src/gallium/drivers/nouveau/codegen/nv50_ir_target_nvc0.cpp

   1 /*
   2  * Copyright 2011 Christoph Bumiller
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  */
  22
  23 #include "codegen/nv50_ir_target_nvc0.h"
  24
  25 namespace nv50_ir {
  26
  27 Target *getTargetNVC0(unsigned int chipset)
  28 {
  29    return new TargetNVC0(chipset);
  30 }
  31
  32 TargetNVC0::TargetNVC0(unsigned int card) :
  33    Target(card < 0x110, false, card >= 0xe4)
  34 {
  35    chipset = card;
  36    initOpInfo();
  37 }
  38
  39 // BULTINS / LIBRARY FUNCTIONS:
  40
  41 // lazyness -> will just hardcode everything for the time being
  42
  43 #include "lib/gf100.asm.h"
  44 #include "lib/gk104.asm.h"
  45 #include "lib/gk110.asm.h"
  46
  47 void
  48 TargetNVC0::getBuiltinCode(const uint32_t **code, uint32_t *size) const
  49 {
  50    switch (chipset & ~0xf) {
  51    case 0xe0:
  52       if (chipset < NVISA_GK20A_CHIPSET) {
  53          *code = (const uint32_t *)&gk104_builtin_code[0];
  54          *size = sizeof(gk104_builtin_code);
  55          break;
  56       }
  57       /* fall-through for GK20A */
  58    case 0xf0:
  59    case 0x100:
  60       *code = (const uint32_t *)&gk110_builtin_code[0];
  61       *size = sizeof(gk110_builtin_code);
  62       break;
  63    default:
  64       *code = (const uint32_t *)&gf100_builtin_code[0];
  65       *size = sizeof(gf100_builtin_code);
  66       break;
  67    }
  68 }
  69
  70 uint32_t
  71 TargetNVC0::getBuiltinOffset(int builtin) const
  72 {
  73    assert(builtin < NVC0_BUILTIN_COUNT);
  74
  75    switch (chipset & ~0xf) {
  76    case 0xe0:
  77       if (chipset < NVISA_GK20A_CHIPSET)
  78          return gk104_builtin_offsets[builtin];
  79       /* fall-through for GK20A */
  80    case 0xf0:
  81    case 0x100:
  82       return gk110_builtin_offsets[builtin];
  83    default:
  84       return gf100_builtin_offsets[builtin];
  85    }
  86 }
  87
  88 struct opProperties
  89 {
  90    operation op;
  91    unsigned int mNeg   : 4;
  92    unsigned int mAbs   : 4;
  93    unsigned int mNot   : 4;
  94    unsigned int mSat   : 4;
  95    unsigned int fConst : 3;
  96    unsigned int fImmd  : 4; // last bit indicates if full immediate is suppoted
  97 };
  98
  99 static const struct opProperties _initProps[] =
 100 {
 101    //           neg  abs  not  sat  c[]  imm
 102    { OP_ADD,    0x3, 0x3, 0x0, 0x8, 0x2, 0x2 | 0x8 },
 103    { OP_SUB,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 | 0x8 },
 104    { OP_MUL,    0x3, 0x0, 0x0, 0x8, 0x2, 0x2 | 0x8 },
 105    { OP_MAX,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
 106    { OP_MIN,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
 107    { OP_MAD,    0x7, 0x0, 0x0, 0x8, 0x6, 0x2 | 0x8 }, // special c[] constraint
 108    { OP_MADSP,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
 109    { OP_ABS,    0x0, 0x0, 0x0, 0x0, 0x1, 0x0 },
 110    { OP_NEG,    0x0, 0x1, 0x0, 0x0, 0x1, 0x0 },
 111    { OP_CVT,    0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
 112    { OP_CEIL,   0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
 113    { OP_FLOOR,  0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
 114    { OP_TRUNC,  0x1, 0x1, 0x0, 0x8, 0x1, 0x0 },
 115    { OP_AND,    0x0, 0x0, 0x3, 0x0, 0x2, 0x2 | 0x8 },
 116    { OP_OR,     0x0, 0x0, 0x3, 0x0, 0x2, 0x2 | 0x8 },
 117    { OP_XOR,    0x0, 0x0, 0x3, 0x0, 0x2, 0x2 | 0x8 },
 118    { OP_SHL,    0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
 119    { OP_SHR,    0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
 120    { OP_SET,    0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
 121    { OP_SLCT,   0x4, 0x0, 0x0, 0x0, 0x6, 0x2 }, // special c[] constraint
 122    { OP_PREEX2, 0x1, 0x1, 0x0, 0x0, 0x1, 0x1 },
 123    { OP_PRESIN, 0x1, 0x1, 0x0, 0x0, 0x1, 0x1 },
 124    { OP_COS,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
 125    { OP_SIN,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
 126    { OP_EX2,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
 127    { OP_LG2,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
 128    { OP_RCP,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
 129    { OP_RSQ,    0x1, 0x1, 0x0, 0x8, 0x0, 0x0 },
 130    { OP_DFDX,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0 },
 131    { OP_DFDY,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0 },
 132    { OP_CALL,   0x0, 0x0, 0x0, 0x0, 0x1, 0x0 },
 133    { OP_POPCNT, 0x0, 0x0, 0x3, 0x0, 0x2, 0x2 },
 134    { OP_INSBF,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
 135    { OP_EXTBF,  0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
 136    { OP_BFIND,  0x0, 0x0, 0x1, 0x0, 0x1, 0x1 },
 137    { OP_PERMT,  0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
 138    { OP_SET_AND, 0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
 139    { OP_SET_OR, 0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
 140    { OP_SET_XOR, 0x3, 0x3, 0x0, 0x0, 0x2, 0x2 },
 141    // saturate only:
 142    { OP_LINTERP, 0x0, 0x0, 0x0, 0x8, 0x0, 0x0 },
 143    { OP_PINTERP, 0x0, 0x0, 0x0, 0x8, 0x0, 0x0 },
 144    // nve4 ops:
 145    { OP_SULDB,   0x0, 0x0, 0x0, 0x0, 0x2, 0x0 },
 146    { OP_SUSTB,   0x0, 0x0, 0x0, 0x0, 0x2, 0x0 },
 147    { OP_SUSTP,   0x0, 0x0, 0x0, 0x0, 0x2, 0x0 },
 148    { OP_SUCLAMP, 0x0, 0x0, 0x0, 0x0, 0x2, 0x2 },
 149    { OP_SUBFM,   0x0, 0x0, 0x0, 0x0, 0x6, 0x2 },
 150    { OP_SUEAU,   0x0, 0x0, 0x0, 0x0, 0x6, 0x2 }
 151 };
 152
 153 void TargetNVC0::initOpInfo()
 154 {
 155    unsigned int i, j;
 156
 157    static const uint32_t commutative[(OP_LAST + 31) / 32] =
 158    {
 159       // ADD, MAD, MUL, AND, OR, XOR, MAX, MIN
 160       0x0670ca00, 0x0000003f, 0x00000000, 0x00000000
 161    };
 162
 163    static const uint32_t shortForm[(OP_LAST + 31) / 32] =
 164    {
 165       // ADD, MAD, MUL, AND, OR, XOR, PRESIN, PREEX2, SFN, CVT, PINTERP, MOV
 166       0x0670ca00, 0x00000000, 0x00000000, 0x00000000
 167    };
 168
 169    static const operation noDest[] =
 170    {
 171       OP_STORE, OP_WRSV, OP_EXPORT, OP_BRA, OP_CALL, OP_RET, OP_EXIT,
 172       OP_DISCARD, OP_CONT, OP_BREAK, OP_PRECONT, OP_PREBREAK, OP_PRERET,
 173       OP_JOIN, OP_JOINAT, OP_BRKPT, OP_MEMBAR, OP_EMIT, OP_RESTART,
 174       OP_QUADON, OP_QUADPOP, OP_TEXBAR, OP_SUSTB, OP_SUSTP, OP_SUREDP,
 175       OP_SUREDB, OP_BAR
 176    };
 177
 178    static const operation noPred[] =
 179    {
 180       OP_CALL, OP_PRERET, OP_QUADON, OP_QUADPOP,
 181       OP_JOINAT, OP_PREBREAK, OP_PRECONT, OP_BRKPT
 182    };
 183
 184    for (i = 0; i < DATA_FILE_COUNT; ++i)
 185       nativeFileMap[i] = (DataFile)i;
 186    nativeFileMap[FILE_ADDRESS] = FILE_GPR;
 187
 188    for (i = 0; i < OP_LAST; ++i) {
 189       opInfo[i].variants = NULL;
 190       opInfo[i].op = (operation)i;
 191       opInfo[i].srcTypes = 1 << (int)TYPE_F32;
 192       opInfo[i].dstTypes = 1 << (int)TYPE_F32;
 193       opInfo[i].immdBits = 0;
 194       opInfo[i].srcNr = operationSrcNr[i];
 195
 196       for (j = 0; j < opInfo[i].srcNr; ++j) {
 197          opInfo[i].srcMods[j] = 0;
 198          opInfo[i].srcFiles[j] = 1 << (int)FILE_GPR;
 199       }
 200       opInfo[i].dstMods = 0;
 201       opInfo[i].dstFiles = 1 << (int)FILE_GPR;
 202
 203       opInfo[i].hasDest = 1;
 204       opInfo[i].vector = (i >= OP_TEX && i <= OP_TEXCSAA);
 205       opInfo[i].commutative = (commutative[i / 32] >> (i % 32)) & 1;
 206       opInfo[i].pseudo = (i < OP_MOV);
 207       opInfo[i].predicate = !opInfo[i].pseudo;
 208       opInfo[i].flow = (i >= OP_BRA && i <= OP_JOIN);
 209       opInfo[i].minEncSize = (shortForm[i / 32] & (1 << (i % 32))) ? 4 : 8;
 210    }
 211    for (i = 0; i < sizeof(noDest) / sizeof(noDest[0]); ++i)
 212       opInfo[noDest[i]].hasDest = 0;
 213    for (i = 0; i < sizeof(noPred) / sizeof(noPred[0]); ++i)
 214       opInfo[noPred[i]].predicate = 0;
 215
 216    for (i = 0; i < sizeof(_initProps) / sizeof(_initProps[0]); ++i) {
 217       const struct opProperties *prop = &_initProps[i];
 218
 219       for (int s = 0; s < 3; ++s) {
 220          if (prop->mNeg & (1 << s))
 221             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NEG;
 222          if (prop->mAbs & (1 << s))
 223             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_ABS;
 224          if (prop->mNot & (1 << s))
 225             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NOT;
 226          if (prop->fConst & (1 << s))
 227             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_CONST;
 228          if (prop->fImmd & (1 << s))
 229             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_IMMEDIATE;
 230          if (prop->fImmd & 8)
 231             opInfo[prop->op].immdBits = 0xffffffff;
 232       }
 233       if (prop->mSat & 8)
 234          opInfo[prop->op].dstMods = NV50_IR_MOD_SAT;
 235    }
 236 }
 237
 238 unsigned int
 239 TargetNVC0::getFileSize(DataFile file) const
 240 {
 241    switch (file) {
 242    case FILE_NULL:          return 0;
 243    case FILE_GPR:           return (chipset >= NVISA_GK20A_CHIPSET) ? 255 : 63;
 244    case FILE_PREDICATE:     return 7;
 245    case FILE_FLAGS:         return 1;
 246    case FILE_ADDRESS:       return 0;
 247    case FILE_IMMEDIATE:     return 0;
 248    case FILE_MEMORY_CONST:  return 65536;
 249    case FILE_SHADER_INPUT:  return 0x400;
 250    case FILE_SHADER_OUTPUT: return 0x400;
 251    case FILE_MEMORY_GLOBAL: return 0xffffffff;
 252    case FILE_MEMORY_SHARED: return 16 << 10;
 253    case FILE_MEMORY_LOCAL:  return 48 << 10;
 254    case FILE_SYSTEM_VALUE:  return 32;
 255    default:
 256       assert(!"invalid file");
 257       return 0;
 258    }
 259 }
 260
 261 unsigned int
 262 TargetNVC0::getFileUnit(DataFile file) const
 263 {
 264    if (file == FILE_GPR || file == FILE_ADDRESS || file == FILE_SYSTEM_VALUE)
 265       return 2;
 266    return 0;
 267 }
 268
 269 uint32_t
 270 TargetNVC0::getSVAddress(DataFile shaderFile, const Symbol *sym) const
 271 {
 272    const int idx = sym->reg.data.sv.index;
 273    const SVSemantic sv = sym->reg.data.sv.sv;
 274
 275    const bool isInput = shaderFile == FILE_SHADER_INPUT;
 276    const bool kepler = getChipset() >= NVISA_GK104_CHIPSET;
 277
 278    switch (sv) {
 279    case SV_POSITION:       return 0x070 + idx * 4;
 280    case SV_INSTANCE_ID:    return 0x2f8;
 281    case SV_VERTEX_ID:      return 0x2fc;
 282    case SV_PRIMITIVE_ID:   return isInput ? 0x060 : 0x040;
 283    case SV_LAYER:          return 0x064;
 284    case SV_VIEWPORT_INDEX: return 0x068;
 285    case SV_POINT_SIZE:     return 0x06c;
 286    case SV_CLIP_DISTANCE:  return 0x2c0 + idx * 4;
 287    case SV_POINT_COORD:    return 0x2e0 + idx * 4;
 288    case SV_FACE:           return 0x3fc;
 289    case SV_TESS_OUTER:     return 0x000 + idx * 4;
 290    case SV_TESS_INNER:     return 0x010 + idx * 4;
 291    case SV_TESS_COORD:     return 0x2f0 + idx * 4;
 292    case SV_NTID:           return kepler ? (0x00 + idx * 4) : ~0;
 293    case SV_NCTAID:         return kepler ? (0x0c + idx * 4) : ~0;
 294    case SV_GRIDID:         return kepler ? 0x18 : ~0;
 295    case SV_SAMPLE_INDEX:   return 0;
 296    case SV_SAMPLE_POS:     return 0;
 297    case SV_SAMPLE_MASK:    return 0;
 298    default:
 299       return 0xffffffff;
 300    }
 301 }
 302
 303 bool
 304 TargetNVC0::insnCanLoad(const Instruction *i, int s,
 305                         const Instruction *ld) const
 306 {
 307    DataFile sf = ld->src(0).getFile();
 308
 309    // immediate 0 can be represented by GPR $r63/$r255
 310    if (sf == FILE_IMMEDIATE && ld->getSrc(0)->reg.data.u64 == 0)
 311       return (!i->isPseudo() &&
 312               !i->asTex() &&
 313               i->op != OP_EXPORT && i->op != OP_STORE);
 314
 315    if (s >= opInfo[i->op].srcNr)
 316       return false;
 317    if (!(opInfo[i->op].srcFiles[s] & (1 << (int)sf)))
 318       return false;
 319
 320    // indirect loads can only be done by OP_LOAD/VFETCH/INTERP on nvc0
 321    if (ld->src(0).isIndirect(0))
 322       return false;
 323
 324    for (int k = 0; i->srcExists(k); ++k) {
 325       if (i->src(k).getFile() == FILE_IMMEDIATE) {
 326          if (k == 2 && i->op == OP_SUCLAMP) // special case
 327             continue;
 328          if (i->getSrc(k)->reg.data.u64 != 0)
 329             return false;
 330       } else
 331       if (i->src(k).getFile() != FILE_GPR &&
 332           i->src(k).getFile() != FILE_PREDICATE) {
 333          return false;
 334       }
 335    }
 336
 337    // not all instructions support full 32 bit immediates
 338    if (sf == FILE_IMMEDIATE) {
 339       Storage &reg = ld->getSrc(0)->asImm()->reg;
 340
 341       if (opInfo[i->op].immdBits != 0xffffffff || typeSizeof(i->sType) > 4) {
 342          switch (i->sType) {
 343          case TYPE_F64:
 344             if (reg.data.u64 & 0x00000fffffffffffULL)
 345                return false;
 346             break;
 347          case TYPE_F32:
 348             if (reg.data.u32 & 0xfff)
 349                return false;
 350             break;
 351          case TYPE_S32:
 352          case TYPE_U32:
 353             // with u32, 0xfffff counts as 0xffffffff as well
 354             if (reg.data.s32 > 0x7ffff || reg.data.s32 < -0x80000)
 355                return false;
 356             break;
 357          case TYPE_U8:
 358          case TYPE_S8:
 359          case TYPE_U16:
 360          case TYPE_S16:
 361          case TYPE_F16:
 362             break;
 363          default:
 364             return false;
 365          }
 366       } else
 367       if (i->op == OP_MAD || i->op == OP_FMA) {
 368          // requires src == dst, cannot decide before RA
 369          // (except if we implement more constraints)
 370          if (ld->getSrc(0)->asImm()->reg.data.u32 & 0xfff)
 371             return false;
 372       } else
 373       if (i->op == OP_ADD && i->sType == TYPE_F32) {
 374          // add f32 LIMM cannot saturate
 375          if (i->saturate && (reg.data.u32 & 0xfff))
 376             return false;
 377       }
 378    }
 379
 380    return true;
 381 }
 382
 383 bool
 384 TargetNVC0::isAccessSupported(DataFile file, DataType ty) const
 385 {
 386    if (ty == TYPE_NONE)
 387       return false;
 388    if (file == FILE_MEMORY_CONST && getChipset() >= 0xe0) // wrong encoding ?
 389       return typeSizeof(ty) <= 8;
 390    if (ty == TYPE_B96)
 391       return false;
 392    return true;
 393 }
 394
 395 bool
 396 TargetNVC0::isOpSupported(operation op, DataType ty) const
 397 {
 398    if ((op == OP_MAD || op == OP_FMA) && (ty != TYPE_F32))
 399       return false;
 400    if (op == OP_SAD && ty != TYPE_S32 && ty != TYPE_U32)
 401       return false;
 402    if (op == OP_POW || op == OP_SQRT || op == OP_DIV || op == OP_MOD)
 403       return false;
 404    return true;
 405 }
 406
 407 bool
 408 TargetNVC0::isModSupported(const Instruction *insn, int s, Modifier mod) const
 409 {
 410    if (!isFloatType(insn->dType)) {
 411       switch (insn->op) {
 412       case OP_ABS:
 413       case OP_NEG:
 414       case OP_CVT:
 415       case OP_CEIL:
 416       case OP_FLOOR:
 417       case OP_TRUNC:
 418       case OP_AND:
 419       case OP_OR:
 420       case OP_XOR:
 421       case OP_POPCNT:
 422       case OP_BFIND:
 423          break;
 424       case OP_SET:
 425          if (insn->sType != TYPE_F32)
 426             return false;
 427          break;
 428       case OP_ADD:
 429          if (mod.abs())
 430             return false;
 431          if (insn->src(s ? 0 : 1).mod.neg())
 432             return false;
 433          break;
 434       case OP_SUB:
 435          if (s == 0)
 436             return insn->src(1).mod.neg() ? false : true;
 437          break;
 438       default:
 439          return false;
 440       }
 441    }
 442    if (s >= opInfo[insn->op].srcNr || s >= 3)
 443       return false;
 444    return (mod & Modifier(opInfo[insn->op].srcMods[s])) == mod;
 445 }
 446
 447 bool
 448 TargetNVC0::mayPredicate(const Instruction *insn, const Value *pred) const
 449 {
 450    if (insn->getPredicate())
 451       return false;
 452    return opInfo[insn->op].predicate;
 453 }
 454
 455 bool
 456 TargetNVC0::isSatSupported(const Instruction *insn) const
 457 {
 458    if (insn->op == OP_CVT)
 459       return true;
 460    if (!(opInfo[insn->op].dstMods & NV50_IR_MOD_SAT))
 461       return false;
 462
 463    if (insn->dType == TYPE_U32)
 464       return (insn->op == OP_ADD) || (insn->op == OP_MAD);
 465
 466    // add f32 LIMM cannot saturate
 467    if (insn->op == OP_ADD && insn->sType == TYPE_F32) {
 468       if (insn->getSrc(1)->asImm() &&
 469           insn->getSrc(1)->reg.data.u32 & 0xfff)
 470          return false;
 471    }
 472
 473    return insn->dType == TYPE_F32;
 474 }
 475
 476 bool
 477 TargetNVC0::isPostMultiplySupported(operation op, float f, int& e) const
 478 {
 479    if (op != OP_MUL)
 480       return false;
 481    f = fabsf(f);
 482    e = static_cast<int>(log2f(f));
 483    if (e < -3 || e > 3)
 484       return false;
 485    return f == exp2f(static_cast<float>(e));
 486 }
 487
 488 // TODO: better values
 489 // this could be more precise, e.g. depending on the issue-to-read/write delay
 490 // of the depending instruction, but it's good enough
 491 int TargetNVC0::getLatency(const Instruction *i) const
 492 {
 493    if (chipset >= 0xe4) {
 494       if (i->dType == TYPE_F64 || i->sType == TYPE_F64)
 495          return 20;
 496       switch (i->op) {
 497       case OP_LINTERP:
 498       case OP_PINTERP:
 499          return 15;
 500       case OP_LOAD:
 501          if (i->src(0).getFile() == FILE_MEMORY_CONST)
 502             return 9;
 503          // fall through
 504       case OP_VFETCH:
 505          return 24;
 506       default:
 507          if (Target::getOpClass(i->op) == OPCLASS_TEXTURE)
 508             return 17;
 509          if (i->op == OP_MUL && i->dType != TYPE_F32)
 510             return 15;
 511          return 9;
 512       }
 513    } else {
 514       if (i->op == OP_LOAD) {
 515          if (i->cache == CACHE_CV)
 516             return 700;
 517          return 48;
 518       }
 519       return 24;
 520    }
 521    return 32;
 522 }
 523
 524 // These are "inverse" throughput values, i.e. the number of cycles required
 525 // to issue a specific instruction for a full warp (32 threads).
 526 //
 527 // Assuming we have more than 1 warp in flight, a higher issue latency results
 528 // in a lower result latency since the MP will have spent more time with other
 529 // warps.
 530 // This also helps to determine the number of cycles between instructions in
 531 // a single warp.
 532 //
 533 int TargetNVC0::getThroughput(const Instruction *i) const
 534 {
 535    // TODO: better values
 536    if (i->dType == TYPE_F32) {
 537       switch (i->op) {
 538       case OP_ADD:
 539       case OP_MUL:
 540       case OP_MAD:
 541       case OP_FMA:
 542          return 1;
 543       case OP_CVT:
 544       case OP_CEIL:
 545       case OP_FLOOR:
 546       case OP_TRUNC:
 547       case OP_SET:
 548       case OP_SLCT:
 549       case OP_MIN:
 550       case OP_MAX:
 551          return 2;
 552       case OP_RCP:
 553       case OP_RSQ:
 554       case OP_LG2:
 555       case OP_SIN:
 556       case OP_COS:
 557       case OP_PRESIN:
 558       case OP_PREEX2:
 559       default:
 560          return 8;
 561       }
 562    } else
 563    if (i->dType == TYPE_U32 || i->dType == TYPE_S32) {
 564       switch (i->op) {
 565       case OP_ADD:
 566       case OP_AND:
 567       case OP_OR:
 568       case OP_XOR:
 569       case OP_NOT:
 570          return 1;
 571       case OP_MUL:
 572       case OP_MAD:
 573       case OP_CVT:
 574       case OP_SET:
 575       case OP_SLCT:
 576       case OP_SHL:
 577       case OP_SHR:
 578       case OP_NEG:
 579       case OP_ABS:
 580       case OP_MIN:
 581       case OP_MAX:
 582       default:
 583          return 2;
 584       }
 585    } else
 586    if (i->dType == TYPE_F64) {
 587       return 2;
 588    } else {
 589       return 1;
 590    }
 591 }
 592
 593 bool TargetNVC0::canDualIssue(const Instruction *a, const Instruction *b) const
 594 {
 595    const OpClass clA = operationClass[a->op];
 596    const OpClass clB = operationClass[b->op];
 597
 598    if (getChipset() >= 0xe4) {
 599       // not texturing
 600       // not if the 2nd instruction isn't necessarily executed
 601       if (clA == OPCLASS_TEXTURE || clA == OPCLASS_FLOW)
 602          return false;
 603       // anything with MOV
 604       if (a->op == OP_MOV || b->op == OP_MOV)
 605          return true;
 606       if (clA == clB) {
 607          // only F32 arith or integer additions
 608          if (clA != OPCLASS_ARITH)
 609             return false;
 610          return (a->dType == TYPE_F32 || a->op == OP_ADD ||
 611                  b->dType == TYPE_F32 || b->op == OP_ADD);
 612       }
 613       // nothing with TEXBAR
 614       if (a->op == OP_TEXBAR || b->op == OP_TEXBAR)
 615          return false;
 616       // no loads and stores accessing the the same space
 617       if ((clA == OPCLASS_LOAD && clB == OPCLASS_STORE) ||
 618           (clB == OPCLASS_LOAD && clA == OPCLASS_STORE))
 619          if (a->src(0).getFile() == b->src(0).getFile())
 620             return false;
 621       // no > 32-bit ops
 622       if (typeSizeof(a->dType) > 4 || typeSizeof(b->dType) > 4 ||
 623           typeSizeof(a->sType) > 4 || typeSizeof(b->sType) > 4)
 624          return false;
 625       return true;
 626    } else {
 627       return false; // info not needed (yet)
 628    }
 629 }
 630
 631 } // namespace nv50_ir