src/util/SSE-Float.h

   1 #ifndef __SSE_FLOAT_H__
   2 #define __SSE_FLOAT_H__
   3
   4 #include<emmintrin.h>
   5 #include<iostream>
   6
   7
   8
   9 class SSEFloat
  10 {
  11
  12    public: __m128 val;
  13
  14
  15    public:
  16
  17            SSEFloat() {}
  18
  19            SSEFloat(float f) { val= _mm_set1_ps(f);}
  20
  21            SSEFloat(float f0, float f1,float f2, float f3) {val = _mm_setr_ps(f0,f1,f2,f3);}
  22
  23            /* Arithmetic Operators*/
  24
  25            friend inline SSEFloat operator -(const SSEFloat &a) {SSEFloat c;c.val=_mm_sub_ps(_mm_setzero_ps(),a.val);return c;}
  26
  27            friend inline SSEFloat operator +(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_add_ps(a.val,b.val);return c;}
  28
  29            friend inline SSEFloat operator -(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_sub_ps(a.val,b.val);return c;}
  30
  31            friend inline SSEFloat operator *(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_mul_ps(a.val,b.val);return c;}
  32
  33            friend inline SSEFloat operator /(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_div_ps(a.val,b.val);return c;}
  34
  35            friend inline SSEFloat sqrt      (const SSEFloat &a)                  { SSEFloat c;c.val= _mm_sqrt_ps(a.val);return c;}
  36
  37
  38           friend inline SSEFloat operator +(float a, const SSEFloat &b) {SSEFloat c;c.val= _mm_add_ps(_mm_set1_ps(a),b.val);return c;}
  39
  40
  41           friend inline SSEFloat operator -(float a, const SSEFloat &b) {SSEFloat c;c.val= _mm_sub_ps(_mm_set1_ps(a),b.val);return c;}
  42
  43           friend inline SSEFloat operator *(float a, const SSEFloat &b) {SSEFloat c;c.val= _mm_mul_ps(_mm_set1_ps(a),b.val);return c;}
  44
  45           friend inline SSEFloat operator /(float a, const SSEFloat &b) {SSEFloat c;c.val= _mm_div_ps(_mm_set1_ps(a),b.val);return c;}
  46
  47            inline SSEFloat& operator +=(const SSEFloat &a) {val= _mm_add_ps(val,a.val);return *this;}
  48
  49            inline SSEFloat& operator -=(const SSEFloat &a) {val= _mm_sub_ps(val,a.val);return *this;}
  50
  51            inline SSEFloat& operator *=(const SSEFloat &a) {val= _mm_mul_ps(val,a.val);return *this;}
  52
  53            inline SSEFloat& operator /=(const SSEFloat &a) {val= _mm_div_ps(val,a.val);return *this;}
  54
  55           /*Logical Operators*/
  56
  57            friend inline SSEFloat operator &(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_and_ps(a.val,b.val);return c;}
  58
  59            friend inline SSEFloat operator |(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_or_ps(a.val,b.val);return c;}
  60
  61            friend inline SSEFloat operator ^(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_xor_ps(a.val,b.val);return c;}
  62
  63            friend inline SSEFloat andnot (const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_andnot_ps(a.val,b.val);return c;}
  64
  65          /*Comparison Operators*/
  66
  67
  68             friend inline SSEFloat operator <(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_cmplt_ps(a.val,b.val);return c;}
  69
  70             friend inline SSEFloat operator >(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_cmpgt_ps(a.val,b.val);return c;}
  71
  72             friend inline SSEFloat operator ==(const SSEFloat &a, const SSEFloat &b) {SSEFloat c;c.val= _mm_cmpeq_ps(a.val,b.val);return c;}
  73
  74             friend inline SSEFloat operator <(const SSEFloat &a, float b) {SSEFloat c;c.val= _mm_cmplt_ps(a.val,_mm_set1_ps(b));return c;}
  75
  76             friend inline SSEFloat operator >(const SSEFloat &a, float b) {SSEFloat c;c.val= _mm_cmpgt_ps(a.val,_mm_set1_ps(b));return c;}
  77
  78             friend inline SSEFloat max (const SSEFloat &a, SSEFloat &b) { SSEFloat c; c.val= _mm_max_ps(a.val,b.val);return c;}
  79
  80
  81         /*Masking Operations */
  82
  83            friend inline int movemask( const SSEFloat &a) {return _mm_movemask_ps(a.val);}
  84
  85
  86         /*Store Operations*/
  87
  88           friend inline void storeu(float *p, const SSEFloat &a) { _mm_storeu_ps(p,a.val);}
  89
  90       //    friend void storeh(float *p, const SSEFloat &a) { _mm_storeh_pd(p,a.val);}
  91
  92
  93         //   void display();
  94
  95
  96
  97
  98 };
  99
 100
 101 /*
 102 void Double::display()
 103 {
 104
 105 storel(z,val);
 106 //_mm_storeh_pd(z,val);
 107 cout<<*z;
 108 }
 109 */
 110
 111 /*
 112 int main()
 113 {
 114
 115   float i[4];
 116   float *p=i;
 117 // __m128d t1=_mm_setr_pd(3.0,0.0); __m128d t2 = _mm_setr_pd(5.0,0.0);
 118
 119   SSEFloat f1(2.0,1.0,4.0,5.0),f2(4.0,7.0,2.0,5.0),f4(25.0);
 120
 121   SSEFloat d3 =   (f1 * f2)  ;
 122
 123
 124   storeu(p,d3);
 125
 126   cout<<*p;
 127
 128   p++;cout<<*p;
 129
 130   p++;cout<<*p;
 131
 132   p++;cout<<*p;
 133
 134 //      d3 = d1 ^ d2;
 135
 136  // d4 = sqrt(d2);
 137
 138  // __m128d t =  _mm_and_pd(t1,t2);
 139
 140 //  cout << movemask(d3);
 141 //   d3.display();
 142
 143   //int i = movemask(d4);
 144
 145   //cout<<i;
 146
 147 }
 148
 149
 150 */
 151
 152 #endif // __SSE_FLOAT_H__