\t" \
            "pfmul %%mm1, %%mm0\n\t" \
            "pfadd %%mm2, %%mm0\n\t" \
            "movd %%mm0, %0\n\t"
            : "=m" (fpDest[iDataLength - 1])
            : "m" (fpSrc[iDataLength - 1])
            : "mm0", "mm1", "mm2", "memory");
    }
    X86_ASM (
        "femms\n\t" \
        "sfence\n\t");
}


void dsp_x86_sse_ma2f (float *fpDest, const float *fpSrc, 
    float fMul, float fAdd, int iDataLength)
{