Esempio n. 1
0
/* A vectorized version of the Voigt function using X86 SSE instructions */
void my_voigt(const float *damping, const float *frequency_offset, float *voigt_value, int N)   
{                                                                   
    // coefficients of the rational approximation formula
    // to the complementary error function
    const __m128 A0 = _mm_set1_ps(122.607931777104326f);
    const __m128 A1 = _mm_set1_ps(214.382388694706425f);
    const __m128 A2 = _mm_set1_ps(181.928533092181549f);
    const __m128 A3 = _mm_set1_ps(93.155580458138441f);
    const __m128 A4 = _mm_set1_ps(30.180142196210589f);
    const __m128 A5 = _mm_set1_ps(5.912626209773153f);
    const __m128 A6 = _mm_set1_ps(0.564189583562615f);
    const __m128 B0 = _mm_set1_ps(122.60793177387535f);
    const __m128 B1 = _mm_set1_ps(352.730625110963558f);
    const __m128 B2 = _mm_set1_ps(457.334478783897737f);
    const __m128 B3 = _mm_set1_ps(348.703917719495792f);
    const __m128 B4 = _mm_set1_ps(170.354001821091472f);
    const __m128 B5 = _mm_set1_ps(53.992906912940207f);
    const __m128 B6 = _mm_set1_ps(10.479857114260399f);

    __m128 ivsigno;
    __m128 V;
    __m128 Z1_real;
    __m128 Z1_imag;
    __m128 Z2_real;
    __m128 Z2_imag;
    __m128 Z3_real;
    __m128 Z3_imag;
    __m128 Z4_real;
    __m128 Z4_imag;
    __m128 Z5_real;
    __m128 Z5_imag;
    __m128 Z6_real;
    __m128 Z6_imag;
    __m128 ZZ1_real;
    __m128 ZZ1_imag;
    __m128 ZZ2_real;
    __m128 ZZ2_imag;
    __m128 ZZ3_real;
    __m128 ZZ3_imag;
    __m128 ZZ4_real;
    __m128 ZZ4_imag;
    __m128 ZZ5_real;
    __m128 ZZ5_imag;
    __m128 ZZ6_real;
    __m128 ZZ6_imag;
    __m128 ZZ7_real;
    __m128 ZZ7_imag;
    __m128 division_factor;
    __m128 ZZZ_real;
    __m128 damp;
    __m128 offs;
    __m128 vval;
    __m128 one = _mm_set1_ps(1.0f); 
    __m128 zero = _mm_set1_ps(0.0f);
    __m128 mone = _mm_set1_ps(-1.0f);
    __m128 half = _mm_set1_ps(-0.5f);
    __m128 mask;

    float *stmp = (float *) _mm_malloc(4*sizeof(float), 16);

    int i;
    for(i=0; i<N; i+=VECLEN){
        _mm_prefetch((const char *)&damping[i+64], _MM_HINT_T0);
        _mm_prefetch((const char *)&frequency_offset[i+64], _MM_HINT_T0);
        damp = _mm_load_ps(&damping[i]);
        offs = _mm_load_ps(&frequency_offset[i]);
        mask = _mm_cmplt_ps(offs, zero);
        ivsigno = _mm_add_ps(_mm_and_ps(mask,mone),_mm_andnot_ps(mask,one));
        V = _mm_mul_ps(ivsigno, offs);       

        Z1_real = _mm_add_ps(_mm_mul_ps(A6, damp), A5);
        Z1_imag = _mm_mul_ps(A6, V);
        Z2_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(Z1_real, damp), _mm_mul_ps(Z1_imag, V)), A4);
        Z2_imag = _mm_add_ps(_mm_mul_ps(Z1_real, V), _mm_mul_ps(Z1_imag, damp));
        Z3_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(Z2_real, damp), _mm_mul_ps(Z2_imag, V)), A3);
        Z3_imag = _mm_add_ps(_mm_mul_ps(Z2_real, V), _mm_mul_ps(Z2_imag, damp));
        Z4_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(Z3_real, damp), _mm_mul_ps(Z3_imag, V)), A2);
        Z4_imag = _mm_add_ps(_mm_mul_ps(Z3_real, V), _mm_mul_ps(Z3_imag, damp));
        Z5_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(Z4_real, damp), _mm_mul_ps(Z4_imag, V)), A1);
        Z5_imag = _mm_add_ps(_mm_mul_ps(Z4_real, V), _mm_mul_ps(Z4_imag, damp));
        Z6_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(Z5_real, damp), _mm_mul_ps(Z5_imag, V)), A0);
        Z6_imag = _mm_add_ps(_mm_mul_ps(Z5_real, V), _mm_mul_ps(Z5_imag, damp));
        ZZ1_real = _mm_add_ps(damp, B6);          
        ZZ1_imag = V;                    
        ZZ2_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(ZZ1_real, damp), _mm_mul_ps(ZZ1_imag, V)), B5); 
        ZZ2_imag = _mm_add_ps(_mm_mul_ps(ZZ1_real, V), _mm_mul_ps(ZZ1_imag, damp)); 
        ZZ3_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(ZZ2_real, damp), _mm_mul_ps(ZZ2_imag, V)), B4); 
        ZZ3_imag = _mm_add_ps(_mm_mul_ps(ZZ2_real, V), _mm_mul_ps(ZZ2_imag, damp)); 
        ZZ4_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(ZZ3_real, damp), _mm_mul_ps(ZZ3_imag, V)), B3); 
        ZZ4_imag = _mm_add_ps(_mm_mul_ps(ZZ3_real, V), _mm_mul_ps(ZZ3_imag, damp)); 
        ZZ5_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(ZZ4_real, damp), _mm_mul_ps(ZZ4_imag, V)), B2); 
        ZZ5_imag = _mm_add_ps(_mm_mul_ps(ZZ4_real, V), _mm_mul_ps(ZZ4_imag, damp)); 
        ZZ6_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(ZZ5_real, damp), _mm_mul_ps(ZZ5_imag, V)), B1); 
        ZZ6_imag = _mm_add_ps(_mm_mul_ps(ZZ5_real, V), _mm_mul_ps(ZZ5_imag, damp)); 
        ZZ7_real = _mm_add_ps(_mm_sub_ps(_mm_mul_ps(ZZ6_real, damp), _mm_mul_ps(ZZ6_imag, V)), B0); 
        ZZ7_imag = _mm_add_ps(_mm_mul_ps(ZZ6_real, V), _mm_mul_ps(ZZ6_imag, damp)); 
        division_factor = _mm_div_ps(one, _mm_add_ps(_mm_mul_ps(ZZ7_real, ZZ7_real), _mm_mul_ps(ZZ7_imag, ZZ7_imag)));
        ZZZ_real = _mm_mul_ps((_mm_add_ps(_mm_mul_ps(Z6_real, ZZ7_real), _mm_mul_ps(Z6_imag, ZZ7_imag))), division_factor); 

        _mm_stream_ps(&voigt_value[i], ZZZ_real);
    }
    _mm_free(stmp);
}
Esempio n. 2
0
CPLErr
GDALGridInverseDistanceToAPower2NoSmoothingNoSearchSSE(
                                        const void *poOptions,
                                        GUInt32 nPoints,
                                        CPL_UNUSED const double *unused_padfX,
                                        CPL_UNUSED const double *unused_padfY,
                                        CPL_UNUSED const double *unused_padfZ,
                                        double dfXPoint, double dfYPoint,
                                        double *pdfValue,
                                        void* hExtraParamsIn )
{
    size_t i = 0;
    GDALGridExtraParameters* psExtraParams =
        static_cast<GDALGridExtraParameters *>(hExtraParamsIn);
    const float* pafX = psExtraParams->pafX;
    const float* pafY = psExtraParams->pafY;
    const float* pafZ = psExtraParams->pafZ;

    const float fEpsilon = 0.0000000000001f;
    const float fXPoint = static_cast<float>(dfXPoint);
    const float fYPoint = static_cast<float>(dfYPoint);
    const __m128 xmm_small = _mm_load1_ps(const_cast<float *>(&fEpsilon));
    const __m128 xmm_x = _mm_load1_ps(const_cast<float*>(&fXPoint));
    const __m128 xmm_y = _mm_load1_ps(const_cast<float*>(&fYPoint));
    __m128 xmm_nominator = _mm_setzero_ps();
    __m128 xmm_denominator = _mm_setzero_ps();
    int mask = 0;

#if defined(__x86_64) || defined(_M_X64)
    // This would also work in 32bit mode, but there are only 8 XMM registers
    // whereas we have 16 for 64bit.
    const size_t LOOP_SIZE = 8;
    size_t nPointsRound = (nPoints / LOOP_SIZE) * LOOP_SIZE;
    for( i = 0; i < nPointsRound; i += LOOP_SIZE )
    {
        // rx = pafX[i] - fXPoint
        __m128 xmm_rx = _mm_sub_ps(_mm_load_ps(pafX + i), xmm_x);
        __m128 xmm_rx_4 = _mm_sub_ps(_mm_load_ps(pafX + i + 4), xmm_x);
        // ry = pafY[i] - fYPoint
        __m128 xmm_ry = _mm_sub_ps(_mm_load_ps(pafY + i), xmm_y);
        __m128 xmm_ry_4 = _mm_sub_ps(_mm_load_ps(pafY + i + 4), xmm_y);
        // r2 = rx * rx + ry * ry
        __m128 xmm_r2 = _mm_add_ps(_mm_mul_ps(xmm_rx, xmm_rx),
                                   _mm_mul_ps(xmm_ry, xmm_ry));
        __m128 xmm_r2_4 = _mm_add_ps(_mm_mul_ps(xmm_rx_4, xmm_rx_4),
                                     _mm_mul_ps(xmm_ry_4, xmm_ry_4));
        // invr2 = 1.0f / r2
        __m128 xmm_invr2 = _mm_rcp_ps(xmm_r2);
        __m128 xmm_invr2_4 = _mm_rcp_ps(xmm_r2_4);
        // nominator += invr2 * pafZ[i]
        xmm_nominator = _mm_add_ps(xmm_nominator,
                            _mm_mul_ps(xmm_invr2, _mm_load_ps(pafZ + i)));
        xmm_nominator = _mm_add_ps(xmm_nominator,
                            _mm_mul_ps(xmm_invr2_4, _mm_load_ps(pafZ + i + 4)));
        // denominator += invr2
        xmm_denominator = _mm_add_ps(xmm_denominator, xmm_invr2);
        xmm_denominator = _mm_add_ps(xmm_denominator, xmm_invr2_4);
        // if( r2 < fEpsilon)
        mask = _mm_movemask_ps(_mm_cmplt_ps(xmm_r2, xmm_small)) |
              (_mm_movemask_ps(_mm_cmplt_ps(xmm_r2_4, xmm_small)) << 4);
        if( mask )
            break;
    }
#else
#define LOOP_SIZE   4
    size_t nPointsRound = (nPoints / LOOP_SIZE) * LOOP_SIZE;
    for( i = 0; i < nPointsRound; i += LOOP_SIZE )
    {
        __m128 xmm_rx = _mm_sub_ps(_mm_load_ps(pafX + i), xmm_x);           /* rx = pafX[i] - fXPoint */
        __m128 xmm_ry = _mm_sub_ps(_mm_load_ps(pafY + i), xmm_y);           /* ry = pafY[i] - fYPoint */
        __m128 xmm_r2 = _mm_add_ps(_mm_mul_ps(xmm_rx, xmm_rx),              /* r2 = rx * rx + ry * ry */
                                   _mm_mul_ps(xmm_ry, xmm_ry));
        __m128 xmm_invr2 = _mm_rcp_ps(xmm_r2);                              /* invr2 = 1.0f / r2 */
        xmm_nominator = _mm_add_ps(xmm_nominator,                           /* nominator += invr2 * pafZ[i] */
                            _mm_mul_ps(xmm_invr2, _mm_load_ps(pafZ + i)));
        xmm_denominator = _mm_add_ps(xmm_denominator, xmm_invr2);           /* denominator += invr2 */
        mask = _mm_movemask_ps(_mm_cmplt_ps(xmm_r2, xmm_small));            /* if( r2 < fEpsilon) */
        if( mask )
            break;
    }
#endif

    // Find which i triggered r2 < fEpsilon.
    if( mask )
    {
        for( size_t j = 0; j < LOOP_SIZE; j++ )
        {
            if( mask & (1 << j) )
            {
                (*pdfValue) = (pafZ)[i + j];
                return CE_None;
            }
        }
    }

    // Get back nominator and denominator values for XMM registers.
    float afNominator[4];
    float afDenominator[4];
    _mm_storeu_ps(afNominator, xmm_nominator);
    _mm_storeu_ps(afDenominator, xmm_denominator);

    float fNominator = afNominator[0] + afNominator[1] +
                       afNominator[2] + afNominator[3];
    float fDenominator = afDenominator[0] + afDenominator[1] +
                         afDenominator[2] + afDenominator[3];

    /* Do the few remaining loop iterations */
    for( ; i < nPoints; i++ )
    {
        const float fRX = pafX[i] - fXPoint;
        const float fRY = pafY[i] - fYPoint;
        const float fR2 =
            fRX * fRX + fRY * fRY;

        // If the test point is close to the grid node, use the point
        // value directly as a node value to avoid singularity.
        if( fR2 < 0.0000000000001 )
        {
            break;
        }
        else
        {
            const float fInvR2 = 1.0f / fR2;
            fNominator += fInvR2 * pafZ[i];
            fDenominator += fInvR2;
        }
    }

    if( i != nPoints )
    {
        (*pdfValue) = pafZ[i];
    }
    else
    if( fDenominator == 0.0 )
    {
        (*pdfValue) =
            static_cast<const GDALGridInverseDistanceToAPowerOptions*>(poOptions)->dfNoDataValue;
    }
    else
    {
        (*pdfValue) = fNominator / fDenominator;
    }

    return CE_None;
}
Esempio n. 3
0
void process(struct dt_iop_module_t *self, dt_dev_pixelpipe_iop_t *piece, void *ivoid, void *ovoid,
             const dt_iop_roi_t *roi_in, const dt_iop_roi_t *roi_out)
{
  const dt_iop_colorout_data_t *const d = (dt_iop_colorout_data_t *)piece->data;
  const int ch = piece->colors;
  const int gamutcheck = (d->softproof_enabled == DT_SOFTPROOF_GAMUTCHECK);

  if(!isnan(d->cmatrix[0]))
  {
// fprintf(stderr,"Using cmatrix codepath\n");
// convert to rgb using matrix
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none) shared(roi_in, roi_out, ivoid, ovoid)
#endif
    for(int j = 0; j < roi_out->height; j++)
    {

      float *in = (float *)ivoid + (size_t)ch * roi_in->width * j;
      float *out = (float *)ovoid + (size_t)ch * roi_out->width * j;
      const __m128 m0 = _mm_set_ps(0.0f, d->cmatrix[6], d->cmatrix[3], d->cmatrix[0]);
      const __m128 m1 = _mm_set_ps(0.0f, d->cmatrix[7], d->cmatrix[4], d->cmatrix[1]);
      const __m128 m2 = _mm_set_ps(0.0f, d->cmatrix[8], d->cmatrix[5], d->cmatrix[2]);

      for(int i = 0; i < roi_out->width; i++, in += ch, out += ch)
      {
        const __m128 xyz = dt_Lab_to_XYZ_SSE(_mm_load_ps(in));
        const __m128 t
            = _mm_add_ps(_mm_mul_ps(m0, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(0, 0, 0, 0))),
                         _mm_add_ps(_mm_mul_ps(m1, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(1, 1, 1, 1))),
                                    _mm_mul_ps(m2, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(2, 2, 2, 2)))));

        _mm_stream_ps(out, t);
      }
    }
    _mm_sfence();
// apply profile
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none) shared(roi_in, roi_out, ivoid, ovoid)
#endif
    for(int j = 0; j < roi_out->height; j++)
    {

      float *in = (float *)ivoid + (size_t)ch * roi_in->width * j;
      float *out = (float *)ovoid + (size_t)ch * roi_out->width * j;

      for(int i = 0; i < roi_out->width; i++, in += ch, out += ch)
      {
        for(int i = 0; i < 3; i++)
          if(d->lut[i][0] >= 0.0f)
          {
            out[i] = (out[i] < 1.0f) ? lerp_lut(d->lut[i], out[i])
                                     : dt_iop_eval_exp(d->unbounded_coeffs[i], out[i]);
          }
      }
    }
  }
  else
  {
    // fprintf(stderr,"Using xform codepath\n");
    const __m128 outofgamutpixel = _mm_set_ps(0.0f, 1.0f, 1.0f, 0.0f);
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none) shared(ivoid, ovoid, roi_out)
#endif
    for(int k = 0; k < roi_out->height; k++)
    {
      const float *in = ((float *)ivoid) + (size_t)ch * k * roi_out->width;
      float *out = ((float *)ovoid) + (size_t)ch * k * roi_out->width;

      if(!gamutcheck)
      {
        cmsDoTransform(d->xform, in, out, roi_out->width);
      }
      else
      {
        void *rgb = dt_alloc_align(16, 4 * sizeof(float) * roi_out->width);
        cmsDoTransform(d->xform, in, rgb, roi_out->width);
        float *rgbptr = (float *)rgb;
        for(int j = 0; j < roi_out->width; j++, rgbptr += 4, out += 4)
        {
          const __m128 pixel = _mm_load_ps(rgbptr);
          __m128 ingamut = _mm_cmplt_ps(pixel, _mm_set_ps(-FLT_MAX, 0.0f, 0.0f, 0.0f));

          ingamut = _mm_or_ps(_mm_unpacklo_ps(ingamut, ingamut), _mm_unpackhi_ps(ingamut, ingamut));
          ingamut = _mm_or_ps(_mm_unpacklo_ps(ingamut, ingamut), _mm_unpackhi_ps(ingamut, ingamut));

          const __m128 result
              = _mm_or_ps(_mm_and_ps(ingamut, outofgamutpixel), _mm_andnot_ps(ingamut, pixel));
          _mm_stream_ps(out, result);
        }
        dt_free_align(rgb);
      }
    }
    _mm_sfence();
  }

  if(piece->pipe->mask_display) dt_iop_alpha_copy(ivoid, ovoid, roi_out->width, roi_out->height);
}
Esempio n. 4
0
void process_sse2(struct dt_iop_module_t *self, dt_dev_pixelpipe_iop_t *piece, const void *const ivoid,
                  void *const ovoid, const dt_iop_roi_t *const roi_in, const dt_iop_roi_t *const roi_out)
{
  const dt_iop_colorout_data_t *const d = (dt_iop_colorout_data_t *)piece->data;
  const int ch = piece->colors;
  const int gamutcheck = (d->mode == DT_PROFILE_GAMUTCHECK);

  if(d->type == DT_COLORSPACE_LAB)
  {
    memcpy(ovoid, ivoid, sizeof(float)*4*roi_out->width*roi_out->height);
  }
  else if(!isnan(d->cmatrix[0]))
  {
// fprintf(stderr,"Using cmatrix codepath\n");
// convert to rgb using matrix
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none)
#endif
    for(int j = 0; j < roi_out->height; j++)
    {

      float *in = (float *)ivoid + (size_t)ch * roi_in->width * j;
      float *out = (float *)ovoid + (size_t)ch * roi_out->width * j;
      const __m128 m0 = _mm_set_ps(0.0f, d->cmatrix[6], d->cmatrix[3], d->cmatrix[0]);
      const __m128 m1 = _mm_set_ps(0.0f, d->cmatrix[7], d->cmatrix[4], d->cmatrix[1]);
      const __m128 m2 = _mm_set_ps(0.0f, d->cmatrix[8], d->cmatrix[5], d->cmatrix[2]);

      for(int i = 0; i < roi_out->width; i++, in += ch, out += ch)
      {
        const __m128 xyz = dt_Lab_to_XYZ_SSE(_mm_load_ps(in));
        const __m128 t
            = _mm_add_ps(_mm_mul_ps(m0, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(0, 0, 0, 0))),
                         _mm_add_ps(_mm_mul_ps(m1, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(1, 1, 1, 1))),
                                    _mm_mul_ps(m2, _mm_shuffle_ps(xyz, xyz, _MM_SHUFFLE(2, 2, 2, 2)))));

        _mm_stream_ps(out, t);
      }
    }
    _mm_sfence();

    process_fastpath_apply_tonecurves(self, piece, ivoid, ovoid, roi_in, roi_out);
  }
  else
  {
    // fprintf(stderr,"Using xform codepath\n");
    const __m128 outofgamutpixel = _mm_set_ps(0.0f, 1.0f, 1.0f, 0.0f);
#ifdef _OPENMP
#pragma omp parallel for schedule(static) default(none)
#endif
    for(int k = 0; k < roi_out->height; k++)
    {
      const float *in = ((float *)ivoid) + (size_t)ch * k * roi_out->width;
      float *out = ((float *)ovoid) + (size_t)ch * k * roi_out->width;

      cmsDoTransform(d->xform, in, out, roi_out->width);

      if(gamutcheck)
      {
        for(int j = 0; j < roi_out->width; j++, out += 4)
        {
          const __m128 pixel = _mm_load_ps(out);
          __m128 ingamut = _mm_cmplt_ps(pixel, _mm_set_ps(-FLT_MAX, 0.0f, 0.0f, 0.0f));

          ingamut = _mm_or_ps(_mm_unpacklo_ps(ingamut, ingamut), _mm_unpackhi_ps(ingamut, ingamut));
          ingamut = _mm_or_ps(_mm_unpacklo_ps(ingamut, ingamut), _mm_unpackhi_ps(ingamut, ingamut));

          const __m128 result
              = _mm_or_ps(_mm_and_ps(ingamut, outofgamutpixel), _mm_andnot_ps(ingamut, pixel));
          _mm_stream_ps(out, result);
        }
      }
    }
    _mm_sfence();
  }

  if(piece->pipe->mask_display) dt_iop_alpha_copy(ivoid, ovoid, roi_out->width, roi_out->height);
}
Esempio n. 5
0
int main()
{
	float *arr = get_arr(); // [4, 3, 2, 1]
	float *uarr = get_uarr(); // [5, 4, 3, 2]
	float *arr2 = get_arr2(); // [4, 3, 2, 1]
	float *uarr2 = get_uarr2(); // [5, 4, 3, 2]
	__m128 a = get_a(); // [8, 6, 4, 2]
	__m128 b = get_b(); // [1, 2, 3, 4]

	// Check that test data is like expected.
	Assert(((uintptr_t)arr & 0xF) == 0); // arr must be aligned by 16.
	Assert(((uintptr_t)uarr & 0xF) != 0); // uarr must be unaligned.
	Assert(((uintptr_t)arr2 & 0xF) == 0); // arr must be aligned by 16.
	Assert(((uintptr_t)uarr2 & 0xF) != 0); // uarr must be unaligned.

	// Test that aeq itself works and does not trivially return true on everything.
	Assert(aeq_("",_mm_load_ps(arr), 4.f, 3.f, 2.f, 0.f, false) == false);
#ifdef TEST_M64
	Assert(aeq64(u64castm64(0x22446688AACCEEFFULL), 0xABABABABABABABABULL, false) == false);
#endif
	// SSE1 Load instructions:	
	aeq(_mm_load_ps(arr), 4.f, 3.f, 2.f, 1.f); // 4-wide load from aligned address.
	aeq(_mm_load_ps1(uarr), 2.f, 2.f, 2.f, 2.f); // Load scalar from unaligned address and populate 4-wide.
	aeq(_mm_load_ss(uarr), 0.f, 0.f, 0.f, 2.f); // Load scalar from unaligned address to lowest, and zero all highest.
	aeq(_mm_load1_ps(uarr), 2.f, 2.f, 2.f, 2.f); // _mm_load1_ps == _mm_load_ps1
	aeq(_mm_loadh_pi(a, (__m64*)uarr), 3.f, 2.f, 4.f, 2.f); // Load two highest addresses, preserve two lowest.
	aeq(_mm_loadl_pi(a, (__m64*)uarr), 8.f, 6.f, 3.f, 2.f); // Load two lowest addresses, preserve two highest.
	aeq(_mm_loadr_ps(arr), 1.f, 2.f, 3.f, 4.f); // 4-wide load from an aligned address, but reverse order.
	aeq(_mm_loadu_ps(uarr), 5.f, 4.f, 3.f, 2.f); // 4-wide load from an unaligned address.

	// SSE1 Set instructions:
	aeq(_mm_set_ps(uarr[3], 2.f, 3.f, 4.f), 5.f, 2.f, 3.f, 4.f); // 4-wide set by specifying four immediate or memory operands.
	aeq(_mm_set_ps1(uarr[3]), 5.f, 5.f, 5.f, 5.f); // 4-wide set by specifying one scalar that is expanded.
	aeq(_mm_set_ss(uarr[3]), 0.f, 0.f, 0.f, 5.f); // Set scalar at lowest index, zero all higher.
	aeq(_mm_set1_ps(uarr[3]), 5.f, 5.f, 5.f, 5.f); // _mm_set1_ps == _mm_set_ps1
	aeq(_mm_setr_ps(uarr[3], 2.f, 3.f, 4.f), 4.f, 3.f, 2.f, 5.f); // 4-wide set by specifying four immediate or memory operands, but reverse order.
	aeq(_mm_setzero_ps(), 0.f, 0.f, 0.f, 0.f); // Returns a new zero register.

	// SSE1 Move instructions:
	aeq(_mm_move_ss(a, b), 8.f, 6.f, 4.f, 4.f); // Copy three highest elements from a, and lowest from b.
	aeq(_mm_movehl_ps(a, b), 8.f, 6.f, 1.f, 2.f); // Copy two highest elements from a, and take two highest from b and place them to the two lowest in output.
	aeq(_mm_movelh_ps(a, b), 3.f, 4.f, 4.f, 2.f); // Copy two lowest elements from a, and take two lowest from b and place them to the two highest in output.

	// SSE1 Store instructions:
#ifdef TEST_M64
	/*M64*/*(uint64_t*)uarr = 0xCDCDCDCDCDCDCDCDULL; _mm_maskmove_si64(u64castm64(0x00EEDDCCBBAA9988ULL), u64castm64(0x0080FF7F01FEFF40ULL), (char*)uarr); Assert(*(uint64_t*)uarr == 0xCDEEDDCDCDAA99CDULL); // _mm_maskmove_si64: Conditionally store bytes of a 64-bit value.
	/*M64*/*(uint64_t*)uarr = 0xABABABABABABABABULL;       _m_maskmovq(u64castm64(0x00EEDDCCBBAA9988ULL), u64castm64(0x0080FF7F01FEFF40ULL), (char*)uarr); Assert(*(uint64_t*)uarr == 0xABEEDDABABAA99ABULL); // _m_maskmovq is an alias to _mm_maskmove_si64.
#endif
	_mm_store_ps(arr2, a); aeq(_mm_load_ps(arr2), 8.f, 6.f, 4.f, 2.f); // _mm_store_ps: 4-wide store to aligned memory address.
	_mm_store_ps1(arr2, a); aeq(_mm_load_ps(arr2), 2.f, 2.f, 2.f, 2.f); // _mm_store_ps1: Store lowest scalar to aligned address, duplicating the element 4 times. 
	_mm_storeu_ps(uarr2, _mm_set1_ps(100.f)); _mm_store_ss(uarr2, b); aeq(_mm_loadu_ps(uarr2), 100.f, 100.f, 100.f, 4.f); // _mm_store_ss: Store lowest scalar to unaligned address. Don't adjust higher addresses in memory.
	_mm_store_ps(arr2, _mm_set1_ps(100.f)); _mm_store1_ps(arr2, a); aeq(_mm_load_ps(arr2), 2.f, 2.f, 2.f, 2.f); // _mm_store1_ps == _mm_store_ps1
	_mm_storeu_ps(uarr2, _mm_set1_ps(100.f)); _mm_storeh_pi((__m64*)uarr2, a); aeq(_mm_loadu_ps(uarr2), 100.f, 100.f, 8.f, 6.f); // _mm_storeh_pi: Store two highest elements to memory.
	_mm_storeu_ps(uarr2, _mm_set1_ps(100.f)); _mm_storel_pi((__m64*)uarr2, a); aeq(_mm_loadu_ps(uarr2), 100.f, 100.f, 4.f, 2.f); // _mm_storel_pi: Store two lowest elements to memory.
	_mm_storer_ps(arr2, a); aeq(_mm_load_ps(arr2), 2.f, 4.f, 6.f, 8.f); // _mm_storer_ps: 4-wide store to aligned memory address, but reverse the elements on output.
	_mm_storeu_ps(uarr2, a); aeq(_mm_loadu_ps(uarr2), 8.f, 6.f, 4.f, 2.f); // _mm_storeu_ps: 4-wide store to unaligned memory address.
#ifdef TEST_M64
	/*M64*/_mm_stream_pi((__m64*)uarr, u64castm64(0x0080FF7F01FEFF40ULL)); Assert(*(uint64_t*)uarr == 0x0080FF7F01FEFF40ULL); // _mm_stream_pi: 2-wide store, but with a non-temporal memory cache hint.
#endif
	_mm_store_ps(arr2, _mm_set1_ps(100.f)); _mm_stream_ps(arr2, a); aeq(_mm_load_ps(arr2), 8.f, 6.f, 4.f, 2.f); // _mm_stream_ps: 4-wide store, but with a non-temporal memory cache hint.

	// SSE1 Arithmetic instructions:
	aeq(_mm_add_ps(a, b), 9.f, 8.f, 7.f, 6.f); // 4-wide add.
	aeq(_mm_add_ss(a, b), 8.f, 6.f, 4.f, 6.f); // Add lowest element, preserve three highest unchanged from a.
	aeq(_mm_div_ps(a, _mm_set_ps(2.f, 3.f, 8.f, 2.f)), 4.f, 2.f, 0.5f, 1.f); // 4-wide div.
	aeq(_mm_div_ss(a, _mm_set_ps(2.f, 3.f, 8.f, 8.f)), 8.f, 6.f, 4.f, 0.25f); // Div lowest element, preserve three highest unchanged from a.
	aeq(_mm_mul_ps(a, b), 8.f, 12.f, 12.f, 8.f); // 4-wide mul.
	aeq(_mm_mul_ss(a, b), 8.f, 6.f, 4.f, 8.f); // Mul lowest element, preserve three highest unchanged from a.
#ifdef TEST_M64
	__m64 m1 = get_m1();
	/*M64*/aeq64(_mm_mulhi_pu16(m1, u64castm64(0x22446688AACCEEFFULL)), 0x002233440B4C33CFULL); // Multiply u16 channels, and store high parts.
	/*M64*/aeq64(    _m_pmulhuw(m1, u64castm64(0x22446688AACCEEFFULL)), 0x002233440B4C33CFULL); // _m_pmulhuw is an alias to _mm_mulhi_pu16.
	__m64 m2 = get_m2();
	/*M64*/aeq64(_mm_sad_pu8(m1, m2), 0x368ULL); // Compute abs. differences of u8 channels, and sum those up to a single 16-bit scalar.
	/*M64*/aeq64(  _m_psadbw(m1, m2), 0x368ULL); // _m_psadbw is an alias to _mm_sad_pu8.
#endif
	aeq(_mm_sub_ps(a, b), 7.f, 4.f, 1.f, -2.f); // 4-wide sub.
	aeq(_mm_sub_ss(a, b), 8.f, 6.f, 4.f, -2.f); // Sub lowest element, preserve three highest unchanged from a.

	// SSE1 Elementary Math functions:
#ifndef __EMSCRIPTEN__ // TODO: Enable support for this to pass.
	aeq(_mm_rcp_ps(a), 0.124969f, 0.166626f, 0.249939f, 0.499878f); // Compute 4-wide 1/x.
	aeq(_mm_rcp_ss(a), 8.f, 6.f, 4.f, 0.499878f); // Compute 1/x of lowest element, pass higher elements unchanged.
	aeq(_mm_rsqrt_ps(a), 0.353455f, 0.408203f, 0.499878f, 0.706909f); // Compute 4-wide 1/sqrt(x).
	aeq(_mm_rsqrt_ss(a), 8.f, 6.f, 4.f, 0.706909f); // Compute 1/sqrt(x) of lowest element, pass higher elements unchanged.
#endif
	aeq(_mm_sqrt_ps(a), 2.82843f, 2.44949f, 2.f, 1.41421f); // Compute 4-wide sqrt(x).
	aeq(_mm_sqrt_ss(a), 8.f, 6.f, 4.f, 1.41421f); // Compute sqrt(x) of lowest element, pass higher elements unchanged.

	__m128 i1 = get_i1();
	__m128 i2 = get_i2();

	// SSE1 Logical instructions:
#ifndef __EMSCRIPTEN__ // TODO: The polyfill currently does NaN canonicalization and breaks these.
	aeqi(_mm_and_ps(i1, i2), 0x83200100, 0x0fecc988, 0x80244021, 0x13458a88); // 4-wide binary AND
	aeqi(_mm_andnot_ps(i1, i2), 0x388a9888, 0xf0021444, 0x7000289c, 0x00121046); // 4-wide binary (!i1) & i2
	aeqi(_mm_or_ps(i1, i2), 0xbfefdba9, 0xffefdfed, 0xf7656bbd, 0xffffdbef); // 4-wide binary OR
	aeqi(_mm_xor_ps(i1, i2), 0x3ccfdaa9, 0xf0031665, 0x77412b9c, 0xecba5167); // 4-wide binary XOR
#endif

	// SSE1 Compare instructions:
	// a = [8, 6, 4, 2], b = [1, 2, 3, 4]
	aeqi(_mm_cmpeq_ps(a, _mm_set_ps(8.f, 0.f, 4.f, 0.f)), 0xFFFFFFFF, 0, 0xFFFFFFFF, 0); // 4-wide cmp ==
	aeqi(_mm_cmpeq_ss(a, _mm_set_ps(8.f, 0.f, 4.f, 2.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0xFFFFFFFF); // scalar cmp ==, pass three highest unchanged.
	aeqi(_mm_cmpge_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0xFFFFFFFF, 0, 0xFFFFFFFF, 0); // 4-wide cmp >=
	aeqi(_mm_cmpge_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 0.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0xFFFFFFFF); // scalar cmp >=, pass three highest unchanged.
	aeqi(_mm_cmpgt_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0, 0, 0xFFFFFFFF, 0); // 4-wide cmp >
	aeqi(_mm_cmpgt_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 2.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0); // scalar cmp >, pass three highest unchanged.
	aeqi(_mm_cmple_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0xFFFFFFFF, 0xFFFFFFFF, 0, 0xFFFFFFFF); // 4-wide cmp <=
	aeqi(_mm_cmple_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 0.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0); // scalar cmp <=, pass three highest unchanged.
	aeqi(_mm_cmplt_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0, 0xFFFFFFFF, 0, 0xFFFFFFFF); // 4-wide cmp <
	aeqi(_mm_cmplt_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 2.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0); // scalar cmp <, pass three highest unchanged.
	aeqi(_mm_cmpneq_ps(a, _mm_set_ps(8.f, 0.f, 4.f, 0.f)), 0, 0xFFFFFFFF, 0, 0xFFFFFFFF); // 4-wide cmp !=
	aeqi(_mm_cmpneq_ss(a, _mm_set_ps(8.f, 0.f, 4.f, 2.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0); // scalar cmp !=, pass three highest unchanged.
	aeqi(_mm_cmpnge_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0, 0xFFFFFFFF, 0, 0xFFFFFFFF); // 4-wide cmp not >=
	aeqi(_mm_cmpnge_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 0.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0); // scalar cmp not >=, pass three highest unchanged.
	aeqi(_mm_cmpngt_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0xFFFFFFFF, 0xFFFFFFFF, 0, 0xFFFFFFFF); // 4-wide cmp not >
	aeqi(_mm_cmpngt_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 2.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0xFFFFFFFF); // scalar cmp not >, pass three highest unchanged.
	aeqi(_mm_cmpnle_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0, 0, 0xFFFFFFFF, 0); // 4-wide cmp not <=
	aeqi(_mm_cmpnle_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 0.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0xFFFFFFFF); // scalar cmp not <=, pass three highest unchanged.
	aeqi(_mm_cmpnlt_ps(a, _mm_set_ps(8.f, 7.f, 3.f, 5.f)), 0xFFFFFFFF, 0, 0xFFFFFFFF, 0); // 4-wide cmp not <
	aeqi(_mm_cmpnlt_ss(a, _mm_set_ps(8.f, 7.f, 3.f, 2.f)), fcastu(8.f), fcastu(6.f), fcastu(4.f), 0xFFFFFFFF); // scalar cmp not <, pass three highest unchanged.

	__m128 nan1 = get_nan1(); // [NAN, 0, 0, NAN]
	__m128 nan2 = get_nan2(); // [NAN, NAN, 0, 0]
	aeqi(_mm_cmpord_ps(nan1, nan2), 0, 0, 0xFFFFFFFF, 0); // 4-wide test if both operands are not nan.
	aeqi(_mm_cmpord_ss(nan1, nan2), fcastu(NAN), 0, 0, 0); // scalar test if both operands are not nan, pass three highest unchanged.
	// Intel Intrinsics Guide documentation is wrong on _mm_cmpunord_ps and _mm_cmpunord_ss. MSDN is right: http://msdn.microsoft.com/en-us/library/khy6fk1t(v=vs.90).aspx
	aeqi(_mm_cmpunord_ps(nan1, nan2), 0xFFFFFFFF, 0xFFFFFFFF, 0, 0xFFFFFFFF); // 4-wide test if one of the operands is nan.
#ifndef __EMSCRIPTEN__ // TODO: The polyfill currently does NaN canonicalization and breaks these.
	aeqi(_mm_cmpunord_ss(nan1, nan2), fcastu(NAN), 0, 0, 0xFFFFFFFF); // scalar test if one of the operands is nan, pass three highest unchanged.
#endif

	Assert(_mm_comieq_ss(a, b) == 0); Assert(_mm_comieq_ss(a, a) == 1); // Scalar cmp == of lowest element, return int.
	Assert(_mm_comige_ss(a, b) == 0); Assert(_mm_comige_ss(a, a) == 1); // Scalar cmp >= of lowest element, return int.
	Assert(_mm_comigt_ss(b, a) == 1); Assert(_mm_comigt_ss(a, a) == 0); // Scalar cmp > of lowest element, return int.
	Assert(_mm_comile_ss(b, a) == 0); Assert(_mm_comile_ss(a, a) == 1); // Scalar cmp <= of lowest element, return int.
	Assert(_mm_comilt_ss(a, b) == 1); Assert(_mm_comilt_ss(a, a) == 0); // Scalar cmp < of lowest element, return int.
	Assert(_mm_comineq_ss(a, b) == 1); Assert(_mm_comineq_ss(a, a) == 0); // Scalar cmp != of lowest element, return int.

	// The ucomi versions are identical to comi, except that ucomi signal a FP exception only if one of the input operands is a SNaN, whereas the comi versions signal a FP
	// exception when one of the input operands is either a QNaN or a SNaN.
#ifndef __EMSCRIPTEN__ // TODO: Fix ucomi support in SSE to treat NaNs properly.
	Assert(_mm_ucomieq_ss(a, b) == 0); Assert(_mm_ucomieq_ss(a, a) == 1); Assert(_mm_ucomieq_ss(a, nan1) == 1);
#endif
	Assert(_mm_ucomige_ss(a, b) == 0); Assert(_mm_ucomige_ss(a, a) == 1); Assert(_mm_ucomige_ss(a, nan1) == 0);
	Assert(_mm_ucomigt_ss(b, a) == 1); Assert(_mm_ucomigt_ss(a, a) == 0); Assert(_mm_ucomigt_ss(a, nan1) == 0);
	Assert(_mm_ucomile_ss(b, a) == 0); Assert(_mm_ucomile_ss(a, a) == 1); Assert(_mm_ucomile_ss(a, nan1) == 1);
	Assert(_mm_ucomilt_ss(a, b) == 1); Assert(_mm_ucomilt_ss(a, a) == 0); Assert(_mm_ucomilt_ss(a, nan1) == 1);
#ifndef __EMSCRIPTEN__ // TODO: Fix ucomi support in SSE to treat NaNs properly.
	Assert(_mm_ucomineq_ss(a, b) == 1); Assert(_mm_ucomineq_ss(a, a) == 0); Assert(_mm_ucomineq_ss(a, nan1) == 0);
#endif

	// SSE1 Convert instructions:
	__m128 c = get_c(); // [1.5, 2.5, 3.5, 4.5]
	__m128 e = get_e(); // [INF, -INF, 2.5, 3.5]
	__m128 f = get_f(); // [-1.5, 1.5, -2.5, -9223372036854775808]
#ifdef TEST_M64
	/*M64*/aeq(_mm_cvt_pi2ps(a, m2), 8.f, 6.f, -19088744.f, 1985229312.f); // 2-way int32 to float conversion to two lowest channels of m128.
	/*M64*/aeq64(_mm_cvt_ps2pi(c), 0x400000004ULL); // 2-way two lowest floats from m128 to integer, return as m64.
#endif
	aeq(_mm_cvtsi32_ss(c, -16777215), 1.5f, 2.5f, 3.5f, -16777215.f); // Convert int to float, store in lowest channel of m128.
	aeq( _mm_cvt_si2ss(c, -16777215), 1.5f, 2.5f, 3.5f, -16777215.f); // _mm_cvt_si2ss is an alias to _mm_cvtsi32_ss.
#ifndef __EMSCRIPTEN__ // TODO: Fix banker's rounding in cvt functions.
	Assert(_mm_cvtss_si32(c) == 4); Assert(_mm_cvtss_si32(e) == 4); // Convert lowest channel of m128 from float to int.
	Assert( _mm_cvt_ss2si(c) == 4); Assert( _mm_cvt_ss2si(e) == 4); // _mm_cvt_ss2si is an alias to _mm_cvtss_si32.
#endif
#ifdef TEST_M64
	/*M64*/aeq(_mm_cvtpi16_ps(m1), 255.f , -32767.f, 4336.f, 14207.f); // 4-way convert int16s to floats, return in a m128.
	/*M64*/aeq(_mm_cvtpi32_ps(a, m1), 8.f, 6.f, 16744449.f, 284178304.f); // 2-way convert int32s to floats, return in two lowest channels of m128, pass two highest unchanged.
	/*M64*/aeq(_mm_cvtpi32x2_ps(m1, m2), -19088744.f, 1985229312.f, 16744449.f, 284178304.f); // 4-way convert int32s from two different m64s to float.
	/*M64*/aeq(_mm_cvtpi8_ps(m1), 16.f, -16.f, 55.f, 127.f); // 4-way convert int8s from lowest end of m64 to float in a m128.
	/*M64*/aeq64(_mm_cvtps_pi16(c), 0x0002000200040004ULL); // 4-way convert floats to int16s in a m64.
	/*M64*/aeq64(_mm_cvtps_pi32(c), 0x0000000400000004ULL); // 2-way convert two lowest floats to int32s in a m64.
	/*M64*/aeq64(_mm_cvtps_pi8(c),  0x0000000002020404ULL); // 4-way convert floats to int8s in a m64, zero higher half of the returned m64.
	/*M64*/aeq(_mm_cvtpu16_ps(m1), 255.f , 32769.f, 4336.f, 14207.f); // 4-way convert uint16s to floats, return in a m128.
	/*M64*/aeq(_mm_cvtpu8_ps(m1), 16.f, 240.f, 55.f, 127.f); // 4-way convert uint8s from lowest end of m64 to float in a m128.
#endif
	aeq(_mm_cvtsi64_ss(c, -9223372036854775808ULL), 1.5f, 2.5f, 3.5f, -9223372036854775808.f); // Convert single int64 to float, store in lowest channel of m128, and pass three higher channel unchanged.
	Assert(_mm_cvtss_f32(c) == 4.5f); // Extract lowest channel of m128 to a plain old float.
	Assert(_mm_cvtss_si64(f) == -9223372036854775808ULL); // Convert lowest channel of m128 from float to int64.
#ifdef TEST_M64
	/*M64*/aeq64(_mm_cvtt_ps2pi(e), 0x0000000200000003ULL); aeq64(_mm_cvtt_ps2pi(f), 0xfffffffe80000000ULL); // Truncating conversion from two lowest floats of m128 to int32s, return in a m64.
#endif
	Assert(_mm_cvttss_si32(e) == 3); // Truncating conversion from the lowest float of a m128 to int32.
	Assert( _mm_cvtt_ss2si(e) == 3); // _mm_cvtt_ss2si is an alias to _mm_cvttss_si32.
#ifdef TEST_M64
	/*M64*/aeq64(_mm_cvttps_pi32(c), 0x0000000300000004ULL); // Truncating conversion from two lowest floats of m128 to m64.
#endif
	Assert(_mm_cvttss_si64(f) == -9223372036854775808ULL); // Truncating conversion from lowest channel of m128 from float to int64.

#ifndef __EMSCRIPTEN__ // TODO: Not implemented.
	// SSE1 General support:
	unsigned int mask = _MM_GET_EXCEPTION_MASK();
	_MM_SET_EXCEPTION_MASK(mask);
	unsigned int flushZeroMode = _MM_GET_FLUSH_ZERO_MODE();
	_MM_SET_FLUSH_ZERO_MODE(flushZeroMode);
	unsigned int roundingMode = _MM_GET_ROUNDING_MODE();
	_MM_SET_ROUNDING_MODE(roundingMode);
	unsigned int csr = _mm_getcsr();
	_mm_setcsr(csr);
	unsigned char dummyData[4096];
	_mm_prefetch(dummyData, _MM_HINT_T0);
	_mm_prefetch(dummyData, _MM_HINT_T1);
	_mm_prefetch(dummyData, _MM_HINT_T2);
	_mm_prefetch(dummyData, _MM_HINT_NTA);
	_mm_sfence();
#endif

	// SSE1 Misc instructions:
#ifdef TEST_M64
	/*M64*/Assert(_mm_movemask_pi8(m1) == 100); // Return int with eight lowest bits set depending on the highest bits of the 8 uint8 input channels of the m64.
	/*M64*/Assert(     _m_pmovmskb(m1) == 100); // _m_pmovmskb is an alias to _mm_movemask_pi8.
#endif
	Assert(_mm_movemask_ps(_mm_set_ps(-1.f, 0.f, 1.f, NAN)) == 8); Assert(_mm_movemask_ps(_mm_set_ps(-INFINITY, -0.f, INFINITY, -INFINITY)) == 13); // Return int with four lowest bits set depending on the highest bits of the 4 m128 input channels.

	// SSE1 Probability/Statistics instructions:
#ifdef TEST_M64
	/*M64*/aeq64(_mm_avg_pu16(m1, m2), 0x7FEE9D4D43A234C8ULL); // 4-way average uint16s.
	/*M64*/aeq64(    _m_pavgw(m1, m2), 0x7FEE9D4D43A234C8ULL); // _m_pavgw is an alias to _mm_avg_pu16.
	/*M64*/aeq64(_mm_avg_pu8(m1, m2),  0x7FEE9D4D43A23548ULL); // 8-way average uint8s.
	/*M64*/aeq64(   _m_pavgb(m1, m2),  0x7FEE9D4D43A23548ULL); // _m_pavgb is an alias to _mm_avg_pu8.

	// SSE1 Special Math instructions:
	/*M64*/aeq64(_mm_max_pi16(m1, m2), 0xFFBA987654377FULL); // 4-way average uint16s.
	/*M64*/aeq64(   _m_pmaxsw(m1, m2), 0xFFBA987654377FULL); // _m_pmaxsw is an alias to _mm_max_pi16.
	/*M64*/aeq64(_mm_max_pu8(m1, m2), 0xFEFFBA9876F0377FULL); // 4-way average uint16s.
	/*M64*/aeq64(  _m_pmaxub(m1, m2), 0xFEFFBA9876F0377FULL); // _m_pmaxub is an alias to _mm_max_pu8.
	/*M64*/aeq64(_mm_min_pi16(m1, m2), 0xFEDC800110F03210ULL); // 4-way average uint16s.
	/*M64*/aeq64(   _m_pminsw(m1, m2), 0xFEDC800110F03210ULL); // is an alias to _mm_min_pi16.
	/*M64*/aeq64(_mm_min_pu8(m1, m2), 0xDC800110543210ULL); // 4-way average uint16s.
	/*M64*/aeq64(  _m_pminub(m1, m2), 0xDC800110543210ULL); // is an alias to _mm_min_pu8.
#endif
	// a = [8, 6, 4, 2], b = [1, 2, 3, 4]
	aeq(_mm_max_ps(a, b), 8.f, 6.f, 4.f, 4.f); // 4-wide max.
	aeq(_mm_max_ss(a, _mm_set1_ps(100.f)), 8.f, 6.f, 4.f, 100.f); // Scalar max, pass three highest unchanged.
	aeq(_mm_min_ps(a, b), 1.f, 2.f, 3.f, 2.f); // 4-wide min.
	aeq(_mm_min_ss(a, _mm_set1_ps(-100.f)), 8.f, 6.f, 4.f, -100.f); // Scalar min, pass three highest unchanged.

	// SSE1 Swizzle instructions:
#ifdef TEST_M64
	/*M64*/Assert(_mm_extract_pi16(m1, 1) == 4336); // Extract the given int16 channel from a m64.
	/*M64*/Assert(       _m_pextrw(m1, 1) == 4336); // _m_pextrw is an alias to _mm_extract_pi16.
	/*M64*/aeq64(_mm_insert_pi16(m1, 0xABCD, 1), 0xFF8001ABCD377FULL); // Insert a int16 to a specific channel of a m64.
	/*M64*/aeq64(      _m_pinsrw(m1, 0xABCD, 1), 0xFF8001ABCD377FULL); // _m_pinsrw is an alias to _mm_insert_pi16.
	/*M64*/aeq64(_mm_shuffle_pi16(m1, _MM_SHUFFLE(1, 0, 3, 2)), 0x10F0377F00FF8001ULL); // Shuffle int16s around in the 4 channels of the m64.
	/*M64*/aeq64(       _m_pshufw(m1, _MM_SHUFFLE(1, 0, 3, 2)), 0x10F0377F00FF8001ULL); // _m_pshufw is an alias to _mm_shuffle_pi16.
#endif
	aeq(_mm_shuffle_ps(a, b, _MM_SHUFFLE(1, 0, 3, 2)), 3.f, 4.f, 8.f, 6.f);
	aeq(_mm_unpackhi_ps(a, b), 1.f , 8.f, 2.f, 6.f);
	aeq(_mm_unpacklo_ps(a, b), 3.f , 4.f, 4.f, 2.f);

	// Transposing a matrix via the xmmintrin.h-provided intrinsic.
	__m128 c0 = a; // [8, 6, 4, 2]
	__m128 c1 = b; // [1, 2, 3, 4]
	__m128 c2 = get_c(); // [1.5, 2.5, 3.5, 4.5]
	__m128 c3 = get_d(); // [8.5, 6.5, 4.5, 2.5]
	_MM_TRANSPOSE4_PS(c0, c1, c2, c3);
	aeq(c0, 2.5f, 4.5f, 4.f, 2.f);
	aeq(c1, 4.5f, 3.5f, 3.f, 4.f);
	aeq(c2, 6.5f, 2.5f, 2.f, 6.f);
	aeq(c3, 8.5f, 1.5f, 1.f, 8.f);

	// All done!
	if (numFailures == 0)
		printf("Success!\n");
	else
		printf("%d tests failed!\n", numFailures);
}
Esempio n. 6
0
inline float4 lt(const float4& a, const float4& b)
{
	return float4(_mm_cmplt_ps(a.data, b.data));
}
Esempio n. 7
0
/*
====================
R_OverlayPointCullSkinned
====================
*/
static void R_OverlayPointCullSkinned( byte* cullBits, halfFloat_t* texCoordS, halfFloat_t* texCoordT, const idPlane* planes, const idDrawVert* verts, const int numVerts, const idJointMat* joints )
{
	assert_16_byte_aligned( cullBits );
	assert_16_byte_aligned( texCoordS );
	assert_16_byte_aligned( texCoordT );
	assert_16_byte_aligned( verts );
	
#if defined(USE_INTRINSICS)
	idODSStreamedArray< idDrawVert, 16, SBT_DOUBLE, 4 > vertsODS( verts, numVerts );
	
	const __m128 vector_float_zero	= { 0.0f, 0.0f, 0.0f, 0.0f };
	const __m128 vector_float_one	= { 1.0f, 1.0f, 1.0f, 1.0f };
	const __m128i vector_int_mask0	= _mm_set1_epi32( 1 << 0 );
	const __m128i vector_int_mask1	= _mm_set1_epi32( 1 << 1 );
	const __m128i vector_int_mask2	= _mm_set1_epi32( 1 << 2 );
	const __m128i vector_int_mask3	= _mm_set1_epi32( 1 << 3 );
	
	const __m128 p0 = _mm_loadu_ps( planes[0].ToFloatPtr() );
	const __m128 p1 = _mm_loadu_ps( planes[1].ToFloatPtr() );
	
	const __m128 p0X = _mm_splat_ps( p0, 0 );
	const __m128 p0Y = _mm_splat_ps( p0, 1 );
	const __m128 p0Z = _mm_splat_ps( p0, 2 );
	const __m128 p0W = _mm_splat_ps( p0, 3 );
	
	const __m128 p1X = _mm_splat_ps( p1, 0 );
	const __m128 p1Y = _mm_splat_ps( p1, 1 );
	const __m128 p1Z = _mm_splat_ps( p1, 2 );
	const __m128 p1W = _mm_splat_ps( p1, 3 );
	
	for( int i = 0; i < numVerts; )
	{
	
		const int nextNumVerts = vertsODS.FetchNextBatch() - 4;
		
		for( ; i <= nextNumVerts; i += 4 )
		{
			const __m128 v0 = LoadSkinnedDrawVertPosition( vertsODS[i + 0], joints );
			const __m128 v1 = LoadSkinnedDrawVertPosition( vertsODS[i + 1], joints );
			const __m128 v2 = LoadSkinnedDrawVertPosition( vertsODS[i + 2], joints );
			const __m128 v3 = LoadSkinnedDrawVertPosition( vertsODS[i + 3], joints );
			
			const __m128 r0 = _mm_unpacklo_ps( v0, v2 );	// v0.x, v2.x, v0.z, v2.z
			const __m128 r1 = _mm_unpackhi_ps( v0, v2 );	// v0.y, v2.y, v0.w, v2.w
			const __m128 r2 = _mm_unpacklo_ps( v1, v3 );	// v1.x, v3.x, v1.z, v3.z
			const __m128 r3 = _mm_unpackhi_ps( v1, v3 );	// v1.y, v3.y, v1.w, v3.w
			
			const __m128 vX = _mm_unpacklo_ps( r0, r2 );	// v0.x, v1.x, v2.x, v3.x
			const __m128 vY = _mm_unpackhi_ps( r0, r2 );	// v0.y, v1.y, v2.y, v3.y
			const __m128 vZ = _mm_unpacklo_ps( r1, r3 );	// v0.z, v1.z, v2.z, v3.z
			
			const __m128 d0 = _mm_madd_ps( vX, p0X, _mm_madd_ps( vY, p0Y, _mm_madd_ps( vZ, p0Z, p0W ) ) );
			const __m128 d1 = _mm_madd_ps( vX, p1X, _mm_madd_ps( vY, p1Y, _mm_madd_ps( vZ, p1Z, p1W ) ) );
			const __m128 d2 = _mm_sub_ps( vector_float_one, d0 );
			const __m128 d3 = _mm_sub_ps( vector_float_one, d1 );
			
			__m128i flt16S = FastF32toF16( __m128c( d0 ) );
			__m128i flt16T = FastF32toF16( __m128c( d1 ) );
			
			_mm_storel_epi64( ( __m128i* )&texCoordS[i], flt16S );
			_mm_storel_epi64( ( __m128i* )&texCoordT[i], flt16T );
			
			__m128i c0 = __m128c( _mm_cmplt_ps( d0, vector_float_zero ) );
			__m128i c1 = __m128c( _mm_cmplt_ps( d1, vector_float_zero ) );
			__m128i c2 = __m128c( _mm_cmplt_ps( d2, vector_float_zero ) );
			__m128i c3 = __m128c( _mm_cmplt_ps( d3, vector_float_zero ) );
			
			c0 = _mm_and_si128( c0, vector_int_mask0 );
			c1 = _mm_and_si128( c1, vector_int_mask1 );
			c2 = _mm_and_si128( c2, vector_int_mask2 );
			c3 = _mm_and_si128( c3, vector_int_mask3 );
			
			c0 = _mm_or_si128( c0, c1 );
			c2 = _mm_or_si128( c2, c3 );
			c0 = _mm_or_si128( c0, c2 );
			
			c0 = _mm_packs_epi32( c0, c0 );
			c0 = _mm_packus_epi16( c0, c0 );
			
			*( unsigned int* )&cullBits[i] = _mm_cvtsi128_si32( c0 );
		}
	}
	
#else
	
	idODSStreamedArray< idDrawVert, 16, SBT_DOUBLE, 1 > vertsODS( verts, numVerts );
	
	for( int i = 0; i < numVerts; )
	{
	
		const int nextNumVerts = vertsODS.FetchNextBatch() - 1;
	
		for( ; i <= nextNumVerts; i++ )
		{
			const idVec3 transformed = Scalar_LoadSkinnedDrawVertPosition( vertsODS[i], joints );
	
			const float d0 = planes[0].Distance( transformed );
			const float d1 = planes[1].Distance( transformed );
			const float d2 = 1.0f - d0;
			const float d3 = 1.0f - d1;
	
			halfFloat_t s = Scalar_FastF32toF16( d0 );
			halfFloat_t t = Scalar_FastF32toF16( d1 );
	
			texCoordS[i] = s;
			texCoordT[i] = t;
	
			byte bits;
			bits  = IEEE_FLT_SIGNBITSET( d0 ) << 0;
			bits |= IEEE_FLT_SIGNBITSET( d1 ) << 1;
			bits |= IEEE_FLT_SIGNBITSET( d2 ) << 2;
			bits |= IEEE_FLT_SIGNBITSET( d3 ) << 3;
	
			cullBits[i] = bits;
		}
	}
	
#endif
}
Esempio n. 8
0
// vectorize the below code using SIMD intrinsics
int *
mandelbrot_vector(float x[SIZE], float y[SIZE]) {
	/*static int ret[SIZE];
	float x1, y1, x2, y2;

	for (int i = 0 ; i < SIZE ; i ++) {
		x1 = y1 = 0.0;

		// Run M_ITER iterations
		for (int j = 0 ; j < M_ITER ; j ++) {
			// Calculate the real part of (x1 + y1 * i)^2 + (x + y * i)
			x2 = (x1 * x1) - (y1 * y1) + x[i];

			// Calculate the imaginary part of (x1 + y1 * i)^2 + (x + y * i)
			y2 = 2 * (x1 * y1) + y[i];

			// Use the new complex number as input for the next iteration
			x1 = x2;
			y1 = y2;
		}

		// caculate the magnitude of the result
		// We could take the square root, but instead we just
		// compare squares
		ret[i] = ((x2 * x2) + (y2 * y2)) < (M_MAG * M_MAG);
	}

	return ret;*/
	static int ret[SIZE];
	float* retf=(float*)ret;
	float x1, y1, x2, y2;
	__m128 X, Y, X1, Y1, X2, Y2;
	__m128 two;
	two=_mm_set1_ps(2.0);
	__m128 mag;
	mag=_mm_set1_ps(M_MAG);
	for (int i = 0 ; i < SIZE-3 ; i +=4) {
		__m128 temp;
		x1 = y1 = 0.0;
		X=_mm_loadu_ps(&x[i]);
		Y=_mm_loadu_ps(&y[i]);
		X1=_mm_set1_ps(x1);
		Y1=_mm_set1_ps(y1);
		X2=_mm_set1_ps(x2);
		Y2=_mm_set1_ps(y2);		

		for (int j = 0 ; j < M_ITER ; j ++) {
			X2=_mm_add_ps(_mm_sub_ps(_mm_mul_ps(X1,X1),_mm_mul_ps(Y1,Y1)),X);

			Y2=_mm_add_ps(_mm_mul_ps(_mm_mul_ps(X1, Y1),two),Y);

			X1=X2;
			Y1=Y2;
		}

		temp=_mm_cmplt_ps(_mm_add_ps(_mm_mul_ps(X2, X2),_mm_mul_ps(Y2,Y2)),_mm_mul_ps(mag,mag));
		_mm_storeu_ps(&retf[i],temp);
	}

	return ret;
}
Esempio n. 9
0
/* natural logarithm computed for 4 simultaneous float
   return NaN for x <= 0
*/
__m128 log_ps(__m128 x) {
    typedef __m128 v4sf;
    typedef __m128i v4si;

    v4si emm0;
    v4sf one = constants::ps_1.ps;

    v4sf invalid_mask = _mm_cmple_ps(x, _mm_setzero_ps());

    x = _mm_max_ps(x, constants::min_norm_pos.ps);  // cut off denormalized stuff

    emm0 = _mm_srli_epi32(_mm_castps_si128(x), 23);
    // keep only the fractional part
    x = _mm_and_ps(x, constants::inv_mant_mask.ps);
    x = _mm_or_ps(x,  constants::ps_0p5.ps);

    emm0 = _mm_sub_epi32(emm0, constants::pi32_0x7f.pi);
    v4sf e = _mm_cvtepi32_ps(emm0);

    e = _mm_add_ps(e, one);

    /* part2:
       if( x < SQRTHF ) {
         e -= 1;
         x = x + x - 1.0;
       } else { x = x - 1.0; }
    */
    v4sf mask = _mm_cmplt_ps(x, constants::cephes_SQRTHF.ps);
    v4sf tmp = _mm_and_ps(x, mask);
    x = _mm_sub_ps(x, one);
    e = _mm_sub_ps(e, _mm_and_ps(one, mask));
    x = _mm_add_ps(x, tmp);

    v4sf z = _mm_mul_ps(x,x);

    v4sf y = constants::cephes_log_p0.ps;
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p1.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p2.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p3.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p4.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p5.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p6.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p7.ps);
    y = _mm_mul_ps(y, x);
    y = _mm_add_ps(y, constants::cephes_log_p8.ps);
    y = _mm_mul_ps(y, x);

    y = _mm_mul_ps(y, z);

    tmp = _mm_mul_ps(e, constants::cephes_log_q1.ps);
    y = _mm_add_ps(y, tmp);

    tmp = _mm_mul_ps(z, constants::ps_0p5.ps);
    y = _mm_sub_ps(y, tmp);

    tmp = _mm_mul_ps(e, constants::cephes_log_q2.ps);
    x = _mm_add_ps(x, y);
    x = _mm_add_ps(x, tmp);
    x = _mm_or_ps(x, invalid_mask); // negative arg will be NAN
    return x;
}
Esempio n. 10
0
__m128 test_mm_cmplt_ps(__m128 __a, __m128 __b) {
  // CHECK-LABEL: @test_mm_cmplt_ps
  // CHECK: @llvm.x86.sse.cmp.ps(<4 x float> %{{.*}}, <4 x float> %{{.*}}, i8 1)
  return _mm_cmplt_ps(__a, __b);
}
//----------------------------------------------------------------
// Transforms the AABB vertices to screen space once every frame
// Also performs a coarse depth pre-test
//----------------------------------------------------------------
PreTestResult TransformedAABBoxAVX::TransformAndPreTestAABBox(__m128 xformedPos[], const __m128 cumulativeMatrix[4], const float *pDepthSummary)
{
	// w ends up being garbage, but it doesn't matter - we ignore it anyway.
	__m128 vCenter = _mm_loadu_ps(&mBBCenter.x);
	__m128 vHalf   = _mm_loadu_ps(&mBBHalf.x);

	__m128 vMin    = _mm_sub_ps(vCenter, vHalf);
	__m128 vMax    = _mm_add_ps(vCenter, vHalf);

	// transforms
	__m128 xRow[2], yRow[2], zRow[2];
	xRow[0] = _mm_shuffle_ps(vMin, vMin, 0x00) * cumulativeMatrix[0];
	xRow[1] = _mm_shuffle_ps(vMax, vMax, 0x00) * cumulativeMatrix[0];
	yRow[0] = _mm_shuffle_ps(vMin, vMin, 0x55) * cumulativeMatrix[1];
	yRow[1] = _mm_shuffle_ps(vMax, vMax, 0x55) * cumulativeMatrix[1];
	zRow[0] = _mm_shuffle_ps(vMin, vMin, 0xaa) * cumulativeMatrix[2];
	zRow[1] = _mm_shuffle_ps(vMax, vMax, 0xaa) * cumulativeMatrix[2];

	__m128 zAllIn = _mm_castsi128_ps(_mm_set1_epi32(~0));
	__m128 screenMin = _mm_set1_ps(FLT_MAX);
	__m128 screenMax = _mm_set1_ps(-FLT_MAX);

	for(UINT i = 0; i < AABB_VERTICES; i++)
	{
		// Transform the vertex
		__m128 vert = cumulativeMatrix[3];
		vert += xRow[sBBxInd[i]];
		vert += yRow[sBByInd[i]];
		vert += zRow[sBBzInd[i]];

		// We have inverted z; z is in front of near plane iff z <= w.
		__m128 vertZ = _mm_shuffle_ps(vert, vert, 0xaa); // vert.zzzz
		__m128 vertW = _mm_shuffle_ps(vert, vert, 0xff); // vert.wwww
		__m128 zIn = _mm_cmple_ps(vertZ, vertW);
		zAllIn = _mm_and_ps(zAllIn, zIn);

		// project
		xformedPos[i] = _mm_div_ps(vert, vertW);
		
	    // update bounds
	    screenMin = _mm_min_ps(screenMin, xformedPos[i]);
	    screenMax = _mm_max_ps(screenMax, xformedPos[i]);
	}

	// if any of the verts are z-clipped, we (conservatively) say the box is in
	if(_mm_movemask_ps(zAllIn) != 0xf)
		return ePT_VISIBLE;

	// Clip against screen bounds
	screenMin = _mm_max_ps(screenMin, _mm_setr_ps(0.0f, 0.0f, 0.0f, -FLT_MAX));
	screenMax = _mm_min_ps(screenMax, _mm_setr_ps((float) (SCREENW - 1), (float) (SCREENH - 1), 1.0f, FLT_MAX));

	// Quick rejection test
	if(_mm_movemask_ps(_mm_cmplt_ps(screenMax, screenMin)))
		return ePT_INVISIBLE;

	// Prepare integer bounds
	__m128 minMaxXY = _mm_shuffle_ps(screenMin, screenMax, 0x44); // minX,minY,maxX,maxY
	__m128i minMaxXYi = _mm_cvtps_epi32(minMaxXY);
	__m128i minMaxXYis = _mm_srai_epi32(minMaxXYi, 3);

	__m128 maxZ = _mm_shuffle_ps(screenMax, screenMax, 0xaa);

	// Traverse all 8x8 blocks covered by 2d screen-space BBox;
	// if we know for sure that this box is behind the geometry we know is there,
	// we can stop.
	int rX0 = minMaxXYis.m128i_i32[0];
	int rY0 = minMaxXYis.m128i_i32[1];
	int rX1 = minMaxXYis.m128i_i32[2];
	int rY1 = minMaxXYis.m128i_i32[3];

	__m128 anyCloser = _mm_setzero_ps();
	for(int by = rY0; by <= rY1; by++)
	{
		const float *srcRow = pDepthSummary + by * (SCREENW/BLOCK_SIZE);

		// If for any 8x8 block, maxZ is not less than (=behind) summarized
		// min Z, box might be visible.
		for(int bx = rX0; bx <= rX1; bx++)
		{
			anyCloser = _mm_or_ps(anyCloser, _mm_cmpnlt_ss(maxZ, _mm_load_ss(&srcRow[bx])));
		}

		if(_mm_movemask_ps(anyCloser))
		{
			return ePT_UNSURE; // okay, box might be in
		}
	}

	// If we get here, we know for sure that the box is fully behind the stuff in the
	// depth buffer.
	return ePT_INVISIBLE;
}
Esempio n. 12
0
bool AABB::IntersectLineAABB_SSE(const float4 &rayPos, const float4 &rayDir, float tNear, float tFar) const
{
	assume(rayDir.IsNormalized4());
	assume(tNear <= tFar && "AABB::IntersectLineAABB: User gave a degenerate line as input for the intersection test!");
	/* For reference, this is the C++ form of the vectorized SSE code below.

	float4 recipDir = rayDir.RecipFast4();
	float4 t1 = (aabbMinPoint - rayPos).Mul(recipDir);
	float4 t2 = (aabbMaxPoint - rayPos).Mul(recipDir);
	float4 near = t1.Min(t2);
	float4 far = t1.Max(t2);
	float4 rayDirAbs = rayDir.Abs();

	if (rayDirAbs.x > 1e-4f) // ray is parallel to plane in question
	{
		tNear = Max(near.x, tNear); // tNear tracks distance to intersect (enter) the AABB.
		tFar = Min(far.x, tFar); // tFar tracks the distance to exit the AABB.
	}
	else if (rayPos.x < aabbMinPoint.x || rayPos.x > aabbMaxPoint.x) // early-out if the ray can't possibly enter the box.
		return false;

	if (rayDirAbs.y > 1e-4f) // ray is parallel to plane in question
	{
		tNear = Max(near.y, tNear); // tNear tracks distance to intersect (enter) the AABB.
		tFar = Min(far.y, tFar); // tFar tracks the distance to exit the AABB.
	}
	else if (rayPos.y < aabbMinPoint.y || rayPos.y > aabbMaxPoint.y) // early-out if the ray can't possibly enter the box.
		return false;

	if (rayDirAbs.z > 1e-4f) // ray is parallel to plane in question
	{
		tNear = Max(near.z, tNear); // tNear tracks distance to intersect (enter) the AABB.
		tFar = Min(far.z, tFar); // tFar tracks the distance to exit the AABB.
	}
	else if (rayPos.z < aabbMinPoint.z || rayPos.z > aabbMaxPoint.z) // early-out if the ray can't possibly enter the box.
		return false;

	return tNear < tFar;
	*/

	__m128 recipDir = _mm_rcp_ps(rayDir.v);
	// Note: The above performs an approximate reciprocal (11 bits of precision).
	// For a full precision reciprocal, perform a div:
//	__m128 recipDir = _mm_div_ps(_mm_set1_ps(1.f), rayDir.v);

	__m128 t1 = _mm_mul_ps(_mm_sub_ps(MinPoint_SSE(), rayPos.v), recipDir);
	__m128 t2 = _mm_mul_ps(_mm_sub_ps(MaxPoint_SSE(), rayPos.v), recipDir);

	__m128 nearD = _mm_min_ps(t1, t2); // [0 n3 n2 n1]
	__m128 farD = _mm_max_ps(t1, t2);  // [0 f3 f2 f1]

	// Check if the ray direction is parallel to any of the cardinal axes, and if so,
	// mask those [near, far] ranges away from the hit test computations.
	__m128 rayDirAbs = abs_ps(rayDir.v);

	const __m128 epsilon = _mm_set1_ps(1e-4f);
	// zeroDirections[i] will be nonzero for each axis i the ray is parallel to.
	__m128 zeroDirections = _mm_cmple_ps(rayDirAbs, epsilon);

	const __m128 floatInf = _mm_set1_ps(FLOAT_INF);
	const __m128 floatNegInf = _mm_set1_ps(-FLOAT_INF);

	// If the ray is parallel to one of the axes, replace the slab range for that axis
	// with [-inf, inf] range instead. (which is a no-op in the comparisons below)
	nearD = cmov_ps(nearD, floatNegInf, zeroDirections);
	farD = cmov_ps(farD , floatInf, zeroDirections);

	// Next, we need to compute horizontally max(nearD[0], nearD[1], nearD[2]) and min(farD[0], farD[1], farD[2])
	// to see if there is an overlap in the hit ranges.
	__m128 v1 = _mm_shuffle_ps(nearD, farD, _MM_SHUFFLE(0, 0, 0, 0)); // [f1 f1 n1 n1]
	__m128 v2 = _mm_shuffle_ps(nearD, farD, _MM_SHUFFLE(1, 1, 1, 1)); // [f2 f2 n2 n2]
	__m128 v3 = _mm_shuffle_ps(nearD, farD, _MM_SHUFFLE(2, 2, 2, 2)); // [f3 f3 n3 n3]
	nearD = _mm_max_ps(v1, _mm_max_ps(v2, v3));
	farD = _mm_min_ps(v1, _mm_min_ps(v2, v3));
	farD = _mm_shuffle_ps(farD, farD, _MM_SHUFFLE(3, 3, 3, 3)); // Unpack the result from high offset in the register.
	nearD = _mm_max_ps(nearD, _mm_set_ss(tNear));
	farD = _mm_min_ps(farD, _mm_set_ss(tFar));

	// Finally, test if the ranges overlap.
	__m128 rangeIntersects = _mm_cmple_ss(nearD, farD);

	// To store out out the interval of intersection, uncomment the following:
	// These are disabled, since without these, the whole function runs without a single memory store,
	// which has been profiled to be very fast! Uncommenting these causes an order-of-magnitude slowdown.
	// For now, using the SSE version only where the tNear and tFar ranges are not interesting.
//	_mm_store_ss(&tNear, nearD);
//	_mm_store_ss(&tFar, farD);

	// To avoid false positives, need to have an additional rejection test for each cardinal axis the ray direction
	// is parallel to.
	__m128 out2 = _mm_cmplt_ps(rayPos.v, MinPoint_SSE());
	__m128 out3 = _mm_cmpgt_ps(rayPos.v, MaxPoint_SSE());
	out2 = _mm_or_ps(out2, out3);
	zeroDirections = _mm_and_ps(zeroDirections, out2);

	__m128 yOut = _mm_shuffle_ps(zeroDirections, zeroDirections, _MM_SHUFFLE(1,1,1,1));
	__m128 zOut = _mm_shuffle_ps(zeroDirections, zeroDirections, _MM_SHUFFLE(2,2,2,2));

	zeroDirections = _mm_or_ps(_mm_or_ps(zeroDirections, yOut), zOut);
	// Intersection occurs if the slab ranges had positive overlap and if the test was not rejected by the ray being
	// parallel to some cardinal axis.
	__m128 intersects = _mm_andnot_ps(zeroDirections, rangeIntersects);
	__m128 epsilonMasked = _mm_and_ps(epsilon, intersects);
	return _mm_comieq_ss(epsilon, epsilonMasked) != 0;
}
Esempio n. 13
0
inline static void bar(float(&inout)[8])
{
	__m128 leftSideElements[6],
		rightSideElements[6],
		leftGERight[6],
		leftLTRight[6],
		leftElementsGE[6],  // swaped elements on the left part of comparison
		leftElementsLT[6],  // not-swaped elements on the left part of comparison
		rightElementsGE[6], // swaped elements on the right part of comparison
		rightElementsLT[6]; // not-swaped elements on the right part of comparison
	float resultLeftElements[6][4], resultRightElements[6][4];

	const size_t idx[][2] = {
			{ 0, 1 }, { 3, 2 }, { 4, 5 }, { 7, 6 },
			{ 0, 2 }, { 1, 3 }, { 6, 4 }, { 7, 5 },
			{ 0, 1 }, { 2, 3 }, { 5, 4 }, { 7, 6 },
			{ 0, 4 }, { 1, 5 }, { 2, 6 }, { 3, 7 },
			{ 0, 2 }, { 1, 3 }, { 4, 6 }, { 5, 7 },
			{ 0, 1 }, { 2, 3 }, { 4, 5 }, { 6, 7 }
	};

	// First row
	leftSideElements[0] = _mm_set_ps(inout[idx[3][0]], inout[idx[2][0]], inout[idx[1][0]], inout[idx[0][0]]);
	rightSideElements[0] = _mm_set_ps(inout[idx[3][1]], inout[idx[2][1]], inout[idx[1][1]], inout[idx[0][1]]);

	leftGERight[0] = _mm_cmpge_ps(leftSideElements[0], rightSideElements[0]); // Something like 0 0 -1 -1.
	leftLTRight[0] = _mm_cmplt_ps(leftSideElements[0], rightSideElements[0]); // Something like -1 -1 0 0.

	// Calculates the values of the elements on the left.
	leftElementsGE[0] = _mm_and_ps(rightSideElements[0], leftGERight[0]); // If the element on left side is bigger or equal to the element on the right side - swaps, so writes the element on the left side to be the element on the right.
	leftElementsLT[0] = _mm_and_ps(leftSideElements[0], leftLTRight[0]);  // If the element on the left side is less than element on the right side - don`t swap and writes the element on left side on it`s place.

	// Calculates the values of the elements on the right
	rightElementsGE[0] = _mm_and_ps(leftSideElements[0], leftGERight[0]);  // If the element on the left side is bigger or equal to the element on the right side - swaps, so writes on the element on the right side to be the element on the left.
	rightElementsLT[0] = _mm_and_ps(rightSideElements[0], leftLTRight[0]); // If the element on the left side is less than element on the right side - don`t swap and writes the element on the right side on it`s place.

	// Now let`s combine the elements, because we have two vectors @leftGERight and @leftLTRight, which are basically inverted, so one OR operation will do it.
	// (in the @leftElemetnsGE will have something like [0, 0, element, element] and in the @leftElemetnsLT will be [element, element, 0, 0]) 
	leftSideElements[0] = _mm_or_ps(leftElementsGE[0], leftElementsLT[0]);
	rightSideElements[0] = _mm_or_ps(rightElementsGE[0], rightElementsLT[0]);

	// Now let`s write them in our array so we can put them in their original places on the given @inout.
	_mm_storeu_ps(resultLeftElements[0], leftSideElements[0]);
	_mm_storeu_ps(resultRightElements[0], rightSideElements[0]);

	// Puts the swaped(if needed) elements on their places.
	inout[idx[0][0]] = resultLeftElements[0][0];
	inout[idx[0][1]] = resultRightElements[0][0];
	inout[idx[1][0]] = resultLeftElements[0][1];
	inout[idx[1][1]] = resultRightElements[0][1];
	inout[idx[2][0]] = resultLeftElements[0][2];
	inout[idx[2][1]] = resultRightElements[0][2];
	inout[idx[3][0]] = resultLeftElements[0][3];
	inout[idx[3][1]] = resultRightElements[0][3];

	// Second row
	leftSideElements[1] = _mm_set_ps(inout[idx[7][0]], inout[idx[6][0]], inout[idx[5][0]], inout[idx[4][0]]);
	rightSideElements[1] = _mm_set_ps(inout[idx[7][1]], inout[idx[6][1]], inout[idx[5][1]], inout[idx[4][1]]);

	leftGERight[1] = _mm_cmpge_ps(leftSideElements[1], rightSideElements[1]);
	leftLTRight[1] = _mm_cmplt_ps(leftSideElements[1], rightSideElements[1]);

	leftElementsGE[1] = _mm_and_ps(rightSideElements[1], leftGERight[1]);
	leftElementsLT[1] = _mm_and_ps(leftSideElements[1], leftLTRight[1]);

	rightElementsGE[1] = _mm_and_ps(leftSideElements[1], leftGERight[1]);
	rightElementsLT[1] = _mm_and_ps(rightSideElements[1], leftLTRight[1]);

	leftSideElements[1] = _mm_or_ps(leftElementsGE[1], leftElementsLT[1]);
	rightSideElements[1] = _mm_or_ps(rightElementsGE[1], rightElementsLT[1]);

	_mm_storeu_ps(resultLeftElements[1], leftSideElements[1]);
	_mm_storeu_ps(resultRightElements[1], rightSideElements[1]);

	inout[idx[4][0]] = resultLeftElements[1][0];
	inout[idx[4][1]] = resultRightElements[1][0];
	inout[idx[5][0]] = resultLeftElements[1][1];
	inout[idx[5][1]] = resultRightElements[1][1];
	inout[idx[6][0]] = resultLeftElements[1][2];
	inout[idx[6][1]] = resultRightElements[1][2];
	inout[idx[7][0]] = resultLeftElements[1][3];
	inout[idx[7][1]] = resultRightElements[1][3];

	// Third row
	leftSideElements[2] = _mm_set_ps(inout[idx[11][0]], inout[idx[10][0]], inout[idx[9][0]], inout[idx[8][0]]);
	rightSideElements[2] = _mm_set_ps(inout[idx[11][1]], inout[idx[10][1]], inout[idx[9][1]], inout[idx[8][1]]);

	leftGERight[2] = _mm_cmpge_ps(leftSideElements[2], rightSideElements[2]);
	leftLTRight[2] = _mm_cmplt_ps(leftSideElements[2], rightSideElements[2]);

	leftElementsGE[2] = _mm_and_ps(rightSideElements[2], leftGERight[2]);
	leftElementsLT[2] = _mm_and_ps(leftSideElements[2], leftLTRight[2]);

	rightElementsGE[2] = _mm_and_ps(leftSideElements[2], leftGERight[2]);
	rightElementsLT[2] = _mm_and_ps(rightSideElements[2], leftLTRight[2]);

	leftSideElements[2] = _mm_or_ps(leftElementsGE[2], leftElementsLT[2]);
	rightSideElements[2] = _mm_or_ps(rightElementsGE[2], rightElementsLT[2]);

	_mm_storeu_ps(resultLeftElements[2], leftSideElements[2]);
	_mm_storeu_ps(resultRightElements[2], rightSideElements[2]);

	inout[idx[8][0]] = resultLeftElements[2][0];
	inout[idx[8][1]] = resultRightElements[2][0];
	inout[idx[9][0]] = resultLeftElements[2][1];
	inout[idx[9][1]] = resultRightElements[2][1];
	inout[idx[10][0]] = resultLeftElements[2][2];
	inout[idx[10][1]] = resultRightElements[2][2];
	inout[idx[11][0]] = resultLeftElements[2][3];
	inout[idx[11][1]] = resultRightElements[2][3];

	// Fourth row
	leftSideElements[3] = _mm_set_ps(inout[idx[15][0]], inout[idx[14][0]], inout[idx[13][0]], inout[idx[12][0]]);
	rightSideElements[3] = _mm_set_ps(inout[idx[15][1]], inout[idx[14][1]], inout[idx[13][1]], inout[idx[12][1]]);

	leftGERight[3] = _mm_cmpge_ps(leftSideElements[3], rightSideElements[3]);
	leftLTRight[3] = _mm_cmplt_ps(leftSideElements[3], rightSideElements[3]);

	leftElementsGE[3] = _mm_and_ps(rightSideElements[3], leftGERight[3]);
	leftElementsLT[3] = _mm_and_ps(leftSideElements[3], leftLTRight[3]);

	rightElementsGE[3] = _mm_and_ps(leftSideElements[3], leftGERight[3]);
	rightElementsLT[3] = _mm_and_ps(rightSideElements[3], leftLTRight[3]);

	leftSideElements[3] = _mm_or_ps(leftElementsGE[3], leftElementsLT[3]);
	rightSideElements[3] = _mm_or_ps(rightElementsGE[3], rightElementsLT[3]);

	_mm_storeu_ps(resultLeftElements[3], leftSideElements[3]);
	_mm_storeu_ps(resultRightElements[3], rightSideElements[3]);

	inout[idx[12][0]] = resultLeftElements[3][0];
	inout[idx[12][1]] = resultRightElements[3][0];
	inout[idx[13][0]] = resultLeftElements[3][1];
	inout[idx[13][1]] = resultRightElements[3][1];
	inout[idx[14][0]] = resultLeftElements[3][2];
	inout[idx[14][1]] = resultRightElements[3][2];
	inout[idx[15][0]] = resultLeftElements[3][3];
	inout[idx[15][1]] = resultRightElements[3][3];

	// Fifth row
	leftSideElements[4] = _mm_set_ps(inout[idx[19][0]], inout[idx[18][0]], inout[idx[17][0]], inout[idx[16][0]]);
	rightSideElements[4] = _mm_set_ps(inout[idx[19][1]], inout[idx[18][1]], inout[idx[17][1]], inout[idx[16][1]]);

	leftGERight[4] = _mm_cmpge_ps(leftSideElements[4], rightSideElements[4]);
	leftLTRight[4] = _mm_cmplt_ps(leftSideElements[4], rightSideElements[4]);

	leftElementsGE[4] = _mm_and_ps(rightSideElements[4], leftGERight[4]);
	leftElementsLT[4] = _mm_and_ps(leftSideElements[4], leftLTRight[4]);

	rightElementsGE[4] = _mm_and_ps(leftSideElements[4], leftGERight[4]);
	rightElementsLT[4] = _mm_and_ps(rightSideElements[4], leftLTRight[4]);

	leftSideElements[4] = _mm_or_ps(leftElementsGE[4], leftElementsLT[4]);
	rightSideElements[4] = _mm_or_ps(rightElementsGE[4], rightElementsLT[4]);

	_mm_storeu_ps(resultLeftElements[4], leftSideElements[4]);
	_mm_storeu_ps(resultRightElements[4], rightSideElements[4]);

	inout[idx[16][0]] = resultLeftElements[4][0];
	inout[idx[16][1]] = resultRightElements[4][0];
	inout[idx[17][0]] = resultLeftElements[4][1];
	inout[idx[17][1]] = resultRightElements[4][1];
	inout[idx[18][0]] = resultLeftElements[4][2];
	inout[idx[18][1]] = resultRightElements[4][2];
	inout[idx[19][0]] = resultLeftElements[4][3];
	inout[idx[19][1]] = resultRightElements[4][3];

	// Sixth row
	leftSideElements[5] = _mm_set_ps(inout[idx[23][0]], inout[idx[22][0]], inout[idx[21][0]], inout[idx[20][0]]);
	rightSideElements[5] = _mm_set_ps(inout[idx[23][1]], inout[idx[22][1]], inout[idx[21][1]], inout[idx[20][1]]);

	leftGERight[5] = _mm_cmpge_ps(leftSideElements[5], rightSideElements[5]);
	leftLTRight[5] = _mm_cmplt_ps(leftSideElements[5], rightSideElements[5]);

	leftElementsGE[5] = _mm_and_ps(rightSideElements[5], leftGERight[5]);
	leftElementsLT[5] = _mm_and_ps(leftSideElements[5], leftLTRight[5]);

	rightElementsGE[5] = _mm_and_ps(leftSideElements[5], leftGERight[5]);
	rightElementsLT[5] = _mm_and_ps(rightSideElements[5], leftLTRight[5]);

	leftSideElements[5] = _mm_or_ps(leftElementsGE[5], leftElementsLT[5]);
	rightSideElements[5] = _mm_or_ps(rightElementsGE[5], rightElementsLT[5]);

	_mm_storeu_ps(resultLeftElements[5], leftSideElements[5]);
	_mm_storeu_ps(resultRightElements[5], rightSideElements[5]);

	inout[idx[20][0]] = resultLeftElements[5][0];
	inout[idx[20][1]] = resultRightElements[5][0];
	inout[idx[21][0]] = resultLeftElements[5][1];
	inout[idx[21][1]] = resultRightElements[5][1];
	inout[idx[22][0]] = resultLeftElements[5][2];
	inout[idx[22][1]] = resultRightElements[5][2];
	inout[idx[23][0]] = resultLeftElements[5][3];
	inout[idx[23][1]] = resultRightElements[5][3];
}
Packet4f plt(const Packet4f& a, Packet4f& b) { return _mm_cmplt_ps(a,b); }
Esempio n. 15
0
 inline vector4fb operator<(const vector4f& lhs, const vector4f& rhs)
 {
     return _mm_cmplt_ps(lhs, rhs);
 }