| 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 |
#include "RegSet.hpp" |
| 10 |
#include "../asserts.hpp" |
| 11 |
#include "Messages.hpp" |
| 12 |
|
| 13 |
#include <sstream> |
| 14 |
|
| 15 |
using namespace iga; |
| 16 |
|
| 17 |
// This module implements what was the original intent for RegDeps. |
| 18 |
// Unfortunately, that module was coupled certain machine state and other |
| 19 |
// dependencies that we don't want for simpler things like DU analysis. |
| 20 |
// |
| 21 |
// FIXME: would be to |
| 22 |
// (1) rename RegDeps in SWSBSetter to something like DepState |
| 23 |
// (2) compose a RegSet within the DepState |
| 24 |
// |
| 25 |
|
| 26 |
static size_t BytesForRegSet(RegName rn, const Model &m) { |
| 27 |
// For platforms lacking a particular register this returns a |
| 28 |
// nonzero size so that m.getRegCount() doesn't assert |
| 29 |
if (m.lookupRegInfoByRegName(rn) == nullptr) |
| 30 |
return 1; |
| 31 |
return (size_t)m.getRegCount(rn) * m.getBytesPerReg(rn); |
| 32 |
} |
| 33 |
|
| 34 |
static const RegName TRACKED[] { |
| 35 |
RegName::GRF_R, RegName::ARF_A, RegName::ARF_ACC, RegName::ARF_F, |
| 36 |
RegName::ARF_S, |
| 37 |
}; |
| 38 |
|
| 39 |
BitSet<> *RegSet::bitSetForPtr(RegName rn) { |
| 40 |
switch (rn) { |
| 41 |
case RegName::GRF_R: |
| 42 |
return &bitsR; |
| 43 |
case RegName::ARF_A: |
| 44 |
return &bitsA; |
| 45 |
case RegName::ARF_ACC: |
| 46 |
return &bitsAcc; |
| 47 |
case RegName::ARF_F: |
| 48 |
return &bitsF; |
| 49 |
case RegName::ARF_S: |
| 50 |
return &bitsS; |
| 51 |
default: |
| 52 |
break; |
| 53 |
} |
| 54 |
return nullptr; |
| 55 |
} |
| 56 |
|
| 57 |
BitSet<> &RegSet::bitSetFor(RegName rn) { |
| 58 |
auto bs = bitSetForPtr(rn); |
| 59 |
IGA_ASSERT(bs, "not a tracked bitset"); |
| 60 |
return *bs; |
| 61 |
} |
| 62 |
const BitSet<> &RegSet::bitSetFor(RegName rn) const { |
| 63 |
return const_cast<RegSet *>(this)->bitSetFor(rn); |
| 64 |
} |
| 65 |
|
| 66 |
bool RegSet::isTrackedReg(RegName rn) const { |
| 67 |
return const_cast<RegSet *>(this)->bitSetForPtr(rn) != nullptr; |
| 68 |
} |
| 69 |
|
| 70 |
static size_t OffsetOf(const Model &m, RegName rn, int reg) { |
| 71 |
return (size_t)reg * m.getBytesPerReg(rn) * 8; |
| 72 |
} |
| 73 |
|
| 74 |
size_t RegSet::offsetOf(RegName rn, int reg) const { |
| 75 |
return OffsetOf(model, rn, reg); |
| 76 |
} |
| 77 |
size_t RegSet::offsetOf(RegName rn, RegRef rr, size_t typeSizeBits) const { |
| 78 |
return OffsetOf(model, rn, rr.regNum) + rr.subRegNum * typeSizeBits; |
| 79 |
} |
| 80 |
size_t RegSet::offsetOf(RegName rn, RegRef rr, Type t) const { |
| 81 |
size_t typeSizeBits = TypeSizeInBitsWithDefault(t, 8); |
| 82 |
return offsetOf(rn, rr, typeSizeBits); |
| 83 |
} |
| 84 |
|
| 85 |
RegSet::RegSet(const Model &m) |
| 86 |
: model(m), bitsR(BytesForRegSet(RegName::GRF_R, m)), |
| 87 |
bitsA(BytesForRegSet(RegName::ARF_A, m)), |
| 88 |
bitsAcc(BytesForRegSet(RegName::ARF_ACC, m)), |
| 89 |
bitsF(BytesForRegSet(RegName::ARF_F, m)), |
| 90 |
bitsS(BytesForRegSet(RegName::ARF_S, m)) |
| 91 |
{ |
| 92 |
} |
| 93 |
RegSet::RegSet(const RegSet &rs) |
| 94 |
: model(rs.model), bitsR(rs.bitsR), bitsA(rs.bitsA), bitsAcc(rs.bitsAcc), |
| 95 |
bitsF(rs.bitsF), bitsS(rs.bitsS) |
| 96 |
{ |
| 97 |
} |
| 98 |
RegSet &RegSet::operator=(const RegSet &rhs) { |
| 99 |
IGA_ASSERT(&model == &rhs.model, "model mismatch"); |
| 100 |
for (RegName rn : TRACKED) { |
| 101 |
bitSetFor(rn) = rhs.bitSetFor(rn); |
| 102 |
} |
| 103 |
return *this; |
| 104 |
} |
| 105 |
|
| 106 |
bool RegSet::operator==(const RegSet &rs) const { |
| 107 |
for (RegName rn : TRACKED) { |
| 108 |
if (bitSetFor(rn) != rs.bitSetFor(rn)) { |
| 109 |
return false; |
| 110 |
} |
| 111 |
} |
| 112 |
return true; |
| 113 |
} |
| 114 |
bool RegSet::empty() const { |
| 115 |
for (RegName rn : TRACKED) |
| 116 |
if (!bitSetFor(rn).empty()) |
| 117 |
return false; |
| 118 |
return true; |
| 119 |
} |
| 120 |
void RegSet::reset() { |
| 121 |
for (RegName rn : TRACKED) |
| 122 |
bitSetFor(rn).reset(); |
| 123 |
} |
| 124 |
|
| 125 |
bool RegSet::intersects(const RegSet &rhs) const { |
| 126 |
for (RegName rn : TRACKED) |
| 127 |
if (bitSetFor(rn).intersects(rhs.bitSetFor(rn))) |
| 128 |
return true; |
| 129 |
return false; |
| 130 |
} |
| 131 |
|
| 132 |
bool RegSet::destructiveUnion(const RegSet &rhs) { |
| 133 |
bool changed = false; |
| 134 |
for (RegName rn : TRACKED) |
| 135 |
changed |= bitSetFor(rn).add(rhs.bitSetFor(rn)); |
| 136 |
return changed; |
| 137 |
} |
| 138 |
|
| 139 |
bool RegSet::destructiveSubtract(const RegSet &rhs) { |
| 140 |
bool changed = false; |
| 141 |
for (RegName rn : TRACKED) |
| 142 |
changed |= bitSetFor(rn).andNot(rhs.bitSetFor(rn)); |
| 143 |
return changed; |
| 144 |
} |
| 145 |
|
| 146 |
bool RegSet::intersectInto(const RegSet &rhs, RegSet &into) const { |
| 147 |
bool notEmpty = false; |
| 148 |
for (RegName rn : TRACKED) { |
| 149 |
notEmpty |= |
| 150 |
bitSetFor(rn).intersectInto(rhs.bitSetFor(rn), into.bitSetFor(rn)); |
| 151 |
} |
| 152 |
return notEmpty; |
| 153 |
} |
| 154 |
|
| 155 |
bool RegSet::addReg(RegName rn, int reg) { return addRegs(rn, reg, 1); } |
| 156 |
|
| 157 |
bool RegSet::addRegs(RegName rn, int reg, int n) { |
| 158 |
size_t bitsPerReg = 8 * (size_t)model.getBytesPerReg(rn); |
| 159 |
return add(rn, bitsPerReg * reg, bitsPerReg * n); |
| 160 |
} |
| 161 |
|
| 162 |
bool RegSet::add(RegName rn, size_t regFileOffBits, size_t numBits) { |
| 163 |
if (!isTrackedReg(rn)) { |
| 164 |
return false; |
| 165 |
} |
| 166 |
// coarsen analysis to byte |
| 167 |
// need std::max since numBits might be <8 (e.g. flag reg SIMD1) |
| 168 |
size_t offBytes = regFileOffBits / 8; |
| 169 |
size_t lenBytes = std::max<size_t>(numBits / 8, 1); |
| 170 |
BitSet<> &bs = bitSetFor(rn); |
| 171 |
if (offBytes + lenBytes <= BytesForRegSet(rn, model)) { |
| 172 |
return bs.set(offBytes, lenBytes); |
| 173 |
} else { |
| 174 |
IGA_ASSERT_FALSE("access is out of bounds for register"); |
| 175 |
return false; |
| 176 |
} |
| 177 |
} |
| 178 |
|
| 179 |
bool RegSet::add(RegName rn, RegRef rr, Type t) { |
| 180 |
size_t typeSizeBits = TypeSizeInBitsWithDefault(t, 8); |
| 181 |
return add(rn, offsetOf(rn, rr, t), typeSizeBits); |
| 182 |
} |
| 183 |
|
| 184 |
bool RegSet::addFlagRegArf(RegRef fr, ExecSize es, ChannelOffset co) { |
| 185 |
size_t execSize = size_t(es); |
| 186 |
size_t execOff = 4 * size_t(co); |
| 187 |
|
| 188 |
// can't use subregisters here because subregisters are only a |
| 189 |
// word boundary. E.g. (f1.1) op (M8|8) touches upper byte of f1.1 |
| 190 |
// which doesn't have a syntax representation |
| 191 |
size_t off = offsetOf(RegName::ARF_F, fr, 16); |
| 192 |
|
| 193 |
RegRef frWithOff = fr; |
| 194 |
frWithOff.subRegNum += (uint16_t)execOff / 16; // 16b per subreg |
| 195 |
return add(RegName::ARF_F, off + execOff, execSize); |
| 196 |
} |
| 197 |
|
| 198 |
bool RegSet::addPredicationInputs(const Instruction &i) { |
| 199 |
// does it read the flag register |
| 200 |
// predication does this |
| 201 |
// conditional modifier on 'sel' does this |
| 202 |
const Predication &pred = i.getPredication(); |
| 203 |
bool readsFlagRegister = |
| 204 |
pred.function != PredCtrl::NONE && i.getOp() != Op::SEL; |
| 205 |
if (readsFlagRegister) { |
| 206 |
return addFlagAccess(i); |
| 207 |
} |
| 208 |
return false; |
| 209 |
} |
| 210 |
|
| 211 |
bool RegSet::addSourceInputs(const Instruction &i) { |
| 212 |
bool added = false; |
| 213 |
|
| 214 |
added |= addSourceDescriptorInputs(i); |
| 215 |
|
| 216 |
added |= addSourceImplicit(i); |
| 217 |
|
| 218 |
added |= addDestinationInputs(i); // e.g. indirect access uses a0 |
| 219 |
|
| 220 |
// check all the source operands |
| 221 |
for (unsigned srcIx = 0; srcIx < i.getSourceCount(); srcIx++) { |
| 222 |
added |= addSourceOperandInput(i, srcIx); |
| 223 |
} |
| 224 |
return added; |
| 225 |
} |
| 226 |
|
| 227 |
bool RegSet::addSendOperand(const Instruction &i, int opIx) { |
| 228 |
IGA_ASSERT(i.getOpSpec().isAnySendFormat(), "Expected a send format"); |
| 229 |
|
| 230 |
bool changed = false; |
| 231 |
const Operand &op = opIx < 0 ? i.getDestination() : i.getSource(opIx); |
| 232 |
if (op.getDirRegName() != RegName::GRF_R) { |
| 233 |
return false; |
| 234 |
} |
| 235 |
|
| 236 |
// new approach |
| 237 |
if (i.getOpSpec().isSendgFormat()) { |
| 238 |
const DecodeResult di = tryDecode(i, nullptr); |
| 239 |
if (di) { |
| 240 |
int bitsAccessed = -1; |
| 241 |
if (opIx < 0) { |
| 242 |
bitsAccessed = di.info.dstLenBytes; |
| 243 |
} else if (opIx == 0) { |
| 244 |
bitsAccessed = di.info.src0LenBytes; |
| 245 |
} else { |
| 246 |
bitsAccessed = di.info.src1LenBytes; |
| 247 |
} |
| 248 |
if (bitsAccessed >= 0) { |
| 249 |
auto regOff = offsetOf(RegName::GRF_R, op.getDirRegRef().regNum); |
| 250 |
changed |= add(RegName::GRF_R, regOff, 8 * bitsAccessed); |
| 251 |
return changed; |
| 252 |
} |
| 253 |
} |
| 254 |
} |
| 255 |
|
| 256 |
// add special handling for send messages |
| 257 |
// e.g. |
| 258 |
// send (4) r10:1 ... untyped 32b read of X |
| 259 |
// reads only 4 dwords |
| 260 |
// send (1) r10:1 OWORD block read |
| 261 |
// reads 128b ... |
| 262 |
// load.ugm.d32x8t.a32.ca.ca (1|M0) ... 8 DW in a single register |
| 263 |
// load.ugm.d32x4.a64 (4|M0) ... // four successive registers with 1 DW |
| 264 |
bool decodeFineGrained = |
| 265 |
!i.getOpSpec().isSendgFormat() || i.getMsgDescriptor().isImm(); |
| 266 |
if (op.getDirRegName() == RegName::GRF_R && decodeFineGrained) { |
| 267 |
const DecodeResult di = tryDecode(i, nullptr); |
| 268 |
if (di && (di.info.isLoad() || di.info.isStore() || di.info.isAtomic())) { |
| 269 |
// the compiler won't necesssarily generate a correct exec size |
| 270 |
// when it's ignored (part of the descriptor) |
| 271 |
auto regOff = offsetOf(RegName::GRF_R, op.getDirRegRef().regNum); |
| 272 |
if (opIx == 0) { |
| 273 |
size_t bitsAccessed = |
| 274 |
(size_t)(di.info.execWidth * di.info.addrSizeBits); |
| 275 |
changed |= add(RegName::GRF_R, regOff, bitsAccessed); |
| 276 |
} else { |
| 277 |
// data type |
| 278 |
if (di.info.isTransposed()) { |
| 279 |
// transpose packs into GRF |
| 280 |
size_t bitsAccessed = |
| 281 |
(size_t)(di.info.elemsPerAddr * di.info.execWidth * |
| 282 |
di.info.elemSizeBitsRegFile); |
| 283 |
changed |= add(RegName::GRF_R, regOff, bitsAccessed); |
| 284 |
} else { |
| 285 |
// non-transpose places elements in successive registers |
| 286 |
size_t bitsAccessed = |
| 287 |
(size_t)(di.info.execWidth * di.info.elemSizeBitsRegFile); |
| 288 |
int elemsPerAddr = di.info.elemsPerAddr; |
| 289 |
if (opIx == 1 && (di.info.op == SendOp::ATOMIC_ICAS || |
| 290 |
di.info.op == SendOp::ATOMIC_FCAS)) { |
| 291 |
// ICAS packs src2 in to src1's payload |
| 292 |
elemsPerAddr *= 2; |
| 293 |
} |
| 294 |
for (int ei = 0; ei < elemsPerAddr; ei++) { |
| 295 |
// each vector element starts at a new GRF |
| 296 |
// vector elements might be <GRF (small SIMD) or |
| 297 |
// multiple GRF (e.g. SIMD32 D64V2) |
| 298 |
changed |= add(RegName::GRF_R, regOff, bitsAccessed); |
| 299 |
regOff += ALIGN_UP_TO(8 * model.getGRFByteSize(), bitsAccessed); |
| 300 |
} |
| 301 |
} |
| 302 |
} |
| 303 |
return changed; // bail out early |
| 304 |
} |
| 305 |
} |
| 306 |
// else: fallthrough on the full register counts |
| 307 |
|
| 308 |
// send source GRF (not null reg) |
| 309 |
int nregs = 0; |
| 310 |
if (opIx < 0) { |
| 311 |
nregs = i.getDstLength(); |
| 312 |
if (nregs < 0) |
| 313 |
nregs = 8; // assume the worst |
| 314 |
} else if (opIx == 0) { // mlen |
| 315 |
nregs = i.getSrc0Length(); |
| 316 |
if (nregs < 0) |
| 317 |
nregs = 4; // assume the worst |
| 318 |
} else { // xlen |
| 319 |
nregs = i.getSrc1Length(); |
| 320 |
if (nregs < 0) |
| 321 |
nregs = 8; // assume the worst |
| 322 |
} |
| 323 |
int regNum = (int)op.getDirRegRef().regNum; |
| 324 |
changed |= addRegs(RegName::GRF_R, regNum, nregs); |
| 325 |
return changed; |
| 326 |
} |
| 327 |
|
| 328 |
// c.f. DepSet::getDPASSrcDepUpBound |
| 329 |
bool RegSet::addDpasOperand(const Instruction &i, int opIx) { |
| 330 |
const Operand &op = opIx < 0 ? i.getDestination() : i.getSource(opIx); |
| 331 |
if (op.getDirRegName() != RegName::GRF_R) { |
| 332 |
return false; // e.g. null |
| 333 |
} |
| 334 |
|
| 335 |
// The operand sizes of the multiplicands dictate how many ops per |
| 336 |
// channel we process. |
| 337 |
size_t mulSize = |
| 338 |
std::max(TypeSizeInBitsWithDefault(i.getSource(1).getType(), 8), |
| 339 |
TypeSizeInBitsWithDefault(i.getSource(2).getType(), 8)); |
| 340 |
size_t opsPerChannel = 32 / mulSize; // e.g. hf is 2 ... |
| 341 |
size_t execSize = size_t(i.getExecSize()); |
| 342 |
size_t R = GetDpasRepeatCount(i.getDpasFc()); |
| 343 |
size_t S = GetDpasSystolicDepth(i.getDpasFc()); |
| 344 |
size_t typeSizeBits = TypeSizeInBitsWithDefault(op.getType(), 32); |
| 345 |
|
| 346 |
size_t startOff = offsetOf(RegName::GRF_R, op.getDirRegRef(), op.getType()); |
| 347 |
bool changed = false; |
| 348 |
if (opIx <= 0) { // dst/src0 |
| 349 |
size_t bitsAccessed = execSize * typeSizeBits * R; |
| 350 |
changed |= add(RegName::GRF_R, startOff, bitsAccessed); |
| 351 |
} else if (opIx == 1) { // src1 |
| 352 |
size_t bitsAccessed = execSize * typeSizeBits * opsPerChannel * S; |
| 353 |
changed |= add(RegName::GRF_R, startOff, bitsAccessed); |
| 354 |
} else if (opIx == 2) { // src2 |
| 355 |
// this is a little subtle |
| 356 |
// given bit precisions a src2 operand can stride across the |
| 357 |
// bottom or top of a GRF without touching the other half |
| 358 |
// |
| 359 |
// e.g. dpas.4x2 (8) r24.0:d r24.0:d r6.0:u2 r14.32:u4 |
| 360 |
// src2 - touches r14[16-31] and r15[16-31] |
| 361 |
// |
| 362 |
// TODO: can something like this happen for src1 |
| 363 |
for (size_t r = 0; r < R; r++) { |
| 364 |
size_t startSubReg = r * opsPerChannel * typeSizeBits * 8; |
| 365 |
size_t repReads = S * opsPerChannel * typeSizeBits; |
| 366 |
changed |= add(RegName::GRF_R, startOff + startSubReg, repReads); |
| 367 |
} |
| 368 |
} |
| 369 |
else { |
| 370 |
// Currently only bdpas has src3/src4, and only byte type is allowed on |
| 371 |
// src3/src4. |
| 372 |
IGA_ASSERT(i.getOp() == Op::BDPAS, "invalid src idx"); |
| 373 |
IGA_ASSERT(typeSizeBits == 8, "invalid src type size"); |
| 374 |
auto isFP4 = [](Type ty) { |
| 375 |
return ty == Type::E2M1; |
| 376 |
}; |
| 377 |
uint16_t passes = |
| 378 |
(isFP4(i.getSource(1).getType()) || isFP4(i.getSource(2).getType())) ? |
| 379 |
2 : 1; |
| 380 |
for (uint16_t pass = 0; pass < passes; pass++) { |
| 381 |
auto tmpRR = op.getDirRegRef(); |
| 382 |
tmpRR.subRegNum += 32 * pass; |
| 383 |
if (opIx == 3) { |
| 384 |
// src3 accesses number of execSize bytes in each pass |
| 385 |
changed |= setSrcRegion(op.getDirRegName(), tmpRR, |
| 386 |
Region::SRC110, execSize, typeSizeBits); |
| 387 |
} else if (opIx == 4) { |
| 388 |
// src4 accesses number of SDepth bytes in each pass |
| 389 |
changed |= setSrcRegion(op.getDirRegName(), tmpRR, |
| 390 |
Region::SRC110, S, typeSizeBits); |
| 391 |
} |
| 392 |
} |
| 393 |
} |
| 394 |
return changed; |
| 395 |
} |
| 396 |
|
| 397 |
bool RegSet::addSourceOperandInput(const Instruction &i, int srcIx) { |
| 398 |
size_t execSize = size_t(i.getExecSize()); |
| 399 |
bool added = false; |
| 400 |
const Operand &op = i.getSource(srcIx); |
| 401 |
auto typeSizeBits = TypeSizeInBitsWithDefault(op.getType(), 32); |
| 402 |
|
| 403 |
// possibly need to fixup the region |
| 404 |
Region rgn = op.getRegion(); |
| 405 |
if (i.getSourceCount() == 3) { |
| 406 |
if (i.isMacro()) { |
| 407 |
// macros and sends are packed access only |
| 408 |
// madm (4) r3.acc1 r7.acc2 r9.acc4 r6.noacc |
| 409 |
rgn = Region::SRC110; |
| 410 |
} else { |
| 411 |
// ternary align 1 has some implicit regions that need filling in |
| 412 |
// "GEN10 Regioning Rules for Align1 Ternary Operations" |
| 413 |
// 1. Width is 1 when Vertical and Horizontal Strides are both zero |
| 414 |
// (broadcast access). |
| 415 |
// 2. Width is equal to Vertical Stride when Horizontal Stride is zero. |
| 416 |
// 3. Width is equal to Vertical Stride/Horizontal Stride when both |
| 417 |
// Strides are non-zero. |
| 418 |
// 4. Vertical Stride must not be zero if Horizontal Stride is non-zero. |
| 419 |
// This implies Vertical Stride is always greater than Horizontal |
| 420 |
// Stride. |
| 421 |
// 5. For Source 2, if Horizontal Stride is non-zero, then Width is the |
| 422 |
// a register's width of elements (e.g. 8 for a 32-bit data type). |
| 423 |
// Otherwise, if Horizontal Stride is 0, then so is the Vertical (and |
| 424 |
// rule 1 applies). This means Vertical Stride is always 'Width' * |
| 425 |
// 'Horizontal Stride'. |
| 426 |
if (srcIx < 2) { |
| 427 |
// <V;H> |
| 428 |
if (rgn.getVt() == Region::Vert::VT_0 && |
| 429 |
rgn.getHz() == Region::Horz::HZ_0) { |
| 430 |
rgn.set(Region::Width::WI_1); // by #1 |
| 431 |
} else if (rgn.getHz() == Region::Horz::HZ_0) { |
| 432 |
rgn.w = rgn.v; // by #2 |
| 433 |
} else if (rgn.getVt() != Region::Vert::VT_0 && |
| 434 |
rgn.getHz() != Region::Horz::HZ_0) { |
| 435 |
rgn.w = rgn.v / rgn.h; // by #3 |
| 436 |
} else { |
| 437 |
// error condition #4, just use vertical stride |
| 438 |
// SPECIFY: should this be an assertion? |
| 439 |
rgn.w = rgn.v; |
| 440 |
} |
| 441 |
} else { // (srcIx == 2) |
| 442 |
// <H> |
| 443 |
// <H> -> <H;1,0> |
| 444 |
// if (rgn.h == Region::HZ_0) { |
| 445 |
// rgn.w = Region::WI_1; // by #5 and #1 |
| 446 |
//} else { |
| 447 |
// rgn.w = |
| 448 |
// static_cast<Region::Width>(GRF_BYTES_PER_REG/typeSize); // |
| 449 |
// by #5 |
| 450 |
//} |
| 451 |
// rgn.v = static_cast<Region::Vert>( |
| 452 |
// static_cast<int>(rgn.w)*static_cast<int>(rgn.h)/typeSize); |
| 453 |
// // by #5 |
| 454 |
rgn.set(Region::Vert(size_t(rgn.h)), Region::Width::WI_1, |
| 455 |
Region::Horz::HZ_0); |
| 456 |
} |
| 457 |
} |
| 458 |
} else if (i.getSourceCount() == 2 && i.isMacro()) { |
| 459 |
// math macro |
| 460 |
rgn = Region::SRC110; |
| 461 |
} // not ternary, not macro ==> regions are good |
| 462 |
|
| 463 |
switch (op.getKind()) { |
| 464 |
case Operand::Kind::DIRECT: |
| 465 |
added = true; |
| 466 |
if (i.getOpSpec().isAnySendFormat()) { |
| 467 |
if (op.getDirRegName() == RegName::GRF_R) { |
| 468 |
addSendOperand(i, srcIx); |
| 469 |
} |
| 470 |
} else if (i.getOpSpec().isDpasFormat()) { |
| 471 |
added |= addDpasOperand(i, srcIx); |
| 472 |
} else { |
| 473 |
setSrcRegion(op.getDirRegName(), op.getDirRegRef(), rgn, execSize, |
| 474 |
typeSizeBits); |
| 475 |
} |
| 476 |
break; |
| 477 |
case Operand::Kind::MACRO: |
| 478 |
added = true; |
| 479 |
setSrcRegion(op.getDirRegName(), op.getDirRegRef(), rgn, execSize, |
| 480 |
typeSizeBits); |
| 481 |
// Don't bother tracking MME registers, technically they are 2 bits |
| 482 |
// of exponent each, but they are mixed in with other ACC creating a |
| 483 |
// royal mess. |
| 484 |
// |
| 485 |
// auto mme = op.getMathMacroExt(); |
| 486 |
// if (mme != MathMacroExt::NOMME && mme != MathMacroExt::INVALID) { |
| 487 |
// int mmeRegNum = int(op.getMathMacroExt()) - |
| 488 |
// int(MathMacroExt::MME0); |
| 489 |
// rs.setSrcRegion( |
| 490 |
// RegName::ARF_ACC, |
| 491 |
// RegRef(mmeRegNum, 0), |
| 492 |
// Region::SRC110, |
| 493 |
// execSize, |
| 494 |
// typeSizeBits); |
| 495 |
// } |
| 496 |
break; |
| 497 |
case Operand::Kind::INDIRECT: { |
| 498 |
if (i.isGatherSend()) { |
| 499 |
// reads one byte per GRF from Src0.Length (rlen in old send) |
| 500 |
int src0Len = i.getSrc0Length(); |
| 501 |
if (src0Len > 0) { |
| 502 |
RegRef rr = i.getSource(0).getIndAddrReg(); |
| 503 |
for (int i = 0; i < src0Len; i++) { |
| 504 |
added |= add(RegName::ARF_S, rr, Type::UB); |
| 505 |
rr.subRegNum++; |
| 506 |
} |
| 507 |
} |
| 508 |
break; |
| 509 |
} |
| 510 |
added = true; |
| 511 |
auto rgn = op.getRegion(); |
| 512 |
if (rgn.getVt() == Region::Vert::VT_VxH) { |
| 513 |
// VxH or Vx1 mode |
| 514 |
// op (K) dst r[a0.0]<W,1>:w (reads K/W elements) |
| 515 |
// op (K) dst r[a0.0]<1,0>:w (reads K elements) |
| 516 |
setSrcRegion(RegName::ARF_A, op.getIndAddrReg(), Region::SRC110, |
| 517 |
execSize / rgn.w, // |
| 518 |
16); // :w is 16-bits a piece |
| 519 |
} else { |
| 520 |
// uniform: consumes one value in a0 |
| 521 |
// op (..) dst r[a0.0]<16,8,1>:w |
| 522 |
setSrcRegion(RegName::ARF_A, op.getIndAddrReg(), Region::SRC110, |
| 523 |
1, // 1 element only |
| 524 |
16); // :w is 16-bits a piece |
| 525 |
} |
| 526 |
// we can't do anything else for this |
| 527 |
break; |
| 528 |
} |
| 529 |
default: |
| 530 |
break; |
| 531 |
} |
| 532 |
return added; |
| 533 |
} |
| 534 |
|
| 535 |
static bool addSendDescriptors(const Instruction &i, RegSet &rs) { |
| 536 |
bool added = false; |
| 537 |
// send register descriptors may touch a0.# |
| 538 |
auto desc = i.getMsgDescriptor(); |
| 539 |
if (desc.isReg()) { |
| 540 |
// send is strange: a0.# is in word offsets but reads 32b |
| 541 |
// fake it by converting a0.4:uw to a0.2:ud and writing the full 32b |
| 542 |
desc.reg.subRegNum /= 2; |
| 543 |
added |= rs.add(RegName::ARF_A, desc.reg, Type::UD); |
| 544 |
} |
| 545 |
auto exDesc = i.getExtMsgDescriptor(); |
| 546 |
if (exDesc.isReg()) { |
| 547 |
// see above |
| 548 |
exDesc.reg.subRegNum /= 2; |
| 549 |
added |= rs.add(RegName::ARF_A, exDesc.reg, Type::UD); |
| 550 |
} |
| 551 |
return added; |
| 552 |
} |
| 553 |
|
| 554 |
bool RegSet::addSourceDescriptorInputs(const Instruction &i) { |
| 555 |
bool added = false; |
| 556 |
|
| 557 |
if (i.getOpSpec().isAnySendFormat()) { |
| 558 |
if (i.getOpSpec().isSendgFormat()) { |
| 559 |
RegRef id0 = i.getSendgIndDesc0Reg(); |
| 560 |
if (id0 != REGREF_INVALID) { |
| 561 |
id0.subRegNum /= 8; // normalize to QW |
| 562 |
added |= add(RegName::ARF_S, id0, Type::UQ); |
| 563 |
} |
| 564 |
RegRef id1 = i.getSendgIndDesc1Reg(); |
| 565 |
if (id1 != REGREF_INVALID) { |
| 566 |
id1.subRegNum /= 8; // normalize to QW |
| 567 |
added |= add(RegName::ARF_S, id1, Type::UQ); |
| 568 |
} |
| 569 |
} else { |
| 570 |
addSendDescriptors(i, *this); |
| 571 |
} |
| 572 |
} |
| 573 |
|
| 574 |
return added; |
| 575 |
} |
| 576 |
|
| 577 |
bool RegSet::addSourceImplicit(const Instruction &i) { |
| 578 |
bool changed = false; |
| 579 |
changed |= addImplicitAccumulatorAccess(i, false); |
| 580 |
if (i.getOp() == Op::SEL && i.hasPredication()) { |
| 581 |
changed |= addFlagAccess(i); |
| 582 |
} |
| 583 |
return changed; |
| 584 |
} |
| 585 |
|
| 586 |
bool RegSet::addDestinationInputs(const Instruction &i) { |
| 587 |
if (!i.getOpSpec().supportsDestination()) { |
| 588 |
return false; |
| 589 |
} |
| 590 |
bool added = false; |
| 591 |
const auto &dst = i.getDestination(); |
| 592 |
if (dst.getKind() == Operand::Kind::INDIRECT) { |
| 593 |
added |= setDstRegion(RegName::ARF_A, dst.getIndAddrReg(), Region::DST1, |
| 594 |
1, // one element only |
| 595 |
16); // :w is 16-bits |
| 596 |
} |
| 597 |
return added; |
| 598 |
} |
| 599 |
|
| 600 |
bool RegSet::addExplicitDestinationOutputs(const Instruction &i) { |
| 601 |
if (!i.getOpSpec().supportsDestination()) { |
| 602 |
return false; |
| 603 |
} |
| 604 |
bool added = false; |
| 605 |
|
| 606 |
int execSize = int(i.getExecSize()); |
| 607 |
const auto &op = i.getDestination(); |
| 608 |
auto typeSizeBits = TypeSizeInBitsWithDefault(op.getType(), 32); |
| 609 |
|
| 610 |
switch (op.getKind()) { |
| 611 |
case Operand::Kind::DIRECT: |
| 612 |
// includes send target (a GRF, not null reg) |
| 613 |
if (i.getOpSpec().isAnySendFormat() && |
| 614 |
op.getDirRegName() == RegName::GRF_R) { |
| 615 |
addSendOperand(i, -1); |
| 616 |
} else if (i.getOpSpec().isDpasFormat()) { |
| 617 |
// destination is mirror of src0 |
| 618 |
added |= addDpasOperand(i, -1); |
| 619 |
} else if (op.getDirRegName() == RegName::ARF_ACC && i.is(Op::MUL) && |
| 620 |
(op.getType() == Type::D || op.getType() == Type::UD)) { |
| 621 |
// special case: |
| 622 |
// mul (..) acc0:d ... // writes acc0[63:0] |
| 623 |
// mach (..) .... // reads acc0[63:0] |
| 624 |
// technically I think mul only writes the bottom 48b |
| 625 |
added |= setDstRegion(op.getDirRegName(), op.getDirRegRef(), |
| 626 |
op.getRegion(), execSize, 64); |
| 627 |
} else { |
| 628 |
// normal GRF target |
| 629 |
added |= setDstRegion(op.getDirRegName(), op.getDirRegRef(), |
| 630 |
op.getRegion(), execSize, typeSizeBits); |
| 631 |
} |
| 632 |
if (i.is(Op::MULLH)) { |
| 633 |
// mullh is an odd duck; it writes out low/high pairs in |
| 634 |
// - a cheap trick is to just increment the GRF and add the new region |
| 635 |
// for MULLH |
| 636 |
// - we increment the GRF twice if it's SIMD32 since the low half will |
| 637 |
// chew up two GRFs instead of one. |
| 638 |
// |
| 639 |
// |
| 640 |
RegRef rr = op.getDirRegRef(); |
| 641 |
rr.regNum++; |
| 642 |
if (execSize > 16) |
| 643 |
rr.regNum++; |
| 644 |
added |= setDstRegion(op.getDirRegName(), rr, op.getRegion(), execSize, |
| 645 |
typeSizeBits); |
| 646 |
} |
| 647 |
break; |
| 648 |
case Operand::Kind::MACRO: { |
| 649 |
// math macro |
| 650 |
// GRF + ACC |
| 651 |
added |= setDstRegion(op.getDirRegName(), op.getDirRegRef(), Region::DST1, |
| 652 |
execSize, typeSizeBits); |
| 653 |
// don't track MME registers (see RegSet::addSourceInputs for why) |
| 654 |
// auto MathMacroReg = op.getMathMacroExt(); |
| 655 |
// if (MathMacroReg != MathMacroExt::NOMME && MathMacroReg != |
| 656 |
// MathMacroExt::INVALID) { |
| 657 |
// // and the math macro register |
| 658 |
// int mmeRegNum = int(MathMacroReg) - int(MathMacroExt::MME0); |
| 659 |
// RegRef mmeRegRef(mmeRegNum, 0); |
| 660 |
// added |= rs.setDstRegion( |
| 661 |
// RegName::ARF_MME, |
| 662 |
// mmeRegRef, |
| 663 |
// Region::DST1, |
| 664 |
// execSize, |
| 665 |
// typeSizeBits); |
| 666 |
// } |
| 667 |
break; |
| 668 |
} |
| 669 |
case Operand::Kind::INDIRECT: |
| 670 |
// this is an input! not an output |
| 671 |
// (use addDestinationInputs or part of addSourceInputs) |
| 672 |
// |
| 673 |
// indirect destinations use a single a0 value |
| 674 |
// (no fancy scattering) |
| 675 |
// added |= setDstRegion( |
| 676 |
// RegName::ARF_A, |
| 677 |
// op.getIndAddrReg(), |
| 678 |
// Region::DST1, |
| 679 |
// 1, // one element only |
| 680 |
// 16); // :w is 16-bits |
| 681 |
break; |
| 682 |
default: |
| 683 |
break; |
| 684 |
} |
| 685 |
return added; |
| 686 |
} |
| 687 |
|
| 688 |
bool RegSet::addDestinationOutputs(const Instruction &i) { |
| 689 |
bool changed = false; |
| 690 |
changed |= addDestinationImplicit(i); |
| 691 |
changed |= addExplicitDestinationOutputs(i); |
| 692 |
return changed; |
| 693 |
} |
| 694 |
|
| 695 |
bool RegSet::addFlagModifierOutputs(const Instruction &i) { |
| 696 |
const FlagModifier fm = i.getFlagModifier(); |
| 697 |
bool writesFlagRegister = // sel uses flag modifier as input |
| 698 |
fm != FlagModifier::NONE && i.getOp() != Op::SEL; |
| 699 |
if (writesFlagRegister) { |
| 700 |
return addFlagAccess(i); |
| 701 |
} |
| 702 |
return false; |
| 703 |
} |
| 704 |
|
| 705 |
bool RegSet::addDestinationImplicit(const Instruction &i) { |
| 706 |
// other special instructions that write things |
| 707 |
return addImplicitAccumulatorAccess(i, true); |
| 708 |
} |
| 709 |
|
| 710 |
bool RegSet::addImplicitAccumulatorAccess(const Instruction &i, bool isDst) { |
| 711 |
bool added = false; |
| 712 |
|
| 713 |
bool accessesAcc = false; |
| 714 |
accessesAcc |= i.is(Op::MACH); // mac and mach uses both input and output |
| 715 |
accessesAcc |= isDst && i.hasInstOpt(InstOpt::ACCWREN); |
| 716 |
accessesAcc |= (isDst && i.is(Op::ADDC)) || i.is(Op::SUBB); |
| 717 |
accessesAcc |= !isDst && i.is(Op::MAC); |
| 718 |
accessesAcc |= i.is(Op::MACL); |
| 719 |
if (accessesAcc) { |
| 720 |
Type type = i.getDestination().getType(); |
| 721 |
if (i.is(Op::MACH)) { |
| 722 |
type = Type::Q; // acc0[63:0] is written |
| 723 |
} |
| 724 |
if (i.is(Op::MACL)) { |
| 725 |
type = Type::Q; // acc0[63:0] is written according to BXML |
| 726 |
} |
| 727 |
auto typeSizeBytes = TypeSizeInBitsWithDefault(type, 32); |
| 728 |
// unlike flag modifier and predicates the channel offset does not |
| 729 |
// impact accumulator |
| 730 |
added |= setDstRegion(RegName::ARF_ACC, RegRef(0, 0), // acc0 |
| 731 |
Region::DST1, size_t(i.getExecSize()), typeSizeBytes); |
| 732 |
} |
| 733 |
return added; |
| 734 |
} |
| 735 |
|
| 736 |
bool RegSet::addFlagAccess(const Instruction &i) { |
| 737 |
const RegRef &fr = i.getFlagReg(); |
| 738 |
return addFlagRegArf(fr, i.getExecSize(), i.getChannelOffset()); |
| 739 |
} |
| 740 |
|
| 741 |
bool RegSet::setDstRegion(RegName rn, RegRef rr, Region rgn, size_t execSize, |
| 742 |
size_t typeSizeBits) { |
| 743 |
if (!isTrackedReg(rn)) |
| 744 |
return false; |
| 745 |
|
| 746 |
size_t hz = size_t(rgn.getHz()); |
| 747 |
// sets a region for a basic operand |
| 748 |
if (rgn.getHz() == Region::Horz::HZ_INVALID) { |
| 749 |
hz = 1; |
| 750 |
} |
| 751 |
|
| 752 |
bool added = false; |
| 753 |
size_t baseOffBits = |
| 754 |
rr.regNum * model.getBytesPerReg(rn) * 8 + rr.subRegNum * typeSizeBits; |
| 755 |
for (size_t ch = 0; ch < execSize; ch++) { |
| 756 |
size_t offsetBits = ch * hz * typeSizeBits; |
| 757 |
added |= add(rn, baseOffBits + offsetBits, typeSizeBits); |
| 758 |
} |
| 759 |
return added; |
| 760 |
} |
| 761 |
|
| 762 |
// sets a region for a basic operand |
| 763 |
bool RegSet::setSrcRegion(RegName rn, RegRef rr, Region rgn, size_t execSize, |
| 764 |
size_t typeSizeBits) { |
| 765 |
if (!isTrackedReg(rn)) |
| 766 |
return false; |
| 767 |
|
| 768 |
bool changed = false; |
| 769 |
size_t v = 1, w = 1, h = 0; // e.g. old-style default regions |
| 770 |
if (rgn != Region::INVALID) { |
| 771 |
v = size_t(rgn.getVt()); |
| 772 |
w = size_t(rgn.getWi()); |
| 773 |
w = w == 0 ? 1 : w; |
| 774 |
h = size_t(rgn.getHz()); |
| 775 |
} |
| 776 |
|
| 777 |
size_t rowBaseBits = offsetOf(rn, rr, typeSizeBits); |
| 778 |
size_t rows = execSize / w; |
| 779 |
rows = (rows != 0) ? rows : 1; |
| 780 |
for (size_t y = 0; y < rows; y++) { |
| 781 |
size_t offsetBits = rowBaseBits; |
| 782 |
for (size_t x = 0; x < w; x++) { |
| 783 |
changed |= add(rn, offsetBits, typeSizeBits); |
| 784 |
offsetBits += h * typeSizeBits; |
| 785 |
} |
| 786 |
rowBaseBits += v * typeSizeBits; |
| 787 |
} |
| 788 |
return changed; |
| 789 |
} |
| 790 |
|
| 791 |
static void strForReg(const Model &model, RegName rn, const RegSet::Bits &bs, |
| 792 |
bool &first, std::ostream &os) { |
| 793 |
const RegInfo *ri = model.lookupRegInfoByRegName(rn); |
| 794 |
if (!ri) // for cases where a platform lacks a register |
| 795 |
return; |
| 796 |
const size_t bytesPerReg = model.getBytesPerReg(rn); |
| 797 |
auto emitReg = [&](size_t i) { |
| 798 |
if (first) |
| 799 |
first = false; |
| 800 |
else |
| 801 |
os << ","; |
| 802 |
os << ri->syntax; |
| 803 |
if (ri->hasRegNum()) { |
| 804 |
os << (int)i; |
| 805 |
} |
| 806 |
}; |
| 807 |
|
| 808 |
for (size_t ri = 0; ri < model.getRegCount(rn); ri++) { |
| 809 |
size_t regOff = OffsetOf(model, rn, (int)ri) / 8; |
| 810 |
if (bs.testAll(regOff, bytesPerReg)) { |
| 811 |
// full reg: e.g. "r13" or "r13:8" |
| 812 |
emitReg(ri); |
| 813 |
size_t len = 1; |
| 814 |
if (rn == RegName::GRF_R || rn == RegName::ARF_ACC) { |
| 815 |
// for GRF and ACC permit payload-form syntax |
| 816 |
// r10:8, acc0:2 |
| 817 |
while (ri + len < model.getRegCount(rn) && |
| 818 |
bs.testAll(OffsetOf(model, rn, (int)ri + (int)len) / 8, |
| 819 |
bytesPerReg)) { |
| 820 |
len++; |
| 821 |
} |
| 822 |
} |
| 823 |
if (len > 1) { |
| 824 |
os << ":" << len; |
| 825 |
ri += len - 1; |
| 826 |
} |
| 827 |
} else if (bs.testAny(regOff, bytesPerReg)) { |
| 828 |
// partial register |
| 829 |
// r13[7] (e.g. r13.7<0;1,0>:b) |
| 830 |
// r13[0-3,8-11,..] (e.g. r13.0<2;1,0>:d) |
| 831 |
emitReg(ri); |
| 832 |
os << '['; |
| 833 |
bool firstSubreg = true; |
| 834 |
for (size_t byte = 0; byte < bytesPerReg; byte++) { |
| 835 |
if (!bs.test(regOff + byte)) { |
| 836 |
continue; |
| 837 |
} |
| 838 |
if (firstSubreg) { |
| 839 |
firstSubreg = false; |
| 840 |
} else { |
| 841 |
os << ','; |
| 842 |
} |
| 843 |
os << byte; |
| 844 |
// find the end |
| 845 |
size_t len = 1; |
| 846 |
while (byte + len < bytesPerReg && bs.test(regOff + byte + len)) |
| 847 |
len++; |
| 848 |
if (len > 1) { |
| 849 |
os << '-' << byte + len - 1; |
| 850 |
byte += len - 1; |
| 851 |
} |
| 852 |
} |
| 853 |
os << ']'; |
| 854 |
} |
| 855 |
} // for |
| 856 |
} |
| 857 |
|
| 858 |
std::string RegSet::str(const Model &m, RegName rn, const RegSet::Bits &bs) { |
| 859 |
bool first = true; |
| 860 |
std::stringstream ss; |
| 861 |
strForReg(m, rn, bs, first, ss); |
| 862 |
return ss.str(); |
| 863 |
} |
| 864 |
|
| 865 |
// emits a shorter description... something like |
| 866 |
// {r0,r1[0-3],r2[16],r4[0-3,8-11,16-19,24-27]} |
| 867 |
void RegSet::str(std::ostream &os) const { |
| 868 |
bool first = true; |
| 869 |
os << "{"; |
| 870 |
for (RegName rn : TRACKED) { |
| 871 |
const BitSet<> &bs = bitSetFor(rn); |
| 872 |
strForReg(model, rn, bs, first, os); |
| 873 |
} |
| 874 |
os << "}"; |
| 875 |
} |
| 876 |
std::string RegSet::str() const { |
| 877 |
std::stringstream ss; |
| 878 |
str(ss); |
| 879 |
return ss.str(); |
| 880 |
} |
| 881 |
|
| 882 |
RegSet::Bits RegSet::emptyPredSet(const Model &model) { |
| 883 |
return RegSet::Bits(BytesForRegSet(RegName::ARF_F, model)); |
| 884 |
} |