src/gallium/drivers/nouveau/codegen/nv50_ir_target_nv50.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_nv50.h"
  24
  25 namespace nv50_ir {
  26
  27 Target *getTargetNV50(unsigned int chipset)
  28 {
  29    return new TargetNV50(chipset);
  30 }
  31
  32 TargetNV50::TargetNV50(unsigned int card) : Target(true, false)
  33 {
  34    chipset = card;
  35
  36    wposMask = 0;
  37    for (unsigned int i = 0; i <= SV_LAST; ++i)
  38       sysvalLocation[i] = ~0;
  39
  40    initOpInfo();
  41 }
  42
  43 #if 0
  44 // BULTINS / LIBRARY FUNCTIONS:
  45
  46 // TODO
  47 static const uint32_t nvc0_builtin_code[] =
  48 {
  49 };
  50
  51 static const uint16_t nvc0_builtin_offsets[NV50_BUILTIN_COUNT] =
  52 {
  53 };
  54 #endif
  55
  56 void
  57 TargetNV50::getBuiltinCode(const uint32_t **code, uint32_t *size) const
  58 {
  59    *code = NULL;
  60    *size = 0;
  61 }
  62
  63 uint32_t
  64 TargetNV50::getBuiltinOffset(int builtin) const
  65 {
  66    return 0;
  67 }
  68
  69 struct opProperties
  70 {
  71    operation op;
  72    unsigned int mNeg    : 4;
  73    unsigned int mAbs    : 4;
  74    unsigned int mNot    : 4;
  75    unsigned int mSat    : 4;
  76    unsigned int fConst  : 3;
  77    unsigned int fShared : 3;
  78    unsigned int fAttrib : 3;
  79    unsigned int fImm    : 3;
  80 };
  81
  82 static const struct opProperties _initProps[] =
  83 {
  84    //           neg  abs  not  sat  c[]  s[], a[], imm
  85    { OP_ADD,    0x3, 0x0, 0x0, 0x8, 0x2, 0x1, 0x1, 0x2 },
  86    { OP_SUB,    0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 },
  87    { OP_MUL,    0x3, 0x0, 0x0, 0x0, 0x2, 0x1, 0x1, 0x2 },
  88    { OP_MAX,    0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
  89    { OP_MIN,    0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
  90    { OP_MAD,    0x7, 0x0, 0x0, 0x0, 0x6, 0x1, 0x1, 0x0 }, // special constraint
  91    { OP_ABS,    0x0, 0x0, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 },
  92    { OP_NEG,    0x0, 0x1, 0x0, 0x0, 0x0, 0x1, 0x1, 0x0 },
  93    { OP_CVT,    0x1, 0x1, 0x0, 0x8, 0x0, 0x1, 0x1, 0x0 },
  94    { OP_AND,    0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
  95    { OP_OR,     0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
  96    { OP_XOR,    0x0, 0x0, 0x3, 0x0, 0x0, 0x0, 0x0, 0x2 },
  97    { OP_SHL,    0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 },
  98    { OP_SHR,    0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x2 },
  99    { OP_SET,    0x3, 0x3, 0x0, 0x0, 0x2, 0x1, 0x1, 0x0 },
 100    { OP_PREEX2, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 101    { OP_PRESIN, 0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 102    { OP_LG2,    0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 103    { OP_RCP,    0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 104    { OP_RSQ,    0x1, 0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 105    { OP_DFDX,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 106    { OP_DFDY,   0x1, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0, 0x0 },
 107 };
 108
 109 void TargetNV50::initOpInfo()
 110 {
 111    unsigned int i, j;
 112
 113    static const uint32_t commutative[(OP_LAST + 31) / 32] =
 114    {
 115       // ADD,MAD,MUL,AND,OR,XOR,MAX,MIN
 116       0x0670ca00, 0x0000003f, 0x00000000, 0x00000000
 117    };
 118    static const uint32_t shortForm[(OP_LAST + 31) / 32] =
 119    {
 120       // MOV,ADD,SUB,MUL,SAD,L/PINTERP,RCP,TEX,TXF
 121       0x00010e40, 0x00000040, 0x00000498, 0x00000000
 122    };
 123    static const operation noDestList[] =
 124    {
 125       OP_STORE, OP_WRSV, OP_EXPORT, OP_BRA, OP_CALL, OP_RET, OP_EXIT,
 126       OP_DISCARD, OP_CONT, OP_BREAK, OP_PRECONT, OP_PREBREAK, OP_PRERET,
 127       OP_JOIN, OP_JOINAT, OP_BRKPT, OP_MEMBAR, OP_EMIT, OP_RESTART,
 128       OP_QUADON, OP_QUADPOP, OP_TEXBAR, OP_SUSTB, OP_SUSTP, OP_SUREDP,
 129       OP_SUREDB, OP_BAR
 130    };
 131    static const operation noPredList[] =
 132    {
 133       OP_CALL, OP_PREBREAK, OP_PRERET, OP_QUADON, OP_QUADPOP, OP_JOINAT
 134    };
 135
 136    for (i = 0; i < DATA_FILE_COUNT; ++i)
 137       nativeFileMap[i] = (DataFile)i;
 138    nativeFileMap[FILE_PREDICATE] = FILE_FLAGS;
 139
 140    for (i = 0; i < OP_LAST; ++i) {
 141       opInfo[i].variants = NULL;
 142       opInfo[i].op = (operation)i;
 143       opInfo[i].srcTypes = 1 << (int)TYPE_F32;
 144       opInfo[i].dstTypes = 1 << (int)TYPE_F32;
 145       opInfo[i].immdBits = 0xffffffff;
 146       opInfo[i].srcNr = operationSrcNr[i];
 147
 148       for (j = 0; j < opInfo[i].srcNr; ++j) {
 149          opInfo[i].srcMods[j] = 0;
 150          opInfo[i].srcFiles[j] = 1 << (int)FILE_GPR;
 151       }
 152       opInfo[i].dstMods = 0;
 153       opInfo[i].dstFiles = 1 << (int)FILE_GPR;
 154
 155       opInfo[i].hasDest = 1;
 156       opInfo[i].vector = (i >= OP_TEX && i <= OP_TEXCSAA);
 157       opInfo[i].commutative = (commutative[i / 32] >> (i % 32)) & 1;
 158       opInfo[i].pseudo = (i < OP_MOV);
 159       opInfo[i].predicate = !opInfo[i].pseudo;
 160       opInfo[i].flow = (i >= OP_BRA && i <= OP_JOIN);
 161       opInfo[i].minEncSize = (shortForm[i / 32] & (1 << (i % 32))) ? 4 : 8;
 162    }
 163    for (i = 0; i < sizeof(noDestList) / sizeof(noDestList[0]); ++i)
 164       opInfo[noDestList[i]].hasDest = 0;
 165    for (i = 0; i < sizeof(noPredList) / sizeof(noPredList[0]); ++i)
 166       opInfo[noPredList[i]].predicate = 0;
 167
 168    for (i = 0; i < sizeof(_initProps) / sizeof(_initProps[0]); ++i) {
 169       const struct opProperties *prop = &_initProps[i];
 170
 171       for (int s = 0; s < 3; ++s) {
 172          if (prop->mNeg & (1 << s))
 173             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NEG;
 174          if (prop->mAbs & (1 << s))
 175             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_ABS;
 176          if (prop->mNot & (1 << s))
 177             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NOT;
 178          if (prop->fConst & (1 << s))
 179             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_CONST;
 180          if (prop->fShared & (1 << s))
 181             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_SHARED;
 182          if (prop->fAttrib & (1 << s))
 183             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_SHADER_INPUT;
 184          if (prop->fImm & (1 << s))
 185             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_IMMEDIATE;
 186       }
 187       if (prop->mSat & 8)
 188          opInfo[prop->op].dstMods = NV50_IR_MOD_SAT;
 189    }
 190 }
 191
 192 unsigned int
 193 TargetNV50::getFileSize(DataFile file) const
 194 {
 195    switch (file) {
 196    case FILE_NULL:          return 0;
 197    case FILE_GPR:           return 256; // in 16-bit units **
 198    case FILE_PREDICATE:     return 0;
 199    case FILE_FLAGS:         return 4;
 200    case FILE_ADDRESS:       return 4;
 201    case FILE_IMMEDIATE:     return 0;
 202    case FILE_MEMORY_CONST:  return 65536;
 203    case FILE_SHADER_INPUT:  return 0x200;
 204    case FILE_SHADER_OUTPUT: return 0x200;
 205    case FILE_MEMORY_GLOBAL: return 0xffffffff;
 206    case FILE_MEMORY_SHARED: return 16 << 10;
 207    case FILE_MEMORY_LOCAL:  return 48 << 10;
 208    case FILE_SYSTEM_VALUE:  return 16;
 209    default:
 210       assert(!"invalid file");
 211       return 0;
 212    }
 213    // ** only first 128 units encodable for 16-bit regs
 214 }
 215
 216 unsigned int
 217 TargetNV50::getFileUnit(DataFile file) const
 218 {
 219    if (file == FILE_GPR || file == FILE_ADDRESS)
 220       return 1;
 221    if (file == FILE_SYSTEM_VALUE)
 222       return 2;
 223    return 0;
 224 }
 225
 226 uint32_t
 227 TargetNV50::getSVAddress(DataFile shaderFile, const Symbol *sym) const
 228 {
 229    switch (sym->reg.data.sv.sv) {
 230    case SV_FACE:
 231       return 0x3fc;
 232    case SV_POSITION:
 233    {
 234       uint32_t addr = sysvalLocation[sym->reg.data.sv.sv];
 235       for (int c = 0; c < sym->reg.data.sv.index; ++c)
 236          if (wposMask & (1 << c))
 237             addr += 4;
 238       return addr;
 239    }
 240    case SV_NCTAID:
 241       return 0x8 + 2 * sym->reg.data.sv.index;
 242    case SV_CTAID:
 243       return 0xc + 2 * sym->reg.data.sv.index;
 244    case SV_NTID:
 245       return 0x2 + 2 * sym->reg.data.sv.index;
 246    case SV_TID:
 247       return 0;
 248    default:
 249       return sysvalLocation[sym->reg.data.sv.sv];
 250    }
 251 }
 252
 253 // long:  rrr, arr, rcr, acr, rrc, arc, gcr, grr
 254 // short: rr, ar, rc, gr
 255 // immd:  ri, gi
 256 bool
 257 TargetNV50::insnCanLoad(const Instruction *i, int s,
 258                         const Instruction *ld) const
 259 {
 260    DataFile sf = ld->src(0).getFile();
 261
 262    if (sf == FILE_IMMEDIATE && (i->predSrc >= 0 || i->flagsDef >= 0))
 263       return false;
 264    if (s >= opInfo[i->op].srcNr)
 265       return false;
 266    if (!(opInfo[i->op].srcFiles[s] & (1 << (int)sf)))
 267       return false;
 268    if (s == 2 && i->src(1).getFile() != FILE_GPR)
 269       return false;
 270
 271    // NOTE: don't rely on flagsDef
 272    for (int d = 0; i->defExists(d); ++d)
 273       if (i->def(d).getFile() == FILE_FLAGS)
 274          return false;
 275
 276    unsigned mode = 0;
 277
 278    for (int z = 0; z < Target::operationSrcNr[i->op]; ++z) {
 279       DataFile zf = (z == s) ? sf : i->src(z).getFile();
 280       switch (zf) {
 281       case FILE_GPR:
 282          break;
 283       case FILE_MEMORY_SHARED:
 284       case FILE_SHADER_INPUT:
 285          mode |= 1 << (z * 2);
 286          break;
 287       case FILE_MEMORY_CONST:
 288          mode |= 2 << (z * 2);
 289          break;
 290       case FILE_IMMEDIATE:
 291          mode |= 3 << (z * 2);
 292       default:
 293          break;
 294       }
 295    }
 296
 297    switch (mode) {
 298    case 0x00:
 299    case 0x01:
 300    case 0x03:
 301    case 0x08:
 302    case 0x09:
 303    case 0x0c:
 304    case 0x20:
 305    case 0x21:
 306       break;
 307    case 0x0d:
 308       if (ld->bb->getProgram()->getType() != Program::TYPE_GEOMETRY)
 309          return false;
 310    default:
 311       return false;
 312    }
 313
 314    uint8_t ldSize;
 315
 316    if ((i->op == OP_MUL || i->op == OP_MAD) && !isFloatType(i->dType)) {
 317       // 32-bit MUL will be split into 16-bit MULs
 318       if (ld->src(0).isIndirect(0))
 319          return false;
 320       if (sf == FILE_IMMEDIATE)
 321          return false;
 322       ldSize = 2;
 323    } else {
 324       ldSize = typeSizeof(ld->dType);
 325    }
 326
 327    if (sf == FILE_IMMEDIATE)
 328       return true;
 329
 330
 331    // Check if memory access is encodable:
 332
 333    if (ldSize < 4 && sf == FILE_SHADER_INPUT) // no < 4-byte aligned a[] access
 334       return false;
 335    if (ld->getSrc(0)->reg.data.offset > (int32_t)(127 * ldSize))
 336       return false;
 337
 338    if (ld->src(0).isIndirect(0)) {
 339       for (int z = 0; i->srcExists(z); ++z)
 340          if (i->src(z).isIndirect(0))
 341             return false;
 342
 343       // s[] access only possible in CP, $aX always applies
 344       if (sf == FILE_MEMORY_SHARED)
 345          return true;
 346       if (!ld->bb) // can't check type ...
 347          return false;
 348       Program::Type pt = ld->bb->getProgram()->getType();
 349
 350       // $aX applies to c[] only in VP, FP, GP if p[] is not accessed
 351       if (pt == Program::TYPE_COMPUTE)
 352          return false;
 353       if (pt == Program::TYPE_GEOMETRY) {
 354          if (sf == FILE_MEMORY_CONST)
 355             return i->src(s).getFile() != FILE_SHADER_INPUT;
 356          return sf == FILE_SHADER_INPUT;
 357       }
 358       return sf == FILE_MEMORY_CONST;
 359    }
 360    return true;
 361 }
 362
 363 bool
 364 TargetNV50::isAccessSupported(DataFile file, DataType ty) const
 365 {
 366    if (ty == TYPE_B96 || ty == TYPE_NONE)
 367       return false;
 368    if (typeSizeof(ty) > 4)
 369       return (file == FILE_MEMORY_LOCAL) || (file == FILE_MEMORY_GLOBAL);
 370    return true;
 371 }
 372
 373 bool
 374 TargetNV50::isOpSupported(operation op, DataType ty) const
 375 {
 376    if (ty == TYPE_F64 && chipset < 0xa0)
 377       return false;
 378
 379    switch (op) {
 380    case OP_PRERET:
 381       return chipset >= 0xa0;
 382    case OP_TXG:
 383       return chipset >= 0xa3;
 384    case OP_POW:
 385    case OP_SQRT:
 386    case OP_DIV:
 387    case OP_MOD:
 388    case OP_SET_AND:
 389    case OP_SET_OR:
 390    case OP_SET_XOR:
 391    case OP_SLCT:
 392    case OP_SELP:
 393    case OP_POPCNT:
 394    case OP_INSBF:
 395    case OP_EXTBF:
 396    case OP_EXIT: // want exit modifier instead (on NOP if required)
 397    case OP_MEMBAR:
 398       return false;
 399    case OP_SAD:
 400       return ty == TYPE_S32;
 401    default:
 402       return true;
 403    }
 404 }
 405
 406 bool
 407 TargetNV50::isModSupported(const Instruction *insn, int s, Modifier mod) const
 408 {
 409    if (!isFloatType(insn->dType)) {
 410       switch (insn->op) {
 411       case OP_ABS:
 412       case OP_NEG:
 413       case OP_CVT:
 414       case OP_CEIL:
 415       case OP_FLOOR:
 416       case OP_TRUNC:
 417       case OP_AND:
 418       case OP_OR:
 419       case OP_XOR:
 420          break;
 421       case OP_ADD:
 422          if (insn->src(s ? 0 : 1).mod.neg())
 423             return false;
 424          break;
 425       case OP_SUB:
 426          if (s == 0)
 427             return insn->src(1).mod.neg() ? false : true;
 428          break;
 429       case OP_SET:
 430          if (insn->sType != TYPE_F32)
 431             return false;
 432          break;
 433       default:
 434          return false;
 435       }
 436    }
 437    if (s > 3)
 438       return false;
 439    return (mod & Modifier(opInfo[insn->op].srcMods[s])) == mod;
 440 }
 441
 442 bool
 443 TargetNV50::mayPredicate(const Instruction *insn, const Value *pred) const
 444 {
 445    if (insn->getPredicate() || insn->flagsSrc >= 0)
 446       return false;
 447    for (int s = 0; insn->srcExists(s); ++s)
 448       if (insn->src(s).getFile() == FILE_IMMEDIATE)
 449          return false;
 450    return opInfo[insn->op].predicate;
 451 }
 452
 453 bool
 454 TargetNV50::isSatSupported(const Instruction *insn) const
 455 {
 456    if (insn->op == OP_CVT)
 457       return true;
 458    if (insn->dType != TYPE_F32)
 459       return false;
 460    return opInfo[insn->op].dstMods & NV50_IR_MOD_SAT;
 461 }
 462
 463 int TargetNV50::getLatency(const Instruction *i) const
 464 {
 465    // TODO: tune these values
 466    if (i->op == OP_LOAD) {
 467       switch (i->src(0).getFile()) {
 468       case FILE_MEMORY_LOCAL:
 469       case FILE_MEMORY_GLOBAL:
 470          return 100; // really 400 to 800
 471       default:
 472          return 22;
 473       }
 474    }
 475    return 22;
 476 }
 477
 478 // These are "inverse" throughput values, i.e. the number of cycles required
 479 // to issue a specific instruction for a full warp (32 threads).
 480 //
 481 // Assuming we have more than 1 warp in flight, a higher issue latency results
 482 // in a lower result latency since the MP will have spent more time with other
 483 // warps.
 484 // This also helps to determine the number of cycles between instructions in
 485 // a single warp.
 486 //
 487 int TargetNV50::getThroughput(const Instruction *i) const
 488 {
 489    // TODO: tune these values
 490    if (i->dType == TYPE_F32) {
 491       switch (i->op) {
 492       case OP_RCP:
 493       case OP_RSQ:
 494       case OP_LG2:
 495       case OP_SIN:
 496       case OP_COS:
 497       case OP_PRESIN:
 498       case OP_PREEX2:
 499          return 16;
 500       default:
 501          return 4;
 502       }
 503    } else
 504    if (i->dType == TYPE_U32 || i->dType == TYPE_S32) {
 505       return 4;
 506    } else
 507    if (i->dType == TYPE_F64) {
 508       return 32;
 509    } else {
 510       return 1;
 511    }
 512 }
 513
 514 static void
 515 recordLocation(uint16_t *locs, uint8_t *masks,
 516                const struct nv50_ir_varying *var)
 517 {
 518    uint16_t addr = var->slot[0] * 4;
 519
 520    switch (var->sn) {
 521    case TGSI_SEMANTIC_POSITION: locs[SV_POSITION] = addr; break;
 522    case TGSI_SEMANTIC_INSTANCEID: locs[SV_INSTANCE_ID] = addr; break;
 523    case TGSI_SEMANTIC_VERTEXID: locs[SV_VERTEX_ID] = addr; break;
 524    case TGSI_SEMANTIC_PRIMID: locs[SV_PRIMITIVE_ID] = addr; break;
 525    case NV50_SEMANTIC_LAYER: locs[SV_LAYER] = addr; break;
 526    case NV50_SEMANTIC_VIEWPORTINDEX: locs[SV_VIEWPORT_INDEX] = addr; break;
 527    default:
 528       break;
 529    }
 530    if (var->sn == TGSI_SEMANTIC_POSITION && masks)
 531       masks[0] = var->mask;
 532 }
 533
 534 void
 535 TargetNV50::parseDriverInfo(const struct nv50_ir_prog_info *info)
 536 {
 537    unsigned int i;
 538    for (i = 0; i < info->numOutputs; ++i)
 539       recordLocation(sysvalLocation, NULL, &info->out[i]);
 540    for (i = 0; i < info->numInputs; ++i)
 541       recordLocation(sysvalLocation, &wposMask, &info->in[i]);
 542    for (i = 0; i < info->numSysVals; ++i)
 543       recordLocation(sysvalLocation, NULL, &info->sv[i]);
 544
 545    if (sysvalLocation[SV_POSITION] >= 0x200) {
 546       // not assigned by driver, but we need it internally
 547       wposMask = 0x8;
 548       sysvalLocation[SV_POSITION] = 0;
 549    }
 550 }
 551
 552 } // namespace nv50_ir