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       OP_EMIT, OP_RESTART
 135    };
 136
 137    for (i = 0; i < DATA_FILE_COUNT; ++i)
 138       nativeFileMap[i] = (DataFile)i;
 139    nativeFileMap[FILE_PREDICATE] = FILE_FLAGS;
 140
 141    for (i = 0; i < OP_LAST; ++i) {
 142       opInfo[i].variants = NULL;
 143       opInfo[i].op = (operation)i;
 144       opInfo[i].srcTypes = 1 << (int)TYPE_F32;
 145       opInfo[i].dstTypes = 1 << (int)TYPE_F32;
 146       opInfo[i].immdBits = 0xffffffff;
 147       opInfo[i].srcNr = operationSrcNr[i];
 148
 149       for (j = 0; j < opInfo[i].srcNr; ++j) {
 150          opInfo[i].srcMods[j] = 0;
 151          opInfo[i].srcFiles[j] = 1 << (int)FILE_GPR;
 152       }
 153       opInfo[i].dstMods = 0;
 154       opInfo[i].dstFiles = 1 << (int)FILE_GPR;
 155
 156       opInfo[i].hasDest = 1;
 157       opInfo[i].vector = (i >= OP_TEX && i <= OP_TEXCSAA);
 158       opInfo[i].commutative = (commutative[i / 32] >> (i % 32)) & 1;
 159       opInfo[i].pseudo = (i < OP_MOV);
 160       opInfo[i].predicate = !opInfo[i].pseudo;
 161       opInfo[i].flow = (i >= OP_BRA && i <= OP_JOIN);
 162       opInfo[i].minEncSize = (shortForm[i / 32] & (1 << (i % 32))) ? 4 : 8;
 163    }
 164    for (i = 0; i < sizeof(noDestList) / sizeof(noDestList[0]); ++i)
 165       opInfo[noDestList[i]].hasDest = 0;
 166    for (i = 0; i < sizeof(noPredList) / sizeof(noPredList[0]); ++i)
 167       opInfo[noPredList[i]].predicate = 0;
 168
 169    for (i = 0; i < sizeof(_initProps) / sizeof(_initProps[0]); ++i) {
 170       const struct opProperties *prop = &_initProps[i];
 171
 172       for (int s = 0; s < 3; ++s) {
 173          if (prop->mNeg & (1 << s))
 174             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NEG;
 175          if (prop->mAbs & (1 << s))
 176             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_ABS;
 177          if (prop->mNot & (1 << s))
 178             opInfo[prop->op].srcMods[s] |= NV50_IR_MOD_NOT;
 179          if (prop->fConst & (1 << s))
 180             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_CONST;
 181          if (prop->fShared & (1 << s))
 182             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_MEMORY_SHARED;
 183          if (prop->fAttrib & (1 << s))
 184             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_SHADER_INPUT;
 185          if (prop->fImm & (1 << s))
 186             opInfo[prop->op].srcFiles[s] |= 1 << (int)FILE_IMMEDIATE;
 187       }
 188       if (prop->mSat & 8)
 189          opInfo[prop->op].dstMods = NV50_IR_MOD_SAT;
 190    }
 191 }
 192
 193 unsigned int
 194 TargetNV50::getFileSize(DataFile file) const
 195 {
 196    switch (file) {
 197    case FILE_NULL:          return 0;
 198    case FILE_GPR:           return 256; // in 16-bit units **
 199    case FILE_PREDICATE:     return 0;
 200    case FILE_FLAGS:         return 4;
 201    case FILE_ADDRESS:       return 4;
 202    case FILE_IMMEDIATE:     return 0;
 203    case FILE_MEMORY_CONST:  return 65536;
 204    case FILE_SHADER_INPUT:  return 0x200;
 205    case FILE_SHADER_OUTPUT: return 0x200;
 206    case FILE_MEMORY_GLOBAL: return 0xffffffff;
 207    case FILE_MEMORY_SHARED: return 16 << 10;
 208    case FILE_MEMORY_LOCAL:  return 48 << 10;
 209    case FILE_SYSTEM_VALUE:  return 16;
 210    default:
 211       assert(!"invalid file");
 212       return 0;
 213    }
 214    // ** only first 128 units encodable for 16-bit regs
 215 }
 216
 217 unsigned int
 218 TargetNV50::getFileUnit(DataFile file) const
 219 {
 220    if (file == FILE_GPR || file == FILE_ADDRESS)
 221       return 1;
 222    if (file == FILE_SYSTEM_VALUE)
 223       return 2;
 224    return 0;
 225 }
 226
 227 uint32_t
 228 TargetNV50::getSVAddress(DataFile shaderFile, const Symbol *sym) const
 229 {
 230    switch (sym->reg.data.sv.sv) {
 231    case SV_FACE:
 232       return 0x3fc;
 233    case SV_POSITION:
 234    {
 235       uint32_t addr = sysvalLocation[sym->reg.data.sv.sv];
 236       for (int c = 0; c < sym->reg.data.sv.index; ++c)
 237          if (wposMask & (1 << c))
 238             addr += 4;
 239       return addr;
 240    }
 241    case SV_NCTAID:
 242       return 0x8 + 2 * sym->reg.data.sv.index;
 243    case SV_CTAID:
 244       return 0xc + 2 * sym->reg.data.sv.index;
 245    case SV_NTID:
 246       return 0x2 + 2 * sym->reg.data.sv.index;
 247    case SV_TID:
 248       return 0;
 249    default:
 250       return sysvalLocation[sym->reg.data.sv.sv];
 251    }
 252 }
 253
 254 // long:  rrr, arr, rcr, acr, rrc, arc, gcr, grr
 255 // short: rr, ar, rc, gr
 256 // immd:  ri, gi
 257 bool
 258 TargetNV50::insnCanLoad(const Instruction *i, int s,
 259                         const Instruction *ld) const
 260 {
 261    DataFile sf = ld->src(0).getFile();
 262
 263    if (sf == FILE_IMMEDIATE && (i->predSrc >= 0 || i->flagsDef >= 0))
 264       return false;
 265    if (s >= opInfo[i->op].srcNr)
 266       return false;
 267    if (!(opInfo[i->op].srcFiles[s] & (1 << (int)sf)))
 268       return false;
 269    if (s == 2 && i->src(1).getFile() != FILE_GPR)
 270       return false;
 271
 272    // NOTE: don't rely on flagsDef
 273    for (int d = 0; i->defExists(d); ++d)
 274       if (i->def(d).getFile() == FILE_FLAGS)
 275          return false;
 276
 277    unsigned mode = 0;
 278
 279    for (int z = 0; z < Target::operationSrcNr[i->op]; ++z) {
 280       DataFile zf = (z == s) ? sf : i->src(z).getFile();
 281       switch (zf) {
 282       case FILE_GPR:
 283          break;
 284       case FILE_MEMORY_SHARED:
 285       case FILE_SHADER_INPUT:
 286          mode |= 1 << (z * 2);
 287          break;
 288       case FILE_MEMORY_CONST:
 289          mode |= 2 << (z * 2);
 290          break;
 291       case FILE_IMMEDIATE:
 292          mode |= 3 << (z * 2);
 293       default:
 294          break;
 295       }
 296    }
 297
 298    switch (mode) {
 299    case 0x00:
 300    case 0x01:
 301    case 0x03:
 302    case 0x08:
 303    case 0x09:
 304    case 0x0c:
 305    case 0x20:
 306    case 0x21:
 307       break;
 308    case 0x0d:
 309       if (ld->bb->getProgram()->getType() != Program::TYPE_GEOMETRY)
 310          return false;
 311    default:
 312       return false;
 313    }
 314
 315    uint8_t ldSize;
 316
 317    if ((i->op == OP_MUL || i->op == OP_MAD) && !isFloatType(i->dType)) {
 318       // 32-bit MUL will be split into 16-bit MULs
 319       if (ld->src(0).isIndirect(0))
 320          return false;
 321       if (sf == FILE_IMMEDIATE)
 322          return false;
 323       ldSize = 2;
 324    } else {
 325       ldSize = typeSizeof(ld->dType);
 326    }
 327
 328    if (sf == FILE_IMMEDIATE)
 329       return true;
 330
 331
 332    // Check if memory access is encodable:
 333
 334    if (ldSize < 4 && sf == FILE_SHADER_INPUT) // no < 4-byte aligned a[] access
 335       return false;
 336    if (ld->getSrc(0)->reg.data.offset > (int32_t)(127 * ldSize))
 337       return false;
 338
 339    if (ld->src(0).isIndirect(0)) {
 340       for (int z = 0; i->srcExists(z); ++z)
 341          if (i->src(z).isIndirect(0))
 342             return false;
 343
 344       // s[] access only possible in CP, $aX always applies
 345       if (sf == FILE_MEMORY_SHARED)
 346          return true;
 347       if (!ld->bb) // can't check type ...
 348          return false;
 349       Program::Type pt = ld->bb->getProgram()->getType();
 350
 351       // $aX applies to c[] only in VP, FP, GP if p[] is not accessed
 352       if (pt == Program::TYPE_COMPUTE)
 353          return false;
 354       if (pt == Program::TYPE_GEOMETRY) {
 355          if (sf == FILE_MEMORY_CONST)
 356             return i->src(s).getFile() != FILE_SHADER_INPUT;
 357          return sf == FILE_SHADER_INPUT;
 358       }
 359       return sf == FILE_MEMORY_CONST;
 360    }
 361    return true;
 362 }
 363
 364 bool
 365 TargetNV50::isAccessSupported(DataFile file, DataType ty) const
 366 {
 367    if (ty == TYPE_B96 || ty == TYPE_NONE)
 368       return false;
 369    if (typeSizeof(ty) > 4)
 370       return (file == FILE_MEMORY_LOCAL) || (file == FILE_MEMORY_GLOBAL);
 371    return true;
 372 }
 373
 374 bool
 375 TargetNV50::isOpSupported(operation op, DataType ty) const
 376 {
 377    if (ty == TYPE_F64 && chipset < 0xa0)
 378       return false;
 379
 380    switch (op) {
 381    case OP_PRERET:
 382       return chipset >= 0xa0;
 383    case OP_TXG:
 384       return chipset >= 0xa3 && chipset != 0xaa && chipset != 0xac;
 385    case OP_POW:
 386    case OP_SQRT:
 387    case OP_DIV:
 388    case OP_MOD:
 389    case OP_SET_AND:
 390    case OP_SET_OR:
 391    case OP_SET_XOR:
 392    case OP_SLCT:
 393    case OP_SELP:
 394    case OP_POPCNT:
 395    case OP_INSBF:
 396    case OP_EXTBF:
 397    case OP_EXIT: // want exit modifier instead (on NOP if required)
 398    case OP_MEMBAR:
 399       return false;
 400    case OP_SAD:
 401       return ty == TYPE_S32;
 402    default:
 403       return true;
 404    }
 405 }
 406
 407 bool
 408 TargetNV50::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          break;
 422       case OP_ADD:
 423          if (insn->src(s ? 0 : 1).mod.neg())
 424             return false;
 425          break;
 426       case OP_SUB:
 427          if (s == 0)
 428             return insn->src(1).mod.neg() ? false : true;
 429          break;
 430       case OP_SET:
 431          if (insn->sType != TYPE_F32)
 432             return false;
 433          break;
 434       default:
 435          return false;
 436       }
 437    }
 438    if (s > 3)
 439       return false;
 440    return (mod & Modifier(opInfo[insn->op].srcMods[s])) == mod;
 441 }
 442
 443 bool
 444 TargetNV50::mayPredicate(const Instruction *insn, const Value *pred) const
 445 {
 446    if (insn->getPredicate() || insn->flagsSrc >= 0)
 447       return false;
 448    for (int s = 0; insn->srcExists(s); ++s)
 449       if (insn->src(s).getFile() == FILE_IMMEDIATE)
 450          return false;
 451    return opInfo[insn->op].predicate;
 452 }
 453
 454 bool
 455 TargetNV50::isSatSupported(const Instruction *insn) const
 456 {
 457    if (insn->op == OP_CVT)
 458       return true;
 459    if (insn->dType != TYPE_F32)
 460       return false;
 461    return opInfo[insn->op].dstMods & NV50_IR_MOD_SAT;
 462 }
 463
 464 int TargetNV50::getLatency(const Instruction *i) const
 465 {
 466    // TODO: tune these values
 467    if (i->op == OP_LOAD) {
 468       switch (i->src(0).getFile()) {
 469       case FILE_MEMORY_LOCAL:
 470       case FILE_MEMORY_GLOBAL:
 471          return 100; // really 400 to 800
 472       default:
 473          return 22;
 474       }
 475    }
 476    return 22;
 477 }
 478
 479 // These are "inverse" throughput values, i.e. the number of cycles required
 480 // to issue a specific instruction for a full warp (32 threads).
 481 //
 482 // Assuming we have more than 1 warp in flight, a higher issue latency results
 483 // in a lower result latency since the MP will have spent more time with other
 484 // warps.
 485 // This also helps to determine the number of cycles between instructions in
 486 // a single warp.
 487 //
 488 int TargetNV50::getThroughput(const Instruction *i) const
 489 {
 490    // TODO: tune these values
 491    if (i->dType == TYPE_F32) {
 492       switch (i->op) {
 493       case OP_RCP:
 494       case OP_RSQ:
 495       case OP_LG2:
 496       case OP_SIN:
 497       case OP_COS:
 498       case OP_PRESIN:
 499       case OP_PREEX2:
 500          return 16;
 501       default:
 502          return 4;
 503       }
 504    } else
 505    if (i->dType == TYPE_U32 || i->dType == TYPE_S32) {
 506       return 4;
 507    } else
 508    if (i->dType == TYPE_F64) {
 509       return 32;
 510    } else {
 511       return 1;
 512    }
 513 }
 514
 515 static void
 516 recordLocation(uint16_t *locs, uint8_t *masks,
 517                const struct nv50_ir_varying *var)
 518 {
 519    uint16_t addr = var->slot[0] * 4;
 520
 521    switch (var->sn) {
 522    case TGSI_SEMANTIC_POSITION: locs[SV_POSITION] = addr; break;
 523    case TGSI_SEMANTIC_INSTANCEID: locs[SV_INSTANCE_ID] = addr; break;
 524    case TGSI_SEMANTIC_VERTEXID: locs[SV_VERTEX_ID] = addr; break;
 525    case TGSI_SEMANTIC_PRIMID: locs[SV_PRIMITIVE_ID] = addr; break;
 526    case NV50_SEMANTIC_LAYER: locs[SV_LAYER] = addr; break;
 527    case NV50_SEMANTIC_VIEWPORTINDEX: locs[SV_VIEWPORT_INDEX] = addr; break;
 528    default:
 529       break;
 530    }
 531    if (var->sn == TGSI_SEMANTIC_POSITION && masks)
 532       masks[0] = var->mask;
 533 }
 534
 535 void
 536 TargetNV50::parseDriverInfo(const struct nv50_ir_prog_info *info)
 537 {
 538    unsigned int i;
 539    for (i = 0; i < info->numOutputs; ++i)
 540       recordLocation(sysvalLocation, NULL, &info->out[i]);
 541    for (i = 0; i < info->numInputs; ++i)
 542       recordLocation(sysvalLocation, &wposMask, &info->in[i]);
 543    for (i = 0; i < info->numSysVals; ++i)
 544       recordLocation(sysvalLocation, NULL, &info->sv[i]);
 545
 546    if (sysvalLocation[SV_POSITION] >= 0x200) {
 547       // not assigned by driver, but we need it internally
 548       wposMask = 0x8;
 549       sysvalLocation[SV_POSITION] = 0;
 550    }
 551 }
 552
 553 } // namespace nv50_ir