mcs/class/Mono.Simd/Mono.Simd/Vector8us.cs

   1 // Vector8us.cs
   2 //
   3 // Author:
   4 //   Rodrigo Kumpera (rkumpera@novell.com)
   5 //
   6 // (C) 2008 Novell, Inc. (http://www.novell.com)
   7 //
   8 // Permission is hereby granted, free of charge, to any person obtaining
   9 // a copy of this software and associated documentation files (the
  10 // "Software"), to deal in the Software without restriction, including
  11 // without limitation the rights to use, copy, modify, merge, publish,
  12 // distribute, sublicense, and/or sell copies of the Software, and to
  13 // permit persons to whom the Software is furnished to do so, subject to
  14 // the following conditions:
  15 //
  16 // The above copyright notice and this permission notice shall be
  17 // included in all copies or substantial portions of the Software.
  18 //
  19 // THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  20 // EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  21 // MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  22 // NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE
  23 // LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION
  24 // OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
  25 // WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
  26 //
  27 using System;
  28 using System.Runtime.InteropServices;
  29
  30 namespace Mono.Simd
  31 {
  32         [StructLayout(LayoutKind.Sequential, Pack = 0, Size = 16)]
  33         [CLSCompliant(false)]
  34         public struct Vector8us
  35         {
  36                 private ushort v0, v1, v2, v3, v4, v5, v6, v7;
  37                 public Vector8us (ushort v0, ushort v1, ushort v2, ushort v3, ushort v4, ushort v5, ushort v6, ushort v7)
  38                 {
  39                         this.v0 = v0;
  40                         this.v1 = v1;
  41                         this.v2 = v2;
  42                         this.v3 = v3;
  43                         this.v4 = v4;
  44                         this.v5 = v5;
  45                         this.v6 = v6;
  46                         this.v7 = v7;
  47                 }
  48
  49                 public ushort V0 { get { return v0; } set { v0 = value; } }
  50                 public ushort V1 { get { return v1; } set { v1 = value; } }
  51                 public ushort V2 { get { return v2; } set { v2 = value; } }
  52                 public ushort V3 { get { return v3; } set { v3 = value; } }
  53                 public ushort V4 { get { return v4; } set { v4 = value; } }
  54                 public ushort V5 { get { return v5; } set { v5 = value; } }
  55                 public ushort V6 { get { return v6; } set { v6 = value; } }
  56                 public ushort V7 { get { return v7; } set { v7 = value; } }
  57
  58                 [Acceleration (AccelMode.SSE2)]
  59                 public static unsafe Vector8us operator + (Vector8us va, Vector8us vb)
  60                 {
  61                         Vector8us res = new Vector8us ();
  62                         ushort *a = &va.v0;
  63                         ushort *b = &vb.v0;
  64                         ushort *c = &res.v0;
  65                         for (int i = 0; i < 8; ++i)
  66                                 *c++ = (ushort)(*a++ + *b++);
  67                         return res;
  68                 }
  69
  70                 [Acceleration (AccelMode.SSE2)]
  71                 public static unsafe Vector8us operator - (Vector8us va, Vector8us vb)
  72                 {
  73                         Vector8us res = new Vector8us ();
  74                         ushort *a = &va.v0;
  75                         ushort *b = &vb.v0;
  76                         ushort *c = &res.v0;
  77                         for (int i = 0; i < 8; ++i)
  78                                 *c++ = (ushort)(*a++ - *b++);
  79                         return res;
  80                 }
  81
  82                 /*
  83                  * NOTE: Thou pmullw states it does signed multiplication, it works for unsigned numbers
  84                  * if only the lower part is considered and the flags disregarded.
  85                  */
  86                 [Acceleration (AccelMode.SSE2)]
  87                 public static unsafe Vector8us operator * (Vector8us va, Vector8us vb)
  88                 {
  89                         Vector8us res = new Vector8us ();
  90                         ushort *a = &va.v0;
  91                         ushort *b = &vb.v0;
  92                         ushort *c = &res.v0;
  93                         for (int i = 0; i < 8; ++i)
  94                                 *c++ = (ushort)(*a++ * (*b++));
  95                         return res;
  96                 }
  97
  98                 [Acceleration (AccelMode.SSE2)]
  99                 public static unsafe Vector8us operator >> (Vector8us va, int amount)
 100                 {
 101                         Vector8us res = new Vector8us ();
 102                         ushort *a = &va.v0;
 103                         ushort *b = &res.v0;
 104                         for (int i = 0; i < 8; ++i)
 105                                 *b++ = (ushort)(*a++ >> amount);
 106                         return res;
 107                 }
 108
 109                 [Acceleration (AccelMode.SSE2)]
 110                 public static unsafe Vector8us operator << (Vector8us va, int amount)
 111                 {
 112                         Vector8us res = new Vector8us ();
 113                         ushort *a = &va.v0;
 114                         ushort *b = &res.v0;
 115                         for (int i = 0; i < 8; ++i)
 116                                 *b++ = (ushort)(*a++ << amount);
 117                         return res;
 118                 }
 119
 120                 [Acceleration (AccelMode.SSE2)]
 121                 public static unsafe Vector8us operator & (Vector8us va, Vector8us vb)
 122                 {
 123                         Vector8us res = new Vector8us ();
 124                         ushort *a = &va.v0;
 125                         ushort *b = &vb.v0;
 126                         ushort *c = &res.v0;
 127                         for (int i = 0; i < 8; ++i)
 128                                 *c++ = (ushort)(*a++ & *b++);
 129                         return res;
 130                 }
 131
 132                 [Acceleration (AccelMode.SSE2)]
 133                 public static unsafe Vector8us operator | (Vector8us va, Vector8us vb)
 134                 {
 135                         Vector8us res = new Vector8us ();
 136                         ushort *a = &va.v0;
 137                         ushort *b = &vb.v0;
 138                         ushort *c = &res.v0;
 139                         for (int i = 0; i < 8; ++i)
 140                                 *c++ = (ushort)(*a++ | *b++);
 141                         return res;
 142                 }
 143
 144                 [Acceleration (AccelMode.SSE2)]
 145                 public static unsafe Vector8us operator ^ (Vector8us va, Vector8us vb)
 146                 {
 147                         Vector8us res = new Vector8us ();
 148                         ushort *a = &va.v0;
 149                         ushort *b = &vb.v0;
 150                         ushort *c = &res.v0;
 151                         for (int i = 0; i < 8; ++i)
 152                                 *c++ = (ushort)(*a++ ^ *b++);
 153                         return res;
 154                 }
 155
 156                 [Acceleration (AccelMode.SSE2)]
 157                 public static unsafe Vector8us UnpackLow (Vector8us va, Vector8us vb)
 158                 {
 159                         return new Vector8us (va.v0, vb.v0, va.v1, vb.v1, va.v2, vb.v2, va.v3, vb.v3);
 160                 }
 161
 162                 [Acceleration (AccelMode.SSE2)]
 163                 public static unsafe Vector8us UnpackHigh (Vector8us va, Vector8us vb)
 164                 {
 165                         return new Vector8us (va.v4, vb.v4, va.v5, vb.v5, va.v6, vb.v6, va.v7, vb.v7);
 166                 }
 167
 168                 [Acceleration (AccelMode.SSE2)]
 169                 public static unsafe Vector8us ShiftRightArithmetic (Vector8us va, int amount)
 170                 {
 171                         Vector8us res = new Vector8us ();
 172                         ushort *a = &va.v0;
 173                         ushort *b = &res.v0;
 174                         for (int i = 0; i < 8; ++i)
 175                                 *b++ = (ushort)((short)(*a++) >> amount);
 176                         return res;
 177                 }
 178
 179                 [Acceleration (AccelMode.SSE2)]
 180                 public static unsafe Vector8us AddWithSaturation (Vector8us va, Vector8us vb) {
 181                         Vector8us res = new Vector8us ();
 182                         ushort *a = &va.v0;
 183                         ushort *b = &vb.v0;
 184                         ushort *c = &res.v0;
 185                         for (int i = 0; i < 8; ++i)
 186                                 *c++ = (ushort) System.Math.Min (*a++ + *b++, ushort.MaxValue);
 187                         return res;
 188                 }
 189
 190                 [Acceleration (AccelMode.SSE2)]
 191                 public static unsafe Vector8us SubWithSaturation (Vector8us va, Vector8us vb) {
 192                         Vector8us res = new Vector8us ();
 193                         ushort *a = &va.v0;
 194                         ushort *b = &vb.v0;
 195                         ushort *c = &res.v0;
 196                         for (int i = 0; i < 8; ++i)
 197                                 *c++ = (ushort) System.Math.Max (*a++ - *b++, 0);
 198                         return res;
 199                 }
 200
 201                 [Acceleration (AccelMode.SSE2)]
 202                 public static unsafe Vector8us Average (Vector8us va, Vector8us vb) {
 203                         Vector8us res = new Vector8us ();
 204                         ushort *a = &va.v0;
 205                         ushort *b = &vb.v0;
 206                         ushort *c = &res.v0;
 207                         for (int i = 0; i < 8; ++i)
 208                                 *c++ = (ushort) ((*a++ + *b++ + 1) >> 1);
 209                         return res;
 210                 }
 211
 212                 [Acceleration (AccelMode.SSE41)]
 213                 public static unsafe Vector8us Max (Vector8us va, Vector8us vb) {
 214                         Vector8us res = new Vector8us ();
 215                         ushort *a = &va.v0;
 216                         ushort *b = &vb.v0;
 217                         ushort *c = &res.v0;
 218                         for (int i = 0; i < 8; ++i)
 219                                 *c++ = (ushort) System.Math.Max (*a++, *b++);
 220                         return res;
 221                 }
 222
 223                 [Acceleration (AccelMode.SSE41)]
 224                 public static unsafe Vector8us Min (Vector8us va, Vector8us vb) {
 225                         Vector8us res = new Vector8us ();
 226                         ushort *a = &va.v0;
 227                         ushort *b = &vb.v0;
 228                         ushort *c = &res.v0;
 229                         for (int i = 0; i < 8; ++i)
 230                                 *c++ = (ushort) System.Math.Min (*a++, *b++);
 231                         return res;
 232                 }
 233
 234                 [Acceleration (AccelMode.SSE2)]
 235                 public static unsafe int ExtractByteMask (Vector8us va) {
 236                         int res = 0;
 237                         byte *a = (byte*)&va;
 238                         for (int i = 0; i < 16; ++i)
 239                                 res |= (*a++ & 0x80) >> 7 << i;
 240                         return res;
 241                 }
 242
 243                 [Acceleration (AccelMode.SSE2)]
 244                 public static unsafe Vector8us ShuffleHigh (Vector8us va, ShuffleSel sel)
 245                 {
 246                         ushort *ptr = ((ushort*)&va) + 4;
 247                         int idx = (int)sel;
 248                         return new Vector8us (va.v0, va.v1, va.v2, va.v3, *(ptr + ((idx >> 0) & 0x3)), *(ptr + ((idx >> 2) & 0x3)), *(ptr + ((idx >> 4) & 0x3)), *(ptr + ((idx >> 6) & 0x3)));
 249                 }
 250
 251                 [Acceleration (AccelMode.SSE2)]
 252                 public static unsafe Vector8us ShuffleLow (Vector8us va, ShuffleSel sel)
 253                 {
 254                         ushort *ptr = ((ushort*)&va);
 255                         int idx = (int)sel;
 256                         return new Vector8us (*(ptr + ((idx >> 0) & 0x3)), *(ptr + ((idx >> 2) & 0x3)), *(ptr + ((idx >> 4) & 0x3)), *(ptr + ((idx >> 6) & 0x3)), va.v4, va.v5, va.v6, va.v7);
 257                 }
 258
 259                 [Acceleration (AccelMode.SSE2)]
 260                 public static unsafe Vector8us CompareEqual (Vector8us va, Vector8us vb) {
 261                         Vector8us res = new Vector8us ();
 262                         ushort *a = &va.v0;
 263                         ushort *b = &vb.v0;
 264                         ushort *c = &res.v0;
 265                         for (int i = 0; i < 8; ++i)
 266                                 *c++ = (ushort) (*a++ == *b++ ? -1 : 0);
 267                         return res;
 268                 }
 269
 270                 [Acceleration (AccelMode.SSE2)]
 271                 public static unsafe Vector8us MultiplyStoreHigh (Vector8us va, Vector8us vb) {
 272                         Vector8us res = new Vector8us ();
 273                         ushort *a = &va.v0;
 274                         ushort *b = &vb.v0;
 275                         ushort *c = &res.v0;
 276                         for (int i = 0; i < 8; ++i)
 277                                 *c++ = (ushort)((uint)*a++ * (uint)*b++ >> 16);
 278                         return res;
 279                 }
 280
 281                 /*This function performs a packuswb, which treats the source as a signed value */
 282                 [Acceleration (AccelMode.SSE2)]
 283                 public static unsafe Vector16b SignedPackWithUnsignedSaturation (Vector8us va, Vector8us vb) {
 284                         Vector16b res = new Vector16b ();
 285                         short *a = (short*)&va;
 286                         short *b = (short*)&vb;
 287                         byte *c = (byte*)&res;
 288                         for (int i = 0; i < 8; ++i)
 289                                 *c++ = (byte)System.Math.Max (0, System.Math.Min ((int)*a++, byte.MaxValue));
 290                         for (int i = 0; i < 8; ++i)
 291                                 *c++ = (byte)System.Math.Max (0, System.Math.Min ((int)*b++, byte.MaxValue));
 292                         return res;
 293                 }
 294
 295                 /*This function performs a packsswb, which treats the source as a signed value */
 296                 [Acceleration (AccelMode.SSE2)]
 297                 public static unsafe Vector16sb SignedPackWithSignedSaturation (Vector8us va, Vector8us vb) {
 298                         Vector16sb res = new Vector16sb ();
 299                         short *a = (short*)&va;
 300                         short *b = (short*)&vb;
 301                         sbyte *c = (sbyte*)&res;
 302                         for (int i = 0; i < 8; ++i)
 303                                 *c++ = (sbyte)System.Math.Max (System.Math.Min ((int)*a++, sbyte.MaxValue), sbyte.MinValue);
 304                         for (int i = 0; i < 8; ++i)
 305                                 *c++ = (sbyte)System.Math.Max (System.Math.Min ((int)*b++, sbyte.MaxValue), sbyte.MinValue);
 306                         return res;
 307                 }
 308
 309                 [Acceleration (AccelMode.SSE1)]
 310                 public static unsafe explicit operator Vector2d (Vector8us v)
 311                 {
 312                         Vector2d* p = (Vector2d*)&v;
 313                         return *p;
 314                 }
 315
 316                 [Acceleration (AccelMode.SSE1)]
 317                 public static unsafe explicit operator Vector4f (Vector8us v)
 318                 {
 319                         Vector4f* p = (Vector4f*)&v;
 320                         return *p;
 321                 }
 322
 323                 [Acceleration (AccelMode.SSE1)]
 324                 public static unsafe explicit operator Vector2l (Vector8us v)
 325                 {
 326                         Vector2l* p = (Vector2l*)&v;
 327                         return *p;
 328                 }
 329
 330                 [Acceleration (AccelMode.SSE1)]
 331                 public static unsafe explicit operator Vector2ul (Vector8us v)
 332                 {
 333                         Vector2ul* p = (Vector2ul*)&v;
 334                         return *p;
 335                 }
 336
 337                 [Acceleration (AccelMode.SSE1)]
 338                 public static unsafe explicit operator Vector4i (Vector8us v)
 339                 {
 340                         Vector4i* p = (Vector4i*)&v;
 341                         return *p;
 342                 }
 343
 344                 [Acceleration (AccelMode.SSE1)]
 345                 public static unsafe explicit operator Vector4ui (Vector8us v)
 346                 {
 347                         Vector4ui* p = (Vector4ui*)&v;
 348                         return *p;
 349                 }
 350
 351                 [Acceleration (AccelMode.SSE1)]
 352                 public static unsafe explicit operator Vector8s (Vector8us v)
 353                 {
 354                         Vector8s* p = (Vector8s*)&v;
 355                         return *p;
 356                 }
 357
 358                 [Acceleration (AccelMode.SSE1)]
 359                 public static unsafe explicit operator Vector16sb (Vector8us v)
 360                 {
 361                         Vector16sb* p = (Vector16sb*)&v;
 362                         return *p;
 363                 }
 364
 365                 [Acceleration (AccelMode.SSE1)]
 366                 public static unsafe explicit operator Vector16b (Vector8us v)
 367                 {
 368                         Vector16b* p = (Vector16b*)&v;
 369                         return *p;
 370                 }
 371
 372
 373                 [Acceleration (AccelMode.SSE1)]
 374                 public static Vector8us LoadAligned (ref Vector8us v)
 375                 {
 376                         return v;
 377                 }
 378
 379                 [Acceleration (AccelMode.SSE1)]
 380                 public static void StoreAligned (ref Vector8us res, Vector8us val)
 381                 {
 382                         res = val;
 383                 }
 384
 385                 [Acceleration (AccelMode.SSE1)]
 386                 public static unsafe Vector8us LoadAligned (Vector8us *v)
 387                 {
 388                         return *v;
 389                 }
 390
 391                 [Acceleration (AccelMode.SSE1)]
 392                 public static unsafe void StoreAligned (Vector8us *res, Vector8us val)
 393                 {
 394                         *res = val;
 395                 }
 396
 397                 [Acceleration (AccelMode.SSE1)]
 398                 [CLSCompliant(false)]
 399                 public static void PrefetchTemporalAllCacheLevels (ref Vector8us res)
 400                 {
 401                 }
 402
 403                 [Acceleration (AccelMode.SSE1)]
 404                 [CLSCompliant(false)]
 405                 public static void PrefetchTemporal1stLevelCache (ref Vector8us res)
 406                 {
 407                 }
 408
 409                 [Acceleration (AccelMode.SSE1)]
 410                 [CLSCompliant(false)]
 411                 public static void PrefetchTemporal2ndLevelCache (ref Vector8us res)
 412                 {
 413                 }
 414
 415                 [Acceleration (AccelMode.SSE1)]
 416                 [CLSCompliant(false)]
 417                 public static void PrefetchNonTemporal (ref Vector8us res)
 418                 {
 419                 }
 420
 421                 [Acceleration (AccelMode.SSE1)]
 422                 [CLSCompliant(false)]
 423                 public static unsafe void PrefetchTemporalAllCacheLevels (Vector8us *res)
 424                 {
 425                 }
 426
 427                 [Acceleration (AccelMode.SSE1)]
 428                 [CLSCompliant(false)]
 429                 public static unsafe void PrefetchTemporal1stLevelCache (Vector8us *res)
 430                 {
 431                 }
 432
 433                 [Acceleration (AccelMode.SSE1)]
 434                 [CLSCompliant(false)]
 435                 public static unsafe void PrefetchTemporal2ndLevelCache (Vector8us *res)
 436                 {
 437                 }
 438
 439                 [Acceleration (AccelMode.SSE1)]
 440                 [CLSCompliant(false)]
 441                 public static unsafe void PrefetchNonTemporal (Vector8us *res)
 442                 {
 443                 }
 444         }
 445 }