| 1 | //===- ArithToAMDGPU.cpp - Arith to AMDGPU dialect conversion ---------===// |
| 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/Conversion/ArithToAMDGPU/ArithToAMDGPU.h" |
| 10 | |
| 11 | #include "mlir/Dialect/AMDGPU/IR/AMDGPUDialect.h" |
| 12 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 13 | #include "mlir/Dialect/Arith/Utils/Utils.h" |
| 14 | #include "mlir/Dialect/LLVMIR/LLVMConstantOps.h" |
| 15 | #include "mlir/Dialect/LLVMIR/ROCDLDialect.h" |
| 16 | #include "mlir/Dialect/LLVMIR/ROCDLTargetInfo.h" |
| 17 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 18 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 19 | #include "mlir/Dialect/Vector/Utils/VectorUtils.h" |
| 20 | #include "mlir/IR/BuiltinTypes.h" |
| 21 | #include "mlir/IR/PatternMatch.h" |
| 22 | #include "mlir/IR/TypeUtilities.h" |
| 23 | #include "mlir/Pass/Pass.h" |
| 24 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| 25 | |
| 26 | namespace mlir { |
| 27 | #define GEN_PASS_DEF_ARITHTOAMDGPUCONVERSIONPASS |
| 28 | #include "mlir/Conversion/Passes.h.inc" |
| 29 | } // namespace mlir |
| 30 | |
| 31 | using namespace mlir; |
| 32 | using namespace mlir::amdgpu; |
| 33 | |
| 34 | namespace { |
| 35 | struct ArithToAMDGPUConversionPass final |
| 36 | : impl::ArithToAMDGPUConversionPassBase<ArithToAMDGPUConversionPass> { |
| 37 | using impl::ArithToAMDGPUConversionPassBase< |
| 38 | ArithToAMDGPUConversionPass>::ArithToAMDGPUConversionPassBase; |
| 39 | |
| 40 | void runOnOperation() override; |
| 41 | }; |
| 42 | |
| 43 | struct ExtFOnFloat8RewritePattern final : OpRewritePattern<arith::ExtFOp> { |
| 44 | using Base::Base; |
| 45 | |
| 46 | ROCDL::TargetInfo target; |
| 47 | ExtFOnFloat8RewritePattern(MLIRContext *ctx, const ROCDL::TargetInfo &target, |
| 48 | PatternBenefit benefit) |
| 49 | : OpRewritePattern::OpRewritePattern(ctx, benefit), target(target) {} |
| 50 | |
| 51 | LogicalResult matchAndRewrite(arith::ExtFOp op, |
| 52 | PatternRewriter &rewriter) const override; |
| 53 | }; |
| 54 | |
| 55 | struct TruncFToFloat8RewritePattern final : OpRewritePattern<arith::TruncFOp> { |
| 56 | bool saturateFP8 = false; |
| 57 | TruncFToFloat8RewritePattern(MLIRContext *ctx, bool saturateFP8, |
| 58 | const ROCDL::TargetInfo &target, |
| 59 | PatternBenefit benefit) |
| 60 | : OpRewritePattern::OpRewritePattern(ctx, benefit), |
| 61 | saturateFP8(saturateFP8), target(target) {} |
| 62 | ROCDL::TargetInfo target; |
| 63 | |
| 64 | LogicalResult matchAndRewrite(arith::TruncFOp op, |
| 65 | PatternRewriter &rewriter) const override; |
| 66 | }; |
| 67 | |
| 68 | struct TruncfToFloat16RewritePattern final |
| 69 | : public OpRewritePattern<arith::TruncFOp> { |
| 70 | |
| 71 | using Base::Base; |
| 72 | |
| 73 | LogicalResult matchAndRewrite(arith::TruncFOp op, |
| 74 | PatternRewriter &rewriter) const override; |
| 75 | }; |
| 76 | |
| 77 | struct ScalingExtFRewritePattern final |
| 78 | : OpRewritePattern<arith::ScalingExtFOp> { |
| 79 | using Base::Base; |
| 80 | |
| 81 | LogicalResult matchAndRewrite(arith::ScalingExtFOp op, |
| 82 | PatternRewriter &rewriter) const override; |
| 83 | }; |
| 84 | |
| 85 | struct ScalingTruncFRewritePattern final |
| 86 | : OpRewritePattern<arith::ScalingTruncFOp> { |
| 87 | using Base::Base; |
| 88 | |
| 89 | LogicalResult matchAndRewrite(arith::ScalingTruncFOp op, |
| 90 | PatternRewriter &rewriter) const override; |
| 91 | }; |
| 92 | } // end namespace |
| 93 | |
| 94 | static bool isSupportedF8(Type elementType, const ROCDL::TargetInfo &target) { |
| 95 | if (target.hasFnuzFp8()) |
| 96 | return isa<Float8E4M3FNUZType, Float8E5M2FNUZType>(Val: elementType); |
| 97 | if (target.hasOcpFp8()) |
| 98 | return isa<Float8E4M3FNType, Float8E5M2Type>(Val: elementType); |
| 99 | return false; |
| 100 | } |
| 101 | |
| 102 | static Value castF32To(Type desType, Value f32, Location loc, |
| 103 | PatternRewriter &rewriter) { |
| 104 | Type elementType = getElementTypeOrSelf(type: desType); |
| 105 | if (elementType.isF32()) |
| 106 | return f32; |
| 107 | if (elementType.getIntOrFloatBitWidth() < 32) |
| 108 | return arith::TruncFOp::create(builder&: rewriter, location: loc, out: desType, in: f32); |
| 109 | if (elementType.getIntOrFloatBitWidth() > 32) |
| 110 | return arith::ExtFOp::create(builder&: rewriter, location: loc, resultTypes: TypeRange{desType}, |
| 111 | operands: ValueRange{f32}, properties: arith::ExtFOp::Properties{}); |
| 112 | llvm_unreachable("The only 32-bit float type is f32" ); |
| 113 | } |
| 114 | |
| 115 | LogicalResult |
| 116 | ExtFOnFloat8RewritePattern::matchAndRewrite(arith::ExtFOp op, |
| 117 | PatternRewriter &rewriter) const { |
| 118 | Type inType = op.getIn().getType(); |
| 119 | auto inVecType = dyn_cast<VectorType>(Val&: inType); |
| 120 | if (inVecType) { |
| 121 | if (inVecType.isScalable()) |
| 122 | return failure(); |
| 123 | inType = inVecType.getElementType(); |
| 124 | } |
| 125 | if (!isSupportedF8(elementType: inType, target)) |
| 126 | return failure(); |
| 127 | |
| 128 | Location loc = op.getLoc(); |
| 129 | Value in = op.getIn(); |
| 130 | Type outElemType = getElementTypeOrSelf(type: op.getOut().getType()); |
| 131 | VectorType extResType = VectorType::get(shape: 2, elementType: rewriter.getF32Type()); |
| 132 | if (!inVecType) { |
| 133 | Value asFloat = amdgpu::ExtPackedFp8Op::create( |
| 134 | builder&: rewriter, location: loc, res: rewriter.getF32Type(), source: in, index: 0); |
| 135 | Value result = castF32To(desType: outElemType, f32: asFloat, loc, rewriter); |
| 136 | rewriter.replaceOp(op, newValues: result); |
| 137 | return success(); |
| 138 | } |
| 139 | int64_t numElements = inVecType.getNumElements(); |
| 140 | |
| 141 | Value zero = arith::ConstantOp::create( |
| 142 | builder&: rewriter, location: loc, result: outElemType, value: rewriter.getFloatAttr(type: outElemType, value: 0.0)); |
| 143 | VectorType outType = cast<VectorType>(Val: op.getOut().getType()); |
| 144 | |
| 145 | if (inVecType.getShape().empty()) { |
| 146 | Value zerodSplat = |
| 147 | rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: outType, args&: zero); |
| 148 | Value scalarIn = |
| 149 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: in, position: ArrayRef<int64_t>{}); |
| 150 | Value scalarExt = arith::ExtFOp::create( |
| 151 | builder&: rewriter, location: loc, resultTypes: TypeRange{outElemType}, operands: ValueRange{scalarIn}, |
| 152 | properties: arith::ExtFOp::Properties{}); |
| 153 | Value result = vector::InsertOp::create(builder&: rewriter, location: loc, valueToStore: scalarExt, |
| 154 | dest: zerodSplat, position: ArrayRef<int64_t>{}); |
| 155 | rewriter.replaceOp(op, newValues: result); |
| 156 | return success(); |
| 157 | } |
| 158 | |
| 159 | VectorType flatTy = VectorType::get(shape: SmallVector<int64_t>{numElements}, |
| 160 | elementType: outType.getElementType()); |
| 161 | Value result = rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: flatTy, args&: zero); |
| 162 | |
| 163 | if (inVecType.getRank() > 1) { |
| 164 | inVecType = VectorType::get(shape: SmallVector<int64_t>{numElements}, |
| 165 | elementType: inVecType.getElementType()); |
| 166 | in = vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: inVecType, source: in); |
| 167 | } |
| 168 | |
| 169 | for (int64_t i = 0; i < numElements; i += 4) { |
| 170 | int64_t elemsThisOp = std::min(a: numElements, b: i + 4) - i; |
| 171 | Value inSlice = vector::ExtractStridedSliceOp::create(builder&: rewriter, location: loc, source: in, offsets: i, |
| 172 | sizes: elemsThisOp, strides: 1); |
| 173 | for (int64_t j = 0; j < elemsThisOp; j += 2) { |
| 174 | if (i + j + 1 < numElements) { // Convert two 8-bit elements |
| 175 | Value asFloats = amdgpu::ExtPackedFp8Op::create( |
| 176 | builder&: rewriter, location: loc, res: extResType, source: inSlice, index: j / 2); |
| 177 | Type desType = VectorType::get(shape: 2, elementType: outElemType); |
| 178 | Value asType = castF32To(desType, f32: asFloats, loc, rewriter); |
| 179 | result = vector::InsertStridedSliceOp::create(builder&: rewriter, location: loc, valueToStore: asType, |
| 180 | dest: result, offsets: i + j, strides: 1); |
| 181 | } else { // Convert a 8-bit element |
| 182 | Value asFloat = amdgpu::ExtPackedFp8Op::create( |
| 183 | builder&: rewriter, location: loc, res: rewriter.getF32Type(), source: inSlice, index: j / 2 * 2); |
| 184 | Value asType = castF32To(desType: outElemType, f32: asFloat, loc, rewriter); |
| 185 | result = vector::InsertOp::create(builder&: rewriter, location: loc, valueToStore: asType, dest: result, position: i + j); |
| 186 | } |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | if (inVecType.getRank() != outType.getRank()) { |
| 191 | result = vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: outType, source: result); |
| 192 | } |
| 193 | |
| 194 | rewriter.replaceOp(op, newValues: result); |
| 195 | return success(); |
| 196 | } |
| 197 | |
| 198 | static Value castToF32(Value value, Location loc, PatternRewriter &rewriter) { |
| 199 | Type type = value.getType(); |
| 200 | if (type.isF32()) |
| 201 | return value; |
| 202 | if (type.getIntOrFloatBitWidth() < 32) |
| 203 | return arith::ExtFOp::create( |
| 204 | builder&: rewriter, location: loc, resultTypes: TypeRange{rewriter.getF32Type()}, operands: ValueRange{value}, |
| 205 | properties: arith::ExtFOp::Properties{}); |
| 206 | if (type.getIntOrFloatBitWidth() > 32) |
| 207 | return arith::TruncFOp::create(builder&: rewriter, location: loc, out: rewriter.getF32Type(), in: value); |
| 208 | llvm_unreachable("The only 32-bit float type is f32" ); |
| 209 | } |
| 210 | |
| 211 | // If `in` is a finite value, clamp it between the maximum and minimum values |
| 212 | // of `outElemType` so that subsequent conversion instructions don't |
| 213 | // overflow those out-of-range values to NaN. These semantics are commonly |
| 214 | // used in machine-learning contexts where failure to clamp would lead to |
| 215 | // excessive NaN production. |
| 216 | static Value clampInput(PatternRewriter &rewriter, Location loc, |
| 217 | Type outElemType, Value source) { |
| 218 | Type sourceType = source.getType(); |
| 219 | const llvm::fltSemantics &sourceSem = |
| 220 | cast<FloatType>(Val: getElementTypeOrSelf(type: sourceType)).getFloatSemantics(); |
| 221 | const llvm::fltSemantics &targetSem = |
| 222 | cast<FloatType>(Val&: outElemType).getFloatSemantics(); |
| 223 | |
| 224 | APFloat min = APFloat::getLargest(Sem: targetSem, /*Negative=*/true); |
| 225 | APFloat max = APFloat::getLargest(Sem: targetSem, /*Negative=*/false); |
| 226 | bool ignoredLosesInfo = false; |
| 227 | // We can ignore conversion failures here because this conversion promotes |
| 228 | // from a smaller type to a larger one - ex. there can be no loss of precision |
| 229 | // when casting fp8 to f16. |
| 230 | (void)min.convert(ToSemantics: sourceSem, RM: APFloat::rmNearestTiesToEven, losesInfo: &ignoredLosesInfo); |
| 231 | (void)max.convert(ToSemantics: sourceSem, RM: APFloat::rmNearestTiesToEven, losesInfo: &ignoredLosesInfo); |
| 232 | |
| 233 | Value minCst = createScalarOrSplatConstant(builder&: rewriter, loc, type: sourceType, value: min); |
| 234 | Value maxCst = createScalarOrSplatConstant(builder&: rewriter, loc, type: sourceType, value: max); |
| 235 | |
| 236 | Value inf = createScalarOrSplatConstant( |
| 237 | builder&: rewriter, loc, type: sourceType, |
| 238 | value: APFloat::getInf(Sem: sourceSem, /*Negative=*/false)); |
| 239 | Value negInf = createScalarOrSplatConstant( |
| 240 | builder&: rewriter, loc, type: sourceType, value: APFloat::getInf(Sem: sourceSem, /*Negative=*/true)); |
| 241 | Value isInf = rewriter.createOrFold<arith::CmpFOp>( |
| 242 | location: loc, args: arith::CmpFPredicate::OEQ, args&: source, args&: inf); |
| 243 | Value isNegInf = rewriter.createOrFold<arith::CmpFOp>( |
| 244 | location: loc, args: arith::CmpFPredicate::OEQ, args&: source, args&: negInf); |
| 245 | Value isNan = rewriter.createOrFold<arith::CmpFOp>( |
| 246 | location: loc, args: arith::CmpFPredicate::UNO, args&: source, args&: source); |
| 247 | Value isNonFinite = arith::OrIOp::create( |
| 248 | builder&: rewriter, location: loc, lhs: arith::OrIOp::create(builder&: rewriter, location: loc, lhs: isInf, rhs: isNegInf), |
| 249 | rhs: isNan); |
| 250 | |
| 251 | Value clampedBelow = arith::MaximumFOp::create(builder&: rewriter, location: loc, lhs: source, rhs: minCst); |
| 252 | Value clamped = |
| 253 | arith::MinimumFOp::create(builder&: rewriter, location: loc, lhs: clampedBelow, rhs: maxCst); |
| 254 | Value res = |
| 255 | arith::SelectOp::create(builder&: rewriter, location: loc, condition: isNonFinite, true_value: source, false_value: clamped); |
| 256 | return res; |
| 257 | } |
| 258 | |
| 259 | LogicalResult |
| 260 | TruncFToFloat8RewritePattern::matchAndRewrite(arith::TruncFOp op, |
| 261 | PatternRewriter &rewriter) const { |
| 262 | // Only supporting default rounding mode as of now. |
| 263 | if (op.getRoundingmodeAttr()) |
| 264 | return failure(); |
| 265 | Type outType = op.getOut().getType(); |
| 266 | auto outVecType = dyn_cast<VectorType>(Val&: outType); |
| 267 | if (outVecType) { |
| 268 | if (outVecType.isScalable()) |
| 269 | return failure(); |
| 270 | outType = outVecType.getElementType(); |
| 271 | } |
| 272 | auto inType = dyn_cast<FloatType>(Val: getElementTypeOrSelf(type: op.getIn().getType())); |
| 273 | if (inType && inType.getWidth() <= 8 && saturateFP8) |
| 274 | // Conversion between 8-bit floats is not supported with truncation enabled. |
| 275 | return failure(); |
| 276 | |
| 277 | if (!isSupportedF8(elementType: outType, target)) |
| 278 | return failure(); |
| 279 | |
| 280 | Location loc = op.getLoc(); |
| 281 | Value in = op.getIn(); |
| 282 | Type outElemType = getElementTypeOrSelf(type: op.getOut().getType()); |
| 283 | if (saturateFP8) |
| 284 | in = clampInput(rewriter, loc, outElemType, source: in); |
| 285 | auto inVectorTy = dyn_cast<VectorType>(Val: in.getType()); |
| 286 | VectorType truncResType = VectorType::get(shape: 4, elementType: outElemType); |
| 287 | if (!inVectorTy) { |
| 288 | Value asFloat = castToF32(value: in, loc, rewriter); |
| 289 | Value asF8s = amdgpu::PackedTrunc2xFp8Op::create( |
| 290 | builder&: rewriter, location: loc, res: truncResType, sourceA: asFloat, /*sourceB=*/nullptr, wordIndex: 0, |
| 291 | /*existing=*/nullptr); |
| 292 | Value result = vector::ExtractOp::create(builder&: rewriter, location: loc, source: asF8s, position: 0); |
| 293 | rewriter.replaceOp(op, newValues: result); |
| 294 | return success(); |
| 295 | } |
| 296 | |
| 297 | int64_t numElements = outVecType.getNumElements(); |
| 298 | Value zero = arith::ConstantOp::create( |
| 299 | builder&: rewriter, location: loc, result: outElemType, value: rewriter.getFloatAttr(type: outElemType, value: 0.0)); |
| 300 | if (outVecType.getShape().empty()) { |
| 301 | Value scalarIn = |
| 302 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: in, position: ArrayRef<int64_t>{}); |
| 303 | // Recurse to send the 0-D vector case to the 1-D vector case |
| 304 | Value scalarTrunc = |
| 305 | arith::TruncFOp::create(builder&: rewriter, location: loc, out: outElemType, in: scalarIn); |
| 306 | Value result = vector::InsertOp::create(builder&: rewriter, location: loc, valueToStore: scalarTrunc, dest: zero, |
| 307 | position: ArrayRef<int64_t>{}); |
| 308 | rewriter.replaceOp(op, newValues: result); |
| 309 | return success(); |
| 310 | } |
| 311 | |
| 312 | VectorType flatTy = VectorType::get(shape: SmallVector<int64_t>{numElements}, |
| 313 | elementType: outVecType.getElementType()); |
| 314 | Value result = rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: flatTy, args&: zero); |
| 315 | |
| 316 | if (inVectorTy.getRank() > 1) { |
| 317 | inVectorTy = VectorType::get(shape: SmallVector<int64_t>{numElements}, |
| 318 | elementType: inVectorTy.getElementType()); |
| 319 | in = vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: inVectorTy, source: in); |
| 320 | } |
| 321 | |
| 322 | for (int64_t i = 0; i < numElements; i += 4) { |
| 323 | int64_t elemsThisOp = std::min(a: numElements, b: i + 4) - i; |
| 324 | Value thisResult = nullptr; |
| 325 | for (int64_t j = 0; j < elemsThisOp; j += 2) { |
| 326 | Value elemA = vector::ExtractOp::create(builder&: rewriter, location: loc, source: in, position: i + j); |
| 327 | Value asFloatA = castToF32(value: elemA, loc, rewriter); |
| 328 | Value asFloatB = nullptr; |
| 329 | if (j + 1 < elemsThisOp) { |
| 330 | Value elemB = vector::ExtractOp::create(builder&: rewriter, location: loc, source: in, position: i + j + 1); |
| 331 | asFloatB = castToF32(value: elemB, loc, rewriter); |
| 332 | } |
| 333 | thisResult = amdgpu::PackedTrunc2xFp8Op::create( |
| 334 | builder&: rewriter, location: loc, res: truncResType, sourceA: asFloatA, sourceB: asFloatB, wordIndex: j / 2, existing: thisResult); |
| 335 | } |
| 336 | if (elemsThisOp < 4) |
| 337 | thisResult = vector::ExtractStridedSliceOp::create( |
| 338 | builder&: rewriter, location: loc, source: thisResult, offsets: 0, sizes: elemsThisOp, strides: 1); |
| 339 | result = vector::InsertStridedSliceOp::create(builder&: rewriter, location: loc, valueToStore: thisResult, |
| 340 | dest: result, offsets: i, strides: 1); |
| 341 | } |
| 342 | |
| 343 | if (inVectorTy.getRank() != outVecType.getRank()) { |
| 344 | result = vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: outVecType, source: result); |
| 345 | } |
| 346 | |
| 347 | rewriter.replaceOp(op, newValues: result); |
| 348 | return success(); |
| 349 | } |
| 350 | |
| 351 | LogicalResult TruncfToFloat16RewritePattern::matchAndRewrite( |
| 352 | arith::TruncFOp op, PatternRewriter &rewriter) const { |
| 353 | Type outType = op.getOut().getType(); |
| 354 | Type inputType = getElementTypeOrSelf(val: op.getIn()); |
| 355 | auto outVecType = dyn_cast<VectorType>(Val&: outType); |
| 356 | if (outVecType) { |
| 357 | if (outVecType.isScalable()) |
| 358 | return failure(); |
| 359 | outType = outVecType.getElementType(); |
| 360 | } |
| 361 | if (!(outType.isF16() && inputType.isF32())) |
| 362 | return failure(); |
| 363 | |
| 364 | Location loc = op.getLoc(); |
| 365 | Value in = op.getIn(); |
| 366 | Type outElemType = getElementTypeOrSelf(type: op.getOut().getType()); |
| 367 | VectorType truncResType = VectorType::get(shape: 2, elementType: outElemType); |
| 368 | auto inVectorTy = dyn_cast<VectorType>(Val: in.getType()); |
| 369 | |
| 370 | // Handle the case where input type is not a vector type |
| 371 | if (!inVectorTy) { |
| 372 | auto sourceB = LLVM::PoisonOp::create(builder&: rewriter, location: loc, res: rewriter.getF32Type()); |
| 373 | Value asF16s = |
| 374 | ROCDL::CvtPkRtz::create(builder&: rewriter, location: loc, res: truncResType, srcA: in, srcB: sourceB); |
| 375 | Value result = vector::ExtractOp::create(builder&: rewriter, location: loc, source: asF16s, position: 0); |
| 376 | rewriter.replaceOp(op, newValues: result); |
| 377 | return success(); |
| 378 | } |
| 379 | int64_t numElements = outVecType.getNumElements(); |
| 380 | Value zero = rewriter.createOrFold<arith::ConstantOp>( |
| 381 | location: loc, args&: outElemType, args: rewriter.getFloatAttr(type: outElemType, value: 0.0)); |
| 382 | Value result = |
| 383 | rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: outVecType, args&: zero); |
| 384 | |
| 385 | if (inVectorTy.getRank() > 1) { |
| 386 | inVectorTy = VectorType::get(shape: SmallVector<int64_t>{numElements}, |
| 387 | elementType: inVectorTy.getElementType()); |
| 388 | in = vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: inVectorTy, source: in); |
| 389 | } |
| 390 | |
| 391 | // Handle the vector case. We also handle the (uncommon) case where the vector |
| 392 | // length is odd |
| 393 | for (int64_t i = 0; i < numElements; i += 2) { |
| 394 | int64_t elemsThisOp = std::min(a: numElements, b: i + 2) - i; |
| 395 | Value thisResult = nullptr; |
| 396 | Value elemA = vector::ExtractOp::create(builder&: rewriter, location: loc, source: in, position: i); |
| 397 | Value elemB = LLVM::PoisonOp::create(builder&: rewriter, location: loc, res: rewriter.getF32Type()); |
| 398 | |
| 399 | if (elemsThisOp == 2) { |
| 400 | elemB = vector::ExtractOp::create(builder&: rewriter, location: loc, source: in, position: i + 1); |
| 401 | } |
| 402 | |
| 403 | thisResult = |
| 404 | ROCDL::CvtPkRtz::create(builder&: rewriter, location: loc, res: truncResType, srcA: elemA, srcB: elemB); |
| 405 | // Place back the truncated result into the possibly larger vector. If we |
| 406 | // are operating on a size 2 vector, these operations should be folded away |
| 407 | thisResult = vector::ExtractStridedSliceOp::create( |
| 408 | builder&: rewriter, location: loc, source: thisResult, offsets: 0, sizes: elemsThisOp, strides: 1); |
| 409 | result = vector::InsertStridedSliceOp::create(builder&: rewriter, location: loc, valueToStore: thisResult, |
| 410 | dest: result, offsets: i, strides: 1); |
| 411 | } |
| 412 | |
| 413 | if (inVectorTy.getRank() != outVecType.getRank()) { |
| 414 | result = vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: outVecType, source: result); |
| 415 | } |
| 416 | |
| 417 | rewriter.replaceOp(op, newValues: result); |
| 418 | return success(); |
| 419 | } |
| 420 | |
| 421 | /// Get the broadcasted / splatted value for a chain of ops. |
| 422 | static Value getOriginalVectorValue(Value value) { |
| 423 | Value current = value; |
| 424 | while (Operation *definingOp = current.getDefiningOp()) { |
| 425 | bool skipOp = llvm::TypeSwitch<Operation *, bool>(definingOp) |
| 426 | .Case(caseFn: [¤t](vector::ShapeCastOp op) { |
| 427 | current = op.getSource(); |
| 428 | return true; |
| 429 | }) |
| 430 | .Case(caseFn: [¤t](vector::BroadcastOp op) { |
| 431 | current = op.getSource(); |
| 432 | return false; |
| 433 | }) |
| 434 | .Default(defaultResult: false); |
| 435 | |
| 436 | if (!skipOp) { |
| 437 | break; |
| 438 | } |
| 439 | } |
| 440 | return current; |
| 441 | } |
| 442 | |
| 443 | LogicalResult |
| 444 | ScalingExtFRewritePattern::matchAndRewrite(arith::ScalingExtFOp op, |
| 445 | PatternRewriter &rewriter) const { |
| 446 | Location loc = op.getLoc(); |
| 447 | constexpr int64_t opOutWidth = 2; |
| 448 | |
| 449 | Value in = op.getIn(); |
| 450 | Value scale = op.getScale(); |
| 451 | Value out = op.getOut(); |
| 452 | |
| 453 | Type f32 = rewriter.getF32Type(); |
| 454 | Type inType = getElementTypeOrSelf(val: in); |
| 455 | Type scaleType = getElementTypeOrSelf(val: scale); |
| 456 | Type outType = getElementTypeOrSelf(val: out); |
| 457 | |
| 458 | int64_t opInWidth = 32 / inType.getIntOrFloatBitWidth(); |
| 459 | |
| 460 | VectorType outVecType = dyn_cast<VectorType>(Val: out.getType()); |
| 461 | VectorType scaleVecType = dyn_cast<VectorType>(Val: scale.getType()); |
| 462 | |
| 463 | if (outVecType && outVecType.isScalable()) |
| 464 | return failure(); |
| 465 | |
| 466 | if (isa<RankedTensorType>(Val: out.getType()) || |
| 467 | isa<RankedTensorType>(Val: in.getType()) || |
| 468 | isa<RankedTensorType>(Val: scale.getType())) |
| 469 | return failure(); |
| 470 | |
| 471 | Type scaleF32Type = |
| 472 | scaleVecType ? VectorType::get(shape: scaleVecType.getShape(), elementType: f32) : f32; |
| 473 | if (scaleType.getIntOrFloatBitWidth() < 32) |
| 474 | scale = |
| 475 | arith::ExtFOp::create(builder&: rewriter, location: loc, resultTypes: TypeRange{scaleF32Type}, |
| 476 | operands: ValueRange{scale}, properties: arith::ExtFOp::Properties{}); |
| 477 | else if (scaleType.getIntOrFloatBitWidth() > 32) |
| 478 | scale = arith::TruncFOp::create(builder&: rewriter, location: loc, out: scaleF32Type, in: scale); |
| 479 | |
| 480 | VectorType extScaleResultType = VectorType::get(shape: opOutWidth, elementType: outType); |
| 481 | |
| 482 | if (!outVecType) { |
| 483 | Value inCast = vector::BroadcastOp::create(builder&: rewriter, location: loc, |
| 484 | result: VectorType::get(shape: 1, elementType: inType), source: in); |
| 485 | // TODO: replace this with non-packed ScaledExtOp |
| 486 | Value scaleExt = amdgpu::ScaledExtPackedOp::create( |
| 487 | builder&: rewriter, location: loc, res: extScaleResultType, source: inCast, scale, index: 0); |
| 488 | scaleExt = rewriter.replaceOpWithNewOp<vector::ExtractOp>(op, args&: scaleExt, args: 0); |
| 489 | return success(); |
| 490 | } |
| 491 | |
| 492 | VectorType inVecType = cast<VectorType>(Val: in.getType()); |
| 493 | Value origScale = getOriginalVectorValue(value: op.getScale()); |
| 494 | VectorType origScaleVecType = dyn_cast<VectorType>(Val: origScale.getType()); |
| 495 | |
| 496 | ArrayRef<int64_t> inShape = inVecType.getShape(); |
| 497 | SmallVector<int64_t> originalScaleShape; |
| 498 | if (origScaleVecType) |
| 499 | llvm::append_range(C&: originalScaleShape, R: origScaleVecType.getShape()); |
| 500 | |
| 501 | originalScaleShape.insert(I: originalScaleShape.end(), |
| 502 | NumToInsert: inShape.size() - originalScaleShape.size(), Elt: 1); |
| 503 | |
| 504 | auto maybeRatio = computeShapeRatio(shape: inShape, subShape: originalScaleShape); |
| 505 | assert(maybeRatio && |
| 506 | "failed to derive block size from broadcast or splat operation" ); |
| 507 | |
| 508 | SmallVector<int64_t> ratio = |
| 509 | maybeRatio.value_or(u: SmallVector<int64_t>(inShape.size(), 1)); |
| 510 | |
| 511 | int64_t blockSize = computeProduct(basis: ratio); |
| 512 | |
| 513 | Value zero = arith::ConstantOp::create(builder&: rewriter, location: loc, result: outType, |
| 514 | value: rewriter.getFloatAttr(type: outType, value: 0.0)); |
| 515 | Value result = |
| 516 | rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: outVecType, args&: zero); |
| 517 | |
| 518 | for (SmallVector<int64_t> offsets : StaticTileOffsetRange(inShape, ratio)) { |
| 519 | SmallVector<int64_t> strides(offsets.size(), 1); |
| 520 | Value block = vector::ExtractStridedSliceOp::create( |
| 521 | builder&: rewriter, location: loc, source: in, offsets, sizes: ratio, strides); |
| 522 | VectorType block1DType = VectorType::get(shape: blockSize, elementType: inType); |
| 523 | Value block1D = |
| 524 | vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: block1DType, source: block); |
| 525 | Value uniformScale = |
| 526 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: scale, position: offsets); |
| 527 | |
| 528 | VectorType blockResultType = VectorType::get(shape: blockSize, elementType: outType); |
| 529 | Value blockResult = |
| 530 | rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: blockResultType, args&: zero); |
| 531 | |
| 532 | for (int64_t i = 0, inSliceWidth = std::min(a: opInWidth, b: blockSize - i); |
| 533 | i < blockSize; |
| 534 | i += inSliceWidth, inSliceWidth = std::min(a: opInWidth, b: blockSize - i)) { |
| 535 | Value inSlice = vector::ExtractStridedSliceOp::create( |
| 536 | builder&: rewriter, location: loc, source: block1D, offsets: i, sizes: inSliceWidth, strides: 1); |
| 537 | for (int64_t j = 0, |
| 538 | outSliceWidth = std::min(a: opOutWidth, b: inSliceWidth - j); |
| 539 | j < inSliceWidth; j += outSliceWidth, |
| 540 | outSliceWidth = std::min(a: opOutWidth, b: inSliceWidth - j)) { |
| 541 | // TODO: replace this with non-packed ScaledExtOp for sliceWidth == 1 |
| 542 | Value scaleExt = amdgpu::ScaledExtPackedOp::create( |
| 543 | builder&: rewriter, location: loc, res: extScaleResultType, source: inSlice, scale: uniformScale, |
| 544 | index: j / opOutWidth); |
| 545 | if (outSliceWidth < opOutWidth) { |
| 546 | scaleExt = vector::ExtractStridedSliceOp::create( |
| 547 | builder&: rewriter, location: loc, source: scaleExt, offsets: 0, sizes: outSliceWidth, strides: 1); |
| 548 | } |
| 549 | blockResult = vector::InsertStridedSliceOp::create( |
| 550 | builder&: rewriter, location: loc, valueToStore: scaleExt, dest: blockResult, offsets: i + j, strides: 1); |
| 551 | } |
| 552 | } |
| 553 | |
| 554 | VectorType resultType = VectorType::get(shape: ratio, elementType: outType); |
| 555 | Value cast = |
| 556 | vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: resultType, source: blockResult); |
| 557 | result = vector::InsertStridedSliceOp::create(builder&: rewriter, location: loc, valueToStore: cast, dest: result, |
| 558 | offsets, strides); |
| 559 | } |
| 560 | |
| 561 | rewriter.replaceOp(op, newValues: result); |
| 562 | |
| 563 | return success(); |
| 564 | } |
| 565 | |
| 566 | LogicalResult |
| 567 | ScalingTruncFRewritePattern::matchAndRewrite(arith::ScalingTruncFOp op, |
| 568 | PatternRewriter &rewriter) const { |
| 569 | Location loc = op.getLoc(); |
| 570 | constexpr int64_t opInWidth = 2; |
| 571 | |
| 572 | Value in = op.getIn(); |
| 573 | Value scale = op.getScale(); |
| 574 | Value out = op.getOut(); |
| 575 | |
| 576 | Type f32 = rewriter.getF32Type(); |
| 577 | Type inType = getElementTypeOrSelf(val: in); |
| 578 | Type scaleType = getElementTypeOrSelf(val: scale); |
| 579 | Type outType = getElementTypeOrSelf(val: out); |
| 580 | |
| 581 | VectorType outVecType = dyn_cast<VectorType>(Val: out.getType()); |
| 582 | VectorType scaleVecType = dyn_cast<VectorType>(Val: scale.getType()); |
| 583 | if (outVecType && outVecType.isScalable()) |
| 584 | return failure(); |
| 585 | |
| 586 | if (isa<RankedTensorType>(Val: out.getType()) || |
| 587 | isa<RankedTensorType>(Val: in.getType()) || |
| 588 | isa<RankedTensorType>(Val: scale.getType())) |
| 589 | return failure(); |
| 590 | |
| 591 | Type scaleF32Type = |
| 592 | scaleVecType ? VectorType::get(shape: scaleVecType.getShape(), elementType: f32) : f32; |
| 593 | if (scaleType.getIntOrFloatBitWidth() < 32) |
| 594 | scale = |
| 595 | arith::ExtFOp::create(builder&: rewriter, location: loc, resultTypes: TypeRange{scaleF32Type}, |
| 596 | operands: ValueRange{scale}, properties: arith::ExtFOp::Properties{}); |
| 597 | else if (scaleType.getIntOrFloatBitWidth() > 32) |
| 598 | scale = arith::TruncFOp::create(builder&: rewriter, location: loc, out: scaleF32Type, in: scale); |
| 599 | |
| 600 | Value zero = arith::ConstantOp::create(builder&: rewriter, location: loc, result: outType, |
| 601 | value: rewriter.getFloatAttr(type: outType, value: 0.0)); |
| 602 | int64_t opOutWidth = 32 / outType.getIntOrFloatBitWidth(); |
| 603 | VectorType truncScaleResultType = VectorType::get(shape: opOutWidth, elementType: outType); |
| 604 | |
| 605 | if (!outVecType) { |
| 606 | Type inVecType = VectorType::get(shape: 1, elementType: inType); |
| 607 | Value inCast = vector::BroadcastOp::create(builder&: rewriter, location: loc, result: inVecType, source: in); |
| 608 | // TODO: replace this with non-packed ScaledTruncOp |
| 609 | Value scaleTrunc = amdgpu::PackedScaledTruncOp::create( |
| 610 | builder&: rewriter, location: loc, res: truncScaleResultType, source: inCast, scale, index: 0, |
| 611 | /*existing=*/nullptr); |
| 612 | scaleTrunc = |
| 613 | rewriter.replaceOpWithNewOp<vector::ExtractOp>(op, args&: scaleTrunc, args: 0); |
| 614 | return success(); |
| 615 | } |
| 616 | |
| 617 | VectorType inVecType = cast<VectorType>(Val: in.getType()); |
| 618 | Value origScale = getOriginalVectorValue(value: op.getScale()); |
| 619 | VectorType origScaleVecType = dyn_cast<VectorType>(Val: origScale.getType()); |
| 620 | |
| 621 | ArrayRef<int64_t> inShape = inVecType.getShape(); |
| 622 | SmallVector<int64_t> scaleShape; |
| 623 | if (origScaleVecType) |
| 624 | llvm::append_range(C&: scaleShape, R: origScaleVecType.getShape()); |
| 625 | |
| 626 | scaleShape.insert(I: scaleShape.end(), NumToInsert: inShape.size() - scaleShape.size(), Elt: 1); |
| 627 | |
| 628 | auto maybeRatio = computeShapeRatio(shape: inShape, subShape: scaleShape); |
| 629 | assert(maybeRatio && |
| 630 | "failed to derive block size from broadcast or splat operation" ); |
| 631 | |
| 632 | SmallVector<int64_t> ratio = |
| 633 | maybeRatio.value_or(u: SmallVector<int64_t>(inShape.size(), 1)); |
| 634 | |
| 635 | int64_t blockSize = computeProduct(basis: ratio); |
| 636 | |
| 637 | Value result = |
| 638 | rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: outVecType, args&: zero); |
| 639 | |
| 640 | for (SmallVector<int64_t> offsets : StaticTileOffsetRange(inShape, ratio)) { |
| 641 | SmallVector<int64_t> strides(offsets.size(), 1); |
| 642 | Value block = vector::ExtractStridedSliceOp::create( |
| 643 | builder&: rewriter, location: loc, source: in, offsets, sizes: ratio, strides); |
| 644 | VectorType block1DType = VectorType::get(shape: blockSize, elementType: inType); |
| 645 | Value block1D = |
| 646 | vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: block1DType, source: block); |
| 647 | Value uniformScale = |
| 648 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: scale, position: offsets); |
| 649 | |
| 650 | VectorType blockResultType = VectorType::get(shape: blockSize, elementType: outType); |
| 651 | Value blockResult = |
| 652 | rewriter.createOrFold<vector::BroadcastOp>(location: loc, args&: blockResultType, args&: zero); |
| 653 | |
| 654 | for (int64_t i = 0, outSliceWidth = std::min(a: opOutWidth, b: blockSize - i); |
| 655 | i < blockSize; i += outSliceWidth, |
| 656 | outSliceWidth = std::min(a: opOutWidth, b: blockSize - i)) { |
| 657 | Value scaleTrunc; |
| 658 | // Case where <= 2 elements are being truncated. |
| 659 | if (outSliceWidth <= opInWidth) { |
| 660 | Value slice = vector::ExtractStridedSliceOp::create( |
| 661 | builder&: rewriter, location: loc, source: block1D, offsets: i, sizes: outSliceWidth, strides: 1); |
| 662 | // TODO: replace this with non-packed ScaledTruncOp for sliceWidth == 1 |
| 663 | scaleTrunc = amdgpu::PackedScaledTruncOp::create( |
| 664 | builder&: rewriter, location: loc, res: truncScaleResultType, source: slice, scale: uniformScale, index: 0, |
| 665 | /*existing=*/nullptr); |
| 666 | } else { |
| 667 | scaleTrunc = vector::BroadcastOp::create(builder&: rewriter, location: loc, |
| 668 | result: truncScaleResultType, source: zero); |
| 669 | for (int64_t j = 0, |
| 670 | inSliceWidth = std::min(a: opInWidth, b: outSliceWidth - j); |
| 671 | j < outSliceWidth; j += opInWidth, |
| 672 | inSliceWidth = std::min(a: opInWidth, b: outSliceWidth - j)) { |
| 673 | Value slice = vector::ExtractStridedSliceOp::create( |
| 674 | builder&: rewriter, location: loc, source: block1D, offsets: i + j, sizes: inSliceWidth, strides: 1); |
| 675 | scaleTrunc = amdgpu::PackedScaledTruncOp::create( |
| 676 | builder&: rewriter, location: loc, res: truncScaleResultType, source: slice, scale: uniformScale, |
| 677 | index: j / opInWidth, existing: scaleTrunc); |
| 678 | } |
| 679 | } |
| 680 | if (outSliceWidth != opOutWidth) { |
| 681 | scaleTrunc = vector::ExtractStridedSliceOp::create( |
| 682 | builder&: rewriter, location: loc, source: scaleTrunc, offsets: 0, sizes: outSliceWidth, strides: 1); |
| 683 | } |
| 684 | blockResult = vector::InsertStridedSliceOp::create( |
| 685 | builder&: rewriter, location: loc, valueToStore: scaleTrunc, dest: blockResult, offsets: i, strides: 1); |
| 686 | } |
| 687 | |
| 688 | VectorType resultType = VectorType::get(shape: ratio, elementType: outType); |
| 689 | Value cast = |
| 690 | vector::ShapeCastOp::create(builder&: rewriter, location: loc, result: resultType, source: blockResult); |
| 691 | result = vector::InsertStridedSliceOp::create(builder&: rewriter, location: loc, valueToStore: cast, dest: result, |
| 692 | offsets, strides); |
| 693 | } |
| 694 | |
| 695 | rewriter.replaceOp(op, newValues: result); |
| 696 | |
| 697 | return success(); |
| 698 | } |
| 699 | |
| 700 | void mlir::arith::populateArithToAMDGPUConversionPatterns( |
| 701 | RewritePatternSet &patterns, bool convertFP8Arithmetic, |
| 702 | bool saturateFP8Truncf, bool allowPackedF16Rtz, bool supportsScaledExtTrunc, |
| 703 | const ROCDL::TargetInfo &target, PatternBenefit benefit) { |
| 704 | |
| 705 | if (convertFP8Arithmetic) { |
| 706 | patterns.add<ExtFOnFloat8RewritePattern>(arg: patterns.getContext(), args: target, |
| 707 | args&: benefit); |
| 708 | patterns.add<TruncFToFloat8RewritePattern>( |
| 709 | arg: patterns.getContext(), args&: saturateFP8Truncf, args: target, args&: benefit); |
| 710 | } |
| 711 | if (allowPackedF16Rtz) |
| 712 | patterns.add<TruncfToFloat16RewritePattern>(arg: patterns.getContext(), args&: benefit); |
| 713 | |
| 714 | if (supportsScaledExtTrunc) { |
| 715 | patterns.add<ScalingExtFRewritePattern>(arg: patterns.getContext(), args&: benefit); |
| 716 | patterns.add<ScalingTruncFRewritePattern>(arg: patterns.getContext(), args&: benefit); |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | void ArithToAMDGPUConversionPass::runOnOperation() { |
| 721 | Operation *op = getOperation(); |
| 722 | MLIRContext *ctx = &getContext(); |
| 723 | RewritePatternSet patterns(op->getContext()); |
| 724 | FailureOr<ROCDL::TargetInfo> targetInfo = ROCDL::TargetInfo::get( |
| 725 | arch: ROCDL::resolveArchOption(arch, deprecatedAlias: chipset), /*waveSize=*/0, |
| 726 | emitError: [&] { return emitError(loc: UnknownLoc::get(context: ctx)); }); |
| 727 | if (failed(Result: targetInfo)) |
| 728 | return signalPassFailure(); |
| 729 | |
| 730 | bool convertFP8Arithmetic = |
| 731 | targetInfo->has(feature: llvm::AMDGPU::FEAT_FP8_CONVERSION_INSTS); |
| 732 | bool supportsScaledExtTrunc = |
| 733 | targetInfo->has(feature: llvm::AMDGPU::FEAT_GFX950_INSTS); |
| 734 | arith::populateArithToAMDGPUConversionPatterns( |
| 735 | patterns, convertFP8Arithmetic, saturateFP8Truncf, allowPackedF16Rtz, |
| 736 | supportsScaledExtTrunc, target: *targetInfo); |
| 737 | if (failed(Result: applyPatternsGreedily(op, patterns: std::move(patterns)))) |
| 738 | return signalPassFailure(); |
| 739 | } |
| 740 | |