src/util/SSE-Double.h

   1 #ifndef __SSE_DOUBLE_H__
   2 #define __SSE_DOUBLE_H__
   3
   4 #if CMK_USE_AVX && defined(__AVX__)
   5
   6 #include <x86intrin.h>
   7
   8 #include<iostream>
   9
  10
  11 class SSEDouble
  12 {
  13
  14    public: __m256d val;
  15
  16
  17    public:
  18
  19            SSEDouble() {}
  20
  21            SSEDouble(double d) { val = _mm256_set1_pd(d); }
  22
  23            SSEDouble(double d0, double d1, double d2, double d3) {  val = _mm256_setr_pd(d0,d1,d2,d3); }
  24
  25            /* Arithmetic Operators*/
  26            friend inline SSEDouble operator -(const SSEDouble &a) {SSEDouble c;c.val=_mm256_sub_pd(_mm256_setzero_pd(),a.val);return c;}
  27
  28            friend inline SSEDouble operator +(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_add_pd(a.val,b.val);return c;}
  29
  30            friend inline SSEDouble operator -(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_sub_pd(a.val,b.val);return c;}
  31
  32            friend inline SSEDouble operator *(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_mul_pd(a.val,b.val);return c;}
  33
  34            friend inline SSEDouble operator /(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_div_pd(a.val,b.val);return c;}
  35
  36            friend inline SSEDouble sqrt      (const SSEDouble &a)                  { SSEDouble c;c.val= _mm256_sqrt_pd(a.val);return c;}
  37
  38
  39           friend inline SSEDouble operator +(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_add_pd(_mm256_set1_pd(a),b.val);return c;}
  40
  41
  42           friend inline SSEDouble operator -(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_sub_pd(_mm256_set1_pd(a),b.val);return c;}
  43
  44           friend inline SSEDouble operator *(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_mul_pd(_mm256_set1_pd(a),b.val);return c;}
  45
  46           friend inline SSEDouble operator /(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_div_pd(_mm256_set1_pd(a),b.val);return c;}
  47
  48            inline SSEDouble& operator +=(const SSEDouble &a) {val= _mm256_add_pd(val,a.val);return *this;}
  49
  50            inline SSEDouble& operator -=(const SSEDouble &a) {val= _mm256_sub_pd(val,a.val);return *this;}
  51
  52            inline SSEDouble& operator *=(const SSEDouble &a) {val= _mm256_mul_pd(val,a.val);return *this;}
  53
  54            inline SSEDouble& operator /=(const SSEDouble &a) {val= _mm256_div_pd(val,a.val);return *this;}
  55
  56            /*Logical Operators*/
  57
  58            friend inline SSEDouble operator &(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_and_pd(a.val,b.val);return c;}
  59
  60            friend inline SSEDouble operator |(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_or_pd(a.val,b.val);return c;}
  61
  62            friend inline SSEDouble operator ^(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_xor_pd(a.val,b.val);return c;}
  63
  64            friend inline SSEDouble andnot (const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_andnot_pd(a.val,b.val);return c;}
  65
  66          /*Comparison Operators*/
  67
  68             //friend inline SSEDouble operator <(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_cmplt_pd(a.val,b.val);return c;}
  69             friend inline SSEDouble operator <(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_cmp_pd(a.val,b.val,_CMP_LT_OS);return c;}
  70
  71             //friend inline SSEDouble operator >(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_cmpgt_pd(a.val,b.val);return c;}
  72             friend inline SSEDouble operator >(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_cmp_pd(a.val,b.val,_CMP_GT_OS);return c;}
  73
  74             //friend inline SSEDouble operator ==(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_cmpeq_pd(a.val,b.val);return c;}
  75             friend inline SSEDouble operator ==(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm256_cmp_pd(a.val,b.val,_CMP_EQ_OQ);return c;}
  76
  77             //friend inline SSEDouble operator <(const SSEDouble &a, double b) {SSEDouble c;c.val= _mm256_cmplt_pd(a.val,_mm256_set1_pd(b));return c;}
  78             friend inline SSEDouble operator <(const SSEDouble &a, double b) {SSEDouble c;c.val= _mm256_cmp_pd(a.val,_mm256_set1_pd(b),_CMP_LT_OS);return c;}
  79
  80             //friend inline SSEDouble operator >(const SSEDouble &a, double b) {SSEDouble c;c.val= _mm256_cmpgt_pd(a.val,_mm256_set1_pd(b));return c;}
  81             friend inline SSEDouble operator >(const SSEDouble &a, double b) {SSEDouble c;c.val= _mm256_cmp_pd(a.val,_mm256_set1_pd(b),_CMP_GT_OS);return c;}
  82
  83             friend inline SSEDouble max (const SSEDouble &a, SSEDouble &b) { SSEDouble c; c.val= _mm256_max_pd(a.val,b.val);return c;}
  84
  85
  86         /*Masking Operations */
  87
  88            friend inline int movemask( const SSEDouble &a) {return _mm256_movemask_pd(a.val);}
  89
  90
  91         /*Store Operations*/
  92
  93           friend inline void storeu(double *p, const SSEDouble &a) { _mm256_storeu_pd(p,a.val);}
  94
  95
  96        //   void display();
  97
  98
  99
 100
 101 };
 102
 103
 104 #else
 105
 106
 107 #include<emmintrin.h>
 108
 109 #include<iostream>
 110
 111
 112
 113 class SSEDouble
 114 {
 115
 116    public: __m128d val;
 117
 118
 119    public:
 120
 121            SSEDouble() {}
 122
 123            SSEDouble(double d) { val= _mm_set1_pd(d);}
 124
 125            SSEDouble(double d0, double d1) {val = _mm_setr_pd(d0,d1);}
 126
 127            /* Arithmetic Operators*/
 128            friend inline SSEDouble operator -(const SSEDouble &a) {SSEDouble c;c.val=_mm_sub_pd(_mm_setzero_pd(),a.val);return c;}
 129
 130            friend inline SSEDouble operator +(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_add_pd(a.val,b.val);return c;}
 131
 132            friend inline SSEDouble operator -(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_sub_pd(a.val,b.val);return c;}
 133
 134            friend inline SSEDouble operator *(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_mul_pd(a.val,b.val);return c;}
 135
 136            friend inline SSEDouble operator /(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_div_pd(a.val,b.val);return c;}
 137
 138            friend inline SSEDouble sqrt      (const SSEDouble &a)                  { SSEDouble c;c.val= _mm_sqrt_pd(a.val);return c;}
 139
 140
 141           friend inline SSEDouble operator +(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm_add_pd(_mm_set1_pd(a),b.val);return c;}
 142
 143
 144           friend inline SSEDouble operator -(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm_sub_pd(_mm_set1_pd(a),b.val);return c;}
 145
 146           friend inline SSEDouble operator *(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm_mul_pd(_mm_set1_pd(a),b.val);return c;}
 147
 148           friend inline SSEDouble operator /(double a, const SSEDouble &b) {SSEDouble c;c.val= _mm_div_pd(_mm_set1_pd(a),b.val);return c;}
 149
 150            inline SSEDouble& operator +=(const SSEDouble &a) {val= _mm_add_pd(val,a.val);return *this;}
 151
 152            inline SSEDouble& operator -=(const SSEDouble &a) {val= _mm_sub_pd(val,a.val);return *this;}
 153
 154            inline SSEDouble& operator *=(const SSEDouble &a) {val= _mm_mul_pd(val,a.val);return *this;}
 155
 156            inline SSEDouble& operator /=(const SSEDouble &a) {val= _mm_div_pd(val,a.val);return *this;}
 157
 158            /*Logical Operators*/
 159
 160            friend inline SSEDouble operator &(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_and_pd(a.val,b.val);return c;}
 161
 162            friend inline SSEDouble operator |(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_or_pd(a.val,b.val);return c;}
 163
 164            friend inline SSEDouble operator ^(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_xor_pd(a.val,b.val);return c;}
 165
 166            friend inline SSEDouble andnot (const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_andnot_pd(a.val,b.val);return c;}
 167
 168          /*Comparison Operators*/
 169
 170
 171             friend inline SSEDouble operator <(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_cmplt_pd(a.val,b.val);return c;}
 172
 173             friend inline SSEDouble operator >(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_cmpgt_pd(a.val,b.val);return c;}
 174
 175             friend inline SSEDouble operator ==(const SSEDouble &a, const SSEDouble &b) {SSEDouble c;c.val= _mm_cmpeq_pd(a.val,b.val);return c;}
 176
 177             friend inline SSEDouble operator <(const SSEDouble &a, double b) {SSEDouble c;c.val= _mm_cmplt_pd(a.val,_mm_set1_pd(b));return c;}
 178
 179             friend inline SSEDouble operator >(const SSEDouble &a, double b) {SSEDouble c;c.val= _mm_cmpgt_pd(a.val,_mm_set1_pd(b));return c;}
 180
 181             friend inline SSEDouble max (const SSEDouble &a, SSEDouble &b) { SSEDouble c; c.val= _mm_max_pd(a.val,b.val);return c;}
 182
 183
 184         /*Masking Operations */
 185
 186            friend inline int movemask( const SSEDouble &a) {return _mm_movemask_pd(a.val);}
 187
 188
 189         /*Store Operations*/
 190
 191           friend inline void storel(double *p, const SSEDouble &a) { _mm_storel_pd(p,a.val);}
 192
 193           friend inline void storeh(double *p, const SSEDouble &a) { _mm_storeh_pd(p,a.val);}
 194
 195
 196        //   void display();
 197
 198
 199
 200
 201 };
 202
 203
 204 /*
 205 void SSEDouble::display()
 206 {
 207
 208 storel(z,val);
 209 //_mm_storeh_pd(z,val);
 210 cout<<*z;
 211 }
 212
 213 int main()
 214 {
 215
 216   double i=1.0;
 217   double *p=&i;
 218 // __m128d t1=_mm_setr_pd(3.0,0.0); __m128d t2 = _mm_setr_pd(5.0,0.0);
 219
 220   SSEDouble d1(2.0),d2(4.0),d4(25.0);
 221
 222   SSEDouble d3 = (25.0/5.0) + (d1 * d2) + d4 ;
 223
 224
 225   storel(p,d3);
 226
 227   cout<<*p;
 228 //      d3 = d1 ^ d2;
 229
 230  // d4 = sqrt(d2);
 231
 232  // __m128d t =  _mm_and_pd(t1,t2);
 233
 234 //  cout << movemask(d3);
 235    d3.display();
 236
 237   //int i = movemask(d4);
 238
 239   //cout<<i;
 240
 241 }
 242
 243 */
 244
 245 #endif
 246
 247 #endif //__SSE_DOUBLE_H__