| 1 | //===- TypeParser.h - Quantization Type Parser ------------------*- C++ -*-===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | |
| 9 | #include "mlir/Dialect/Quant/IR/Quant.h" |
| 10 | #include "mlir/Dialect/Quant/IR/QuantTypes.h" |
| 11 | #include "mlir/IR/BuiltinTypes.h" |
| 12 | #include "mlir/IR/DialectImplementation.h" |
| 13 | #include "mlir/IR/Location.h" |
| 14 | #include "mlir/IR/Types.h" |
| 15 | #include "llvm/ADT/APFloat.h" |
| 16 | #include "llvm/Support/Format.h" |
| 17 | #include "llvm/Support/MathExtras.h" |
| 18 | #include "llvm/Support/SourceMgr.h" |
| 19 | #include "llvm/Support/raw_ostream.h" |
| 20 | |
| 21 | using namespace mlir; |
| 22 | using namespace quant; |
| 23 | |
| 24 | static IntegerType parseStorageType(DialectAsmParser &parser, bool &isSigned) { |
| 25 | auto typeLoc = parser.getCurrentLocation(); |
| 26 | IntegerType type; |
| 27 | |
| 28 | // Parse storage type (alpha_ident, integer_literal). |
| 29 | StringRef identifier; |
| 30 | unsigned storageTypeWidth = 0; |
| 31 | OptionalParseResult result = parser.parseOptionalType(result&: type); |
| 32 | if (result.has_value()) { |
| 33 | if (!succeeded(Result: *result)) |
| 34 | return nullptr; |
| 35 | isSigned = !type.isUnsigned(); |
| 36 | storageTypeWidth = type.getWidth(); |
| 37 | } else if (succeeded(Result: parser.parseKeyword(keyword: &identifier))) { |
| 38 | // Otherwise, this must be an unsigned integer (`u` integer-literal). |
| 39 | if (!identifier.consume_front(Prefix: "u" )) { |
| 40 | parser.emitError(loc: typeLoc, message: "illegal storage type prefix" ); |
| 41 | return nullptr; |
| 42 | } |
| 43 | if (identifier.getAsInteger(Radix: 10, Result&: storageTypeWidth)) { |
| 44 | parser.emitError(loc: typeLoc, message: "expected storage type width" ); |
| 45 | return nullptr; |
| 46 | } |
| 47 | isSigned = false; |
| 48 | type = parser.getBuilder().getIntegerType(storageTypeWidth); |
| 49 | } else { |
| 50 | return nullptr; |
| 51 | } |
| 52 | |
| 53 | if (storageTypeWidth == 0 || |
| 54 | storageTypeWidth > QuantizedType::MaxStorageBits) { |
| 55 | parser.emitError(loc: typeLoc, message: "illegal storage type size: " ) |
| 56 | << storageTypeWidth; |
| 57 | return nullptr; |
| 58 | } |
| 59 | |
| 60 | return type; |
| 61 | } |
| 62 | |
| 63 | static ParseResult parseStorageRange(DialectAsmParser &parser, |
| 64 | IntegerType storageType, bool isSigned, |
| 65 | int64_t &storageTypeMin, |
| 66 | int64_t &storageTypeMax) { |
| 67 | int64_t defaultIntegerMin = QuantizedType::getDefaultMinimumForInteger( |
| 68 | isSigned, integralWidth: storageType.getWidth()); |
| 69 | int64_t defaultIntegerMax = QuantizedType::getDefaultMaximumForInteger( |
| 70 | isSigned, integralWidth: storageType.getWidth()); |
| 71 | if (failed(Result: parser.parseOptionalLess())) { |
| 72 | storageTypeMin = defaultIntegerMin; |
| 73 | storageTypeMax = defaultIntegerMax; |
| 74 | return success(); |
| 75 | } |
| 76 | |
| 77 | // Explicit storage min and storage max. |
| 78 | SMLoc minLoc = parser.getCurrentLocation(), maxLoc; |
| 79 | if (parser.parseInteger(result&: storageTypeMin) || parser.parseColon() || |
| 80 | parser.getCurrentLocation(loc: &maxLoc) || |
| 81 | parser.parseInteger(result&: storageTypeMax) || parser.parseGreater()) |
| 82 | return failure(); |
| 83 | if (storageTypeMin < defaultIntegerMin) { |
| 84 | return parser.emitError(loc: minLoc, message: "illegal storage type minimum: " ) |
| 85 | << storageTypeMin; |
| 86 | } |
| 87 | if (storageTypeMax > defaultIntegerMax) { |
| 88 | return parser.emitError(loc: maxLoc, message: "illegal storage type maximum: " ) |
| 89 | << storageTypeMax; |
| 90 | } |
| 91 | return success(); |
| 92 | } |
| 93 | |
| 94 | static FloatType parseExpressedTypeAndRange(DialectAsmParser &parser, |
| 95 | double &min, double &max) { |
| 96 | auto typeLoc = parser.getCurrentLocation(); |
| 97 | FloatType type; |
| 98 | |
| 99 | if (failed(parser.parseType(type))) { |
| 100 | parser.emitError(loc: typeLoc, message: "expecting float expressed type" ); |
| 101 | return nullptr; |
| 102 | } |
| 103 | |
| 104 | // Calibrated min and max values. |
| 105 | if (parser.parseLess() || parser.parseFloat(result&: min) || parser.parseColon() || |
| 106 | parser.parseFloat(result&: max) || parser.parseGreater()) { |
| 107 | parser.emitError(loc: typeLoc, message: "calibrated values must be present" ); |
| 108 | return nullptr; |
| 109 | } |
| 110 | return type; |
| 111 | } |
| 112 | |
| 113 | /// Parses an AnyQuantizedType. |
| 114 | /// |
| 115 | /// any ::= `any<` storage-spec (expressed-type-spec)?`>` |
| 116 | /// storage-spec ::= storage-type (`<` storage-range `>`)? |
| 117 | /// storage-range ::= integer-literal `:` integer-literal |
| 118 | /// storage-type ::= (`i` | `u`) integer-literal |
| 119 | /// expressed-type-spec ::= `:` `f` integer-literal |
| 120 | static Type parseAnyType(DialectAsmParser &parser) { |
| 121 | IntegerType storageType; |
| 122 | FloatType expressedType; |
| 123 | unsigned typeFlags = 0; |
| 124 | int64_t storageTypeMin; |
| 125 | int64_t storageTypeMax; |
| 126 | |
| 127 | // Type specification. |
| 128 | if (parser.parseLess()) |
| 129 | return nullptr; |
| 130 | |
| 131 | // Storage type. |
| 132 | bool isSigned = false; |
| 133 | storageType = parseStorageType(parser, isSigned); |
| 134 | if (!storageType) { |
| 135 | return nullptr; |
| 136 | } |
| 137 | if (isSigned) { |
| 138 | typeFlags |= QuantizationFlags::Signed; |
| 139 | } |
| 140 | |
| 141 | // Storage type range. |
| 142 | if (parseStorageRange(parser, storageType, isSigned, storageTypeMin, |
| 143 | storageTypeMax)) { |
| 144 | return nullptr; |
| 145 | } |
| 146 | |
| 147 | // Optional expressed type. |
| 148 | if (succeeded(Result: parser.parseOptionalColon())) { |
| 149 | if (parser.parseType(expressedType)) { |
| 150 | return nullptr; |
| 151 | } |
| 152 | } |
| 153 | |
| 154 | if (parser.parseGreater()) { |
| 155 | return nullptr; |
| 156 | } |
| 157 | |
| 158 | return parser.getChecked<AnyQuantizedType>( |
| 159 | typeFlags, storageType, expressedType, storageTypeMin, storageTypeMax); |
| 160 | } |
| 161 | |
| 162 | /// Checks if the given scale value is within the valid range of the expressed |
| 163 | /// type. The `expressedType` argument is the floating-point type used for |
| 164 | /// expressing the quantized values, and `scale` is the double value to check. |
| 165 | LogicalResult |
| 166 | isScaleInExpressedTypeRange(function_ref<InFlightDiagnostic()> emitError, |
| 167 | Type expressedType, double scale) { |
| 168 | auto floatType = cast<FloatType>(expressedType); |
| 169 | double minScale = |
| 170 | APFloat::getSmallest(Sem: floatType.getFloatSemantics()).convertToDouble(); |
| 171 | double maxScale = |
| 172 | APFloat::getLargest(Sem: floatType.getFloatSemantics()).convertToDouble(); |
| 173 | if (scale < minScale || scale > maxScale) |
| 174 | return emitError() << "scale " << scale << " out of expressed type range [" |
| 175 | << minScale << ", " << maxScale << "]" ; |
| 176 | return success(); |
| 177 | } |
| 178 | |
| 179 | /// Parses a quantization parameter, which is either a scale value (float) or a |
| 180 | /// scale-zero point pair (float:integer). `expressedType`, expressing the type |
| 181 | /// of scale values, is used to validate the scale. The parsed scale and zero |
| 182 | /// point (if any) are stored in `scale` and `zeroPoint`. |
| 183 | static ParseResult parseQuantParams(DialectAsmParser &parser, |
| 184 | Type expressedType, double &scale, |
| 185 | int64_t &zeroPoint) { |
| 186 | |
| 187 | if (parser.parseFloat(result&: scale)) { |
| 188 | return failure(); |
| 189 | } |
| 190 | |
| 191 | if (failed(Result: isScaleInExpressedTypeRange( |
| 192 | emitError: [&]() { return parser.emitError(loc: parser.getCurrentLocation()); }, |
| 193 | expressedType, scale))) { |
| 194 | return failure(); |
| 195 | } |
| 196 | |
| 197 | zeroPoint = 0; |
| 198 | if (failed(Result: parser.parseOptionalColon())) { |
| 199 | return success(); |
| 200 | } |
| 201 | |
| 202 | return parser.parseInteger(result&: zeroPoint); |
| 203 | } |
| 204 | |
| 205 | /// Parses block size information for sub-channel quantization, assuming the |
| 206 | /// leading '{' has already been parsed. The block size information is provided |
| 207 | /// as a comma-separated list of "Axis:BlockSize" pairs, terminated by a '}'. |
| 208 | /// |
| 209 | /// The parsed axis indices are stored in `quantizedDimensions`, and the |
| 210 | /// corresponding block sizes are stored in `blockSizes`. |
| 211 | static ParseResult |
| 212 | parseBlockSizeInfoUntilRBrace(DialectAsmParser &parser, |
| 213 | SmallVectorImpl<int32_t> &quantizedDimensions, |
| 214 | SmallVectorImpl<int64_t> &blockSizes) { |
| 215 | // Empty block-sizes info. |
| 216 | if (succeeded(Result: parser.parseOptionalRBrace())) { |
| 217 | return success(); |
| 218 | } |
| 219 | |
| 220 | auto parseBlockSizeElements = [&]() -> ParseResult { |
| 221 | quantizedDimensions.resize(N: quantizedDimensions.size() + 1); |
| 222 | blockSizes.resize(N: blockSizes.size() + 1); |
| 223 | if (parser.parseInteger(result&: quantizedDimensions.back()) || |
| 224 | parser.parseColon() || parser.parseInteger(result&: blockSizes.back())) |
| 225 | return failure(); |
| 226 | return success(); |
| 227 | }; |
| 228 | |
| 229 | if (parser.parseCommaSeparatedList(parseElementFn: parseBlockSizeElements) || |
| 230 | parser.parseRBrace()) { |
| 231 | return failure(); |
| 232 | } |
| 233 | |
| 234 | return success(); |
| 235 | } |
| 236 | |
| 237 | /// Parses a bracketed list of quantization parameters, returning the dimensions |
| 238 | /// of the parsed sub-tensors in `dims`. The dimension of the list is prepended |
| 239 | /// to the dimensions of the sub-tensors. This function assumes that the initial |
| 240 | /// left brace has already been parsed. For example: |
| 241 | /// |
| 242 | /// parseQuantParamListUntilRBrace(1.0:1, 2.0:4, 3.0:4}) -> Success, |
| 243 | /// dims = [3], scales = [1.0, 2.0, 3.0], zeroPoints = [1, 4, 4] |
| 244 | /// |
| 245 | /// parseQuantParamListUntilRBrace({1.0, 2.0}, {3.0:1, 4.0:9}}) -> Success, |
| 246 | /// dims = [2, 2], scales = [1.0, 2.0, 3.0, 4.0], zeroPoints = [0, 0, 1, |
| 247 | /// 9] |
| 248 | /// |
| 249 | /// This function expects all sub-tensors to have the same rank. |
| 250 | static ParseResult |
| 251 | parseQuantParamListUntilRBrace(DialectAsmParser &parser, Type expressedType, |
| 252 | SmallVectorImpl<double> &scales, |
| 253 | SmallVectorImpl<int64_t> &zeroPoints, |
| 254 | SmallVectorImpl<int64_t> &dims) { |
| 255 | auto checkDims = [&](const SmallVectorImpl<int64_t> &prevDims, |
| 256 | const SmallVectorImpl<int64_t> &newDims) -> ParseResult { |
| 257 | if (prevDims == newDims) |
| 258 | return success(); |
| 259 | return parser.emitError(loc: parser.getCurrentLocation()) |
| 260 | << "tensor literal is invalid; ranks are not consistent " |
| 261 | "between elements" ; |
| 262 | }; |
| 263 | |
| 264 | bool first = true; |
| 265 | SmallVector<int64_t, 4> newDims; |
| 266 | unsigned size = 0; |
| 267 | |
| 268 | auto parseOneElement = [&]() -> ParseResult { |
| 269 | SmallVector<int64_t, 4> thisDims; |
| 270 | if (succeeded(Result: parser.parseOptionalLBrace())) { |
| 271 | if (parseQuantParamListUntilRBrace(parser, expressedType, scales, |
| 272 | zeroPoints, dims&: thisDims)) |
| 273 | return failure(); |
| 274 | } else { |
| 275 | zeroPoints.resize(N: zeroPoints.size() + 1); |
| 276 | scales.resize(N: scales.size() + 1); |
| 277 | if (parseQuantParams(parser, expressedType, scale&: scales.back(), |
| 278 | zeroPoint&: zeroPoints.back())) { |
| 279 | return failure(); |
| 280 | } |
| 281 | } |
| 282 | ++size; |
| 283 | if (!first) |
| 284 | return checkDims(newDims, thisDims); |
| 285 | newDims = thisDims; |
| 286 | first = false; |
| 287 | return success(); |
| 288 | }; |
| 289 | |
| 290 | if (parser.parseCommaSeparatedList(parseElementFn: parseOneElement) || parser.parseRBrace()) { |
| 291 | return failure(); |
| 292 | } |
| 293 | |
| 294 | // Return the sublists' dimensions with 'size' prepended. |
| 295 | dims.clear(); |
| 296 | dims.push_back(Elt: size); |
| 297 | dims.append(in_start: newDims.begin(), in_end: newDims.end()); |
| 298 | |
| 299 | return success(); |
| 300 | } |
| 301 | |
| 302 | /// Parses a UniformQuantizedType. |
| 303 | /// |
| 304 | /// uniform_type ::= uniform_per_layer |
| 305 | /// | uniform_per_axis |
| 306 | /// | uniform_sub_channel |
| 307 | /// uniform_per_layer ::= `uniform<` storage-spec expressed-type-spec |
| 308 | /// `,` scale-zero `>` |
| 309 | /// uniform_per_axis ::= `uniform<` storage-spec expressed-type-spec |
| 310 | /// axis-spec `,` `{` scale-zero-list `}` `>` |
| 311 | /// uniform_sub_channel ::= `uniform<` storage-spec expressed-type-spec |
| 312 | /// block-size-info `,` scale-zero-tensor `>` |
| 313 | /// storage-spec ::= storage-type (`<` storage-range `>`)? |
| 314 | /// storage-range ::= integer-literal `:` integer-literal |
| 315 | /// storage-type ::= (`i` | `u`) integer-literal |
| 316 | /// expressed-type-spec ::= `:` `f` integer-literal |
| 317 | /// axis-spec ::= `:` integer-literal |
| 318 | /// scale-zero ::= scale (`:` zero-point)? |
| 319 | /// scale ::= float-literal |
| 320 | /// zero-point ::= integer-literal |
| 321 | /// scale-zero-list ::= scale-zero (`,` scale-zero)* |
| 322 | /// block-size-info ::= `{` `}` | `{` axis-block `:` (`,` axis-block)* `}` |
| 323 | /// axis-block ::= axis-spec `:` block-size-spec |
| 324 | /// block-size-spec ::= integer-literal |
| 325 | /// scale-zero-tensor ::= scale-zero-dense-exp | scale-zero-list |
| 326 | /// scale-zero-dense-exp ::= `{` |
| 327 | /// scale-zero-tensor (`,` scale-zero-tensor)* |
| 328 | /// `}` |
| 329 | static Type parseUniformType(DialectAsmParser &parser) { |
| 330 | IntegerType storageType; |
| 331 | FloatType expressedType; |
| 332 | unsigned typeFlags = 0; |
| 333 | int64_t storageTypeMin; |
| 334 | int64_t storageTypeMax; |
| 335 | bool isPerAxis = false; |
| 336 | bool isSubChannel = false; |
| 337 | SmallVector<int32_t, 1> quantizedDimensions; |
| 338 | SmallVector<int64_t, 1> blockSizes; |
| 339 | SmallVector<double, 1> scales; |
| 340 | SmallVector<int64_t, 1> zeroPoints; |
| 341 | |
| 342 | // Type specification. |
| 343 | if (parser.parseLess()) { |
| 344 | return nullptr; |
| 345 | } |
| 346 | |
| 347 | // Storage type. |
| 348 | bool isSigned = false; |
| 349 | storageType = parseStorageType(parser, isSigned); |
| 350 | if (!storageType) { |
| 351 | return nullptr; |
| 352 | } |
| 353 | if (isSigned) { |
| 354 | typeFlags |= QuantizationFlags::Signed; |
| 355 | } |
| 356 | |
| 357 | // Storage type range. |
| 358 | if (parseStorageRange(parser, storageType, isSigned, storageTypeMin, |
| 359 | storageTypeMax)) { |
| 360 | return nullptr; |
| 361 | } |
| 362 | |
| 363 | // Expressed type. |
| 364 | if (parser.parseColon() || parser.parseType(expressedType)) { |
| 365 | return nullptr; |
| 366 | } |
| 367 | |
| 368 | // Optionally parse quantized dimension for per-axis or sub-channel |
| 369 | // quantization. |
| 370 | if (succeeded(Result: parser.parseOptionalColon())) { |
| 371 | if (succeeded(Result: parser.parseOptionalLBrace())) { |
| 372 | isSubChannel = true; |
| 373 | if (parseBlockSizeInfoUntilRBrace(parser, quantizedDimensions, |
| 374 | blockSizes)) { |
| 375 | return nullptr; |
| 376 | } |
| 377 | } else { |
| 378 | isPerAxis = true; |
| 379 | quantizedDimensions.resize(N: 1); |
| 380 | if (parser.parseInteger(result&: quantizedDimensions.back())) { |
| 381 | return nullptr; |
| 382 | } |
| 383 | } |
| 384 | } |
| 385 | |
| 386 | // Comma leading into range_spec. |
| 387 | if (parser.parseComma()) { |
| 388 | return nullptr; |
| 389 | } |
| 390 | |
| 391 | // Quantization parameter (scales/zeroPoints) specification. |
| 392 | bool isPerTensor = !isPerAxis && !isSubChannel; |
| 393 | SmallVector<int64_t> dims; |
| 394 | if (isPerTensor) { |
| 395 | zeroPoints.resize(N: zeroPoints.size() + 1); |
| 396 | scales.resize(N: scales.size() + 1); |
| 397 | if (parseQuantParams(parser, expressedType, scales.back(), |
| 398 | zeroPoints.back())) { |
| 399 | return nullptr; |
| 400 | } |
| 401 | |
| 402 | } else { |
| 403 | if (parser.parseLBrace() || |
| 404 | parseQuantParamListUntilRBrace(parser, expressedType, scales, |
| 405 | zeroPoints, dims)) { |
| 406 | return nullptr; |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | if (parser.parseGreater()) { |
| 411 | return nullptr; |
| 412 | } |
| 413 | |
| 414 | if (isPerAxis) { |
| 415 | return parser.getChecked<UniformQuantizedPerAxisType>( |
| 416 | typeFlags, storageType, expressedType, scales, zeroPoints, |
| 417 | quantizedDimensions[0], storageTypeMin, storageTypeMax); |
| 418 | } else if (isSubChannel) { |
| 419 | SmallVector<APFloat> apFloatScales = |
| 420 | llvm::to_vector(Range: llvm::map_range(C&: scales, F: [&](double scale) -> APFloat { |
| 421 | APFloat apFloatScale(scale); |
| 422 | bool unused; |
| 423 | apFloatScale.convert(ToSemantics: expressedType.getFloatSemantics(), |
| 424 | RM: APFloat::rmNearestTiesToEven, losesInfo: &unused); |
| 425 | return apFloatScale; |
| 426 | })); |
| 427 | SmallVector<APInt> apIntZeroPoints = llvm::to_vector( |
| 428 | Range: llvm::map_range(C&: zeroPoints, F: [&](int64_t zeroPoint) -> APInt { |
| 429 | return APInt(storageType.getIntOrFloatBitWidth(), zeroPoint); |
| 430 | })); |
| 431 | auto scalesRef = mlir::DenseElementsAttr::get( |
| 432 | RankedTensorType::get(dims, expressedType), apFloatScales); |
| 433 | auto zeroPointsRef = mlir::DenseElementsAttr::get( |
| 434 | RankedTensorType::get(dims, storageType), apIntZeroPoints); |
| 435 | return parser.getChecked<UniformQuantizedSubChannelType>( |
| 436 | typeFlags, storageType, expressedType, scalesRef, zeroPointsRef, |
| 437 | quantizedDimensions, blockSizes, storageTypeMin, storageTypeMax); |
| 438 | } |
| 439 | |
| 440 | return parser.getChecked<UniformQuantizedType>( |
| 441 | typeFlags, storageType, expressedType, scales.front(), zeroPoints.front(), |
| 442 | storageTypeMin, storageTypeMax); |
| 443 | } |
| 444 | |
| 445 | /// Parses an CalibratedQuantizedType. |
| 446 | /// |
| 447 | /// calibrated ::= `calibrated<` expressed-spec `>` |
| 448 | /// expressed-spec ::= expressed-type `<` calibrated-range `>` |
| 449 | /// expressed-type ::= `f` integer-literal |
| 450 | /// calibrated-range ::= float-literal `:` float-literal |
| 451 | static Type parseCalibratedType(DialectAsmParser &parser) { |
| 452 | FloatType expressedType; |
| 453 | double min; |
| 454 | double max; |
| 455 | |
| 456 | // Type specification. |
| 457 | if (parser.parseLess()) |
| 458 | return nullptr; |
| 459 | |
| 460 | // Expressed type. |
| 461 | expressedType = parseExpressedTypeAndRange(parser, min, max); |
| 462 | if (!expressedType) { |
| 463 | return nullptr; |
| 464 | } |
| 465 | |
| 466 | if (parser.parseGreater()) { |
| 467 | return nullptr; |
| 468 | } |
| 469 | |
| 470 | return parser.getChecked<CalibratedQuantizedType>(expressedType, min, max); |
| 471 | } |
| 472 | |
| 473 | /// Parse a type registered to this dialect. |
| 474 | Type QuantDialect::parseType(DialectAsmParser &parser) const { |
| 475 | // All types start with an identifier that we switch on. |
| 476 | StringRef typeNameSpelling; |
| 477 | if (failed(parser.parseKeyword(&typeNameSpelling))) |
| 478 | return nullptr; |
| 479 | |
| 480 | if (typeNameSpelling == "uniform" ) |
| 481 | return parseUniformType(parser); |
| 482 | if (typeNameSpelling == "any" ) |
| 483 | return parseAnyType(parser); |
| 484 | if (typeNameSpelling == "calibrated" ) |
| 485 | return parseCalibratedType(parser); |
| 486 | |
| 487 | parser.emitError(parser.getNameLoc(), |
| 488 | "unknown quantized type " + typeNameSpelling); |
| 489 | return nullptr; |
| 490 | } |
| 491 | |
| 492 | static void printStorageType(QuantizedType type, DialectAsmPrinter &out) { |
| 493 | // storage type |
| 494 | unsigned storageWidth = type.getStorageTypeIntegralWidth(); |
| 495 | bool isSigned = type.isSigned(); |
| 496 | if (isSigned) { |
| 497 | out << "i" << storageWidth; |
| 498 | } else { |
| 499 | out << "u" << storageWidth; |
| 500 | } |
| 501 | |
| 502 | // storageTypeMin and storageTypeMax if not default. |
| 503 | if (type.hasStorageTypeBounds()) { |
| 504 | out << "<" << type.getStorageTypeMin() << ":" << type.getStorageTypeMax() |
| 505 | << ">" ; |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | static void printQuantParams(double scale, int64_t zeroPoint, |
| 510 | DialectAsmPrinter &out) { |
| 511 | out << scale; |
| 512 | if (zeroPoint != 0) { |
| 513 | out << ":" << zeroPoint; |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | static void |
| 518 | printBlockSizeInfo(ArrayRef<std::pair<int32_t, int64_t>> blockSizeInfo, |
| 519 | DialectAsmPrinter &out) { |
| 520 | out << "{" ; |
| 521 | llvm::interleaveComma( |
| 522 | c: llvm::seq<size_t>(Begin: 0, End: blockSizeInfo.size()), os&: out, each_fn: [&](size_t index) { |
| 523 | out << blockSizeInfo[index].first << ":" << blockSizeInfo[index].second; |
| 524 | }); |
| 525 | out << "}" ; |
| 526 | } |
| 527 | |
| 528 | /// Helper that prints a AnyQuantizedType. |
| 529 | static void printAnyQuantizedType(AnyQuantizedType type, |
| 530 | DialectAsmPrinter &out) { |
| 531 | out << "any<" ; |
| 532 | printStorageType(type, out); |
| 533 | if (Type expressedType = type.getExpressedType()) { |
| 534 | out << ":" << expressedType; |
| 535 | } |
| 536 | out << ">" ; |
| 537 | } |
| 538 | |
| 539 | /// Helper that prints a UniformQuantizedType. |
| 540 | static void printUniformQuantizedType(UniformQuantizedType type, |
| 541 | DialectAsmPrinter &out) { |
| 542 | out << "uniform<" ; |
| 543 | printStorageType(type, out); |
| 544 | out << ":" << type.getExpressedType() << ", " ; |
| 545 | |
| 546 | // scheme specific parameters |
| 547 | printQuantParams(scale: type.getScale(), zeroPoint: type.getZeroPoint(), out); |
| 548 | out << ">" ; |
| 549 | } |
| 550 | |
| 551 | /// Helper that prints a UniformQuantizedPerAxisType. |
| 552 | static void printUniformQuantizedPerAxisType(UniformQuantizedPerAxisType type, |
| 553 | DialectAsmPrinter &out) { |
| 554 | out << "uniform<" ; |
| 555 | printStorageType(type, out); |
| 556 | out << ":" << type.getExpressedType() << ":" ; |
| 557 | out << type.getQuantizedDimension(); |
| 558 | out << ", " ; |
| 559 | |
| 560 | // scheme specific parameters |
| 561 | ArrayRef<double> scales = type.getScales(); |
| 562 | ArrayRef<int64_t> zeroPoints = type.getZeroPoints(); |
| 563 | out << "{" ; |
| 564 | llvm::interleave( |
| 565 | c: llvm::seq<size_t>(Begin: 0, End: scales.size()), os&: out, |
| 566 | each_fn: [&](size_t index) { |
| 567 | printQuantParams(scale: scales[index], zeroPoint: zeroPoints[index], out); |
| 568 | }, |
| 569 | separator: "," ); |
| 570 | out << "}>" ; |
| 571 | } |
| 572 | |
| 573 | /// Prints quantization parameters as a nested list of `scale`[:`zero_point`] |
| 574 | /// elements. The nesting corresponds to the `shape` dimensions. |
| 575 | /// |
| 576 | /// Elements are delimited by commas, and the inner dimensions are enclosed in |
| 577 | /// braces. `zero_point` is only printed if it is non-zero. For example: |
| 578 | /// |
| 579 | /// printDenseQuantizationParameters(scales=[1.0, 2.0, 3.0, 4.0], |
| 580 | /// zeroPoints=[0, 0, 1, 9], |
| 581 | /// shape=[2, 2]) |
| 582 | /// |
| 583 | /// would print: |
| 584 | /// |
| 585 | /// {{1.0, 2.0}, {3.0:1, 4.0:9}} |
| 586 | void printDenseQuantizationParameters(ArrayRef<APFloat> scales, |
| 587 | ArrayRef<APInt> zeroPoints, |
| 588 | ArrayRef<int64_t> shape, |
| 589 | DialectAsmPrinter &out) { |
| 590 | int64_t rank = shape.size(); |
| 591 | SmallVector<unsigned, 4> counter(rank, 0); |
| 592 | unsigned openBrackets = 0; |
| 593 | |
| 594 | auto incrementCounterAndDelimit = [&]() { |
| 595 | ++counter[rank - 1]; |
| 596 | for (unsigned i = rank - 1; i > 0; --i) { |
| 597 | if (counter[i] >= shape[i]) { |
| 598 | counter[i] = 0; |
| 599 | ++counter[i - 1]; |
| 600 | --openBrackets; |
| 601 | out << '}'; |
| 602 | } |
| 603 | } |
| 604 | }; |
| 605 | |
| 606 | for (unsigned idx = 0, e = scales.size(); idx < e; ++idx) { |
| 607 | if (idx != 0) |
| 608 | out << ", " ; |
| 609 | while (openBrackets++ < rank) |
| 610 | out << '{'; |
| 611 | openBrackets = rank; |
| 612 | out << scales[idx]; |
| 613 | if (zeroPoints[idx] != 0) { |
| 614 | out << ":" << zeroPoints[idx]; |
| 615 | } |
| 616 | incrementCounterAndDelimit(); |
| 617 | } |
| 618 | while (openBrackets-- > 0) |
| 619 | out << '}'; |
| 620 | } |
| 621 | |
| 622 | /// Helper that prints a UniformQuantizedSubChannelType. |
| 623 | static void |
| 624 | printUniformQuantizedSubChannelType(UniformQuantizedSubChannelType type, |
| 625 | DialectAsmPrinter &out) { |
| 626 | out << "uniform<" ; |
| 627 | printStorageType(type, out); |
| 628 | out << ":" << type.getExpressedType() << ":" ; |
| 629 | printBlockSizeInfo(blockSizeInfo: type.getBlockSizeInfo(), out); |
| 630 | out << ", " ; |
| 631 | |
| 632 | auto scalesItr = type.getScales().getValues<APFloat>(); |
| 633 | auto zeroPointsItr = type.getZeroPoints().getValues<APInt>(); |
| 634 | SmallVector<APFloat> scales(scalesItr.begin(), scalesItr.end()); |
| 635 | SmallVector<APInt> zeroPoints(zeroPointsItr.begin(), zeroPointsItr.end()); |
| 636 | printDenseQuantizationParameters(scales, zeroPoints, |
| 637 | type.getScales().getType().getShape(), out); |
| 638 | out << ">" ; |
| 639 | } |
| 640 | |
| 641 | /// Helper that prints a CalibratedQuantizedType. |
| 642 | static void printCalibratedQuantizedType(CalibratedQuantizedType type, |
| 643 | DialectAsmPrinter &out) { |
| 644 | out << "calibrated<" << type.getExpressedType(); |
| 645 | out << "<" << type.getMin() << ":" << type.getMax() << ">" ; |
| 646 | out << ">" ; |
| 647 | } |
| 648 | |
| 649 | /// Print a type registered to this dialect. |
| 650 | void QuantDialect::printType(Type type, DialectAsmPrinter &os) const { |
| 651 | if (auto anyType = llvm::dyn_cast<AnyQuantizedType>(type)) |
| 652 | printAnyQuantizedType(anyType, os); |
| 653 | else if (auto uniformType = llvm::dyn_cast<UniformQuantizedType>(type)) |
| 654 | printUniformQuantizedType(uniformType, os); |
| 655 | else if (auto perAxisType = llvm::dyn_cast<UniformQuantizedPerAxisType>(type)) |
| 656 | printUniformQuantizedPerAxisType(perAxisType, os); |
| 657 | else if (auto perAxisType = |
| 658 | llvm::dyn_cast<UniformQuantizedSubChannelType>(type)) |
| 659 | printUniformQuantizedSubChannelType(perAxisType, os); |
| 660 | else if (auto calibratedType = llvm::dyn_cast<CalibratedQuantizedType>(type)) |
| 661 | printCalibratedQuantizedType(calibratedType, os); |
| 662 | else |
| 663 | llvm_unreachable("Unhandled quantized type" ); |
| 664 | } |
| 665 | |