| 1 |
/*========================== begin_copyright_notice ============================ |
| 2 |
|
| 3 |
Copyright (C) 2020-2021 Intel Corporation |
| 4 |
|
| 5 |
SPDX-License-Identifier: MIT |
| 6 |
|
| 7 |
============================= end_copyright_notice ===========================*/ |
| 8 |
|
| 9 |
#include "MessageDecoder.hpp" |
| 10 |
|
| 11 |
#include <utility> |
| 12 |
|
| 13 |
using namespace iga; |
| 14 |
|
| 15 |
enum LscOp : uint32_t { |
| 16 |
LSC_LOAD = 0x00, |
| 17 |
LSC_LOAD_STRIDED = 0x01, |
| 18 |
LSC_LOAD_QUAD = 0x02, // aka load_cmask |
| 19 |
LSC_LOAD_BLOCK2D = 0x03, |
| 20 |
LSC_STORE = 0x04, |
| 21 |
LSC_STORE_STRIDED = 0x05, |
| 22 |
LSC_STORE_QUAD = 0x06, // aka store_cmask |
| 23 |
LSC_STORE_BLOCK2D = 0x07, |
| 24 |
// |
| 25 |
LSC_ATOMIC_IINC = 0x08, |
| 26 |
LSC_ATOMIC_IDEC = 0x09, |
| 27 |
LSC_ATOMIC_LOAD = 0x0A, |
| 28 |
LSC_ATOMIC_STORE = 0x0B, |
| 29 |
LSC_ATOMIC_IADD = 0x0C, |
| 30 |
LSC_ATOMIC_ISUB = 0x0D, |
| 31 |
LSC_ATOMIC_SMIN = 0x0E, |
| 32 |
LSC_ATOMIC_SMAX = 0x0F, |
| 33 |
LSC_ATOMIC_UMIN = 0x10, |
| 34 |
LSC_ATOMIC_UMAX = 0x11, |
| 35 |
LSC_ATOMIC_ICAS = 0x12, |
| 36 |
LSC_ATOMIC_FADD = 0x13, |
| 37 |
LSC_ATOMIC_FSUB = 0x14, |
| 38 |
LSC_ATOMIC_FMIN = 0x15, |
| 39 |
LSC_ATOMIC_FMAX = 0x16, |
| 40 |
LSC_ATOMIC_FCAS = 0x17, |
| 41 |
LSC_ATOMIC_AND = 0x18, |
| 42 |
LSC_ATOMIC_OR = 0x19, |
| 43 |
LSC_ATOMIC_XOR = 0x1A, |
| 44 |
// |
| 45 |
LSC_LOAD_STATUS = 0x1B, |
| 46 |
LSC_STORE_UNCOMPRESSED = 0x1C, |
| 47 |
LSC_CCS = 0x1D, |
| 48 |
// |
| 49 |
LSC_RSI = 0x1E, |
| 50 |
LSC_FENCE = 0x1F, |
| 51 |
// |
| 52 |
LSC_STORE_UNCOMPRESSED_QUAD = 0x20, |
| 53 |
// |
| 54 |
LSC_ATOMIC_APNDADD = 0x28, |
| 55 |
LSC_ATOMIC_APNDSUB = 0x29, |
| 56 |
LSC_ATOMIC_APNDSTORE = 0x2A, |
| 57 |
LSC_LOAD_QUAD_MSRT = 0x31, |
| 58 |
LSC_STORE_QUAD_MSRT = 0x32, |
| 59 |
// |
| 60 |
LSC_INVALID = 0xFFFFFFFF, |
| 61 |
}; |
| 62 |
|
| 63 |
static const uint32_t LSC_AT_FLAT = 0x0; |
| 64 |
static const uint32_t LSC_AT_BSS = 0x1; |
| 65 |
static const uint32_t LSC_AT_SS = 0x2; |
| 66 |
static const uint32_t LSC_AT_BTI = 0x3; |
| 67 |
|
| 68 |
static const uint32_t LSC_A16 = 0x1; |
| 69 |
static const uint32_t LSC_A32 = 0x2; |
| 70 |
static const uint32_t LSC_A64 = 0x3; |
| 71 |
|
| 72 |
static const uint32_t LSC_D8 = 0x0; |
| 73 |
static const uint32_t LSC_D16 = 0x1; |
| 74 |
static const uint32_t LSC_D32 = 0x2; |
| 75 |
static const uint32_t LSC_D64 = 0x3; |
| 76 |
static const uint32_t LSC_D8U32 = 0x4; |
| 77 |
static const uint32_t LSC_D16U32 = 0x5; |
| 78 |
static const uint32_t LSC_D16U32H = 0x6; |
| 79 |
|
| 80 |
static const uint32_t LSC_V1 = 0x0; |
| 81 |
static const uint32_t LSC_V2 = 0x1; |
| 82 |
static const uint32_t LSC_V3 = 0x2; |
| 83 |
static const uint32_t LSC_V4 = 0x3; |
| 84 |
static const uint32_t LSC_V8 = 0x4; |
| 85 |
static const uint32_t LSC_V16 = 0x5; |
| 86 |
static const uint32_t LSC_V32 = 0x6; |
| 87 |
static const uint32_t LSC_V64 = 0x7; |
| 88 |
|
| 89 |
/////////////////////////////////////////////////////// |
| 90 |
// Cache Opt |
| 91 |
// Value for bits[19:17] |
| 92 |
static const uint32_t LSC_DF_DF = 0x0; |
| 93 |
// |
| 94 |
static const uint32_t LSC_UC_UC = 0x1; |
| 95 |
// |
| 96 |
static const uint32_t LSC_UC_CA = 0x2; |
| 97 |
static const uint32_t LSC_UC_WB = 0x2; |
| 98 |
// |
| 99 |
static const uint32_t LSC_CA_UC = 0x3; |
| 100 |
static const uint32_t LSC_WT_UC = 0x3; |
| 101 |
// |
| 102 |
static const uint32_t LSC_CA_CA = 0x4; |
| 103 |
static const uint32_t LSC_WT_WB = 0x4; |
| 104 |
// |
| 105 |
static const uint32_t LSC_ST_UC = 0x5; |
| 106 |
// |
| 107 |
static const uint32_t LSC_ST_CA = 0x6; |
| 108 |
static const uint32_t LSC_ST_WB = 0x6; |
| 109 |
// |
| 110 |
static const uint32_t LSC_RI_CA = 0x7; |
| 111 |
static const uint32_t LSC_WB_WB = 0x7; |
| 112 |
|
| 113 |
// Value for bits [19:16] |
| 114 |
static const uint32_t LSC_UC_CC = 0x5; |
| 115 |
static const uint32_t LSC_CA_CC = 0x9; |
| 116 |
static const uint32_t LSC_RI_RI = 0xE; |
| 117 |
|
| 118 |
// This handles LSC messages only |
| 119 |
struct MessageDecoderLSC : MessageDecoder { |
| 120 |
MessageDecoderLSC(Platform _platform, SFID _sfid, ExecSize _execSize, |
| 121 |
uint32_t exImmOffDesc, SendDesc _exDesc, SendDesc _desc, |
| 122 |
DecodeResult &_result) |
| 123 |
: MessageDecoder(_platform, _sfid, _execSize, exImmOffDesc, _exDesc, |
| 124 |
_desc, _result) {} |
| 125 |
|
| 126 |
// used by decodeLscMessage and subchildren |
| 127 |
std::string dataTypePrefixSyntax; // e.g. d32 or d16 or d32 |
| 128 |
std::string vectorSuffixSyntax; // e.g. x16t (for d16x16t) or .yzw |
| 129 |
std::string addrSizeSyntax; // e.g. a32 |
| 130 |
std::string cacheControlSyntax; // e.g. ca.ca |
| 131 |
|
| 132 |
SendOp op = SendOp::INVALID; |
| 133 |
// |
| 134 |
int expectedExecSize = 1; |
| 135 |
// |
| 136 |
int addrSizeBits = 0; |
| 137 |
int dataSizeRegBits = 0, dataSizeMemBits = 0; |
| 138 |
int vectorSize = 1; |
| 139 |
MessageInfo::Attr extraAttrs = MessageInfo::Attr::NONE; |
| 140 |
|
| 141 |
// the symbol to return in the MessageInfo structure |
| 142 |
std::string symbolFromSyntax() const { |
| 143 |
std::stringstream sym; |
| 144 |
sym << result.syntax.mnemonic; |
| 145 |
if (!result.syntax.controls.empty()) |
| 146 |
sym << result.syntax.controls; |
| 147 |
sym << " "; |
| 148 |
if (!result.syntax.surface.empty()) |
| 149 |
sym << result.syntax.surface; |
| 150 |
sym << "["; |
| 151 |
if (!result.syntax.scale.empty()) { |
| 152 |
sym << result.syntax.scale; |
| 153 |
} |
| 154 |
sym << "A"; |
| 155 |
if (!result.syntax.immOffset.empty()) { |
| 156 |
sym << result.syntax.immOffset; |
| 157 |
} |
| 158 |
sym << "]"; |
| 159 |
return sym.str(); |
| 160 |
} |
| 161 |
|
| 162 |
/////////////////////////////////////////////////////////////////////////// |
| 163 |
void setCacheOpts(std::stringstream &sym, std::stringstream &descs, |
| 164 |
CacheOpt &l1, CacheOpt &l3, CacheOpt _l1, CacheOpt _l3) { |
| 165 |
l1 = _l1; |
| 166 |
l3 = _l3; |
| 167 |
if (_l1 == CacheOpt::DEFAULT && _l3 == CacheOpt::DEFAULT) { |
| 168 |
descs << "use state settings for both L1 and L3"; |
| 169 |
return; |
| 170 |
} |
| 171 |
auto emitCacheOpt = [&](CacheOpt c) { |
| 172 |
sym << '.'; |
| 173 |
switch (c) { |
| 174 |
case CacheOpt::DEFAULT: |
| 175 |
sym << "df"; |
| 176 |
descs << " uses default state settings"; |
| 177 |
break; |
| 178 |
case CacheOpt::READINVALIDATE: |
| 179 |
sym << "ri"; |
| 180 |
descs << " read-invalidate (last use)"; |
| 181 |
break; |
| 182 |
case CacheOpt::CACHED: |
| 183 |
sym << "ca"; |
| 184 |
descs << " cached"; |
| 185 |
break; |
| 186 |
case CacheOpt::STREAMING: |
| 187 |
sym << "st"; |
| 188 |
descs << " streaming"; |
| 189 |
break; |
| 190 |
case CacheOpt::UNCACHED: |
| 191 |
sym << "uc"; |
| 192 |
descs << " uncached (bypass)"; |
| 193 |
break; |
| 194 |
case CacheOpt::WRITETHROUGH: |
| 195 |
sym << "wt"; |
| 196 |
descs << " writethrough"; |
| 197 |
break; |
| 198 |
case CacheOpt::WRITEBACK: |
| 199 |
sym << "wb"; |
| 200 |
descs << " writeback"; |
| 201 |
break; |
| 202 |
case CacheOpt::CONSTCACHED: |
| 203 |
sym << "cc"; |
| 204 |
descs << " constant-cached"; |
| 205 |
break; |
| 206 |
default: |
| 207 |
sym << "?"; |
| 208 |
descs << " invalid"; |
| 209 |
break; |
| 210 |
} |
| 211 |
}; |
| 212 |
descs << "L1"; |
| 213 |
emitCacheOpt(_l1); |
| 214 |
descs << "; L3"; |
| 215 |
emitCacheOpt(_l3); |
| 216 |
descs << ""; |
| 217 |
} |
| 218 |
|
| 219 |
// Descriptor Bits[19:16]: 4 bits of cache control |
| 220 |
bool decodeLscCacheControlBits16_19(SendOp sop, CacheOpt &l1, CacheOpt &l3) { |
| 221 |
// check if it's the 3 cases added for 4-bits Cache Control at [19:16]: |
| 222 |
// L1UC_L3CC, L1CA_L3CC, L1RI_L3RI |
| 223 |
std::stringstream sym, descs; |
| 224 |
l1 = l3 = CacheOpt::DEFAULT; |
| 225 |
bool isLoad = lookupSendOp(sop).isLoad(); |
| 226 |
auto ccBits = getDescBits(16, 4); |
| 227 |
auto setCacheOptsWrapper = [&](CacheOpt _l1, CacheOpt _l3) { |
| 228 |
return setCacheOpts(sym, descs, l1, l3, _l1, _l3); |
| 229 |
}; |
| 230 |
|
| 231 |
switch (ccBits) { |
| 232 |
case LSC_UC_CC: |
| 233 |
if (isLoad) { |
| 234 |
setCacheOptsWrapper(CacheOpt::UNCACHED, CacheOpt::CONSTCACHED); |
| 235 |
break; |
| 236 |
} |
| 237 |
case LSC_CA_CC: |
| 238 |
if (isLoad) { |
| 239 |
setCacheOptsWrapper(CacheOpt::CACHED, CacheOpt::CONSTCACHED); |
| 240 |
break; |
| 241 |
} |
| 242 |
case LSC_RI_RI: |
| 243 |
if (isLoad) { |
| 244 |
setCacheOptsWrapper(CacheOpt::READINVALIDATE, CacheOpt::READINVALIDATE); |
| 245 |
break; |
| 246 |
} |
| 247 |
default: |
| 248 |
return decodeLscCacheControlBits17_19(sop, l1, l3); |
| 249 |
} |
| 250 |
// |
| 251 |
cacheControlSyntax = sym.str(); |
| 252 |
// |
| 253 |
addField("Caching", 16, 4, ccBits, descs.str()); |
| 254 |
return true; |
| 255 |
} |
| 256 |
|
| 257 |
void decodeCacheControl(SendOp sop, CacheOpt &l1, CacheOpt &l3) { |
| 258 |
if (platform() >= Platform::XE2) { |
| 259 |
if (!decodeLscCacheControlBits16_19(sop, l1, l3)) |
| 260 |
error(16, 4, "invalid cache options"); |
| 261 |
return; |
| 262 |
} |
| 263 |
|
| 264 |
if (!decodeLscCacheControlBits17_19(sop, l1, l3)) |
| 265 |
error(17, 3, "invalid cache options"); |
| 266 |
} |
| 267 |
|
| 268 |
// Descriptor Bits[19:17]: 3 bits of cache control |
| 269 |
bool decodeLscCacheControlBits17_19(SendOp sop, CacheOpt &l1, CacheOpt &l3) { |
| 270 |
std::stringstream sym, descs; |
| 271 |
l1 = l3 = CacheOpt::DEFAULT; |
| 272 |
bool isLoad = lookupSendOp(sop).isLoad(); |
| 273 |
auto ccBits = getDescBits(17, 3); |
| 274 |
auto setCacheOptsWrapper = [&](CacheOpt _l1, CacheOpt _l3) { |
| 275 |
return setCacheOpts(sym, descs, l1, l3, _l1, _l3); |
| 276 |
}; |
| 277 |
switch (ccBits) { |
| 278 |
case LSC_DF_DF: |
| 279 |
setCacheOptsWrapper(CacheOpt::DEFAULT, CacheOpt::DEFAULT); |
| 280 |
break; |
| 281 |
case LSC_UC_UC: |
| 282 |
setCacheOptsWrapper(CacheOpt::UNCACHED, CacheOpt::UNCACHED); |
| 283 |
break; |
| 284 |
case LSC_UC_CA: // == LSC_UC_WB |
| 285 |
if (isLoad) |
| 286 |
setCacheOptsWrapper(CacheOpt::UNCACHED, CacheOpt::CACHED); |
| 287 |
else |
| 288 |
setCacheOptsWrapper(CacheOpt::UNCACHED, CacheOpt::WRITEBACK); |
| 289 |
break; |
| 290 |
case LSC_CA_UC: // == LSC_WT_UC |
| 291 |
if (isLoad) |
| 292 |
setCacheOptsWrapper(CacheOpt::CACHED, CacheOpt::UNCACHED); |
| 293 |
else |
| 294 |
setCacheOptsWrapper(CacheOpt::WRITETHROUGH, CacheOpt::UNCACHED); |
| 295 |
break; |
| 296 |
case LSC_CA_CA: // == LSC_WT_WB |
| 297 |
if (isLoad) |
| 298 |
setCacheOptsWrapper(CacheOpt::CACHED, CacheOpt::CACHED); |
| 299 |
else |
| 300 |
setCacheOptsWrapper(CacheOpt::WRITETHROUGH, CacheOpt::WRITEBACK); |
| 301 |
break; |
| 302 |
case LSC_ST_UC: |
| 303 |
setCacheOptsWrapper(CacheOpt::STREAMING, CacheOpt::UNCACHED); |
| 304 |
break; |
| 305 |
case LSC_ST_CA: // == LSC_ST_WB |
| 306 |
if (isLoad) |
| 307 |
setCacheOptsWrapper(CacheOpt::STREAMING, CacheOpt::CACHED); |
| 308 |
else |
| 309 |
setCacheOptsWrapper(CacheOpt::STREAMING, CacheOpt::WRITEBACK); |
| 310 |
break; |
| 311 |
case LSC_RI_CA: |
| 312 |
if (isLoad) { |
| 313 |
// atomic follows store semantics, so compare against load |
| 314 |
setCacheOptsWrapper(CacheOpt::READINVALIDATE, CacheOpt::CACHED); |
| 315 |
} else { |
| 316 |
setCacheOptsWrapper(CacheOpt::WRITEBACK, CacheOpt::WRITEBACK); |
| 317 |
} |
| 318 |
break; |
| 319 |
default: |
| 320 |
return false; |
| 321 |
} |
| 322 |
// |
| 323 |
cacheControlSyntax = sym.str(); |
| 324 |
// |
| 325 |
addField("Caching", 17, 3, ccBits, descs.str()); |
| 326 |
return true; |
| 327 |
} |
| 328 |
|
| 329 |
void decodeLscImmOff(uint32_t atBits) { |
| 330 |
bool exDescHasImmOff = platform() >= Platform::XE2; |
| 331 |
if (!exDescHasImmOff) |
| 332 |
return; |
| 333 |
bool isBlock2d = op == SendOp::LOAD_BLOCK2D || op == SendOp::STORE_BLOCK2D; |
| 334 |
// immediate offset bits go in SendDesc if ExDesc is imm |
| 335 |
// if ExDesc is Reg then they go in dedicated exImmOffDesc |
| 336 |
uint32_t baseOffBits = exDesc.isReg() ? exImmOffDesc : exDesc.imm; |
| 337 |
if (atBits == LSC_AT_FLAT) { |
| 338 |
// [31:12] are all S20 |
| 339 |
if (isBlock2d) { |
| 340 |
addDoc(DocRef::EXDESC, "DP_EXT_DESC_2D", "67780"); |
| 341 |
// ExDescImm[31:22] = XOff (S10) |
| 342 |
// ExDescImm[21:12] = YOff (S10) |
| 343 |
result.info.immediateOffsetBlock2dX = |
| 344 |
(int)getSignedBits((int)baseOffBits, 12, 10); |
| 345 |
result.info.immediateOffsetBlock2dY = |
| 346 |
(int)getSignedBits((int)baseOffBits, 22, 10); |
| 347 |
addField("ImmediateOffsetBlock2dX", 22 + 32, 10, |
| 348 |
(uint32_t)result.info.immediateOffsetBlock2dX, |
| 349 |
fmtHexSigned(result.info.immediateOffsetBlock2dX)); |
| 350 |
addField("ImmediateOffsetBlock2dY", 12 + 32, 10, |
| 351 |
(uint32_t)result.info.immediateOffsetBlock2dY, |
| 352 |
fmtHexSigned(result.info.immediateOffsetBlock2dY)); |
| 353 |
if (result.info.immediateOffsetBlock2dX != 0 || |
| 354 |
result.info.immediateOffsetBlock2dY != 0) { |
| 355 |
result.syntax.immOffset = |
| 356 |
std::string("+(") + |
| 357 |
fmtHexSigned(result.info.immediateOffsetBlock2dX) + "," + |
| 358 |
fmtHexSigned(result.info.immediateOffsetBlock2dY) + ")"; |
| 359 |
} |
| 360 |
} else { |
| 361 |
addDoc(DocRef::EXDESC, "EXDESC_FLAT", "64012"); |
| 362 |
result.info.immediateOffset = getSignedBits((int)baseOffBits, 12, 20); |
| 363 |
if (result.info.immediateOffset > 0) { |
| 364 |
result.syntax.immOffset = |
| 365 |
"+" + fmtHexSigned(result.info.immediateOffset); |
| 366 |
} else if (result.info.immediateOffset < 0) { |
| 367 |
result.syntax.immOffset = fmtHexSigned(result.info.immediateOffset); |
| 368 |
} |
| 369 |
addField("ImmediateOffset", 32 + 12, 20, |
| 370 |
(uint32_t)result.info.immediateOffset, |
| 371 |
result.syntax.immOffset); |
| 372 |
} |
| 373 |
|
| 374 |
} else if ((atBits == LSC_AT_BSS || atBits == LSC_AT_SS) && |
| 375 |
exDesc.isReg() && sfid == SFID::TGM) { |
| 376 |
addDoc(DocRef::EXDESC, "EXDESC_SURFACE", "64017"); |
| 377 |
} else if ((atBits == LSC_AT_BSS || atBits == LSC_AT_SS) && |
| 378 |
exDesc.isReg() && sfid == SFID::UGM) { |
| 379 |
addDoc(DocRef::EXDESC, "EXDESCIMM_BSSO_SSO", "70586"); |
| 380 |
// SS/BSS 16b or 17b (UGM) |
| 381 |
// ExDescImm[31:19] are Offset[16:4] |
| 382 |
// ExDescImm[18:16] are unmapped (addr reg) |
| 383 |
// ExDescImm[15:12] are Offset[3:0] |
| 384 |
if (sfid != SFID::UGM && baseOffBits != 0) { |
| 385 |
error(32, 32, "bss/ss: immediate offset forbidden for non-ugm"); |
| 386 |
} |
| 387 |
uint32_t packed = ((getBits(baseOffBits, 19, 31 - 19 + 1) << 4) | |
| 388 |
getBits(baseOffBits, 12, 3 - 0 + 1)); |
| 389 |
result.info.immediateOffset = getSignedBits(packed, 0, 17); |
| 390 |
if (result.info.immediateOffset > 0) { |
| 391 |
result.syntax.immOffset = |
| 392 |
"+" + fmtHexSigned(result.info.immediateOffset); |
| 393 |
} else if (result.info.immediateOffset < 0) { |
| 394 |
result.syntax.immOffset = fmtHexSigned(result.info.immediateOffset); |
| 395 |
} |
| 396 |
addField("ImmediateOffset[16:4]", 32 + 19, 16 - 4 + 1, |
| 397 |
getBits(packed, 4, 16 - 4 + 1), result.syntax.immOffset); |
| 398 |
addField("ImmediateOffset[3:0]", 32 + 12, 3 - 0 + 1, |
| 399 |
getBits(packed, 0, 3 - 0 + 1), result.syntax.immOffset); |
| 400 |
addField("Reserved", 32 + 16, 3, getBits(baseOffBits, 16, 3), |
| 401 |
"a0.subreg"); |
| 402 |
if (getBits(baseOffBits, 16, 3)) { |
| 403 |
error(32 + 16, 3, "ExDesc[19:16] must be zeros"); |
| 404 |
} |
| 405 |
} else if (atBits == LSC_AT_BSS || atBits == LSC_AT_SS) { // bss/ss imm |
| 406 |
addDoc(DocRef::EXDESC, "EXDESC_SURFACE", "64017"); |
| 407 |
} else if (atBits == LSC_AT_BTI && exDesc.isReg()) { |
| 408 |
if (exImmOffDesc != 0) |
| 409 |
error(32 + 12, 12, "immediate offset forbidden for BTI reg"); |
| 410 |
} else if (atBits == LSC_AT_BTI) { |
| 411 |
addDoc(DocRef::EXDESC, "EXDESC_BTI", "63997"); |
| 412 |
// ExDesc[23:12] is immm off |
| 413 |
result.info.immediateOffset = |
| 414 |
getSignedBits((int32_t)baseOffBits, 12, 23 - 12 + 1); |
| 415 |
if (result.info.immediateOffset > 0) { |
| 416 |
result.syntax.immOffset = |
| 417 |
"+" + fmtHexSigned(result.info.immediateOffset); |
| 418 |
} else if (result.info.immediateOffset < 0) { |
| 419 |
result.syntax.immOffset = fmtHexSigned(result.info.immediateOffset); |
| 420 |
} |
| 421 |
addField("ImmediateOffset[11:0]", 32 + 12, 23 - 12 + 1, |
| 422 |
result.info.immediateOffset, result.syntax.immOffset); |
| 423 |
} // else: BSS/SS with imm ex desc |
| 424 |
} |
| 425 |
|
| 426 |
static const int ADDRTYPE_LOC = 29; |
| 427 |
|
| 428 |
AddrType decodeLscAddrType(SendDesc &surfId, bool allowsFlat = true) { |
| 429 |
surfId = 0; |
| 430 |
AddrType addrType = AddrType::FLAT; |
| 431 |
// |
| 432 |
const char *addrTypeMeaning = "?"; |
| 433 |
// |
| 434 |
const auto atBits = getDescBits(ADDRTYPE_LOC, 2); |
| 435 |
// |
| 436 |
std::stringstream surfSyntax; |
| 437 |
switch (atBits) { |
| 438 |
case LSC_AT_FLAT: |
| 439 |
addrTypeMeaning = "Flat"; |
| 440 |
addrType = AddrType::FLAT; |
| 441 |
if (!allowsFlat) |
| 442 |
error(ADDRTYPE_LOC, 2, "this message may not use FLAT address type"); |
| 443 |
break; |
| 444 |
case LSC_AT_BSS: |
| 445 |
case LSC_AT_SS: |
| 446 |
if (atBits == LSC_AT_BSS) { |
| 447 |
addrTypeMeaning = "BSS"; |
| 448 |
addrType = AddrType::BSS; |
| 449 |
surfSyntax << "bss"; |
| 450 |
} else { |
| 451 |
addrTypeMeaning = "SS"; |
| 452 |
addrType = AddrType::SS; |
| 453 |
surfSyntax << "ss"; |
| 454 |
} |
| 455 |
if (exDesc.isImm()) { |
| 456 |
// XeHPG/XeHPC: we can pull this value out of ExDesc[31:11] |
| 457 |
int exDescOff = 11, len = 31 - exDescOff + 1; |
| 458 |
surfId = getDescBits(32 + exDescOff, len) << exDescOff; |
| 459 |
addField("SurfaceStateOffset", exDescOff, len, surfId.imm, |
| 460 |
"immediate surface state offset"); |
| 461 |
surfSyntax << "[" << iga::fmtHex(surfId.imm) << "]"; |
| 462 |
} else { |
| 463 |
// XeHPG/XeHPC with reg surface state offset |
| 464 |
surfSyntax << "[a0." << (int)exDesc.reg.subRegNum << "]"; |
| 465 |
surfId = exDesc; |
| 466 |
} |
| 467 |
break; |
| 468 |
case LSC_AT_BTI: |
| 469 |
addrTypeMeaning = "BTI"; |
| 470 |
addrType = AddrType::BTI; |
| 471 |
if (exDesc.isImm()) { |
| 472 |
uint32_t bti = decodeExDescField( |
| 473 |
"BTI", 24, 8, [&](std::stringstream &ss, uint32_t bti) { |
| 474 |
ss << "bti[" << bti << "]"; |
| 475 |
}); |
| 476 |
surfSyntax << "bti[" << bti << "]"; |
| 477 |
surfId = bti; |
| 478 |
} else { |
| 479 |
surfSyntax << "bti[a0." << (int)exDesc.reg.subRegNum << "]"; |
| 480 |
surfId = exDesc; |
| 481 |
} |
| 482 |
break; |
| 483 |
default: |
| 484 |
addrTypeMeaning = "INVALID AddrType"; |
| 485 |
addrType = AddrType::FLAT; |
| 486 |
surfSyntax << "?"; |
| 487 |
error(ADDRTYPE_LOC, 2, "invalid address type"); |
| 488 |
break; |
| 489 |
} |
| 490 |
result.syntax.surface = surfSyntax.str(); |
| 491 |
|
| 492 |
///////////////////////////// |
| 493 |
// immediate offset |
| 494 |
decodeLscImmOff(atBits); |
| 495 |
// |
| 496 |
addField("AddrType", ADDRTYPE_LOC, 2, atBits, addrTypeMeaning); |
| 497 |
// |
| 498 |
return addrType; |
| 499 |
} |
| 500 |
|
| 501 |
void decodeLscAddrSize() { |
| 502 |
int addrSzBits = getDescBits(7, 2); // [8:7] |
| 503 |
std::stringstream asym; |
| 504 |
const char *aDesc = ""; |
| 505 |
switch (addrSzBits) { |
| 506 |
case 1: |
| 507 |
asym << "a16"; |
| 508 |
aDesc = "addresses are 16b"; |
| 509 |
addrSizeBits = 16; |
| 510 |
break; |
| 511 |
case 2: |
| 512 |
asym << "a32"; |
| 513 |
aDesc = "addresses are 32b"; |
| 514 |
addrSizeBits = 32; |
| 515 |
break; |
| 516 |
case 3: |
| 517 |
asym << "a64"; |
| 518 |
aDesc = "addresses are 64b"; |
| 519 |
addrSizeBits = 64; |
| 520 |
break; |
| 521 |
default: |
| 522 |
asym << "a???"; |
| 523 |
aDesc = "address size is invalid"; |
| 524 |
error(7, 2, "invalid address size"); |
| 525 |
break; |
| 526 |
} |
| 527 |
// result.syntax.addressType = ":" + asym.str(); |
| 528 |
|
| 529 |
addrSizeSyntax = asym.str(); |
| 530 |
// |
| 531 |
addField("AddrSize", 7, 2, addrSzBits, aDesc); |
| 532 |
} |
| 533 |
|
| 534 |
void decodeLscDataSize() { |
| 535 |
std::stringstream dsym; |
| 536 |
dataSizeRegBits = dataSizeMemBits = 0; |
| 537 |
std::string meaning; |
| 538 |
auto dtBits = getDescBits(9, 3); |
| 539 |
switch (dtBits) { // dat size [11:9] |
| 540 |
case LSC_D8: |
| 541 |
dataSizeRegBits = dataSizeMemBits = 8; |
| 542 |
dsym << "d8"; |
| 543 |
meaning = "8b per data element"; |
| 544 |
break; |
| 545 |
case LSC_D16: |
| 546 |
dataSizeRegBits = dataSizeMemBits = 16; |
| 547 |
meaning = "16b per data element"; |
| 548 |
dsym << "d16"; |
| 549 |
break; |
| 550 |
case LSC_D32: |
| 551 |
dataSizeRegBits = dataSizeMemBits = 32; |
| 552 |
dsym << "d32"; |
| 553 |
meaning = "32b per data element"; |
| 554 |
break; |
| 555 |
case LSC_D64: |
| 556 |
dataSizeRegBits = dataSizeMemBits = 64; |
| 557 |
dsym << "d64"; |
| 558 |
meaning = "64b per data element"; |
| 559 |
break; |
| 560 |
case LSC_D8U32: |
| 561 |
dataSizeRegBits = 32; |
| 562 |
dataSizeMemBits = 8; |
| 563 |
dsym << "d8u32"; |
| 564 |
meaning = "load 8b into the low 8b of 32b register elements " |
| 565 |
"(upper bits are undefined)"; |
| 566 |
if (platform() >= Platform::XE2) |
| 567 |
meaning = "load 8b and zero-extend to 32b per data element"; |
| 568 |
break; |
| 569 |
case LSC_D16U32: |
| 570 |
dataSizeRegBits = 32; |
| 571 |
dataSizeMemBits = 16; |
| 572 |
dsym << "d16u32"; |
| 573 |
meaning = "load 16b into the low 16b of 32b register elements " |
| 574 |
"(upper bits are undefined)"; |
| 575 |
if (platform() >= Platform::XE2) |
| 576 |
meaning = "load 16b and zero-extend to 32b per data element"; |
| 577 |
break; |
| 578 |
case LSC_D16U32H: |
| 579 |
dataSizeRegBits = 32; |
| 580 |
dataSizeMemBits = 16; |
| 581 |
extraAttrs |= MessageInfo::Attr::EXPAND_HIGH; |
| 582 |
dsym << "d16u32h"; |
| 583 |
meaning = "load 16b into the high half of 32b register elements"; |
| 584 |
break; |
| 585 |
default: |
| 586 |
dsym << "0x" << std::uppercase << std::hex << dtBits; |
| 587 |
meaning = "???"; |
| 588 |
} |
| 589 |
// |
| 590 |
// result.syntax.dataType = ":" + dsym.str(); |
| 591 |
dataTypePrefixSyntax = dsym.str(); |
| 592 |
|
| 593 |
addField("DataSize", 9, 3, dtBits, meaning); |
| 594 |
} |
| 595 |
|
| 596 |
void decodeLscVecSize() { |
| 597 |
if (lookupSendOp(op).hasChMask()) { |
| 598 |
decodeLscVecSizeQuad(); |
| 599 |
} else { |
| 600 |
decodeLscVecSizeNormal(); |
| 601 |
} |
| 602 |
} |
| 603 |
|
| 604 |
void decodeLscVecSizeNormal() { |
| 605 |
std::stringstream vsym; |
| 606 |
|
| 607 |
uint32_t vecSzEncd = getDescBits(12, 3); // [14:12] |
| 608 |
switch (vecSzEncd) { |
| 609 |
case LSC_V1: |
| 610 |
vectorSize = 1; |
| 611 |
break; |
| 612 |
case LSC_V2: |
| 613 |
vectorSize = 2; |
| 614 |
break; |
| 615 |
case LSC_V3: |
| 616 |
vectorSize = 3; |
| 617 |
break; |
| 618 |
case LSC_V4: |
| 619 |
vectorSize = 4; |
| 620 |
break; |
| 621 |
case LSC_V8: |
| 622 |
vectorSize = 8; |
| 623 |
break; |
| 624 |
case LSC_V16: |
| 625 |
vectorSize = 16; |
| 626 |
break; |
| 627 |
case LSC_V32: |
| 628 |
vectorSize = 32; |
| 629 |
break; |
| 630 |
case LSC_V64: |
| 631 |
vectorSize = 64; |
| 632 |
break; |
| 633 |
default: |
| 634 |
vsym << "x?"; |
| 635 |
} |
| 636 |
bool opIsBlock2d = |
| 637 |
op == SendOp::LOAD_BLOCK2D || op == SendOp::STORE_BLOCK2D; |
| 638 |
auto transposed = decodeDescBitField( |
| 639 |
"DataOrder", 15, |
| 640 |
"non-transposed (vector elements are in successive registers)", |
| 641 |
"transposed (vector elements are in the same register)"); |
| 642 |
if (vectorSize > 1 || (transposed && !opIsBlock2d)) { |
| 643 |
vsym << 'x' << vectorSize; |
| 644 |
} |
| 645 |
if (transposed && op == SendOp::LOAD_STATUS) { |
| 646 |
error(15, 1, "data order must be non-transposed for this op"); |
| 647 |
} |
| 648 |
std::stringstream vdesc; |
| 649 |
vdesc << "each address accesses " << vectorSize << " element"; |
| 650 |
if (vectorSize != 1) |
| 651 |
vdesc << "s"; |
| 652 |
if (!opIsBlock2d) |
| 653 |
addField("VecSize", 12, 3, vecSzEncd, vdesc.str()); |
| 654 |
if (transposed) { |
| 655 |
vsym << 't'; |
| 656 |
extraAttrs |= MessageInfo::Attr::TRANSPOSED; |
| 657 |
expectedExecSize = 1; // all transpose messages are SIMD1 |
| 658 |
} |
| 659 |
|
| 660 |
if (op == SendOp::LOAD_BLOCK2D) { |
| 661 |
bool vnni = |
| 662 |
decodeDescBitField("Block2dVnniTransform", 7, "disabled", "enabled"); |
| 663 |
if (vnni) |
| 664 |
vsym << 'v'; |
| 665 |
} |
| 666 |
|
| 667 |
vectorSuffixSyntax = vsym.str(); |
| 668 |
} |
| 669 |
|
| 670 |
void decodeLscVecSizeQuad() { |
| 671 |
// LSC channels *enabled* is the inverse of the old messages |
| 672 |
// because the old ChMask used in untyped old (scatter4/gather4) |
| 673 |
// was really a channel "disable" mask |
| 674 |
auto chEn = getDescBits(12, 4); |
| 675 |
vectorSize = 0; |
| 676 |
for (int i = 0; i < 4; ++i) { |
| 677 |
if ((1 << i) & chEn) { |
| 678 |
vectorSize++; |
| 679 |
} |
| 680 |
} |
| 681 |
extraAttrs |= MessageInfo::Attr::HAS_CHMASK; |
| 682 |
|
| 683 |
std::stringstream vsym; |
| 684 |
vsym << "."; |
| 685 |
if (chEn & 1) |
| 686 |
vsym << "x"; |
| 687 |
if (chEn & 2) |
| 688 |
vsym << "y"; |
| 689 |
if (chEn & 4) |
| 690 |
vsym << "z"; |
| 691 |
if (chEn & 8) |
| 692 |
vsym << "w"; |
| 693 |
vectorSuffixSyntax = vsym.str(); |
| 694 |
|
| 695 |
addField("CompEn", 12, 4, chEn, vsym.str()); |
| 696 |
} |
| 697 |
|
| 698 |
/////////////////////////////////////////////////////////////////////////// |
| 699 |
void decodeLscMessage(const std::string& msgDesc, SendOp lscOp) { |
| 700 |
const std::string symbol = ToSyntax(lscOp); |
| 701 |
op = lscOp; |
| 702 |
|
| 703 |
const SendOpDefinition &opInfo = lookupSendOp(lscOp); |
| 704 |
bool opSupportsUvr = lscOp == SendOp::LOAD_QUAD || |
| 705 |
lscOp == SendOp::STORE_QUAD || |
| 706 |
lscOp == SendOp::LOAD_BLOCK2D || |
| 707 |
lscOp == SendOp::STORE_BLOCK2D || |
| 708 |
lscOp == SendOp::LOAD_QUAD_MSRT || |
| 709 |
lscOp == SendOp::STORE_QUAD_MSRT || |
| 710 |
lscOp == SendOp::STORE_UNCOMPRESSED_QUAD || |
| 711 |
opInfo.isAtomic(); |
| 712 |
if (sfid == SFID::TGM && opSupportsUvr) { |
| 713 |
extraAttrs |= MessageInfo::Attr::HAS_UVRLOD; |
| 714 |
} |
| 715 |
|
| 716 |
addField("Opcode", 0, 6, getDescBits(0, 6), symbol); |
| 717 |
|
| 718 |
if (exDesc.isImm() && (exDesc.imm & 0x7FF)) { |
| 719 |
// bit 11 may or may not be available |
| 720 |
error(0, 12, "ExDesc[11:0] must be 0 on this platform"); |
| 721 |
} |
| 722 |
// |
| 723 |
SendDesc surfaceId(0x0); |
| 724 |
AddrType addrType = decodeLscAddrType(surfaceId); |
| 725 |
// |
| 726 |
if (op == SendOp::LOAD_BLOCK2D || op == SendOp::STORE_BLOCK2D) { |
| 727 |
addrSizeBits = 64; |
| 728 |
addrSizeSyntax = "a64"; |
| 729 |
} else if (op == SendOp::ATOMIC_ACADD || op == SendOp::ATOMIC_ACSUB || |
| 730 |
op == SendOp::ATOMIC_ACSTORE) { |
| 731 |
addrSizeBits = 32; |
| 732 |
addrSizeSyntax = "a32"; |
| 733 |
} else { |
| 734 |
decodeLscAddrSize(); |
| 735 |
} |
| 736 |
// |
| 737 |
decodeLscDataSize(); |
| 738 |
// |
| 739 |
expectedExecSize = op == SendOp::LOAD_BLOCK2D || op == SendOp::STORE_BLOCK2D |
| 740 |
? 1 |
| 741 |
: sfid == SFID::TGM ? DEFAULT_EXEC_SIZE / 2 |
| 742 |
: DEFAULT_EXEC_SIZE; |
| 743 |
decodeLscVecSize(); |
| 744 |
// |
| 745 |
if (sfid == SFID::TGM) |
| 746 |
extraAttrs |= MessageInfo::Attr::TYPED; |
| 747 |
if (sfid == SFID::SLM) |
| 748 |
extraAttrs |= MessageInfo::Attr::SLM; |
| 749 |
// |
| 750 |
CacheOpt l1 = CacheOpt::DEFAULT, l3 = CacheOpt::DEFAULT; |
| 751 |
bool hasCc = opInfo.isLoad() || opInfo.isStore() || opInfo.isAtomic(); |
| 752 |
if (sfid != SFID::SLM && hasCc) { |
| 753 |
decodeCacheControl(op, l1, l3); |
| 754 |
} |
| 755 |
// |
| 756 |
result.syntax.mnemonic = symbol; |
| 757 |
// |
| 758 |
result.syntax.controls += '.'; |
| 759 |
result.syntax.controls += dataTypePrefixSyntax; |
| 760 |
result.syntax.controls += vectorSuffixSyntax; |
| 761 |
if (!addrSizeSyntax.empty()) { |
| 762 |
result.syntax.controls += '.'; |
| 763 |
result.syntax.controls += addrSizeSyntax; |
| 764 |
} |
| 765 |
if (!cacheControlSyntax.empty()) { |
| 766 |
result.syntax.controls += cacheControlSyntax; |
| 767 |
} |
| 768 |
// |
| 769 |
setScatterGatherOpX(symbolFromSyntax(), msgDesc, op, addrType, surfaceId, |
| 770 |
l1, l3, addrSizeBits, dataSizeRegBits, dataSizeMemBits, |
| 771 |
vectorSize, int(instExecSize), extraAttrs); |
| 772 |
if (lookupSendOp(op).hasChMask()) { |
| 773 |
result.info.channelsEnabled = getDescBits(12, 4); |
| 774 |
if (result.info.channelsEnabled == 0) |
| 775 |
error(12, 4, "no channels enabled on quad message"); |
| 776 |
} |
| 777 |
} |
| 778 |
|
| 779 |
void decodeLscMessageTypeBlock2D(std::string msgDesc, SendOp lscOp) { |
| 780 |
const std::string symbol = ToSyntax(lscOp); |
| 781 |
op = lscOp; |
| 782 |
extraAttrs |= MessageInfo::Attr::HAS_UVRLOD; |
| 783 |
expectedExecSize = 1; |
| 784 |
addField("Opcode", 0, 6, getDescBits(0, 6), symbol); |
| 785 |
|
| 786 |
SendDesc surfaceId(0x0); |
| 787 |
AddrType addrType = decodeLscAddrType(surfaceId); |
| 788 |
if (addrType != AddrType::BSS && addrType != AddrType::BTI) { |
| 789 |
error(ADDRTYPE_LOC, 2, "addr surface type must be BSS or BTI"); |
| 790 |
} |
| 791 |
|
| 792 |
addrSizeBits = 32; |
| 793 |
addrSizeSyntax = format("a", addrSizeBits); |
| 794 |
|
| 795 |
dataSizeRegBits = dataSizeMemBits = 32; // ? |
| 796 |
dataTypePrefixSyntax = format("d", dataSizeMemBits); |
| 797 |
|
| 798 |
CacheOpt l1 = CacheOpt::DEFAULT, l3 = CacheOpt::DEFAULT; |
| 799 |
decodeCacheControl(op, l1, l3); |
| 800 |
|
| 801 |
result.syntax.mnemonic = symbol; |
| 802 |
result.syntax.controls += '.'; |
| 803 |
result.syntax.controls += dataTypePrefixSyntax; |
| 804 |
result.syntax.controls += '.'; |
| 805 |
result.syntax.controls += addrSizeSyntax; |
| 806 |
if (!cacheControlSyntax.empty()) { |
| 807 |
result.syntax.controls += cacheControlSyntax; |
| 808 |
} |
| 809 |
setScatterGatherOpX(symbolFromSyntax(), msgDesc, op, addrType, surfaceId, |
| 810 |
l1, l3, addrSizeBits, dataSizeRegBits, dataSizeMemBits, |
| 811 |
vectorSize, expectedExecSize, extraAttrs); |
| 812 |
} |
| 813 |
|
| 814 |
void decodeLscAtomicMessage(std::string msgDesc, SendOp atOp) { |
| 815 |
extraAttrs |= getDescBits(20, 5) != 0 ? MessageInfo::Attr::ATOMIC_RETURNS |
| 816 |
: MessageInfo::Attr::NONE; |
| 817 |
if (sfid == SFID::TGM) |
| 818 |
extraAttrs |= MessageInfo::Attr::HAS_UVRLOD; |
| 819 |
decodeLscMessage(msgDesc, atOp); |
| 820 |
addDocRefsVecCmask(); |
| 821 |
} |
| 822 |
|
| 823 |
// similar to decodeLscAtomicMessage, but we disable syntax since |
| 824 |
// we don't wish to support it yet |
| 825 |
void decodeLscAtomicAppendMessage(const char *msgDesc, SendOp atOp) { |
| 826 |
decodeLscAtomicMessage(msgDesc, atOp); |
| 827 |
result.syntax.layout = MessageSyntax::Layout::INVALID; |
| 828 |
} |
| 829 |
|
| 830 |
// load/store/atomic etc... (non-cmask) |
| 831 |
void decodeLscVectorCmask(const char *name, SendOp op) { |
| 832 |
decodeLscMessage(name, op); |
| 833 |
addDocRefsVecCmask(); |
| 834 |
} |
| 835 |
void addDocRefsVecCmask() { |
| 836 |
if (!result.errors.empty()) { |
| 837 |
return; |
| 838 |
} |
| 839 |
///// data payloads |
| 840 |
auto addDataPayload = [&](DocRef::Kind kind) { |
| 841 |
const auto nullaryAtomic = result.info.op == SendOp::ATOMIC_LOAD || |
| 842 |
result.info.op == SendOp::ATOMIC_IINC || |
| 843 |
result.info.op == SendOp::ATOMIC_IDEC; |
| 844 |
if (nullaryAtomic && kind == DocRef::SRC1) |
| 845 |
return; // no payload |
| 846 |
|
| 847 |
const auto binaryAtomicSrc = |
| 848 |
kind == DocRef::SRC1 && (result.info.op == SendOp::ATOMIC_ICAS || |
| 849 |
result.info.op == SendOp::ATOMIC_FCAS); |
| 850 |
|
| 851 |
if (result.info.elemSizeBitsRegFile == 32) { |
| 852 |
if (binaryAtomicSrc) { |
| 853 |
if (result.info.execWidth <= 8 && platform() <= Platform::XE_HPG) { |
| 854 |
addDocXe(kind, "D32_2SRC_ATM_PAYLOAD_SIMT8", "55499", nullptr); |
| 855 |
} else if (result.info.execWidth <= 16) { |
| 856 |
addDocXe(kind, "D32_2SRC_ATM_PAYLOAD_SIMT16", "54696", "64003"); |
| 857 |
} else if (result.info.execWidth <= 32) { |
| 858 |
addDocXe(kind, "D32_2SRC_ATM_PAYLOAD", "54693", "64002"); |
| 859 |
} |
| 860 |
} else if (result.info.execWidth <= 8 && |
| 861 |
platform() <= Platform::XE_HPG) { |
| 862 |
addDocXe(kind, "D32_2SRC_ATM_PAYLOAD_SIMT8", "55499", nullptr); |
| 863 |
} else if (result.info.execWidth <= 8 && |
| 864 |
platform() <= Platform::XE_HPG) { |
| 865 |
addDocXe(kind, "D32_PAYLOAD_SIMT8", "55497", nullptr); |
| 866 |
} else if (result.info.execWidth <= 16) { |
| 867 |
addDocXe(kind, "D32_PAYLOAD_SIMT16", "54146", "64000"); |
| 868 |
} else { |
| 869 |
addDocXe(kind, "D32_PAYLOAD", "53574", "63999"); |
| 870 |
} |
| 871 |
} else if (result.info.elemSizeBitsRegFile == 64) { |
| 872 |
if (binaryAtomicSrc) { |
| 873 |
if (result.info.execWidth <= 8 && platform() <= Platform::XE_HPG) { |
| 874 |
addDocXe(kind, "D64_2SRC_ATM_PAYLOAD_SIMT8", "55498", nullptr); |
| 875 |
} else if (result.info.execWidth <= 16) { |
| 876 |
addDocXe(kind, "D64_2SRC_ATM_PAYLOAD_SIMT16", "54695", "64008"); |
| 877 |
} else if (result.info.execWidth <= 32) { |
| 878 |
addDocXe(kind, "D64_2SRC_ATM_PAYLOAD", "54692", "64007"); |
| 879 |
} |
| 880 |
} else if (result.info.execWidth <= 8 && |
| 881 |
platform() <= Platform::XE_HPG) { |
| 882 |
addDocXe(kind, "D64_PAYLOAD_SIMT8", "75252", nullptr); |
| 883 |
} else if (result.info.execWidth <= 16) { |
| 884 |
addDocXe(kind, "D64_PAYLOAD_SIMT16", "54147", "64006"); |
| 885 |
} else { |
| 886 |
addDocXe(kind, "D64_PAYLOAD", "53575", "64005"); |
| 887 |
} |
| 888 |
} |
| 889 |
}; |
| 890 |
|
| 891 |
if (result.info.isLoad() || result.info.isAtomic()) { |
| 892 |
addDataPayload(DocRef::DST); |
| 893 |
} |
| 894 |
|
| 895 |
///// address payloads |
| 896 |
if (op == SendOp::LOAD_STRIDED || op == SendOp::STORE_STRIDED) { |
| 897 |
addDocXe(DocRef::SRC0, "ABLOCK_PAYLOAD", "53571", "63996"); |
| 898 |
} else if (result.info.isTransposed()) { |
| 899 |
addDocXe(DocRef::SRC0, "A64_PAYLOAD_SIMT1", "55190", "63994"); |
| 900 |
} else if (result.info.addrSizeBits == 32) { |
| 901 |
if (result.info.execWidth <= 16 && platform() <= Platform::XE_HPG) { |
| 902 |
addDocXe(DocRef::SRC0, "A32_PAYLOAD_SIMT8", "55496", nullptr); |
| 903 |
} else if (result.info.execWidth <= 16) { |
| 904 |
addDocXe(DocRef::SRC0, "A32_PAYLOAD_SIMT16", "54148", "63991"); |
| 905 |
} else { |
| 906 |
addDocXe(DocRef::SRC0, "A32_PAYLOAD", "53569", "63990"); |
| 907 |
} |
| 908 |
} else if (result.info.addrSizeBits == 64) { |
| 909 |
if (result.info.execWidth <= 16 && platform() <= Platform::XE_HPG) { |
| 910 |
// pending an update |
| 911 |
addDocXe(DocRef::SRC0, "A64_PAYLOAD_SIMT8", "75251", nullptr); |
| 912 |
} else if (result.info.execWidth <= 16) { |
| 913 |
addDocXe(DocRef::SRC0, "A64_PAYLOAD_SIMT16", "54149", "63995"); |
| 914 |
} else { |
| 915 |
addDocXe(DocRef::SRC0, "A64_PAYLOAD", "53570", "63993"); |
| 916 |
} |
| 917 |
} |
| 918 |
|
| 919 |
// src1 |
| 920 |
if (result.info.isAtomic() || result.info.isStore()) { |
| 921 |
addDataPayload(DocRef::SRC1); |
| 922 |
} |
| 923 |
} // addDocRefsVecCmask |
| 924 |
|
| 925 |
void tryDecodeLsc() { |
| 926 |
int lscOp = getDescBits(0, 6); // Opcode[5:0] |
| 927 |
switch (lscOp) { |
| 928 |
case LSC_LOAD: |
| 929 |
addDocRefDESC("DP_LOAD", "53523", "63970"); |
| 930 |
decodeLscVectorCmask("gathering load", SendOp::LOAD); |
| 931 |
break; |
| 932 |
case LSC_STORE: |
| 933 |
addDocRefDESC("DP_STORE", "53524", "63980"); |
| 934 |
decodeLscVectorCmask("scattering store", SendOp::STORE); |
| 935 |
break; |
| 936 |
case LSC_STORE_UNCOMPRESSED: |
| 937 |
addDocRefDESC("DP_STORE_UNCOMPRESSED", "53532", "63984"); |
| 938 |
decodeLscMessage("scattering store uncompressed", |
| 939 |
SendOp::STORE_UNCOMPRESSED); |
| 940 |
break; |
| 941 |
case LSC_STORE_UNCOMPRESSED_QUAD: |
| 942 |
addDocRefDESC("DP_STORE_UC_CMASK", "55224", "63985"); |
| 943 |
decodeLscVectorCmask("store quad uncompressed", |
| 944 |
SendOp::STORE_UNCOMPRESSED_QUAD); |
| 945 |
break; |
| 946 |
case LSC_LOAD_QUAD: |
| 947 |
addDocRefDESC("DP_LOAD_CMASK", "53527", "63977"); |
| 948 |
decodeLscVectorCmask("quad load (a.k.a. load_cmask)", SendOp::LOAD_QUAD); |
| 949 |
break; |
| 950 |
case LSC_STORE_QUAD: |
| 951 |
addDocRefDESC("DP_STORE_CMASK", "53527", "63983"); |
| 952 |
decodeLscVectorCmask("quad store (a.k.a. store_cmask)", |
| 953 |
SendOp::STORE_QUAD); |
| 954 |
break; |
| 955 |
case LSC_LOAD_STRIDED: |
| 956 |
addDocRefDESC("DP_LOAD_BLOCK", "53525", "63976"); |
| 957 |
decodeLscVectorCmask("strided load (a.k.a load_block)", |
| 958 |
SendOp::LOAD_STRIDED); |
| 959 |
break; |
| 960 |
case LSC_STORE_STRIDED: |
| 961 |
addDocRefDESC("DP_STORE_BLOCK", "53526", "63982"); |
| 962 |
decodeLscVectorCmask("strided store (a.k.a store_block)", |
| 963 |
SendOp::STORE_STRIDED); |
| 964 |
break; |
| 965 |
case LSC_LOAD_BLOCK2D: |
| 966 |
if (sfid == SFID::TGM) { |
| 967 |
addDocRefDESCXE2("DP_LOAD_TYPED_2DBLOCK", "65280"); |
| 968 |
addDocXe(DocRef::SRC0, "Typed_2DBlock_Payload", nullptr, "65282"); |
| 969 |
decodeLscMessageTypeBlock2D("block2d load", SendOp::LOAD_BLOCK2D); |
| 970 |
} else { |
| 971 |
addDocRefDESC("DP_LOAD_2DBLOCK_ARRAY", "53680", "63972"); |
| 972 |
addDocXe(DocRef::SRC0, "A2DBLOCK_PAYLOAD", "53567", "63986"); |
| 973 |
decodeLscMessage("block2d load", SendOp::LOAD_BLOCK2D); |
| 974 |
} |
| 975 |
break; |
| 976 |
case LSC_STORE_BLOCK2D: |
| 977 |
if (sfid == SFID::TGM) { |
| 978 |
addDocRefDESCXE2("DP_STORE_TYPED_2DBLOCK", "65281"); |
| 979 |
decodeLscMessageTypeBlock2D("block2d store", SendOp::STORE_BLOCK2D); |
| 980 |
break; |
| 981 |
} else { |
| 982 |
addDocRefDESC("DP_STORE_2DBLOCK", "53530", "63981"); |
| 983 |
decodeLscMessage("block2d store", SendOp::STORE_BLOCK2D); |
| 984 |
} |
| 985 |
break; |
| 986 |
case LSC_ATOMIC_IINC: |
| 987 |
addDocRefDESC("DP_ATOMIC_INC", "53538", "63955"); |
| 988 |
decodeLscAtomicMessage("atomic integer increment", SendOp::ATOMIC_IINC); |
| 989 |
break; |
| 990 |
case LSC_ATOMIC_IDEC: |
| 991 |
addDocRefDESC("DP_ATOMIC_DEC", "53539", "63949"); |
| 992 |
decodeLscAtomicMessage("atomic integer decrement", SendOp::ATOMIC_IDEC); |
| 993 |
break; |
| 994 |
case LSC_ATOMIC_LOAD: |
| 995 |
addDocRefDESC("DP_ATOMIC_LOAD", "53540", "63956"); |
| 996 |
decodeLscAtomicMessage("atomic load", SendOp::ATOMIC_LOAD); |
| 997 |
break; |
| 998 |
case LSC_ATOMIC_STORE: |
| 999 |
addDocRefDESC("DP_ATOMIC_STORE", "53541", "63960"); |
| 1000 |
decodeLscAtomicMessage("atomic store", SendOp::ATOMIC_STORE); |
| 1001 |
break; |
| 1002 |
case LSC_ATOMIC_IADD: |
| 1003 |
addDocRefDESC("DP_ATOMIC_ADD", "53542", "63946"); |
| 1004 |
decodeLscAtomicMessage("atomic integer add", SendOp::ATOMIC_IADD); |
| 1005 |
break; |
| 1006 |
case LSC_ATOMIC_ISUB: |
| 1007 |
addDocRefDESC("DP_ATOMIC_SUB", "53543", "63961"); |
| 1008 |
decodeLscAtomicMessage("atomic integer subtract", SendOp::ATOMIC_ISUB); |
| 1009 |
break; |
| 1010 |
case LSC_ATOMIC_SMIN: |
| 1011 |
addDocRefDESC("DP_ATOMIC_MIN", "53544", "63958"); |
| 1012 |
decodeLscAtomicMessage("atomic signed-integer minimum", SendOp::ATOMIC_SMIN); |
| 1013 |
break; |
| 1014 |
case LSC_ATOMIC_SMAX: |
| 1015 |
addDocRefDESC("DP_ATOMIC_MAX", "53545", "63957"); |
| 1016 |
decodeLscAtomicMessage("atomic signed-integer maximum", SendOp::ATOMIC_SMAX); |
| 1017 |
break; |
| 1018 |
case LSC_ATOMIC_UMIN: |
| 1019 |
addDocRefDESC("DP_ATOMIC_UMIN", "53546", "63963"); |
| 1020 |
decodeLscAtomicMessage("atomic unsigned-integer minimum", |
| 1021 |
SendOp::ATOMIC_UMIN); |
| 1022 |
break; |
| 1023 |
case LSC_ATOMIC_UMAX: |
| 1024 |
addDocRefDESC("DP_ATOMIC_UMAX", "53547", "63962"); |
| 1025 |
decodeLscAtomicMessage("atomic unsigned-integer maximum", |
| 1026 |
SendOp::ATOMIC_UMAX); |
| 1027 |
break; |
| 1028 |
case LSC_ATOMIC_ICAS: |
| 1029 |
addDocRefDESC("DP_ATOMIC_CMPXCHG", "53555", "63948"); |
| 1030 |
decodeLscAtomicMessage("atomic integer compare and swap", |
| 1031 |
SendOp::ATOMIC_ICAS); |
| 1032 |
break; |
| 1033 |
case LSC_ATOMIC_FADD: |
| 1034 |
addDocRefDESC("DP_ATOMIC_FADD", "53548", "63950"); |
| 1035 |
decodeLscAtomicMessage("atomic float add", SendOp::ATOMIC_FADD); |
| 1036 |
break; |
| 1037 |
case LSC_ATOMIC_FSUB: |
| 1038 |
addDocRefDESC("DP_ATOMIC_FSUB", "53549", "63954"); |
| 1039 |
decodeLscAtomicMessage("atomic float subtract", SendOp::ATOMIC_FSUB); |
| 1040 |
break; |
| 1041 |
case LSC_ATOMIC_FMIN: |
| 1042 |
addDocRefDESC("DP_ATOMIC_FMIN", "53550", "63953"); |
| 1043 |
decodeLscAtomicMessage("atomic float minimum", SendOp::ATOMIC_FMIN); |
| 1044 |
break; |
| 1045 |
case LSC_ATOMIC_FMAX: |
| 1046 |
addDocRefDESC("DP_ATOMIC_FMAX", "53551", "63952"); |
| 1047 |
decodeLscAtomicMessage("atomic float maximum", SendOp::ATOMIC_FMAX); |
| 1048 |
break; |
| 1049 |
case LSC_ATOMIC_FCAS: |
| 1050 |
addDocRefDESC("DP_ATOMIC_FCMPXCHG", "DP_XXX", "63951"); |
| 1051 |
decodeLscAtomicMessage("atomic float compare and swap", SendOp::ATOMIC_FCAS); |
| 1052 |
break; |
| 1053 |
case LSC_ATOMIC_AND: |
| 1054 |
addDocRefDESC("DP_ATOMIC_AND", "53552", "63947"); |
| 1055 |
decodeLscAtomicMessage("atomic logical and", SendOp::ATOMIC_AND); |
| 1056 |
break; |
| 1057 |
case LSC_ATOMIC_OR: |
| 1058 |
addDocRefDESC("DP_ATOMIC_OR", "53553", "63959"); |
| 1059 |
decodeLscAtomicMessage("atomic logical or", SendOp::ATOMIC_OR); |
| 1060 |
break; |
| 1061 |
case LSC_ATOMIC_XOR: |
| 1062 |
addDocRefDESC("DP_ATOMIC_XOR", "53554", "63964"); |
| 1063 |
decodeLscAtomicMessage("atomic logical xor", SendOp::ATOMIC_XOR); |
| 1064 |
break; |
| 1065 |
case LSC_ATOMIC_APNDADD: |
| 1066 |
addDocRefDESCXE2("DP_APPENDCTR_ATOMIC_ADD", "68353"); |
| 1067 |
decodeLscAtomicAppendMessage("atomic counter append add", |
| 1068 |
SendOp::ATOMIC_ACADD); |
| 1069 |
break; |
| 1070 |
case LSC_ATOMIC_APNDSUB: |
| 1071 |
addDocRefDESCXE2("DP_APPENDCTR_ATOMIC_SUB", "68353"); |
| 1072 |
decodeLscAtomicAppendMessage("atomic counter append sub", |
| 1073 |
SendOp::ATOMIC_ACSUB); |
| 1074 |
break; |
| 1075 |
case LSC_ATOMIC_APNDSTORE: |
| 1076 |
addDocRefDESCXE2("DP_APPENDCTR_ATOMIC_ST", "68354"); |
| 1077 |
decodeLscAtomicAppendMessage("atomic counter append store", |
| 1078 |
SendOp::ATOMIC_ACSTORE); |
| 1079 |
break; |
| 1080 |
case LSC_LOAD_QUAD_MSRT: |
| 1081 |
addDocRefDESCXE2("DP_LOAD_CMASK_MSRT", "73733"); |
| 1082 |
decodeLscMessage("load quad from multi-sample render target", |
| 1083 |
SendOp::LOAD_QUAD_MSRT); |
| 1084 |
break; |
| 1085 |
case LSC_STORE_QUAD_MSRT: |
| 1086 |
addDocRefDESCXE2("DP_STORE_CMASK_MSRT", "73734"); |
| 1087 |
decodeLscMessage("store quad to multi-sample render target", |
| 1088 |
SendOp::STORE_QUAD_MSRT); |
| 1089 |
break; |
| 1090 |
case LSC_CCS: |
| 1091 |
decodeLscCcs(); |
| 1092 |
break; |
| 1093 |
case LSC_RSI: { |
| 1094 |
addDocRefDESC("DP_RSI", "54000", "63979"); |
| 1095 |
addDocXe(DocRef::DST, "DP_STATE_INFO_PAYLOAD", "54152", "64015"); |
| 1096 |
addDocXe(DocRef::SRC0, "DP_ASTATE_INFO_PAYLOAD", "55018", "64014"); |
| 1097 |
|
| 1098 |
addField("Opcode", 0, 6, getDescBits(0, 6), "read_state"); |
| 1099 |
// |
| 1100 |
std::stringstream descs; |
| 1101 |
descs << "read state information"; |
| 1102 |
result.syntax.mnemonic = "read_state"; |
| 1103 |
// |
| 1104 |
SendDesc surfId = 0; |
| 1105 |
auto at = decodeLscAddrType(surfId, false); |
| 1106 |
// |
| 1107 |
// XeHPG returns 2 GRF, XeHPC+ only 1 |
| 1108 |
// #54152 |
| 1109 |
int rlen = platform() == Platform::XE_HPG ? 2 : 1; |
| 1110 |
setSpecialOpX(result.syntax.mnemonic, descs.str(), SendOp::READ_STATE, at, |
| 1111 |
surfId, |
| 1112 |
1, // mlen = 1 (U,V,R,LOD) |
| 1113 |
rlen); |
| 1114 |
result.info.addrSizeBits = 64; |
| 1115 |
result.info.execWidth = 1; |
| 1116 |
result.info.attributeSet |= MessageInfo::Attr::HAS_UVRLOD; |
| 1117 |
result.info.attributeSet |= MessageInfo::Attr::TRANSPOSED; |
| 1118 |
break; |
| 1119 |
} |
| 1120 |
case LSC_FENCE: |
| 1121 |
addDocRefDESC("DP_FENCE", "53533", "63969"); |
| 1122 |
decodeLscFence(); |
| 1123 |
break; |
| 1124 |
case LSC_LOAD_STATUS: |
| 1125 |
addDocRefDESC("DP_LOAD_STATUS", "53531", "63978"); |
| 1126 |
addDocXe(DocRef::DST, "DP_STATUS_PAYLOAD", "55018", "64016"); |
| 1127 |
if (getDescBit(15)) { |
| 1128 |
error(15, 1, "transpose forbidden on load_status"); |
| 1129 |
} |
| 1130 |
if (getDescBits(20, 5) != 1) { |
| 1131 |
error(20, 5, "load_status must have rlen (Desc[24:20] == 1)"); |
| 1132 |
} |
| 1133 |
decodeLscMessage("load status", SendOp::LOAD_STATUS); |
| 1134 |
break; |
| 1135 |
default: |
| 1136 |
addField("Opcode", 0, 6, getDescBits(0, 6), "invalid message opcode"); |
| 1137 |
error(0, 6, "unsupported message opcode"); |
| 1138 |
return; |
| 1139 |
} |
| 1140 |
} |
| 1141 |
|
| 1142 |
void decodeLscCcs() { |
| 1143 |
addField("Opcode", 0, 6, static_cast<uint32_t>(LSC_CCS), |
| 1144 |
"compression-state control"); |
| 1145 |
// |
| 1146 |
std::stringstream descs; |
| 1147 |
result.syntax.mnemonic = "ccs"; |
| 1148 |
descs << "compression-state control"; |
| 1149 |
auto ccsOpBits = getDescBits(17, 3); |
| 1150 |
SendOp sop = SendOp::INVALID; |
| 1151 |
std::string opDesc; |
| 1152 |
switch (ccsOpBits) { |
| 1153 |
case 0: |
| 1154 |
sop = SendOp::CCS_PC; |
| 1155 |
result.syntax.mnemonic += "_pc"; |
| 1156 |
opDesc = " page clear (64k)"; |
| 1157 |
addDocRefDESC("DP_CCS_PAGE_CLEAR", "53536", "63965"); |
| 1158 |
break; |
| 1159 |
case 1: |
| 1160 |
sop = SendOp::CCS_SC; |
| 1161 |
result.syntax.mnemonic += "_sc"; |
| 1162 |
opDesc = " sector clear (2-cachelines)"; |
| 1163 |
addDocRefDESC("DP_CCS_SEC_CLEAR", "53534", "63967"); |
| 1164 |
result.syntax.controls += vectorSuffixSyntax; |
| 1165 |
break; |
| 1166 |
case 2: |
| 1167 |
sop = SendOp::CCS_PU; |
| 1168 |
result.syntax.mnemonic += "_pu"; |
| 1169 |
opDesc = " page uncompress (64k)"; |
| 1170 |
addDocRefDESC("DP_CCS_PAGE_UNCOMPRESS", "53537", "63966"); |
| 1171 |
break; |
| 1172 |
case 3: |
| 1173 |
sop = SendOp::CCS_SU; |
| 1174 |
result.syntax.mnemonic += "_su"; |
| 1175 |
opDesc = " sector uncompress (2-cachelines)"; |
| 1176 |
addDocRefDESC("DP_CCS_SEC_UNCOMPRESS", "53535", "63968"); |
| 1177 |
result.syntax.controls += vectorSuffixSyntax; |
| 1178 |
break; |
| 1179 |
default: { |
| 1180 |
std::stringstream ss; |
| 1181 |
ss << ".0x" << std::hex << std::uppercase << ccsOpBits; |
| 1182 |
result.syntax.controls += ss.str(); |
| 1183 |
opDesc = "invalid ccs sop"; |
| 1184 |
error(17, 3, "invalid ccs sop"); |
| 1185 |
} |
| 1186 |
} // switch |
| 1187 |
descs << opDesc; |
| 1188 |
// |
| 1189 |
addField("CcsOp", 17, 3, ccsOpBits, opDesc); |
| 1190 |
// |
| 1191 |
SendDesc surfId = 0; |
| 1192 |
auto at = decodeLscAddrType(surfId); |
| 1193 |
if (ccsOpBits == 0 || ccsOpBits == 2) { |
| 1194 |
// page operations: pc, pu |
| 1195 |
if (at != AddrType::FLAT) |
| 1196 |
error(29, 2, "ccs_{pcc,pcu} requires FLAT address type"); |
| 1197 |
std::stringstream dummy; |
| 1198 |
decodeLscAddrSize(); |
| 1199 |
if (addrSizeBits != 64) |
| 1200 |
error(7, 2, "AddrSize must be A64"); |
| 1201 |
result.info.execWidth = 1; |
| 1202 |
expectedExecSize = 1; |
| 1203 |
// sector uncompress has addresses |
| 1204 |
// FIXME: I could derive this via exec size and a64 |
| 1205 |
int mlen = |
| 1206 |
ccsOpBits == 1 || ccsOpBits == 3 ? 4 : // A64_PAYLOAD_SIMT32 = 4 regs |
| 1207 |
1; // A64_PAYLOAD_SIMT1 = 1 reg |
| 1208 |
int rlen = 0; // always 0 |
| 1209 |
setSpecialOpX(symbolFromSyntax(), descs.str(), sop, at, surfId, mlen, |
| 1210 |
rlen); |
| 1211 |
} else { |
| 1212 |
// sector operations |
| 1213 |
/// |
| 1214 |
// these are vector messages |
| 1215 |
expectedExecSize = DEFAULT_EXEC_SIZE; |
| 1216 |
// const int SECTOR_SIZE_BITS = 128*8; |
| 1217 |
// result.syntax.controls += ".d1024"; |
| 1218 |
result.syntax.controls += vectorSuffixSyntax; |
| 1219 |
result.syntax.controls += addrSizeBits == 64 ? ".a64" : ".a32"; |
| 1220 |
// |
| 1221 |
setScatterGatherOp(symbolFromSyntax(), descs.str(), sop, at, surfId, |
| 1222 |
addrSizeBits, |
| 1223 |
0, // dateSize = 0; nothing returned |
| 1224 |
vectorSize, DEFAULT_EXEC_SIZE, extraAttrs); |
| 1225 |
} |
| 1226 |
} |
| 1227 |
|
| 1228 |
void decodeLscFence() { |
| 1229 |
addField("Opcode", 0, 6, getDescBits(0, 6), "fence"); |
| 1230 |
// |
| 1231 |
std::stringstream descs; |
| 1232 |
result.syntax.mnemonic = "fence"; |
| 1233 |
descs << "fence"; |
| 1234 |
// |
| 1235 |
std::stringstream fenceOpts; |
| 1236 |
addLscFenceFields(fenceOpts, descs); |
| 1237 |
result.syntax.controls += fenceOpts.str(); |
| 1238 |
// |
| 1239 |
setSpecialOpX(symbolFromSyntax(), descs.str(), SendOp::FENCE, |
| 1240 |
AddrType::FLAT, |
| 1241 |
0, // no surface |
| 1242 |
1, // mlen = 1 |
| 1243 |
0); // rlen = 0 |
| 1244 |
} |
| 1245 |
}; // MessageDecoderLSC |
| 1246 |
|
| 1247 |
void iga::decodeDescriptorsLSC(Platform platform, SFID sfid, ExecSize execSize, |
| 1248 |
uint32_t exImmOffDesc, SendDesc exDesc, |
| 1249 |
SendDesc desc, DecodeResult &result) { |
| 1250 |
MessageDecoderLSC md(platform, sfid, execSize, exImmOffDesc, exDesc, desc, |
| 1251 |
result); |
| 1252 |
md.tryDecodeLsc(); |
| 1253 |
} |
| 1254 |
|
| 1255 |
// descriptor bits [19:17]: cache control |
| 1256 |
static bool encLdStVecCachingBits17_19(SendOp op, CacheOpt cachingL1, |
| 1257 |
CacheOpt cachingL3, SendDesc &desc) { |
| 1258 |
const auto &opInfo = lookupSendOp(op); |
| 1259 |
bool isLd = opInfo.isLoad(); |
| 1260 |
bool isSt = opInfo.isStore(); |
| 1261 |
bool isAt = opInfo.isAtomic(); |
| 1262 |
bool isStAt = isSt || isAt; |
| 1263 |
auto ccMatches = [&](CacheOpt l1, CacheOpt l3, uint32_t enc) { |
| 1264 |
if (l1 == cachingL1 && l3 == cachingL3) { |
| 1265 |
desc.imm |= enc << 17; |
| 1266 |
return true; |
| 1267 |
} |
| 1268 |
return false; |
| 1269 |
}; |
| 1270 |
bool matched = |
| 1271 |
ccMatches(CacheOpt::DEFAULT, CacheOpt::DEFAULT, LSC_DF_DF) || |
| 1272 |
// |
| 1273 |
ccMatches(CacheOpt::UNCACHED, CacheOpt::UNCACHED, LSC_UC_UC) || |
| 1274 |
// |
| 1275 |
(isLd && ccMatches(CacheOpt::UNCACHED, CacheOpt::CACHED, LSC_UC_CA)) || |
| 1276 |
(isStAt && |
| 1277 |
ccMatches(CacheOpt::UNCACHED, CacheOpt::WRITEBACK, LSC_UC_WB)) || |
| 1278 |
// |
| 1279 |
(isLd && ccMatches(CacheOpt::CACHED, CacheOpt::UNCACHED, LSC_CA_UC)) || |
| 1280 |
(isSt && |
| 1281 |
ccMatches(CacheOpt::WRITETHROUGH, CacheOpt::UNCACHED, LSC_WT_UC)) || |
| 1282 |
// |
| 1283 |
(isLd && ccMatches(CacheOpt::CACHED, CacheOpt::CACHED, LSC_CA_CA)) || |
| 1284 |
(isSt && |
| 1285 |
ccMatches(CacheOpt::WRITETHROUGH, CacheOpt::WRITEBACK, LSC_WT_WB)) || |
| 1286 |
// |
| 1287 |
ccMatches(CacheOpt::STREAMING, CacheOpt::UNCACHED, LSC_ST_UC) || |
| 1288 |
// |
| 1289 |
(isLd && ccMatches(CacheOpt::STREAMING, CacheOpt::CACHED, LSC_ST_CA)) || |
| 1290 |
(isSt && |
| 1291 |
ccMatches(CacheOpt::STREAMING, CacheOpt::WRITEBACK, LSC_ST_WB)) || |
| 1292 |
// |
| 1293 |
(isLd && |
| 1294 |
ccMatches(CacheOpt::READINVALIDATE, CacheOpt::CACHED, LSC_RI_CA)) || |
| 1295 |
(isSt && ccMatches(CacheOpt::WRITEBACK, CacheOpt::WRITEBACK, LSC_WB_WB)); |
| 1296 |
return matched; |
| 1297 |
} |
| 1298 |
|
| 1299 |
// encLdStVecCachingBits17_19 - Encode 4-bits cache opt at descriptor[19:16] |
| 1300 |
// Cache opt on XE2+ is 4 bits field comparing to 3 bits on pre-XE2 |
| 1301 |
// XE2 is [19:16] and preXE2 is [19:17] |
| 1302 |
static bool encLdStVecCachingBits16_19(SendOp op, CacheOpt cachingL1, |
| 1303 |
CacheOpt cachingL3, SendDesc &desc) { |
| 1304 |
// special handle L1RI_L3CA which is not allowed in 4-bits-cache-opt mode |
| 1305 |
if (cachingL1 == CacheOpt::READINVALIDATE && |
| 1306 |
cachingL3 == CacheOpt::CONSTCACHED) |
| 1307 |
return false; |
| 1308 |
|
| 1309 |
// bits in [19:17] are the same as encLdStVecCachingBits17_19 |
| 1310 |
if (encLdStVecCachingBits17_19(op, cachingL1, cachingL3, desc)) |
| 1311 |
return true; |
| 1312 |
|
| 1313 |
// otherwise, check the new combination (all for load only) |
| 1314 |
// L1UC_L3CC, L1CA_L3CC, L1RI_L3RI |
| 1315 |
const auto &opInfo = lookupSendOp(op); |
| 1316 |
bool isLd = opInfo.isLoad(); |
| 1317 |
auto ccMatches = [&](CacheOpt l1, CacheOpt l3, uint32_t enc) { |
| 1318 |
if (l1 == cachingL1 && l3 == cachingL3) { |
| 1319 |
desc.imm |= enc << 16; |
| 1320 |
return true; |
| 1321 |
} |
| 1322 |
return false; |
| 1323 |
}; |
| 1324 |
bool matched = |
| 1325 |
(isLd && |
| 1326 |
ccMatches(CacheOpt::UNCACHED, CacheOpt::CONSTCACHED, LSC_UC_CC)) || |
| 1327 |
(isLd && ccMatches(CacheOpt::CACHED, CacheOpt::CONSTCACHED, LSC_CA_CC)) || |
| 1328 |
(isLd && ccMatches(CacheOpt::READINVALIDATE, CacheOpt::READINVALIDATE, |
| 1329 |
LSC_RI_RI)); |
| 1330 |
return matched; |
| 1331 |
} |
| 1332 |
|
| 1333 |
static bool encLdStVecImmOffset(Platform p, const VectorMessageArgs &vma, |
| 1334 |
uint32_t &exImmOffDesc, SendDesc &exDesc, |
| 1335 |
std::string &err) { |
| 1336 |
const bool isBlock2d = |
| 1337 |
vma.op == SendOp::LOAD_BLOCK2D || vma.op == SendOp::STORE_BLOCK2D; |
| 1338 |
auto fitsInBitSize = [](int bits, int offset) { |
| 1339 |
return offset >= -(1 << (bits - 1)) && offset <= (1 << (bits - 1)) - 1; |
| 1340 |
}; |
| 1341 |
uint32_t encodedOffset = 0; |
| 1342 |
if (isBlock2d && vma.addrOffsetX * vma.dataSizeMem % 32 != 0) { |
| 1343 |
err = "address offset-y must be 32b aligned"; |
| 1344 |
return false; |
| 1345 |
} else if (isBlock2d && vma.addrOffsetY * vma.dataSizeMem % 32 != 0) { |
| 1346 |
err = "address offset-x must be 32b aligned"; |
| 1347 |
return false; |
| 1348 |
} else if (!isBlock2d && vma.addrOffset % 4 != 0) { |
| 1349 |
err = "address offset must be 32b aligned"; |
| 1350 |
return false; |
| 1351 |
} else if (vma.addrType == AddrType::FLAT) { |
| 1352 |
if (isBlock2d) { |
| 1353 |
if (vma.sfid != SFID::UGM) { |
| 1354 |
err = "address offset only defined for flat ugm"; |
| 1355 |
return false; |
| 1356 |
} else if (!fitsInBitSize(10, vma.addrOffsetX)) { |
| 1357 |
err = "address offset-x only exceeds 10b"; |
| 1358 |
return false; |
| 1359 |
} else if (!fitsInBitSize(10, vma.addrOffsetY)) { |
| 1360 |
err = "address offset-y only exceeds 10b"; |
| 1361 |
return false; |
| 1362 |
} |
| 1363 |
// ExDescImm[31:22] = Y-offset |
| 1364 |
// ExDescImm[21:12] = X-offset |
| 1365 |
encodedOffset |= ((uint32_t)vma.addrOffsetY & 0x3FF) << 22; |
| 1366 |
encodedOffset |= ((uint32_t)vma.addrOffsetX & 0x3FF) << 12; |
| 1367 |
} else if (!fitsInBitSize(20, vma.addrOffset)) { |
| 1368 |
err = "address offset exceeds 20b"; |
| 1369 |
return false; |
| 1370 |
} else { |
| 1371 |
// ExDescImm[31:12] = base offset |
| 1372 |
encodedOffset = (uint32_t)vma.addrOffset << 12; |
| 1373 |
} |
| 1374 |
} else if (isBlock2d) { |
| 1375 |
err = "block2d immediate offset only valid for flat"; |
| 1376 |
return false; |
| 1377 |
} else if (vma.addrType == AddrType::BTI && !vma.addrSurface.isReg()) { |
| 1378 |
if (!fitsInBitSize(12, vma.addrOffset)) { |
| 1379 |
err = "bti address offset exceeds 12b"; |
| 1380 |
return false; |
| 1381 |
} |
| 1382 |
encodedOffset = ((uint32_t)(vma.addrOffset & 0x00FFF) << 12); |
| 1383 |
} else if (vma.addrSurface.isReg()) { |
| 1384 |
if (vma.sfid != SFID::UGM) { |
| 1385 |
err = "unsupported SFID for ExDescReg + BaseOff"; |
| 1386 |
return false; |
| 1387 |
} else if (!fitsInBitSize(17, vma.addrOffset)) { |
| 1388 |
err = "address offset of ss/bss (.ugm) exceeds 17b"; |
| 1389 |
return false; |
| 1390 |
} |
| 1391 |
// |
| 1392 |
// ExDesc[31:19] = BaseOffset[16:4] |
| 1393 |
// ExDesc[15:12] = BaseOffset[3:0] |
| 1394 |
encodedOffset |= getBits((uint32_t)vma.addrOffset, 4, 13) << 19; |
| 1395 |
encodedOffset |= getBits((uint32_t)vma.addrOffset, 0, 4) << 12; |
| 1396 |
} else { |
| 1397 |
// BSS/SS with imm offset |
| 1398 |
err = "address offset forbidden bti/ss/bss with imm ExDesc surface"; |
| 1399 |
return false; |
| 1400 |
} |
| 1401 |
// needs to write to exImmOffDesc only if ExDesc.IsReg |
| 1402 |
if (vma.addrSurface.isReg()) { |
| 1403 |
exImmOffDesc = encodedOffset; |
| 1404 |
} else { |
| 1405 |
exDesc = encodedOffset; |
| 1406 |
} |
| 1407 |
return true; |
| 1408 |
} |
| 1409 |
|
| 1410 |
static bool encLdStVecCaching(const Platform &p, SendOp op, CacheOpt cachingL1, |
| 1411 |
CacheOpt cachingL3, SendDesc &desc) { |
| 1412 |
if (p >= Platform::XE2) |
| 1413 |
return encLdStVecCachingBits16_19(op, cachingL1, cachingL3, desc); |
| 1414 |
|
| 1415 |
return encLdStVecCachingBits17_19(op, cachingL1, cachingL3, desc); |
| 1416 |
} |
| 1417 |
|
| 1418 |
static bool encLdStVec(Platform p, const VectorMessageArgs &vma, |
| 1419 |
uint32_t &exImmOffDesc, |
| 1420 |
SendDesc &exDesc, SendDesc &desc, std::string &err) { |
| 1421 |
desc = 0x0; |
| 1422 |
exDesc = 0x0; |
| 1423 |
// |
| 1424 |
bool hasCMask = false; |
| 1425 |
switch (vma.op) { |
| 1426 |
case SendOp::LOAD: |
| 1427 |
desc.imm |= LSC_LOAD; |
| 1428 |
break; |
| 1429 |
case SendOp::LOAD_STRIDED: |
| 1430 |
desc.imm |= LSC_LOAD_STRIDED; |
| 1431 |
break; |
| 1432 |
case SendOp::LOAD_QUAD: |
| 1433 |
desc.imm |= LSC_LOAD_QUAD; |
| 1434 |
hasCMask = true; |
| 1435 |
break; |
| 1436 |
case SendOp::LOAD_BLOCK2D: |
| 1437 |
desc.imm |= LSC_LOAD_BLOCK2D; |
| 1438 |
break; |
| 1439 |
// |
| 1440 |
case SendOp::STORE: |
| 1441 |
desc.imm |= LSC_STORE; |
| 1442 |
break; |
| 1443 |
case SendOp::STORE_STRIDED: |
| 1444 |
desc.imm |= LSC_STORE_STRIDED; |
| 1445 |
break; |
| 1446 |
case SendOp::STORE_QUAD: |
| 1447 |
desc.imm |= LSC_STORE_QUAD; |
| 1448 |
hasCMask = true; |
| 1449 |
break; |
| 1450 |
case SendOp::STORE_UNCOMPRESSED: |
| 1451 |
desc.imm |= LSC_STORE_UNCOMPRESSED; |
| 1452 |
break; |
| 1453 |
case SendOp::STORE_UNCOMPRESSED_QUAD: |
| 1454 |
desc.imm |= LSC_STORE_UNCOMPRESSED_QUAD; |
| 1455 |
hasCMask = true; |
| 1456 |
break; |
| 1457 |
case SendOp::STORE_BLOCK2D: |
| 1458 |
desc.imm |= LSC_STORE_BLOCK2D; |
| 1459 |
break; |
| 1460 |
// |
| 1461 |
case SendOp::ATOMIC_AND: |
| 1462 |
desc.imm |= LSC_ATOMIC_AND; |
| 1463 |
break; |
| 1464 |
case SendOp::ATOMIC_FADD: |
| 1465 |
desc.imm |= LSC_ATOMIC_FADD; |
| 1466 |
break; |
| 1467 |
case SendOp::ATOMIC_FCAS: |
| 1468 |
desc.imm |= LSC_ATOMIC_FCAS; |
| 1469 |
break; |
| 1470 |
case SendOp::ATOMIC_FMAX: |
| 1471 |
desc.imm |= LSC_ATOMIC_FMAX; |
| 1472 |
break; |
| 1473 |
case SendOp::ATOMIC_FMIN: |
| 1474 |
desc.imm |= LSC_ATOMIC_FMIN; |
| 1475 |
break; |
| 1476 |
case SendOp::ATOMIC_FSUB: |
| 1477 |
desc.imm |= LSC_ATOMIC_FSUB; |
| 1478 |
break; |
| 1479 |
case SendOp::ATOMIC_IADD: |
| 1480 |
desc.imm |= LSC_ATOMIC_IADD; |
| 1481 |
break; |
| 1482 |
case SendOp::ATOMIC_ICAS: |
| 1483 |
desc.imm |= LSC_ATOMIC_ICAS; |
| 1484 |
break; |
| 1485 |
case SendOp::ATOMIC_IDEC: |
| 1486 |
desc.imm |= LSC_ATOMIC_IDEC; |
| 1487 |
break; |
| 1488 |
case SendOp::ATOMIC_IINC: |
| 1489 |
desc.imm |= LSC_ATOMIC_IINC; |
| 1490 |
break; |
| 1491 |
case SendOp::ATOMIC_ISUB: |
| 1492 |
desc.imm |= LSC_ATOMIC_ISUB; |
| 1493 |
break; |
| 1494 |
case SendOp::ATOMIC_LOAD: |
| 1495 |
desc.imm |= LSC_ATOMIC_LOAD; |
| 1496 |
break; |
| 1497 |
case SendOp::ATOMIC_OR: |
| 1498 |
desc.imm |= LSC_ATOMIC_OR; |
| 1499 |
break; |
| 1500 |
case SendOp::ATOMIC_SMAX: |
| 1501 |
desc.imm |= LSC_ATOMIC_SMAX; |
| 1502 |
break; |
| 1503 |
case SendOp::ATOMIC_SMIN: |
| 1504 |
desc.imm |= LSC_ATOMIC_SMIN; |
| 1505 |
break; |
| 1506 |
case SendOp::ATOMIC_STORE: |
| 1507 |
desc.imm |= LSC_ATOMIC_STORE; |
| 1508 |
break; |
| 1509 |
case SendOp::ATOMIC_UMAX: |
| 1510 |
desc.imm |= LSC_ATOMIC_UMAX; |
| 1511 |
break; |
| 1512 |
case SendOp::ATOMIC_UMIN: |
| 1513 |
desc.imm |= LSC_ATOMIC_UMIN; |
| 1514 |
break; |
| 1515 |
case SendOp::ATOMIC_XOR: |
| 1516 |
desc.imm |= LSC_ATOMIC_XOR; |
| 1517 |
break; |
| 1518 |
case SendOp::ATOMIC_ACADD: |
| 1519 |
desc.imm |= LSC_ATOMIC_APNDADD; |
| 1520 |
break; |
| 1521 |
case SendOp::ATOMIC_ACSUB: |
| 1522 |
desc.imm |= LSC_ATOMIC_APNDSUB; |
| 1523 |
break; |
| 1524 |
case SendOp::ATOMIC_ACSTORE: |
| 1525 |
desc.imm |= LSC_ATOMIC_APNDSTORE; |
| 1526 |
break; |
| 1527 |
case SendOp::LOAD_QUAD_MSRT: |
| 1528 |
desc.imm |= LSC_LOAD_QUAD_MSRT; |
| 1529 |
break; |
| 1530 |
case SendOp::STORE_QUAD_MSRT: |
| 1531 |
desc.imm |= LSC_STORE_QUAD_MSRT; |
| 1532 |
break; |
| 1533 |
default: |
| 1534 |
err = "unsupported op"; |
| 1535 |
return false; |
| 1536 |
} |
| 1537 |
bool isBlock2d = |
| 1538 |
vma.op == SendOp::LOAD_BLOCK2D || vma.op == SendOp::STORE_BLOCK2D; |
| 1539 |
bool isBlock2dTyped = isBlock2d && vma.sfid == SFID::TGM; |
| 1540 |
bool isBlock2dUntyped = isBlock2d && vma.sfid != SFID::TGM; |
| 1541 |
bool hasAddrSizeField = !isBlock2d; |
| 1542 |
bool isAtomicAddCounter = vma.op == SendOp::ATOMIC_ACADD || |
| 1543 |
vma.op == SendOp::ATOMIC_ACSUB || |
| 1544 |
vma.op == SendOp::ATOMIC_ACSTORE; |
| 1545 |
hasAddrSizeField &= !isAtomicAddCounter; |
| 1546 |
|
| 1547 |
// |
| 1548 |
//////////////////////////////////////// |
| 1549 |
// data size |
| 1550 |
uint32_t dszEnc = LSC_D8; |
| 1551 |
if (isBlock2dTyped && (vma.dataSizeReg != 32 || vma.dataSizeMem != 32)) { |
| 1552 |
err = "block2d.tgm must be d32"; |
| 1553 |
return false; |
| 1554 |
} |
| 1555 |
if (vma.dataSizeMem == vma.dataSizeReg) { |
| 1556 |
switch (vma.dataSizeMem) { |
| 1557 |
case 8: |
| 1558 |
dszEnc = LSC_D8; |
| 1559 |
break; |
| 1560 |
case 16: |
| 1561 |
dszEnc = LSC_D16; |
| 1562 |
break; |
| 1563 |
case 32: |
| 1564 |
dszEnc = LSC_D32; |
| 1565 |
break; |
| 1566 |
case 64: |
| 1567 |
dszEnc = LSC_D64; |
| 1568 |
break; |
| 1569 |
default: |
| 1570 |
err = "invalid data size"; |
| 1571 |
return false; |
| 1572 |
} |
| 1573 |
} else if (vma.dataSizeMem == 8 && vma.dataSizeReg == 32) { |
| 1574 |
dszEnc = LSC_D8U32; |
| 1575 |
} else if (vma.dataSizeMem == 16 && vma.dataSizeReg == 32) { |
| 1576 |
if (vma.dataSizeExpandHigh) { |
| 1577 |
dszEnc = LSC_D16U32H; |
| 1578 |
} else { |
| 1579 |
dszEnc = LSC_D16U32; |
| 1580 |
} |
| 1581 |
} else { |
| 1582 |
err = "invalid data type"; |
| 1583 |
return false; |
| 1584 |
} |
| 1585 |
if (!isBlock2dTyped) |
| 1586 |
desc.imm |= dszEnc << 9; |
| 1587 |
// |
| 1588 |
//////////////////////////////////////// |
| 1589 |
// vector size |
| 1590 |
if (hasCMask) { |
| 1591 |
if (vma.dataComponentMask & ~0xF) { |
| 1592 |
err = "invalid component mask"; |
| 1593 |
return false; |
| 1594 |
} |
| 1595 |
desc.imm |= vma.dataComponentMask << 12; |
| 1596 |
} else if (isBlock2d) { |
| 1597 |
if (isBlock2dTyped && vma.dataVnni) { |
| 1598 |
err = "block2d.tgm forbids VNNI"; |
| 1599 |
return false; |
| 1600 |
} else if (isBlock2dTyped && vma.dataTranspose) { |
| 1601 |
err = "block2d.tgm forbids transpose data order"; |
| 1602 |
return false; |
| 1603 |
} |
| 1604 |
if (vma.dataVnni) |
| 1605 |
desc.imm |= 1 << 7; |
| 1606 |
if (vma.dataTranspose) |
| 1607 |
desc.imm |= 1 << 15; |
| 1608 |
} else { |
| 1609 |
uint32_t vecEnc = LSC_V1; |
| 1610 |
switch (vma.dataVectorSize) { |
| 1611 |
case 1: |
| 1612 |
vecEnc = LSC_V1; |
| 1613 |
break; |
| 1614 |
case 2: |
| 1615 |
vecEnc = LSC_V2; |
| 1616 |
break; |
| 1617 |
case 3: |
| 1618 |
vecEnc = LSC_V3; |
| 1619 |
break; |
| 1620 |
case 4: |
| 1621 |
vecEnc = LSC_V4; |
| 1622 |
break; |
| 1623 |
case 8: |
| 1624 |
vecEnc = LSC_V8; |
| 1625 |
break; |
| 1626 |
case 16: |
| 1627 |
vecEnc = LSC_V16; |
| 1628 |
break; |
| 1629 |
case 32: |
| 1630 |
vecEnc = LSC_V32; |
| 1631 |
break; |
| 1632 |
case 64: |
| 1633 |
vecEnc = LSC_V64; |
| 1634 |
break; |
| 1635 |
default: |
| 1636 |
err = "invalid vector size"; |
| 1637 |
break; |
| 1638 |
} |
| 1639 |
if (vma.isAtomic() && vma.dataVectorSize != 1) { |
| 1640 |
err = "atomics do not support vector operations"; |
| 1641 |
return false; |
| 1642 |
} |
| 1643 |
if (vma.dataVnni) { |
| 1644 |
err = "vnni only valid on block2d operations"; |
| 1645 |
return false; |
| 1646 |
} |
| 1647 |
// |
| 1648 |
desc.imm |= vecEnc << 12; |
| 1649 |
// |
| 1650 |
if (vma.dataTranspose) { |
| 1651 |
desc.imm |= 1 << 15; |
| 1652 |
// |
| 1653 |
if (vma.isAtomic()) { |
| 1654 |
err = "atomics do not support transpose operations"; |
| 1655 |
return false; |
| 1656 |
} |
| 1657 |
} |
| 1658 |
} // end vec non-cmask case |
| 1659 |
// |
| 1660 |
//////////////////////////////////////// |
| 1661 |
// caching options |
| 1662 |
if (vma.isAtomic() && vma.cachingL1 != CacheOpt::DEFAULT && |
| 1663 |
vma.cachingL1 != CacheOpt::UNCACHED) { |
| 1664 |
err = "atomic L1 must be an uncached option"; |
| 1665 |
return false; |
| 1666 |
} else { |
| 1667 |
if (!encLdStVecCaching(p, vma.op, vma.cachingL1, vma.cachingL3, desc)) { |
| 1668 |
err = "invalid cache-control combination"; |
| 1669 |
return false; |
| 1670 |
} |
| 1671 |
} |
| 1672 |
// |
| 1673 |
//////////////////////////////////////// |
| 1674 |
// address size |
| 1675 |
uint32_t asEnc = 0x0; |
| 1676 |
switch (vma.addrSize) { |
| 1677 |
case 16: |
| 1678 |
asEnc = LSC_A16; |
| 1679 |
break; |
| 1680 |
case 32: |
| 1681 |
asEnc = LSC_A32; |
| 1682 |
break; |
| 1683 |
case 64: |
| 1684 |
asEnc = LSC_A64; |
| 1685 |
break; |
| 1686 |
default: |
| 1687 |
err = "unsupported address size"; |
| 1688 |
return false; |
| 1689 |
} |
| 1690 |
if (isBlock2dTyped && vma.addrSize != 32) { |
| 1691 |
err = "block2d.typed address size must be A32"; |
| 1692 |
return false; |
| 1693 |
} |
| 1694 |
if (isBlock2dUntyped && vma.addrSize != 64) { |
| 1695 |
err = "block2d untyped address size must be A64"; |
| 1696 |
return false; |
| 1697 |
} |
| 1698 |
if (hasAddrSizeField) { |
| 1699 |
desc.imm |= asEnc << 7; |
| 1700 |
} |
| 1701 |
// |
| 1702 |
//////////////////////////////////////// |
| 1703 |
// address type |
| 1704 |
uint32_t atEnc = 0x0; |
| 1705 |
switch (vma.addrType) { |
| 1706 |
case AddrType::FLAT: |
| 1707 |
atEnc = LSC_AT_FLAT; |
| 1708 |
break; |
| 1709 |
case AddrType::BSS: |
| 1710 |
atEnc = LSC_AT_BSS; |
| 1711 |
break; |
| 1712 |
case AddrType::SS: |
| 1713 |
atEnc = LSC_AT_SS; |
| 1714 |
break; |
| 1715 |
case AddrType::BTI: |
| 1716 |
atEnc = LSC_AT_BTI; |
| 1717 |
break; |
| 1718 |
default: |
| 1719 |
err = "unsupported address type"; |
| 1720 |
return false; |
| 1721 |
} |
| 1722 |
if (isBlock2dTyped && vma.addrType == AddrType::FLAT) { |
| 1723 |
err = "block2d.typed forbids flat address"; |
| 1724 |
return false; |
| 1725 |
} |
| 1726 |
desc.imm |= atEnc << 29; |
| 1727 |
// |
| 1728 |
// store the surface |
| 1729 |
if (vma.addrType != AddrType::FLAT) { |
| 1730 |
// use exDesc |
| 1731 |
if (vma.addrType == AddrType::BTI && !vma.addrSurface.isReg()) { |
| 1732 |
exDesc = vma.addrSurface.imm << 24; |
| 1733 |
} else { |
| 1734 |
exDesc = vma.addrSurface; |
| 1735 |
} |
| 1736 |
} |
| 1737 |
// |
| 1738 |
if (vma.addrType != AddrType::FLAT && vma.sfid == SFID::SLM) { |
| 1739 |
err = "SLM requires flat address type"; |
| 1740 |
return false; |
| 1741 |
} |
| 1742 |
// |
| 1743 |
//////////////////////////////////////// |
| 1744 |
// address immediate offset |
| 1745 |
bool hasAddrImmOffset = vma.addrOffset != 0; |
| 1746 |
hasAddrImmOffset |= vma.addrOffsetX != 0; |
| 1747 |
hasAddrImmOffset |= vma.addrOffsetY != 0; |
| 1748 |
if (hasAddrImmOffset) { |
| 1749 |
if (p < Platform::XE2) { |
| 1750 |
err = "address immediate offset not supported on this platform"; |
| 1751 |
return false; |
| 1752 |
} |
| 1753 |
|
| 1754 |
if (!encLdStVecImmOffset(p, vma, exImmOffDesc, exDesc, err)) |
| 1755 |
return false; |
| 1756 |
} // else: addrOffset == 0 |
| 1757 |
|
| 1758 |
//////////////////////////////////////// |
| 1759 |
// set the surface object |
| 1760 |
if (vma.addrType == AddrType::FLAT) { |
| 1761 |
// IR normalization |
| 1762 |
if (!vma.addrSurface.isImm() || vma.addrSurface.imm != 0) { |
| 1763 |
err = "malformed IR: flat address model must have surface = 0"; |
| 1764 |
return false; |
| 1765 |
} |
| 1766 |
} |
| 1767 |
|
| 1768 |
// XeHPG+ have surface in ExDesc |
| 1769 |
if (vma.addrType == AddrType::BTI && vma.addrSurface.isImm()) { |
| 1770 |
// BTI takes the high byte |
| 1771 |
if (vma.addrSurface.imm > 0xFF) { |
| 1772 |
err = "surface index too large for BTI"; |
| 1773 |
return false; |
| 1774 |
} |
| 1775 |
exDesc.imm |= vma.addrSurface.imm << 24; |
| 1776 |
} else if (vma.addrType != AddrType::FLAT) { |
| 1777 |
uint32_t ZERO_MASK = 0xFFF; |
| 1778 |
std::string highBit = "11"; |
| 1779 |
if (p >= Platform::XE2) { |
| 1780 |
ZERO_MASK = 0x7FF; |
| 1781 |
highBit = "10"; |
| 1782 |
} |
| 1783 |
|
| 1784 |
// if BTI reg or BSS/SS reg/imm with just copy |
| 1785 |
// BSS/SS with imm, value is already aligned |
| 1786 |
if (vma.addrType != AddrType::BTI && vma.addrSurface.isImm() && |
| 1787 |
(vma.addrSurface.imm & ZERO_MASK) != 0) { |
| 1788 |
err = "BSS/SS with immediate descriptor require " |
| 1789 |
"ExDesc[" + |
| 1790 |
highBit + ":0] to be 0"; |
| 1791 |
return false; |
| 1792 |
} |
| 1793 |
exDesc = vma.addrSurface; |
| 1794 |
} |
| 1795 |
// |
| 1796 |
return true; |
| 1797 |
} |
| 1798 |
|
| 1799 |
bool iga::encodeDescriptorsLSC(Platform p, const VectorMessageArgs &vma, |
| 1800 |
uint32_t &exImmOffDesc, SendDesc &exDesc, |
| 1801 |
SendDesc &desc, std::string &err) { |
| 1802 |
if (!sendOpSupportsSyntax(p, vma.op, vma.sfid)) { |
| 1803 |
err = "unsupported message for SFID"; |
| 1804 |
return false; |
| 1805 |
} |
| 1806 |
return encLdStVec(p, vma, exImmOffDesc, exDesc, desc, err); |
| 1807 |
} |