P4_F32vec4.h

Go to the documentation of this file.

#ifndef L1Algo_F32vec4P4_H
#define L1Algo_F32vec4P4_H

#include <iostream>
#include <cmath>
#include "xmmintrin.h"
#include "vec_arithmetic.h"

/**********************************
 *
 *   Vector of four single floats
 *
 **********************************/

//#pragma pack(push,16)/* Must ensure class & union 16-B aligned */

// typedef __m128 VectorFloat __attribute__ ((aligned(16)));

const union {
  float f;
  unsigned int i;
} __f_one = {1.f};

const union {
  unsigned int i[4];
  __m128 m;
} __f32vec4_abs_mask_cheat = {{0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff}}, __f32vec4_sgn_mask_cheat = {{0x80000000, 0x80000000, 0x80000000, 0x80000000}},
  __f32vec4_zero_cheat = {{0, 0, 0, 0}}, __f32vec4_one_cheat = {{__f_one.i, __f_one.i, __f_one.i, __f_one.i}},
  __f32vec4_true_cheat = {{0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF}}, __f32vec4_false_cheat = {{0x00000000, 0x00000000, 0x00000000, 0x00000000}};

#define _f32vec4_abs_mask (static_cast<F32vec4>(__f32vec4_abs_mask_cheat.m))
#define _f32vec4_sgn_mask (static_cast<F32vec4>(__f32vec4_sgn_mask_cheat.m))
#define _f32vec4_zero (static_cast<F32vec4>(__f32vec4_zero_cheat.m))
#define _f32vec4_one (static_cast<F32vec4>(__f32vec4_one_cheat.m))
#define _f32vec4_true (static_cast<F32vec4>(__f32vec4_true_cheat.m))
#define _f32vec4_false (static_cast<F32vec4>(__f32vec4_false_cheat.m))

class F32vec4 {
 public:
  __m128 v;

  float &operator[](int i) { return (reinterpret_cast<float *>(&v))[i]; }
  float operator[](int i) const { return (reinterpret_cast<const float *>(&v))[i]; }

  F32vec4() : v(_mm_set_ps1(0)) {}
  F32vec4(const __m128 &a) : v(a) {}
  F32vec4(const float &a) : v(_mm_set_ps1(a)) {}

  F32vec4(const float &f0, const float &f1, const float &f2, const float &f3) : v(_mm_set_ps(f3, f2, f1, f0)) {}

  /* Conversion function */
  operator __m128() const { return v; } /* Convert to __m128 */

  /* Arithmetic Operators */
  friend F32vec4 operator+(const F32vec4 &a, const F32vec4 &b) { return _mm_add_ps(a, b); }
  friend F32vec4 operator-(const F32vec4 &a, const F32vec4 &b) { return _mm_sub_ps(a, b); }
  friend F32vec4 operator*(const F32vec4 &a, const F32vec4 &b) { return _mm_mul_ps(a, b); }
  friend F32vec4 operator/(const F32vec4 &a, const F32vec4 &b) { return _mm_div_ps(a, b); }

  /* Functions */
  friend F32vec4 min(const F32vec4 &a, const F32vec4 &b) { return _mm_min_ps(a, b); }
  friend F32vec4 max(const F32vec4 &a, const F32vec4 &b) { return _mm_max_ps(a, b); }

  /* Square Root */
  friend F32vec4 sqrt(const F32vec4 &a) { return _mm_sqrt_ps(a); }

  /* Reciprocal( inverse) Square Root */
  friend F32vec4 rsqrt(const F32vec4 &a) { return _mm_rsqrt_ps(a); }

  /* Reciprocal (inversion) */
  // friend F32vec4 rcp  ( const F32vec4 &a ){ return _mm_rcp_ps  (a); }
  /* Reciprocal (inversion) */
  // friend F32vec4 rcp  ( const F32vec4 &a ){ return 1. / a; }
  /* NewtonRaphson Reciprocal
    [2 * rcpps(x) - (x * rcpps(x) * rcpps(x))] */
  friend F32vec4 rcp(const F32vec4 &a)
  {
    F32vec4 Ra0 = _mm_rcp_ps(a);
    return _mm_sub_ps(_mm_add_ps(Ra0, Ra0), _mm_mul_ps(_mm_mul_ps(Ra0, a), Ra0));
  }

  /* Absolute value */
  friend F32vec4 fabs(const F32vec4 &a) { return _mm_and_ps(a, _f32vec4_abs_mask); }

  /* Sign */
  friend F32vec4 sgn(const F32vec4 &a) { return _mm_or_ps(_mm_and_ps(a, _f32vec4_sgn_mask), _f32vec4_one); }
  friend F32vec4 asgnb(const F32vec4 &a, const F32vec4 &b) { return _mm_or_ps(_mm_and_ps(b, _f32vec4_sgn_mask), a); }

  /* Logical */

  friend F32vec4 operator&(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_and_ps(a, b);
  }
  friend F32vec4 operator|(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_or_ps(a, b);
  }
  friend F32vec4 operator^(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_xor_ps(a, b);
  }
  friend F32vec4 operator!(const F32vec4 &a)
  { // mask returned
    return _mm_xor_ps(a, _f32vec4_true);
  }
  // friend F32vec4 operator||( const F32vec4 &a, const F32vec4 &b ){ // mask returned
  //   return _mm_or_ps(a, b);
  // }

  /* Comparison */

  friend F32vec4 operator<(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_cmplt_ps(a, b);
  }
  friend F32vec4 operator<=(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_cmple_ps(a, b);
  }
  friend F32vec4 operator>(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_cmpgt_ps(a, b);
  }
  friend F32vec4 operator>=(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_cmpge_ps(a, b);
  }
  friend F32vec4 operator==(const F32vec4 &a, const F32vec4 &b)
  { // mask returned
    return _mm_cmpeq_ps(a, b);
  }

#define if3(a, b, c) ((a) & (b)) | ((!(a)) & (c)) // analog (a) ? b : c

#define NotEmpty(a) bool((a)[0]) | bool((a)[1]) | bool((a)[2]) | bool((a)[3])
#define Empty(a) !(bool((a)[0]) | bool((a)[1]) | bool((a)[2]) | bool((a)[3]))
  // bool NotEmpty(const F32vec4 &a) { return a[0]||a[1]||a[2]||a[3]; }
  // bool    Empty(const F32vec4 &a) { return !(a[0]||a[1]||a[2]||a[3]); } // optimize
  friend F32vec4 bool2int(const F32vec4 &a)
  { // mask returned
    return if3(a, 1, 0);
  }

  /* Define all operators for consistensy */

  vec_arithmetic(F32vec4, float);

  /* Non intrinsic functions */

#define _f1(A, F) F32vec4(F(A[0]), F(A[1]), F(A[2]), F(A[3]))

  friend F32vec4 exp(const F32vec4 &a) { return _f1(a, exp); }
  friend F32vec4 log(const F32vec4 &a) { return _f1(a, log); }
  friend F32vec4 sin(const F32vec4 &a) { return _f1(a, sin); }
  friend F32vec4 cos(const F32vec4 &a) { return _f1(a, cos); }
  friend F32vec4 acos(const F32vec4 &a) { return _f1(a, acos); }

#undef _f1

  friend F32vec4 atan2(const F32vec4 &y, const F32vec4 &x)
  {
    const F32vec4 pi(3.1415926535897932);
    const F32vec4 pi_2 = pi / 2;
    const F32vec4 zero(0);

    const F32vec4 &xZero = F32vec4(x == zero);
    const F32vec4 &yZero = F32vec4(y == zero);
    const F32vec4 &xNeg = F32vec4(x < zero);
    const F32vec4 &yNeg = F32vec4(y < zero);

    const F32vec4 &absX = fabs(x);
    const F32vec4 &absY = fabs(y);

    F32vec4 a = absY / absX;
    const F32vec4 pi_4 = pi / 4;
    const F32vec4 &gt_tan_3pi_8 = F32vec4(a > F32vec4(2.414213562373095));
    const F32vec4 &gt_tan_pi_8 = F32vec4(a > F32vec4(0.4142135623730950)) & F32vec4(!gt_tan_3pi_8);
    const F32vec4 minusOne(-1);
    F32vec4 b(zero);
    b = (pi_2 & gt_tan_3pi_8) + (F32vec4(!gt_tan_3pi_8) & b);
    b = (pi_4 & gt_tan_pi_8) + (F32vec4(!gt_tan_pi_8) & b);
    a = (gt_tan_3pi_8 & (minusOne / a)) + (F32vec4(!gt_tan_3pi_8) & a);
    a = (gt_tan_pi_8 & ((absY - absX) / (absY + absX))) + (F32vec4(!gt_tan_pi_8) & a);
    const F32vec4 &a2 = a * a;
    b += (((8.05374449538e-2 * a2 - 1.38776856032E-1) * a2 + 1.99777106478E-1) * a2 - 3.33329491539E-1) * a2 * a + a;
    F32vec4 xyNeg = F32vec4(xNeg ^ yNeg);
    b = (xyNeg & (-b)) + (F32vec4(!xyNeg) & b);
    xyNeg = F32vec4(xNeg & !yNeg);
    b = (xyNeg & (b + pi)) + (F32vec4(!xyNeg) & b);
    xyNeg = F32vec4(xNeg & yNeg);
    b = (xyNeg & (b - pi)) + (F32vec4(!xyNeg) & b);
    xyNeg = F32vec4(xZero & yZero);
    b = (xyNeg & zero) + (F32vec4(!xyNeg) & b);
    xyNeg = F32vec4(xZero & yNeg);
    b = (xyNeg & (-pi_2)) + (F32vec4(!xyNeg) & b);
    return b;
  }

  friend std::ostream &operator<<(std::ostream &strm, const F32vec4 &a)
  {
    strm << "[" << a[0] << " " << a[1] << " " << a[2] << " " << a[3] << "]";
    return strm;
  }

  friend std::istream &operator>>(std::istream &strm, F32vec4 &a)
  {
    float tmp;
    strm >> tmp;
    a = tmp;
    return strm;
  }

} __attribute__((aligned(16)));

typedef F32vec4 fvec;
typedef float fscal;
const int fvecLen = 4;
//#define fvec_true  _f32vec4_true
//#define fvec_false _f32vec4_false
#define _fvecalignment __attribute__((aligned(16)))

#include "std_alloc.h"

#endif