mcs/class/Mono.Simd/Mono.Simd/Vector8s.cs

   1 // Vector8s.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         public struct Vector8s
  34         {
  35                 private short v0, v1, v2, v3, v4, v5, v6, v7;
  36                 public Vector8s (short v0, short v1, short v2, short v3, short v4, short v5, short v6, short v7)
  37                 {
  38                         this.v0 = v0;
  39                         this.v1 = v1;
  40                         this.v2 = v2;
  41                         this.v3 = v3;
  42                         this.v4 = v4;
  43                         this.v5 = v5;
  44                         this.v6 = v6;
  45                         this.v7 = v7;
  46                 }
  47
  48                 public short V0 { get { return v0; } set { v0 = value; } }
  49                 public short V1 { get { return v1; } set { v1 = value; } }
  50                 public short V2 { get { return v2; } set { v2 = value; } }
  51                 public short V3 { get { return v3; } set { v3 = value; } }
  52                 public short V4 { get { return v4; } set { v4 = value; } }
  53                 public short V5 { get { return v5; } set { v5 = value; } }
  54                 public short V6 { get { return v6; } set { v6 = value; } }
  55                 public short V7 { get { return v7; } set { v7 = value; } }
  56
  57                 [System.Runtime.CompilerServices.IndexerName ("Component")]
  58                 public unsafe short this [int index]
  59                 {
  60                         get {
  61                                 if ((index | 0x7) != 0x7) //index < 0 || index > 7
  62                                         throw new ArgumentOutOfRangeException ("index");
  63                                 fixed (short *v = &v0) {
  64                                         return * (v + index);
  65                                 }
  66                         }
  67                         set {
  68                                 if ( (index | 0x7) != 0x7) //index < 0 || index > 7
  69                                         throw new ArgumentOutOfRangeException ("index");
  70                                 fixed (short *v = &v0) {
  71                                         * (v + index) = value;
  72                                 }
  73                         }
  74                 }
  75
  76                 [Acceleration (AccelMode.SSE2)]
  77                 public static unsafe Vector8s operator + (Vector8s va, Vector8s vb)
  78                 {
  79                         Vector8s res = new Vector8s ();
  80                         short *a = &va.v0;
  81                         short *b = &vb.v0;
  82                         short *c = &res.v0;
  83                         for (int i = 0; i < 8; ++i)
  84                                 *c++ = (short)(*a++ + *b++);
  85                         return res;
  86                 }
  87
  88                 [Acceleration (AccelMode.SSE2)]
  89                 public static unsafe Vector8s operator - (Vector8s va, Vector8s vb)
  90                 {
  91                         Vector8s res = new Vector8s ();
  92                         short *a = &va.v0;
  93                         short *b = &vb.v0;
  94                         short *c = &res.v0;
  95                         for (int i = 0; i < 8; ++i)
  96                                 *c++ = (short)(*a++ - *b++);
  97                         return res;
  98                 }
  99
 100                 [Acceleration (AccelMode.SSE2)]
 101                 public static unsafe Vector8s operator * (Vector8s va, Vector8s vb)
 102                 {
 103                         Vector8s res = new Vector8s ();
 104                         short *a = &va.v0;
 105                         short *b = &vb.v0;
 106                         short *c = &res.v0;
 107                         for (int i = 0; i < 8; ++i)
 108                                 *c++ = (short)(*a++ * (*b++));
 109                         return res;
 110                 }
 111
 112                 [Acceleration (AccelMode.SSE2)]
 113                 public static unsafe Vector8s operator >> (Vector8s va, int amount)
 114                 {
 115                         Vector8s res = new Vector8s ();
 116                         short *a = &va.v0;
 117                         short *b = &res.v0;
 118                         for (int i = 0; i < 8; ++i)
 119                                 *b++ = (short)(*a++ >> amount);
 120                         return res;
 121                 }
 122
 123                 [Acceleration (AccelMode.SSE2)]
 124                 public static unsafe Vector8s operator << (Vector8s va, int amount)
 125                 {
 126                         Vector8s res = new Vector8s ();
 127                         short *a = &va.v0;
 128                         short *b = &res.v0;
 129                         for (int i = 0; i < 8; ++i)
 130                                 *b++ = (short)(*a++ << amount);
 131                         return res;
 132                 }
 133
 134                 [Acceleration (AccelMode.SSE2)]
 135                 public static unsafe Vector8s operator & (Vector8s va, Vector8s vb)
 136                 {
 137                         Vector8s res = new Vector8s ();
 138                         ulong *a = (ulong*) &va.v0;
 139                         ulong *b = (ulong*) &vb.v0;
 140                         ulong *c = (ulong*) &res.v0;
 141                         *c++ = (ulong)(*a++ & *b++);
 142                         *c = (ulong)(*a & *b);
 143                         return res;
 144                 }
 145
 146                 [Acceleration (AccelMode.SSE2)]
 147                 public static unsafe Vector8s operator | (Vector8s va, Vector8s vb)
 148                 {
 149                         Vector8s res = new Vector8s ();
 150                         ulong *a = (ulong*) &va.v0;
 151                         ulong *b = (ulong*) &vb.v0;
 152                         ulong *c = (ulong*) &res.v0;
 153                         *c++ = (ulong)(*a++ | *b++);
 154                         *c = (ulong)(*a | *b);
 155                         return res;
 156                 }
 157
 158                 [Acceleration (AccelMode.SSE2)]
 159                 public static unsafe Vector8s operator ^ (Vector8s va, Vector8s vb)
 160                 {
 161                         Vector8s res = new Vector8s ();
 162                         ulong *a = (ulong*) &va.v0;
 163                         ulong *b = (ulong*) &vb.v0;
 164                         ulong *c = (ulong*) &res.v0;
 165                         *c++ = (ulong)(*a++ ^ *b++);
 166                         *c = (ulong)(*a ^ *b);
 167                         return res;
 168                 }
 169
 170                 [Acceleration (AccelMode.SSE2)]
 171                 public static unsafe Vector8s UnpackLow (Vector8s va, Vector8s vb)
 172                 {
 173                         return new Vector8s (va.v0, vb.v0, va.v1, vb.v1, va.v2, vb.v2, va.v3, vb.v3);
 174                 }
 175
 176                 [Acceleration (AccelMode.SSE2)]
 177                 public static unsafe Vector8s UnpackHigh (Vector8s va, Vector8s vb)
 178                 {
 179                         return new Vector8s (va.v4, vb.v4, va.v5, vb.v5, va.v6, vb.v6, va.v7, vb.v7);
 180                 }
 181
 182                 [Acceleration (AccelMode.SSE2)]
 183                 public static unsafe Vector8s LogicalRightShift (Vector8s va, int amount)
 184                 {
 185                         Vector8s res = new Vector8s ();
 186                         short *a = &va.v0;
 187                         short *b = &res.v0;
 188                         for (int i = 0; i < 8; ++i)
 189                                 *b++ = (short)((ushort)(*a++) >> amount);
 190                         return res;
 191                 }
 192
 193                 [Acceleration (AccelMode.SSE2)]
 194                 public static unsafe Vector8s AddWithSaturation (Vector8s va, Vector8s vb) {
 195                         Vector8s res = new Vector8s ();
 196                         short *a = &va.v0;
 197                         short *b = &vb.v0;
 198                         short *c = &res.v0;
 199                         for (int i = 0; i < 8; ++i)
 200                                 *c++ = (short) System.Math.Max (System.Math.Min (*a++ + *b++, short.MaxValue), short.MinValue);
 201                         return res;
 202                 }
 203
 204                 [Acceleration (AccelMode.SSE2)]
 205                 public static unsafe Vector8s SubtractWithSaturation (Vector8s va, Vector8s vb) {
 206                         Vector8s res = new Vector8s ();
 207                         short *a = &va.v0;
 208                         short *b = &vb.v0;
 209                         short *c = &res.v0;
 210                         for (int i = 0; i < 8; ++i)
 211                                 *c++ = (short) System.Math.Max (System.Math.Min (*a++ - *b++, short.MaxValue), short.MinValue); ;
 212                         return res;
 213                 }
 214
 215                 [Acceleration (AccelMode.SSE2)]
 216                 public static unsafe Vector8s Max (Vector8s va, Vector8s vb) {
 217                         Vector8s res = new Vector8s ();
 218                         short *a = &va.v0;
 219                         short *b = &vb.v0;
 220                         short *c = &res.v0;
 221                         for (int i = 0; i < 8; ++i)
 222                                 *c++ = (short) System.Math.Max (*a++, *b++);
 223                         return res;
 224                 }
 225
 226                 [Acceleration (AccelMode.SSE2)]
 227                 public static unsafe Vector8s Min (Vector8s va, Vector8s vb) {
 228                         Vector8s res = new Vector8s ();
 229                         short *a = &va.v0;
 230                         short *b = &vb.v0;
 231                         short *c = &res.v0;
 232                         for (int i = 0; i < 8; ++i)
 233                                 *c++ = (short) System.Math.Min (*a++, *b++);
 234                         return res;
 235                 }
 236
 237                 [Acceleration (AccelMode.SSE2)]
 238                 public static unsafe int ExtractByteMask (Vector8s va) {
 239                         int res = 0;
 240                         byte *a = (byte*)&va;
 241                         for (int i = 0; i < 16; ++i)
 242                                 res |= (*a++ & 0x80) >> 7 << i;
 243                         return res;
 244                 }
 245
 246                 [Acceleration (AccelMode.SSE2)]
 247                 public static unsafe Vector8s ShuffleHigh (Vector8s va, ShuffleSel sel)
 248                 {
 249                         short *ptr = ((short*)&va) + 4;
 250                         int idx = (int)sel;
 251                         return new Vector8s (va.v0, va.v1, va.v2, va.v3, *(ptr + ((idx >> 0) & 0x3)), *(ptr + ((idx >> 2) & 0x3)), *(ptr + ((idx >> 4) & 0x3)), *(ptr + ((idx >> 6) & 0x3)));
 252                 }
 253
 254                 [Acceleration (AccelMode.SSE2)]
 255                 public static unsafe Vector8s ShuffleLow (Vector8s va, ShuffleSel sel)
 256                 {
 257                         short *ptr = ((short*)&va);
 258                         int idx = (int)sel;
 259                         return new Vector8s (*(ptr + ((idx >> 0) & 0x3)), *(ptr + ((idx >> 2) & 0x3)), *(ptr + ((idx >> 4) & 0x3)), *(ptr + ((idx >> 6) & 0x3)), va.v4, va.v5, va.v6, va.v7);
 260                 }
 261
 262                 [Acceleration (AccelMode.SSE2)]
 263                 public static unsafe Vector8s CompareEqual (Vector8s va, Vector8s vb) {
 264                         Vector8s res = new Vector8s ();
 265                         short *a = &va.v0;
 266                         short *b = &vb.v0;
 267                         short *c = &res.v0;
 268                         for (int i = 0; i < 8; ++i)
 269                                 *c++ = (short) (*a++ == *b++ ? -1 : 0);
 270                         return res;
 271                 }
 272
 273                 [Acceleration (AccelMode.SSE2)]
 274                 public static unsafe Vector8s CompareGreaterThan (Vector8s va, Vector8s vb) {
 275                         Vector8s res = new Vector8s ();
 276                         short *a = &va.v0;
 277                         short *b = &vb.v0;
 278                         short *c = &res.v0;
 279                         for (int i = 0; i < 8; ++i)
 280                                 *c++ = (short) (*a++ > *b++ ? -1 : 0);
 281                         return res;
 282                 }
 283
 284                 [Acceleration (AccelMode.SSE2)]
 285                 public static unsafe Vector8s MultiplyStoreHigh (Vector8s va, Vector8s vb) {
 286                         Vector8s res = new Vector8s ();
 287                         short *a = &va.v0;
 288                         short *b = &vb.v0;
 289                         short *c = &res.v0;
 290                         for (int i = 0; i < 8; ++i)
 291                                 *c++ = (short)((int)*a++ * (int)*b++ >> 16);
 292                         return res;
 293                 }
 294
 295                 [Acceleration (AccelMode.SSE2)]
 296                 public static unsafe Vector16b PackWithUnsignedSaturation (Vector8s va, Vector8s vb) {
 297                         Vector16b res = new Vector16b ();
 298                         short *a = (short*)&va;
 299                         short *b = (short*)&vb;
 300                         byte *c = (byte*)&res;
 301                         for (int i = 0; i < 8; ++i)
 302                                 *c++ = (byte)System.Math.Max (0, System.Math.Min ((int)*a++, byte.MaxValue));
 303                         for (int i = 0; i < 8; ++i)
 304                                 *c++ = (byte)System.Math.Max (0, System.Math.Min ((int)*b++, byte.MaxValue));
 305                         return res;
 306                 }
 307
 308                 [CLSCompliant(false)]
 309                 [Acceleration (AccelMode.SSE2)]
 310                 public static unsafe Vector16sb PackWithSignedSaturation (Vector8s va, Vector8s vb) {
 311                         Vector16sb res = new Vector16sb ();
 312                         short *a = (short*)&va;
 313                         short *b = (short*)&vb;
 314                         sbyte *c = (sbyte*)&res;
 315                         for (int i = 0; i < 8; ++i)
 316                                 *c++ = (sbyte)System.Math.Max (System.Math.Min ((int)*a++, sbyte.MaxValue), sbyte.MinValue);
 317                         for (int i = 0; i < 8; ++i)
 318                                 *c++ = (sbyte)System.Math.Max (System.Math.Min ((int)*b++, sbyte.MaxValue), sbyte.MinValue);
 319                         return res;
 320                 }
 321
 322                 [Acceleration (AccelMode.SSE1)]
 323                 public static unsafe explicit operator Vector2d (Vector8s v)
 324                 {
 325                         Vector2d* p = (Vector2d*)&v;
 326                         return *p;
 327                 }
 328
 329                 [Acceleration (AccelMode.SSE1)]
 330                 public static unsafe explicit operator Vector4f (Vector8s v)
 331                 {
 332                         Vector4f* p = (Vector4f*)&v;
 333                         return *p;
 334                 }
 335
 336                 [Acceleration (AccelMode.SSE1)]
 337                 public static unsafe explicit operator Vector2l (Vector8s v)
 338                 {
 339                         Vector2l* p = (Vector2l*)&v;
 340                         return *p;
 341                 }
 342
 343                 [Acceleration (AccelMode.SSE1)]
 344                 [CLSCompliant(false)]
 345                 public static unsafe explicit operator Vector2ul (Vector8s v)
 346                 {
 347                         Vector2ul* p = (Vector2ul*)&v;
 348                         return *p;
 349                 }
 350
 351                 [Acceleration (AccelMode.SSE1)]
 352                 public static unsafe explicit operator Vector4i (Vector8s v)
 353                 {
 354                         Vector4i* p = (Vector4i*)&v;
 355                         return *p;
 356                 }
 357
 358                 [Acceleration (AccelMode.SSE1)]
 359                 [CLSCompliant(false)]
 360                 public static unsafe explicit operator Vector4ui (Vector8s v)
 361                 {
 362                         Vector4ui* p = (Vector4ui*)&v;
 363                         return *p;
 364                 }
 365
 366                 [Acceleration (AccelMode.SSE1)]
 367                 [CLSCompliant(false)]
 368                 public static unsafe explicit operator Vector8us (Vector8s v)
 369                 {
 370                         Vector8us* p = (Vector8us*)&v;
 371                         return *p;
 372                 }
 373
 374                 [Acceleration (AccelMode.SSE1)]
 375                 [CLSCompliant(false)]
 376                 public static unsafe explicit operator Vector16sb (Vector8s v)
 377                 {
 378                         Vector16sb* p = (Vector16sb*)&v;
 379                         return *p;
 380                 }
 381
 382                 [Acceleration (AccelMode.SSE1)]
 383                 public static unsafe explicit operator Vector16b (Vector8s v)
 384                 {
 385                         Vector16b* p = (Vector16b*)&v;
 386                         return *p;
 387                 }
 388
 389
 390                 [Acceleration (AccelMode.SSE1)]
 391                 public static Vector8s LoadAligned (ref Vector8s v)
 392                 {
 393                         return v;
 394                 }
 395
 396                 [Acceleration (AccelMode.SSE1)]
 397                 public static void StoreAligned (ref Vector8s res, Vector8s val)
 398                 {
 399                         res = val;
 400                 }
 401
 402                 [CLSCompliant(false)]
 403                 [Acceleration (AccelMode.SSE1)]
 404                 public static unsafe Vector8s LoadAligned (Vector8s *v)
 405                 {
 406                         return *v;
 407                 }
 408
 409                 [CLSCompliant(false)]
 410                 [Acceleration (AccelMode.SSE1)]
 411                 public static unsafe void StoreAligned (Vector8s *res, Vector8s val)
 412                 {
 413                         *res = val;
 414                 }
 415
 416                 [Acceleration (AccelMode.SSE1)]
 417                 [CLSCompliant(false)]
 418                 public static void PrefetchTemporalAllCacheLevels (ref Vector8s res)
 419                 {
 420                 }
 421
 422                 [Acceleration (AccelMode.SSE1)]
 423                 [CLSCompliant(false)]
 424                 public static void PrefetchTemporal1stLevelCache (ref Vector8s res)
 425                 {
 426                 }
 427
 428                 [Acceleration (AccelMode.SSE1)]
 429                 [CLSCompliant(false)]
 430                 public static void PrefetchTemporal2ndLevelCache (ref Vector8s res)
 431                 {
 432                 }
 433
 434                 [Acceleration (AccelMode.SSE1)]
 435                 [CLSCompliant(false)]
 436                 public static void PrefetchNonTemporal (ref Vector8s res)
 437                 {
 438                 }
 439
 440                 [Acceleration (AccelMode.SSE1)]
 441                 [CLSCompliant(false)]
 442                 public static unsafe void PrefetchTemporalAllCacheLevels (Vector8s *res)
 443                 {
 444                 }
 445
 446                 [Acceleration (AccelMode.SSE1)]
 447                 [CLSCompliant(false)]
 448                 public static unsafe void PrefetchTemporal1stLevelCache (Vector8s *res)
 449                 {
 450                 }
 451
 452                 [Acceleration (AccelMode.SSE1)]
 453                 [CLSCompliant(false)]
 454                 public static unsafe void PrefetchTemporal2ndLevelCache (Vector8s *res)
 455                 {
 456                 }
 457
 458                 [Acceleration (AccelMode.SSE1)]
 459                 [CLSCompliant(false)]
 460                 public static unsafe void PrefetchNonTemporal (Vector8s *res)
 461                 {
 462                 }
 463         }
 464 }