| 1 |
/*========================== begin_copyright_notice ============================ |
| 2 |
|
| 3 |
Copyright (C) 2017-2021 Intel Corporation |
| 4 |
|
| 5 |
SPDX-License-Identifier: MIT |
| 6 |
|
| 7 |
============================= end_copyright_notice ===========================*/ |
| 8 |
|
| 9 |
#ifndef _IGA_BACKEND_GED_ENCODER_H_ |
| 10 |
#define _IGA_BACKEND_GED_ENCODER_H_ |
| 11 |
|
| 12 |
#include "../../ErrorHandler.hpp" |
| 13 |
#include "../../IR/Block.hpp" |
| 14 |
#include "../../IR/Instruction.hpp" |
| 15 |
#include "../../IR/Kernel.hpp" |
| 16 |
#include "../../Timer/Timer.hpp" |
| 17 |
#include "../EncoderOpts.hpp" |
| 18 |
#include "GEDBitProcessor.hpp" |
| 19 |
#include "IGAToGEDTranslation.hpp" |
| 20 |
|
| 21 |
#include <list> |
| 22 |
#include <map> |
| 23 |
#include <vector> |
| 24 |
|
| 25 |
namespace iga { |
| 26 |
#define GED_ENCODE(FUNC, ARG) GED_ENCODE_TO(FUNC, ARG, &m_gedInst) |
| 27 |
#if defined(GED_TIMER) || defined(_DEBUG) |
| 28 |
#define START_GED_TIMER() startIGATimer(TIMER_GED) |
| 29 |
#define STOP_GED_TIMER() stopIGATimer(TIMER_GED) |
| 30 |
#else |
| 31 |
#define START_GED_TIMER() |
| 32 |
#define STOP_GED_TIMER() |
| 33 |
#endif |
| 34 |
|
| 35 |
#if defined(TOTAL_ENCODE_TIMER) || defined(_DEBUG) |
| 36 |
#define START_ENCODER_TIMER() startIGATimer(TIMER_TOTAL) |
| 37 |
#define STOP_ENCODER_TIMER() stopIGATimer(TIMER_TOTAL) |
| 38 |
#else |
| 39 |
#define START_ENCODER_TIMER() |
| 40 |
#define STOP_ENCODER_TIMER() |
| 41 |
#endif |
| 42 |
|
| 43 |
// uncomment to emit all the GED calls |
| 44 |
// #define TRACE_GED_CALLS |
| 45 |
|
| 46 |
#ifdef TRACE_GED_CALLS |
| 47 |
#define TRACE_GED_SETTER(FIELD, ARG, STATUS) \ |
| 48 |
std::cerr << "Encoder.cpp:" << __LINE__ << ": GED_Set" << #FIELD << "(..," \ |
| 49 |
<< (ARG) << ") results in " << (STATUS) << "\n" |
| 50 |
#else |
| 51 |
#define TRACE_GED_SETTER(FIELD, ARG, STATUS) |
| 52 |
#endif |
| 53 |
|
| 54 |
#define GED_ENCODE_TO(FIELD, ARG, GED) \ |
| 55 |
do { \ |
| 56 |
GED_RETURN_VALUE _status; \ |
| 57 |
START_GED_TIMER(); \ |
| 58 |
_status = GED_Set##FIELD(GED, ARG); \ |
| 59 |
STOP_GED_TIMER(); \ |
| 60 |
TRACE_GED_SETTER(FIELD, ARG, _status); \ |
| 61 |
if (_status != GED_RETURN_VALUE_SUCCESS) { \ |
| 62 |
handleGedError(__LINE__, #FIELD, _status); \ |
| 63 |
} \ |
| 64 |
} while (0) |
| 65 |
|
| 66 |
using BlockList = std::list<Block *, std_arena_based_allocator<Block *>>; |
| 67 |
|
| 68 |
class Encoder : protected GEDBitProcessor { |
| 69 |
public: |
| 70 |
Encoder(const Model &model, ErrorHandler &eh, const EncoderOpts &eos); |
| 71 |
|
| 72 |
void encodeKernel(Kernel &k, MemManager &m, void *&bits, uint32_t &bitsLen); |
| 73 |
|
| 74 |
size_t getNumInstructionsEncoded() const; |
| 75 |
|
| 76 |
/////////////////////////////////////////////////////////////////////// |
| 77 |
// PROFILING API FOR TESTING (must be compiled into) |
| 78 |
// |
| 79 |
// compile with MEASURE_COMPILATION_TIME |
| 80 |
double getElapsedTimeMicros(unsigned int idx); |
| 81 |
int64_t getElapsedTimeTicks(unsigned int idx); |
| 82 |
std::string getTimerName(unsigned int idx); |
| 83 |
|
| 84 |
protected: |
| 85 |
// TODO: phase these out |
| 86 |
void encodeKernelPreProcess(Kernel &k); |
| 87 |
void doEncodeKernelPreProcess(Kernel &k); |
| 88 |
void encodeFC(const Instruction &i); |
| 89 |
void encodeTernaryInstruction(const Instruction &inst, |
| 90 |
GED_ACCESS_MODE accessMode); |
| 91 |
void encodeTernaryAlign1Instruction(const Instruction &inst); |
| 92 |
void encodeTernaryAlign16Instruction(const Instruction &inst); |
| 93 |
void encodeQuinaryInstruction(const Instruction &inst); |
| 94 |
void encodeSendDirectDestination(const Operand &dst); |
| 95 |
void encodeSendDestinationDataType(const Operand &dst); |
| 96 |
void encodeOptionsThreadControl(const Instruction &inst); |
| 97 |
|
| 98 |
protected: |
| 99 |
void encodeDstReg(RegName regName, uint16_t regNum); // just the regnum |
| 100 |
|
| 101 |
void encodeImmVal(const ImmVal &val, Type type); |
| 102 |
|
| 103 |
template <SourceIndex S> void encodeSrcRegFile(GED_REG_FILE rf); |
| 104 |
template <SourceIndex S> void encodeSrcRegionVert(Region::Vert v); |
| 105 |
template <SourceIndex S> void encodeSrcType(Type t); |
| 106 |
template <SourceIndex S> void encodeSrcAddrMode(GED_ADDR_MODE x); |
| 107 |
template <SourceIndex S> void encodeSrcModifier(SrcModifier x); |
| 108 |
template <SourceIndex S> void encodeSrcSubRegNum(uint32_t subRegInByte); |
| 109 |
|
| 110 |
template <SourceIndex S> void encodeSrcMathMacroReg(MathMacroExt a); |
| 111 |
|
| 112 |
template <SourceIndex S> void encodeSrcReg(RegName regName, uint16_t regNum); |
| 113 |
template <SourceIndex S> void encodeSrcAddrImm(int32_t addrImm); |
| 114 |
template <SourceIndex S> void encodeSrcAddrSubRegNum(uint32_t addrSubReg); |
| 115 |
template <SourceIndex S> |
| 116 |
void encodeSrcRegion(const Region &r, bool hasRgnWi = true); |
| 117 |
template <SourceIndex S> void encodeSrcRegionWidth(Region::Width w); |
| 118 |
template <SourceIndex S> void encodeSrcReducedRegion(const Region& rgn); |
| 119 |
template <SourceIndex S> void encodeSrcRepCtrl(GED_REP_CTRL rep); |
| 120 |
template <SourceIndex S> |
| 121 |
void encodeSrcChanSel(GED_SWIZZLE chSelX, GED_SWIZZLE chSelY, |
| 122 |
GED_SWIZZLE chSelZ, GED_SWIZZLE chSelW); |
| 123 |
|
| 124 |
template <SourceIndex S> |
| 125 |
void encodeTernarySourceAlign16(const Instruction &inst); |
| 126 |
void encodeTernaryDestinationAlign16(const Instruction &inst); |
| 127 |
|
| 128 |
template <SourceIndex S> |
| 129 |
void encodeTernarySourceAlign1(const Instruction &inst); |
| 130 |
template <SourceIndex S> |
| 131 |
void encodeQuinarySource(const Instruction &inst); |
| 132 |
void encodeTernaryDestinationAlign1(const Instruction &inst); |
| 133 |
|
| 134 |
void encodeTernarySrcRegionVert(SourceIndex S, Region::Vert v); |
| 135 |
|
| 136 |
template <SourceIndex S> |
| 137 |
void encodeTernaryImmVal(const ImmVal &val, Type type); |
| 138 |
|
| 139 |
template <SourceIndex S> void encodeSrcRegionHorz(Region::Horz s); |
| 140 |
|
| 141 |
void handleGedError(int line, const char *setter, GED_RETURN_VALUE status); |
| 142 |
|
| 143 |
// state that is valid over instance life |
| 144 |
EncoderOpts m_opts; |
| 145 |
|
| 146 |
// state that is valid over encodeInst() |
| 147 |
ged_ins_t m_gedInst; |
| 148 |
bool m_encodeAlign16 = false; |
| 149 |
Op m_opcode = Op::INVALID; |
| 150 |
size_t m_numberInstructionsEncoded; |
| 151 |
|
| 152 |
private: |
| 153 |
void operator delete(void *, MemManager *){}; |
| 154 |
void *operator new(size_t sz, MemManager *m) { return m->alloc(sz); }; |
| 155 |
|
| 156 |
void encodeBlock(Kernel &k, Block *blk); |
| 157 |
void encodeInstruction(Instruction &inst); |
| 158 |
void patchJumpOffsets(); |
| 159 |
|
| 160 |
/////////////////////////////////////////////////////////////////////// |
| 161 |
// BASIC INSTRUCTIONS |
| 162 |
/////////////////////////////////////////////////////////////////////// |
| 163 |
void encodeBasicInstruction(const Instruction &inst, |
| 164 |
GED_ACCESS_MODE accessMode); |
| 165 |
void |
| 166 |
encodeBasicDestination(const Instruction &inst, const Operand &dst, |
| 167 |
GED_ACCESS_MODE accessMode = GED_ACCESS_MODE_Align1); |
| 168 |
template <SourceIndex S> |
| 169 |
void encodeBasicSource(const Instruction &inst, const Operand &src, |
| 170 |
GED_ACCESS_MODE accessMode = GED_ACCESS_MODE_Align1); |
| 171 |
/////////////////////////////////////////////////////////////////////// |
| 172 |
// BRANCH INSTRUCTIONS |
| 173 |
/////////////////////////////////////////////////////////////////////// |
| 174 |
void encodeBranchingInstruction(const Instruction &inst); |
| 175 |
void encodeBranchingInstructionSimplified(const Instruction &inst); |
| 176 |
void encodeBranchDestination(const Operand &dst); |
| 177 |
void encodeBranchSource(const Operand &src); |
| 178 |
/////////////////////////////////////////////////////////////////////// |
| 179 |
// SEND INSTRUCTIONS |
| 180 |
/////////////////////////////////////////////////////////////////////// |
| 181 |
void encodeSendInstruction(const Instruction &inst); |
| 182 |
void encodeSendSource0(const Operand &src); |
| 183 |
void encodeSendDestination(const Operand &dst); |
| 184 |
void encodeSendgDestinationXe3p(const Operand &dst); |
| 185 |
void encodeSendgxDestinationXe3p(const Operand &dst); |
| 186 |
void encodeSendgxSource0Xe3p(const Operand &src); |
| 187 |
void encodeSendgxSource1Xe3p(const Operand &src); |
| 188 |
void encodeSendsSource0(const Operand &src); |
| 189 |
void encodeSendsSource1(const Operand &src); |
| 190 |
void encodeSendsDestination(const Operand &dst); |
| 191 |
|
| 192 |
void encodeSendDescs(const Instruction &inst); |
| 193 |
void encodeSendDescsPreXe(const Instruction &inst); |
| 194 |
void encodeSendDescsXe(const Instruction &inst); |
| 195 |
void encodeSendDescsXeHP(const Instruction &inst); |
| 196 |
void encodeSendDescsXeHPG(const Instruction &inst); |
| 197 |
void encodeSendDescsXe2(const Instruction &inst); |
| 198 |
void encodeSendgDescsXe3(const Instruction &inst); |
| 199 |
|
| 200 |
void encodeInlineBinaryInst(Instruction &inst); |
| 201 |
/////////////////////////////////////////////////////////////////////// |
| 202 |
// SYNC INSTRUCTIONS |
| 203 |
/////////////////////////////////////////////////////////////////////// |
| 204 |
void encodeSyncInstruction(const Instruction &inst); |
| 205 |
|
| 206 |
/////////////////////////////////////////////////////////////////////// |
| 207 |
// OTHER HELPER FUNCTIONS |
| 208 |
/////////////////////////////////////////////////////////////////////// |
| 209 |
void setEncodedPC(Instruction *inst, int32_t encodePC); |
| 210 |
int32_t getEncodedPC(const Instruction *inst) const; |
| 211 |
|
| 212 |
bool getBlockOffset(const Block *b, uint32_t &pc); |
| 213 |
|
| 214 |
// handles encoding ARF registers as well as the easy GRF case |
| 215 |
// caller actually pulls the trigger and encodes the bits, but this |
| 216 |
// call can raise the encoding error |
| 217 |
uint32_t translateRegNum(int opIx, RegName reg, uint16_t regNum); |
| 218 |
uint32_t mathMacroRegToBits(int src, MathMacroExt mme); // ChSel and SubReg |
| 219 |
GED_DST_CHAN_EN mathMacroRegToChEn(MathMacroExt mme); |
| 220 |
|
| 221 |
void applyGedWorkarounds(const Kernel &k, size_t bitsLen); |
| 222 |
void encodeOptions(const Instruction &inst); |
| 223 |
|
| 224 |
|
| 225 |
////////////////////////////////////////////////////////////////////// |
| 226 |
// platform specific queries *but sometimes need the instruction too) |
| 227 |
// |
| 228 |
// GEN7p5 implicitly scales PC offsets by QW except for a few instructions |
| 229 |
bool arePcsInQWords(const OpSpec &os) const; |
| 230 |
|
| 231 |
// Call need to have src0 region be set to: |
| 232 |
// SKL and before: <2;2,1> |
| 233 |
// ICL: <2;4,1> |
| 234 |
// Later GENs ignore the region completely |
| 235 |
bool callNeedsSrc0Region221(const Instruction &inst) const; |
| 236 |
bool callNeedsSrc0Region241(const Instruction &inst) const; |
| 237 |
|
| 238 |
///////////////////////////////////////////////////////////// |
| 239 |
// state valid over encodeKernel() |
| 240 |
MemManager *m_mem; |
| 241 |
uint8_t *m_instBuf = nullptr; // the output bits |
| 242 |
struct JumpPatch { // JIP and UIP label patching |
| 243 |
Instruction *inst; // the instruction |
| 244 |
ged_ins_t gedInst; // the partially constructed GED instruction |
| 245 |
uint8_t *bits; // where to encode it in the heap |
| 246 |
JumpPatch(Instruction *i, const ged_ins_t &gi, uint8_t *bs) |
| 247 |
: inst(i), gedInst(gi), bits(bs) {} |
| 248 |
}; |
| 249 |
std::vector<JumpPatch> m_needToPatch; |
| 250 |
std::map<const Block *, int32_t> m_blockToOffsetMap; |
| 251 |
|
| 252 |
public: |
| 253 |
//////////////////////////////////////////////////////////////// |
| 254 |
static uint64_t typeConvesionHelper(const ImmVal &val, Type type) { |
| 255 |
uint64_t value = 0; |
| 256 |
switch (type) { |
| 257 |
case Type::UD: |
| 258 |
case Type::F: |
| 259 |
case Type::V: |
| 260 |
case Type::UV: |
| 261 |
case Type::VF: |
| 262 |
value = (uint64_t)val.u32; |
| 263 |
break; |
| 264 |
case Type::D: |
| 265 |
value = (uint64_t)val.s32; |
| 266 |
break; |
| 267 |
case Type::W: |
| 268 |
value = (uint64_t)val.s16; |
| 269 |
break; |
| 270 |
case Type::UW: |
| 271 |
case Type::HF: |
| 272 |
case Type::BF: |
| 273 |
value = (uint64_t)val.u16; |
| 274 |
break; |
| 275 |
case Type::DF: |
| 276 |
case Type::UQ: |
| 277 |
case Type::Q: |
| 278 |
value = val.u64; |
| 279 |
break; |
| 280 |
case Type::B: |
| 281 |
case Type::UB: |
| 282 |
// technically not reachable since we don't permit byte moves |
| 283 |
// from immediates |
| 284 |
value = val.u64; |
| 285 |
break; |
| 286 |
case Type::BF8: |
| 287 |
case Type::HF8: |
| 288 |
value = (uint64_t)val.u8; |
| 289 |
break; |
| 290 |
case Type::TF32: |
| 291 |
value = (uint64_t)val.u32; |
| 292 |
break; |
| 293 |
default: |
| 294 |
break; |
| 295 |
} |
| 296 |
|
| 297 |
return value; |
| 298 |
} |
| 299 |
|
| 300 |
protected: |
| 301 |
//////////////////////////////////////////////////////////////// |
| 302 |
// allowable types for ternary Align1 |
| 303 |
static bool isTernaryAlign1Floating(Type t) { |
| 304 |
return lowerExecDataType(t) == |
| 305 |
GED_EXECUTION_DATA_TYPE::GED_EXECUTION_DATA_TYPE_Float; |
| 306 |
} |
| 307 |
|
| 308 |
// allowable types for ternary Align1 |
| 309 |
static bool isTernaryAlign1Integral(Type t) { |
| 310 |
return lowerExecDataType(t) == |
| 311 |
GED_EXECUTION_DATA_TYPE::GED_EXECUTION_DATA_TYPE_Integer; |
| 312 |
} |
| 313 |
}; // end: class definition Encoder |
| 314 |
|
| 315 |
template <SourceIndex S> void Encoder::encodeSrcRegFile(GED_REG_FILE rf) { |
| 316 |
if (S == SourceIndex::SRC0) { |
| 317 |
GED_ENCODE(Src0RegFile, rf); |
| 318 |
} else if (S == SourceIndex::SRC1) { |
| 319 |
GED_ENCODE(Src1RegFile, rf); |
| 320 |
} else if (S == SourceIndex::SRC2) { |
| 321 |
GED_ENCODE(Src2RegFile, rf); |
| 322 |
} else if (S == SourceIndex::SRC3) { |
| 323 |
GED_ENCODE(Src3RegFile, rf); |
| 324 |
} else if (S == SourceIndex::SRC4) { |
| 325 |
GED_ENCODE(Src4RegFile, rf); |
| 326 |
} |
| 327 |
} |
| 328 |
|
| 329 |
template <SourceIndex S> void Encoder::encodeSrcRegionVert(Region::Vert v) { |
| 330 |
if (S == SourceIndex::SRC0) { |
| 331 |
GED_ENCODE(Src0VertStride, lowerRegionVert(v)); |
| 332 |
} else { // (S == SourceIndex::SRC1) |
| 333 |
GED_ENCODE(Src1VertStride, lowerRegionVert(v)); |
| 334 |
} // S != SRC2 since ternary Align1 doesn't have bits for that |
| 335 |
} |
| 336 |
|
| 337 |
template <SourceIndex S> void Encoder::encodeSrcType(Type t) { |
| 338 |
if (S == SourceIndex::SRC0) { |
| 339 |
GED_ENCODE(Src0DataType, lowerDataType(t)); |
| 340 |
} else if (S == SourceIndex::SRC1) { |
| 341 |
GED_ENCODE(Src1DataType, lowerDataType(t)); |
| 342 |
} else { |
| 343 |
GED_ENCODE(Src2DataType, lowerDataType(t)); |
| 344 |
} |
| 345 |
} |
| 346 |
|
| 347 |
template <SourceIndex S> void Encoder::encodeSrcAddrMode(GED_ADDR_MODE x) { |
| 348 |
if (S == SourceIndex::SRC0) { |
| 349 |
GED_ENCODE(Src0AddrMode, x); |
| 350 |
} else { |
| 351 |
GED_ENCODE(Src1AddrMode, x); |
| 352 |
} |
| 353 |
} |
| 354 |
|
| 355 |
template <SourceIndex S> void Encoder::encodeSrcModifier(SrcModifier x) { |
| 356 |
if (S == SourceIndex::SRC0) { |
| 357 |
GED_ENCODE(Src0SrcMod, lowerSrcMod(x)); |
| 358 |
} else if (S == SourceIndex::SRC1) { |
| 359 |
GED_ENCODE(Src1SrcMod, lowerSrcMod(x)); |
| 360 |
} else { |
| 361 |
GED_ENCODE(Src2SrcMod, lowerSrcMod(x)); |
| 362 |
} |
| 363 |
} |
| 364 |
|
| 365 |
template <SourceIndex S> |
| 366 |
void Encoder::encodeSrcSubRegNum(uint32_t subRegInByte) { |
| 367 |
if (S == SourceIndex::SRC0) { |
| 368 |
GED_ENCODE(Src0SubRegNum, subRegInByte); |
| 369 |
} else if (S == SourceIndex::SRC1) { |
| 370 |
GED_ENCODE(Src1SubRegNum, subRegInByte); |
| 371 |
} else if (S == SourceIndex::SRC2) { |
| 372 |
GED_ENCODE(Src2SubRegNum, subRegInByte); |
| 373 |
} else if (S == SourceIndex::SRC3) { |
| 374 |
GED_ENCODE(Src3SubRegNum, subRegInByte); |
| 375 |
} else if (S == SourceIndex::SRC4) { |
| 376 |
GED_ENCODE(Src4SubRegNum, subRegInByte); |
| 377 |
} |
| 378 |
} |
| 379 |
|
| 380 |
template <SourceIndex S> void Encoder::encodeSrcMathMacroReg(MathMacroExt a) { |
| 381 |
if (S == SourceIndex::SRC0) { |
| 382 |
GED_ENCODE(Src0MathMacroExt, lowerSpecialAcc(a)); |
| 383 |
} else if (S == SourceIndex::SRC1) { |
| 384 |
GED_ENCODE(Src1MathMacroExt, lowerSpecialAcc(a)); |
| 385 |
} else { |
| 386 |
GED_ENCODE(Src2MathMacroExt, lowerSpecialAcc(a)); |
| 387 |
} |
| 388 |
} |
| 389 |
template <SourceIndex S> |
| 390 |
void Encoder::encodeSrcReg(RegName regName, uint16_t regNum) { |
| 391 |
uint32_t regBits = 0; |
| 392 |
if (regName == RegName::GRF_R) { |
| 393 |
regBits = regNum; // GRF fast path |
| 394 |
} else { // ARF slower path |
| 395 |
const RegInfo *ri = m_model.lookupRegInfoByRegName(regName); |
| 396 |
if (!ri) { |
| 397 |
errorT("src", (int)S, ": unexpected register on this platform"); |
| 398 |
} else { |
| 399 |
uint16_t reg; |
| 400 |
ri->encode((int)regNum, reg); |
| 401 |
regBits = reg; // widen for GED |
| 402 |
} |
| 403 |
} |
| 404 |
if (S == SourceIndex::SRC0) { |
| 405 |
GED_ENCODE(Src0RegNum, regBits); |
| 406 |
} else if (S == SourceIndex::SRC1) { |
| 407 |
GED_ENCODE(Src1RegNum, regBits); |
| 408 |
} else if (S == SourceIndex::SRC2) { |
| 409 |
GED_ENCODE(Src2RegNum, regBits); |
| 410 |
} else if (S == SourceIndex::SRC3) { |
| 411 |
GED_ENCODE(Src3RegNum, regBits); |
| 412 |
} else if (S == SourceIndex::SRC4) { |
| 413 |
GED_ENCODE(Src4RegNum, regBits); |
| 414 |
} |
| 415 |
} |
| 416 |
|
| 417 |
template <SourceIndex S> void Encoder::encodeSrcAddrImm(int32_t addrImm) { |
| 418 |
if (S == SourceIndex::SRC0) { |
| 419 |
GED_ENCODE(Src0AddrImm, addrImm); |
| 420 |
} else { |
| 421 |
GED_ENCODE(Src1AddrImm, addrImm); |
| 422 |
} |
| 423 |
} |
| 424 |
template <SourceIndex S> |
| 425 |
void Encoder::encodeSrcAddrSubRegNum(uint32_t addrSubReg) { |
| 426 |
if (S == SourceIndex::SRC0) { |
| 427 |
GED_ENCODE(Src0AddrSubRegNum, addrSubReg); |
| 428 |
} else { |
| 429 |
GED_ENCODE(Src1AddrSubRegNum, addrSubReg); |
| 430 |
} |
| 431 |
} |
| 432 |
template <SourceIndex S> |
| 433 |
void Encoder::encodeSrcRegion(const Region &rgn, bool hasRgnWi) { |
| 434 |
uint32_t v = 0; |
| 435 |
if (rgn.getVt() == Region::Vert::VT_VxH) { |
| 436 |
v = 0x3; |
| 437 |
} else if (rgn.getVt() != Region::Vert::VT_INVALID) { |
| 438 |
v = static_cast<uint32_t>(rgn.v); |
| 439 |
} else { |
| 440 |
errorT(S == SourceIndex::SRC0 ? "invalid region vertical stride on src0" |
| 441 |
: "invalid region vertical stride on src1"); |
| 442 |
} |
| 443 |
|
| 444 |
uint32_t w = static_cast<uint32_t>(rgn.getWi()); |
| 445 |
if (rgn.getWi() == Region::Width::WI_INVALID) { |
| 446 |
errorT(S == SourceIndex::SRC0 ? "invalid region width on src0" |
| 447 |
: "invalid region width on src1"); |
| 448 |
} |
| 449 |
|
| 450 |
uint32_t h = static_cast<uint32_t>(rgn.getHz()); |
| 451 |
if (rgn.getHz() == Region::Horz::HZ_INVALID) { |
| 452 |
h = 1; |
| 453 |
errorT(S == SourceIndex::SRC0 ? "invalid region horizontal stride on src0" |
| 454 |
: "invalid region horizontal stride on src1"); |
| 455 |
} |
| 456 |
|
| 457 |
if (S == SourceIndex::SRC0) { |
| 458 |
GED_ENCODE(Src0VertStride, v); |
| 459 |
if (hasRgnWi) { |
| 460 |
GED_ENCODE(Src0Width, w); |
| 461 |
} else { |
| 462 |
// some ops have an implicit width region |
| 463 |
// (e.g. some specialized instructions poaches Src0.RgnWi) |
| 464 |
// |
| 465 |
// Within the IR we use 1 so logic that depends on regioning |
| 466 |
// gets the correct behavior (hardware assumes w=1). |
| 467 |
w = 1; |
| 468 |
} |
| 469 |
GED_ENCODE(Src0HorzStride, h); |
| 470 |
} else if (S == SourceIndex::SRC1) { |
| 471 |
GED_ENCODE(Src1VertStride, v); |
| 472 |
GED_ENCODE(Src1Width, w); |
| 473 |
GED_ENCODE(Src1HorzStride, h); |
| 474 |
} else { |
| 475 |
IGA_ASSERT_FALSE("Encoder::encodeSrcRegion: only works on src0 and src1"); |
| 476 |
} |
| 477 |
} |
| 478 |
|
| 479 |
template <SourceIndex S> |
| 480 |
void Encoder::encodeSrcReducedRegion(const Region& rgn) { |
| 481 |
IGA_ASSERT(S == SourceIndex::SRC1, "Invalid SourceIndex"); |
| 482 |
if (rgn.isScalar()) |
| 483 |
GED_ENCODE(Src1ScalarReg, 1); |
| 484 |
else |
| 485 |
GED_ENCODE(Src1ScalarReg, 0); |
| 486 |
return; |
| 487 |
IGA_ASSERT_FALSE("Unreachable: Unsupported region"); |
| 488 |
} |
| 489 |
|
| 490 |
template <SourceIndex S> void Encoder::encodeSrcRegionWidth(Region::Width w) { |
| 491 |
if (S == SourceIndex::SRC0) { |
| 492 |
GED_ENCODE(Src0Width, lowerRegionWidth(w)); |
| 493 |
} else { // (S == SourceIndex::SRC1) |
| 494 |
GED_ENCODE(Src1Width, lowerRegionWidth(w)); |
| 495 |
} // S != SRC2 since ternary Align1 doesn't have bits for that |
| 496 |
} |
| 497 |
|
| 498 |
template <SourceIndex S> |
| 499 |
void Encoder::encodeTernaryImmVal(const ImmVal &val, Type type) { |
| 500 |
if (S == SourceIndex::SRC0) { |
| 501 |
GED_ENCODE(Src0TernaryImm, typeConvesionHelper(val, type)); |
| 502 |
} else if (S == SourceIndex::SRC2) { |
| 503 |
GED_ENCODE(Src2TernaryImm, typeConvesionHelper(val, type)); |
| 504 |
} else { |
| 505 |
errorT("immediate operands not supported in src1 of ternary formats"); |
| 506 |
} |
| 507 |
} |
| 508 |
|
| 509 |
template <SourceIndex S> void Encoder::encodeSrcRegionHorz(Region::Horz s) { |
| 510 |
if (S == SourceIndex::SRC0) { |
| 511 |
GED_ENCODE(Src0HorzStride, lowerRegionHorz(s)); |
| 512 |
} else if (S == SourceIndex::SRC1) { |
| 513 |
GED_ENCODE(Src1HorzStride, lowerRegionHorz(s)); |
| 514 |
} else { |
| 515 |
GED_ENCODE(Src2HorzStride, lowerRegionHorz(s)); |
| 516 |
} |
| 517 |
} |
| 518 |
} // namespace iga |
| 519 |
|
| 520 |
namespace iga { |
| 521 |
typedef Encoder Encoder; |
| 522 |
} |
| 523 |
#endif |