| 1 | //===- XeGPUPeepHoleOptimizer.cpp - XeGPU optimize block loads -*- 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/Arith/IR/Arith.h" |
| 10 | #include "mlir/Dialect/GPU/IR/GPUDialect.h" |
| 11 | #include "mlir/Dialect/MemRef/IR/MemRef.h" |
| 12 | #include "mlir/Dialect/SCF/Transforms/Patterns.h" |
| 13 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
| 14 | #include "mlir/Dialect/Utils/StaticValueUtils.h" |
| 15 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 16 | #include "mlir/Dialect/XeGPU/IR/XeGPU.h" |
| 17 | #include "mlir/Dialect/XeGPU/Transforms/Passes.h" |
| 18 | #include "mlir/Dialect/XeGPU/Transforms/Transforms.h" |
| 19 | #include "mlir/Dialect/XeGPU/Transforms/XeGPULayoutImpl.h" |
| 20 | #include "mlir/Dialect/XeGPU/Utils/XeGPUUtils.h" |
| 21 | #include "mlir/Dialect/XeGPU/uArch/uArchBase.h" |
| 22 | #include "mlir/Dialect/XeGPU/uArch/uArchCommon.h" |
| 23 | #include "mlir/IR/BuiltinTypes.h" |
| 24 | #include "mlir/IR/OpDefinition.h" |
| 25 | #include "mlir/IR/Types.h" |
| 26 | #include "mlir/IR/Value.h" |
| 27 | #include "mlir/Transforms/DialectConversion.h" |
| 28 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
| 29 | #include "llvm/ADT/STLExtras.h" |
| 30 | #include "llvm/ADT/SmallVector.h" |
| 31 | #include <optional> |
| 32 | |
| 33 | namespace mlir { |
| 34 | namespace xegpu { |
| 35 | #define GEN_PASS_DEF_XEGPUPEEPHOLEOPTIMIZER |
| 36 | #include "mlir/Dialect/XeGPU/Transforms/Passes.h.inc" |
| 37 | } // namespace xegpu |
| 38 | } // namespace mlir |
| 39 | |
| 40 | #define DEBUG_TYPE "xegpu-optimize-peephole" |
| 41 | #define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE "]: ") |
| 42 | |
| 43 | using namespace mlir; |
| 44 | |
| 45 | namespace { |
| 46 | |
| 47 | /// Get the 2D lane data from a tensor desc type if it exists. |
| 48 | static std::optional<SmallVector<int64_t>> |
| 49 | getMaybeLaneData(xegpu::TensorDescType tdescType) { |
| 50 | auto layout = tdescType.getLayoutAttr(); |
| 51 | if (!layout) |
| 52 | return std::nullopt; |
| 53 | return xegpu::getInner2DIfUnitLeadingDims(vals: layout.getEffectiveLaneDataAsInt()); |
| 54 | } |
| 55 | |
| 56 | /// Get the 2D lane layout from a tensor desc type if it exists. |
| 57 | static std::optional<SmallVector<int64_t>> |
| 58 | getMaybeLaneLayout(xegpu::TensorDescType tdescType) { |
| 59 | auto layout = tdescType.getLayoutAttr(); |
| 60 | if (!layout) |
| 61 | return std::nullopt; |
| 62 | return xegpu::getInner2DIfUnitLeadingDims( |
| 63 | vals: layout.getEffectiveLaneLayoutAsInt()); |
| 64 | } |
| 65 | |
| 66 | /// A layout can be optimized if its lane layout is transposed (lane[0] != 1 && |
| 67 | /// lane[1] == 1), but inner lane data is not equal to [1, 1]. |
| 68 | /// Example: |
| 69 | /// !xegpu.tensor_desc<16x16xf16, |
| 70 | /// #xegpu.layout<lane_layout = [16, 1], lane_data = [1, 2]>> |
| 71 | /// In this case, lane layout is transposed (from the usual [1, SG_SIZE] form) |
| 72 | /// indicating that this is a load that requires transpose effect. However, |
| 73 | /// lane data is [1, 2], meaning that each lane must grab 2 f16 elements from |
| 74 | /// the inner dimension. We convert this to a optimized form by converting the |
| 75 | /// tensor_desc to i32 type such that lane data becomes [1, 1]. This makes the |
| 76 | /// later lowering easily use the load with transpose instruction. |
| 77 | static bool canBeOptimizedForTranspose(ArrayRef<int64_t> laneLayout, |
| 78 | ArrayRef<int64_t> laneData) { |
| 79 | if (laneLayout.size() != 2 || laneData.size() != 2) |
| 80 | return false; |
| 81 | if (laneLayout[0] == 1 || laneLayout[1] != 1) |
| 82 | return false; |
| 83 | if (laneData[0] != 1 || laneData[1] == 1) |
| 84 | return false; |
| 85 | return true; |
| 86 | } |
| 87 | |
| 88 | /// A tensor desc type can be optimized if its element type is less than 32 bits |
| 89 | /// and its layout can be optimized. |
| 90 | static bool canBeOptimizedForTranspose(xegpu::TensorDescType tdescType) { |
| 91 | // If the dtype is greater or equal to 32 bits, layout must be valid. |
| 92 | int elementTyBitwidth = tdescType.getElementType().getIntOrFloatBitWidth(); |
| 93 | if (elementTyBitwidth >= 32) |
| 94 | return false; |
| 95 | auto maybeLaneLayout = getMaybeLaneLayout(tdescType); |
| 96 | auto maybeLaneData = getMaybeLaneData(tdescType); |
| 97 | if (!maybeLaneData || !maybeLaneLayout) |
| 98 | return false; |
| 99 | return canBeOptimizedForTranspose(laneLayout: *maybeLaneLayout, laneData: *maybeLaneData); |
| 100 | } |
| 101 | |
| 102 | /// Check if a tensor desc type can be optimized for transpose, if so return the |
| 103 | /// new optimized tensor desc type with a valid transpose layout. |
| 104 | static xegpu::TensorDescType |
| 105 | tryOptimize(xegpu::TensorDescType tdescType, |
| 106 | const xegpu::uArch::uArch *targetuArch) { |
| 107 | if (!canBeOptimizedForTranspose(tdescType)) |
| 108 | return tdescType; |
| 109 | auto laneData = getMaybeLaneData(tdescType) |
| 110 | .value(); // Lane data must exist if we reach here. |
| 111 | int64_t innerLaneData = laneData[1]; |
| 112 | int elementTyBitwidth = tdescType.getElementType().getIntOrFloatBitWidth(); |
| 113 | // Required shape is total shape of the vector result that this tensor desc |
| 114 | // must eventually load after adjusting for the new bitwidth and array |
| 115 | // length. |
| 116 | SmallVector<int64_t> requiredShape(tdescType.getShape()); |
| 117 | requiredShape.back() = |
| 118 | requiredShape.back() * tdescType.getArrayLength() / innerLaneData; |
| 119 | int newBitWidth = elementTyBitwidth * innerLaneData; |
| 120 | Type newElemTy = IntegerType::get(context: tdescType.getContext(), width: newBitWidth); |
| 121 | // Supported shape is the max transpose shape that can be supported by |
| 122 | // hardware that is less than or equal to required shape. |
| 123 | auto *blockLoadTarget = |
| 124 | dyn_cast<xegpu::uArch::Subgroup2DBlockLoadInstruction>( |
| 125 | Val: targetuArch->getInstruction( |
| 126 | instKind: xegpu::uArch::InstructionKind::Subgroup2DBlockLoad)); |
| 127 | auto maybeHWParams = blockLoadTarget->getBlockWidthHeightCount( |
| 128 | elemTy: newElemTy, /** has transform */ hasTransform: false, /** has transpose */ hasTranspose: true); |
| 129 | // If no HW params found, return the original type. |
| 130 | if (!maybeHWParams) |
| 131 | return tdescType; |
| 132 | auto [widths, heights, counts] = maybeHWParams.value(); |
| 133 | // TODO: Currently we expect array length to be 1 for transpose case. |
| 134 | if (counts.size() != 1 || counts[0] != 1) |
| 135 | return tdescType; |
| 136 | int arrayLen = counts[0]; |
| 137 | // The transpose acts on the innermost 2 dims; any leading dims are unit |
| 138 | // batch. |
| 139 | int64_t rank = requiredShape.size(); |
| 140 | int supportedHeight = xegpu::getLargestDivisor( |
| 141 | dim: static_cast<int>(requiredShape[rank - 2]), candidates: heights); |
| 142 | int supportedWidth = xegpu::getLargestDivisor( |
| 143 | dim: static_cast<int>(requiredShape[rank - 1]), candidates: widths); |
| 144 | // If no supported height or width found, return the original type. |
| 145 | if (supportedHeight == -1 || supportedWidth == -1) |
| 146 | return tdescType; |
| 147 | |
| 148 | // Preserve leading (unit) batch dims; only the inner 2 dims are reshaped. |
| 149 | SmallVector<int64_t> supportedShape(requiredShape.begin(), |
| 150 | requiredShape.end() - 2); |
| 151 | supportedShape.push_back(Elt: supportedHeight); |
| 152 | supportedShape.push_back(Elt: supportedWidth); |
| 153 | auto ctx = tdescType.getContext(); |
| 154 | auto origLayout = tdescType.getLayoutAttr(); |
| 155 | auto laneLayoutI64 = origLayout.getEffectiveLaneLayoutAsInt(); |
| 156 | SmallVector<int32_t> laneLayoutI32(laneLayoutI64.begin(), |
| 157 | laneLayoutI64.end()); |
| 158 | |
| 159 | xegpu::LayoutAttr newLayout = xegpu::LayoutAttr::get( |
| 160 | context: ctx, /*lane_layout=*/DenseI32ArrayAttr::get(context: ctx, content: laneLayoutI32), |
| 161 | /*lane_data=*/ |
| 162 | DenseI32ArrayAttr::get(context: ctx, |
| 163 | content: SmallVector<int32_t>(laneLayoutI32.size(), 1)), |
| 164 | /*order=*/origLayout.getOrder()); |
| 165 | |
| 166 | // Array length can not be larger than 1 for transpose case. |
| 167 | return xegpu::TensorDescType::get(shape: supportedShape, elementType: newElemTy, array_length: arrayLen, |
| 168 | boundary_check: tdescType.getBoundaryCheck(), |
| 169 | memory_space: tdescType.getMemorySpace(), layout: newLayout); |
| 170 | } |
| 171 | |
| 172 | /// Helper to convert an OpFoldResult to Value. |
| 173 | static Value convertToValue(ConversionPatternRewriter &rewriter, Location loc, |
| 174 | OpFoldResult ofr) { |
| 175 | std::optional<int64_t> mayBeInt = getConstantIntValue(ofr); |
| 176 | if (mayBeInt) |
| 177 | return arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: *mayBeInt).getResult(); |
| 178 | return llvm::cast<Value>(Val&: ofr); |
| 179 | } |
| 180 | |
| 181 | /// Helper to divide a Value by a constant integer. |
| 182 | static Value divideByConstant(ConversionPatternRewriter &rewriter, Location loc, |
| 183 | Value val, int64_t constant) { |
| 184 | // If the constant is a power of 2, use right shift for division. |
| 185 | if (llvm::isPowerOf2_64(Value: constant)) { |
| 186 | int64_t shiftAmount = llvm::Log2_64(Value: constant); |
| 187 | return arith::ShRUIOp::create( |
| 188 | builder&: rewriter, location: loc, lhs: val, |
| 189 | rhs: arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: shiftAmount) |
| 190 | .getResult()) |
| 191 | .getResult(); |
| 192 | } |
| 193 | auto constantOp = |
| 194 | arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: constant).getResult(); |
| 195 | return arith::DivUIOp::create(builder&: rewriter, location: loc, lhs: val, rhs: constantOp).getResult(); |
| 196 | } |
| 197 | |
| 198 | /// This function takes a larger register block `data` and generates multiple |
| 199 | /// smaller loads (size given by `newTensorDesc`) to fill in the `data` block |
| 200 | /// starting from `offsets`. |
| 201 | static Value generateLoads(ConversionPatternRewriter &rewriter, |
| 202 | TypedValue<VectorType> data, |
| 203 | SmallVector<OpFoldResult> offsets, |
| 204 | TypedValue<xegpu::TensorDescType> newTensorDesc, |
| 205 | xegpu::LoadNdOp origLoadOp) { |
| 206 | Location loc = data.getLoc(); |
| 207 | assert(offsets.size() >= 2 && "Expecting at least 2 offsets for 2D LoadNdOp" ); |
| 208 | int64_t rank = data.getType().getRank(); |
| 209 | Value offsetDim0 = convertToValue(rewriter, loc, ofr: offsets[offsets.size() - 2]); |
| 210 | Value offsetDim1 = convertToValue(rewriter, loc, ofr: offsets[offsets.size() - 1]); |
| 211 | SmallVector<int64_t> supportedShape(newTensorDesc.getType().getShape()); |
| 212 | // Compute the ratio between original shape and supported shape. We need to |
| 213 | // generate loads in this ratio arrangement. |
| 214 | auto shapeRatio = computeShapeRatio(shape: data.getType().getShape(), |
| 215 | subShape: supportedShape) |
| 216 | .value(); // `ratio` must be defined if we reach here. |
| 217 | // The loop below only walks the last 2 entries of `shapeRatio`; leading |
| 218 | // (batch) dims are unit, so their ratio is 1 and their offsets pass through |
| 219 | // unchanged. |
| 220 | int64_t suppDim0 = supportedShape[rank - 2]; |
| 221 | int64_t suppDim1 = supportedShape[rank - 1]; |
| 222 | for (int64_t h = 0; h < shapeRatio[rank - 2]; ++h) { |
| 223 | for (int64_t w = 0; w < shapeRatio[rank - 1]; ++w) { |
| 224 | int64_t localOffsetDim0 = h * suppDim0; |
| 225 | int64_t localOffsetDim1 = w * suppDim1; |
| 226 | Value loadOffsetX = arith::AddIOp::create( |
| 227 | builder&: rewriter, location: loc, lhs: offsetDim0, |
| 228 | rhs: arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: localOffsetDim0) |
| 229 | .getResult()); |
| 230 | Value loadOffsetY = arith::AddIOp::create( |
| 231 | builder&: rewriter, location: loc, lhs: offsetDim1, |
| 232 | rhs: arith::ConstantIndexOp::create(builder&: rewriter, location: loc, value: localOffsetDim1) |
| 233 | .getResult()); |
| 234 | // Keep the leading (batch) offsets; replace only the inner 2. |
| 235 | SmallVector<OpFoldResult> loadOffsets(offsets.begin(), offsets.end()); |
| 236 | loadOffsets[loadOffsets.size() - 2] = loadOffsetX; |
| 237 | loadOffsets[loadOffsets.size() - 1] = loadOffsetY; |
| 238 | auto loadOp = xegpu::LoadNdOp::create( |
| 239 | builder&: rewriter, location: loc, |
| 240 | value: VectorType::get(shape: supportedShape, elementType: data.getType().getElementType()), |
| 241 | TensorDesc: newTensorDesc, offsets: loadOffsets, packed: origLoadOp.getPackedAttr(), |
| 242 | transpose: origLoadOp.getTransposeAttr(), l1_hint: origLoadOp.getL1HintAttr(), |
| 243 | l2_hint: origLoadOp.getL2HintAttr(), l3_hint: origLoadOp.getL3HintAttr(), |
| 244 | layout: origLoadOp.getLayoutAttr()); |
| 245 | // Set the layout for the loadOp. |
| 246 | auto layoutAttr = newTensorDesc.getType().getLayoutAttr(); |
| 247 | loadOp.setAnchorLayout(layoutAttr); |
| 248 | // Insert the loaded block into the right position in data (leading dims |
| 249 | // at 0, inner 2 dims at the local tile offset). |
| 250 | SmallVector<int64_t> insertPos(rank, 0); |
| 251 | insertPos[rank - 2] = localOffsetDim0; |
| 252 | insertPos[rank - 1] = localOffsetDim1; |
| 253 | SmallVector<int64_t> insertStrides(rank, 1); |
| 254 | auto insertOp = vector::InsertStridedSliceOp::create( |
| 255 | builder&: rewriter, location: loc, valueToStore: loadOp.getResult(), dest: data, offsets: insertPos, strides: insertStrides); |
| 256 | // InsertOp must have the same layout as newTensorDesc. |
| 257 | xegpu::setTemporaryLayout(operandOrResult: insertOp->getOpResult(idx: 0), layout: layoutAttr); |
| 258 | data = insertOp.getResult(); |
| 259 | } |
| 260 | } |
| 261 | return data; |
| 262 | } |
| 263 | |
| 264 | /// Checks if a CreateNdDescOp can be optimized for transpose, if so creates a |
| 265 | /// new CreateNdDescOp with optimized tensor desc type. This involves extracting |
| 266 | /// the base pointer from the original memory source and adjusting the shape and |
| 267 | /// strides of the tensor desc to fit with the new optimized transpose layout. |
| 268 | class XeGPUCreateNdDescOpPattern final |
| 269 | : public OpConversionPattern<xegpu::CreateNdDescOp> { |
| 270 | public: |
| 271 | using OpConversionPattern<xegpu::CreateNdDescOp>::OpConversionPattern; |
| 272 | LogicalResult |
| 273 | matchAndRewrite(xegpu::CreateNdDescOp createNdOp, OpAdaptor adaptor, |
| 274 | ConversionPatternRewriter &rewriter) const override { |
| 275 | auto tdescTy = createNdOp.getType(); |
| 276 | // Get the target uArch info. |
| 277 | auto chipStr = xegpu::getChipStr(op: createNdOp); |
| 278 | // Check if the chip is supported. |
| 279 | assert(chipStr && |
| 280 | (chipStr.value() == "pvc" || chipStr.value() == "bmg" || |
| 281 | chipStr.value() == "cri" ) && |
| 282 | "Expecting target chip to be pvc, bmg or cri for transpose " |
| 283 | "optimization." ); |
| 284 | const auto *targetuArch = xegpu::uArch::getUArch(archName: chipStr.value()); |
| 285 | |
| 286 | auto convertType = tryOptimize(tdescType: tdescTy, targetuArch); |
| 287 | if (convertType == tdescTy) |
| 288 | return failure(); |
| 289 | Location loc = createNdOp.getLoc(); |
| 290 | Value source = createNdOp.getSource(); |
| 291 | auto memrefType = dyn_cast<MemRefType>(Val: source.getType()); |
| 292 | |
| 293 | SmallVector<OpFoldResult> mixedSizes; |
| 294 | SmallVector<OpFoldResult> mixedStrides; |
| 295 | memref::ExtractStridedMetadataOp meta; |
| 296 | if (memrefType) { |
| 297 | // extract_aligned_pointer_as_index returns the base pointer without the |
| 298 | // memref offset, so always extract strided metadata to fold it in. |
| 299 | meta = memref::ExtractStridedMetadataOp::create(builder&: rewriter, location: loc, source); |
| 300 | mixedSizes = meta.getConstifiedMixedSizes(); |
| 301 | mixedStrides = meta.getConstifiedMixedStrides(); |
| 302 | } else { |
| 303 | mixedSizes = createNdOp.getMixedSizes(); |
| 304 | mixedStrides = createNdOp.getMixedStrides(); |
| 305 | } |
| 306 | |
| 307 | auto maybeConstInnerStride = getConstantIntValue(ofr: mixedStrides.back()); |
| 308 | // Only row-major memrefs are expected for now. |
| 309 | if (!maybeConstInnerStride || *maybeConstInnerStride != 1) |
| 310 | return rewriter.notifyMatchFailure( |
| 311 | arg&: createNdOp, msg: "Expecting row-major memref for transpose optimization." ); |
| 312 | auto optionalLaneData = getMaybeLaneData(tdescType: tdescTy); |
| 313 | assert(optionalLaneData && "Expected 2D lane data" ); |
| 314 | auto laneData = optionalLaneData.value(); |
| 315 | int64_t innerLaneData = laneData[1]; |
| 316 | // Inner dimension of the shape must be adjusted based on innerLaneData. |
| 317 | SmallVector<OpFoldResult> modifiedShape(mixedSizes); |
| 318 | modifiedShape.back() = divideByConstant( |
| 319 | rewriter, loc, val: convertToValue(rewriter, loc, ofr: modifiedShape.back()), |
| 320 | constant: innerLaneData); |
| 321 | // Repacking to a wider element rescales every stride but the innermost. |
| 322 | assert(mixedStrides.size() >= 2 && |
| 323 | "Expected at least 2 strides for CreateNdDescOp" ); |
| 324 | SmallVector<OpFoldResult> modifiedStrides(mixedStrides); |
| 325 | for (size_t i = 0; i + 1 < modifiedStrides.size(); ++i) |
| 326 | modifiedStrides[i] = divideByConstant( |
| 327 | rewriter, loc, val: convertToValue(rewriter, loc, ofr: modifiedStrides[i]), |
| 328 | constant: innerLaneData); |
| 329 | |
| 330 | if (memrefType) { |
| 331 | // Compute base = aligned pointer + offset (in bytes). |
| 332 | Value alignedPtr = memref::ExtractAlignedPointerAsIndexOp::create( |
| 333 | builder&: rewriter, location: loc, source: meta.getBaseBuffer()); |
| 334 | Value baseIdx = xegpu::addElemOffsetInBytes(builder&: rewriter, loc, baseAddr: alignedPtr, |
| 335 | elemOffset: meta.getOffset(), memrefTy: memrefType); |
| 336 | source = arith::IndexCastOp::create(builder&: rewriter, location: loc, out: rewriter.getI64Type(), |
| 337 | in: baseIdx); |
| 338 | } |
| 339 | // Create a new CreateNdDescOp with the modified shape and converted type. |
| 340 | auto newCreateNdDescOp = xegpu::CreateNdDescOp::create( |
| 341 | builder&: rewriter, location: loc, tdesc: convertType, source, shape: modifiedShape, strides: modifiedStrides); |
| 342 | rewriter.replaceOp(op: createNdOp, newValues: newCreateNdDescOp.getResult()); |
| 343 | return success(); |
| 344 | } |
| 345 | }; |
| 346 | |
| 347 | /// Checks if a LoadNdOp consumes a tensor desc type that was rewritten for |
| 348 | /// tranpose optimization. If so, rewrites the LoadNdOp to to align with the |
| 349 | /// adjusted tensor desc type. This can result in multiple LoadNdOps being |
| 350 | /// generated to fill in the original load shape. |
| 351 | class XeGPULoadNdDescOpPattern final |
| 352 | : public OpConversionPattern<xegpu::LoadNdOp> { |
| 353 | public: |
| 354 | using OpConversionPattern<xegpu::LoadNdOp>::OpConversionPattern; |
| 355 | LogicalResult |
| 356 | matchAndRewrite(xegpu::LoadNdOp loadNdOp, OpAdaptor adaptor, |
| 357 | ConversionPatternRewriter &rewriter) const override { |
| 358 | auto origTensorDescType = loadNdOp.getTensorDescType(); |
| 359 | auto adaptorType = |
| 360 | cast<xegpu::TensorDescType>(Val: adaptor.getTensorDesc().getType()); |
| 361 | if (adaptorType == origTensorDescType) |
| 362 | return failure(); |
| 363 | // Offsets must be adjusted based on innerLaneData. |
| 364 | auto laneData = getMaybeLaneData(tdescType: loadNdOp.getTensorDescType()).value(); |
| 365 | int64_t innerLaneData = laneData[1]; |
| 366 | auto offsets = loadNdOp.getMixedOffsets(); |
| 367 | if (offsets.empty()) |
| 368 | return rewriter.notifyMatchFailure(arg&: loadNdOp, |
| 369 | msg: "Expecting offsets in LoadNd" ); |
| 370 | SmallVector<OpFoldResult> modifiedOffsets(offsets); |
| 371 | modifiedOffsets.back() = divideByConstant( |
| 372 | rewriter, loc: loadNdOp.getLoc(), |
| 373 | val: convertToValue(rewriter, loc: loadNdOp.getLoc(), ofr: modifiedOffsets.back()), |
| 374 | constant: innerLaneData); |
| 375 | // Get the 2D data shape of this loadNdOp in its original type including |
| 376 | // array length. |
| 377 | SmallVector<int64_t> origDataShape(origTensorDescType.getShape()); |
| 378 | // Adjust the data shape based on innerLaneData. |
| 379 | origDataShape.back() /= innerLaneData; |
| 380 | // HW supported shape is the new tensor desc shape after conversion. |
| 381 | SmallVector<int64_t> hwSupportedShape(adaptorType.getShape()); |
| 382 | VectorType origVectorType = |
| 383 | VectorType::get(shape: origDataShape, elementType: adaptorType.getElementType()); |
| 384 | Value data; |
| 385 | // Orig data shape is 3D for the array length case. |
| 386 | if (origTensorDescType.getArrayLength() > 1) { |
| 387 | SmallVector<Value> arraySlices; |
| 388 | for (int64_t i = 0; i < origTensorDescType.getArrayLength(); ++i) { |
| 389 | Value slice = arith::ConstantOp::create( |
| 390 | builder&: rewriter, location: loadNdOp->getLoc(), result: origVectorType, |
| 391 | value: rewriter.getZeroAttr(type: origVectorType)); |
| 392 | // Increase the Y offset for each array slice. |
| 393 | Value offsetY = convertToValue(rewriter, loc: loadNdOp->getLoc(), |
| 394 | ofr: modifiedOffsets.back()); |
| 395 | modifiedOffsets.back() = |
| 396 | arith::AddIOp::create( |
| 397 | builder&: rewriter, location: loadNdOp->getLoc(), lhs: offsetY, |
| 398 | rhs: arith::ConstantIndexOp::create(builder&: rewriter, location: loadNdOp->getLoc(), |
| 399 | value: i * origDataShape[1]) |
| 400 | .getResult()) |
| 401 | .getResult(); |
| 402 | slice = generateLoads( |
| 403 | rewriter, data: cast<TypedValue<VectorType>>(Val&: slice), offsets: modifiedOffsets, |
| 404 | newTensorDesc: cast<TypedValue<xegpu::TensorDescType>>(Val: adaptor.getTensorDesc()), |
| 405 | origLoadOp: loadNdOp); |
| 406 | // BitCast back to original load shape without array length. |
| 407 | auto bitcastType = VectorType::get(shape: origTensorDescType.getShape(), |
| 408 | elementType: origTensorDescType.getElementType()); |
| 409 | auto bitCastOp = vector::BitCastOp::create(builder&: rewriter, location: loadNdOp->getLoc(), |
| 410 | result: bitcastType, source: slice); |
| 411 | // BitCastOp must have the same layout as the original loadNdOp. |
| 412 | xegpu::setTemporaryLayout(operandOrResult: bitCastOp->getOpResult(idx: 0), |
| 413 | layout: origTensorDescType.getLayoutAttr()); |
| 414 | arraySlices.push_back(Elt: bitCastOp.getResult()); |
| 415 | } |
| 416 | rewriter.replaceOpWithMultiple(op: loadNdOp, newValues: {arraySlices}); |
| 417 | return success(); |
| 418 | } |
| 419 | data = arith::ConstantOp::create( |
| 420 | builder&: rewriter, location: loadNdOp->getLoc(), |
| 421 | result: VectorType::get(shape: origDataShape, elementType: adaptorType.getElementType()), |
| 422 | value: rewriter.getZeroAttr(type: origVectorType)); |
| 423 | data = generateLoads( |
| 424 | rewriter, data: cast<TypedValue<VectorType>>(Val&: data), offsets: modifiedOffsets, |
| 425 | newTensorDesc: cast<TypedValue<xegpu::TensorDescType>>(Val: adaptor.getTensorDesc()), |
| 426 | origLoadOp: loadNdOp); |
| 427 | auto bitCastOp = vector::BitCastOp::create(builder&: rewriter, location: loadNdOp->getLoc(), |
| 428 | result: loadNdOp.getType(), source: data); |
| 429 | // BitCastOp must have the same layout as the original loadNdOp. |
| 430 | xegpu::setTemporaryLayout(operandOrResult: bitCastOp->getOpResult(idx: 0), |
| 431 | layout: origTensorDescType.getLayoutAttr()); |
| 432 | rewriter.replaceOp(op: loadNdOp, newOp: bitCastOp); |
| 433 | return success(); |
| 434 | } |
| 435 | }; |
| 436 | |
| 437 | /// Vector ExtractOp must be processed if the original tensor desc type has |
| 438 | /// array length greater than 1. In this case, the LoadNdOp is replaced with |
| 439 | /// multiple LoadNdOps for each array slice making the extraction unnecessary. |
| 440 | /// In this case, we simply remove the ExtractOp. |
| 441 | class final |
| 442 | : public OpConversionPattern<vector::ExtractOp> { |
| 443 | public: |
| 444 | using OpConversionPattern<vector::ExtractOp>::OpConversionPattern; |
| 445 | LogicalResult |
| 446 | matchAndRewrite(vector::ExtractOp , OneToNOpAdaptor adaptor, |
| 447 | ConversionPatternRewriter &rewriter) const override { |
| 448 | // Check if the source of the extraction is split to multiple values. |
| 449 | if (adaptor.getSource().size() == 1) |
| 450 | return failure(); |
| 451 | auto mixedPos = extractOp.getMixedPosition(); |
| 452 | if (mixedPos.size() != 1) |
| 453 | return failure(); |
| 454 | auto mayBeInt = getConstantIntValue(ofr: mixedPos[0]); |
| 455 | if (!mayBeInt) |
| 456 | return failure(); |
| 457 | rewriter.replaceOp(op: extractOp, newValues: adaptor.getSource()[*mayBeInt]); |
| 458 | return success(); |
| 459 | } |
| 460 | }; |
| 461 | |
| 462 | /// Performs a reduction over 2 dimensions by decomposing it into two 1D |
| 463 | /// reductions ordered based on layout to minimize cross-lane communication. |
| 464 | class MultiRed2dOpPattern |
| 465 | : public OpConversionPattern<vector::MultiDimReductionOp> { |
| 466 | using Base::Base; |
| 467 | LogicalResult |
| 468 | matchAndRewrite(vector::MultiDimReductionOp reductionOp, OpAdaptor adaptor, |
| 469 | ConversionPatternRewriter &rewriter) const override { |
| 470 | auto sourceVecType = reductionOp.getSourceVectorType(); |
| 471 | if (reductionOp.getReductionDims().size() != 2) |
| 472 | return rewriter.notifyMatchFailure(arg&: reductionOp, msg: "Expected 2D reduction" ); |
| 473 | auto resLayout = xegpu::getDistributeLayoutAttr(value: reductionOp.getResult()); |
| 474 | // Retrieve and order dims for 1D decomposition (prefer intra-lane first). |
| 475 | auto dims = llvm::to_vector(Range: reductionOp.getReductionDims()); |
| 476 | auto [intraLaneDim, crossLaneDim] = getReductionDimOrder(reductionDims: dims, layout: resLayout); |
| 477 | // Order does not matter |
| 478 | if (intraLaneDim == -1 || crossLaneDim == -1) { |
| 479 | intraLaneDim = dims[0]; |
| 480 | crossLaneDim = dims[1]; |
| 481 | } |
| 482 | auto loc = reductionOp.getLoc(); |
| 483 | auto acc = reductionOp.getAcc(); |
| 484 | |
| 485 | // The decomposition below splits the 2D reduction into an intra-lane |
| 486 | // then a cross-lane 1D reduction. The natural result layout of the |
| 487 | // decomposed sequence (a doubly-sliced layout) differs from the |
| 488 | // original 2D reduction's result layout that the rest of the IR was |
| 489 | // written/propagated against. To keep the post-peephole IR |
| 490 | // self-consistent without depending on a follow-up layout |
| 491 | // propagation pass, we always insert a bridge xegpu.convert_layout |
| 492 | // from the natural post-decomposition layout to the original |
| 493 | // reduction's result layout. Trivial bridges fold away in |
| 494 | // canonicalization. |
| 495 | xegpu::DistributeLayoutAttr postDecompLayout; |
| 496 | if (resLayout) { |
| 497 | // Derive the source vector's layout. |
| 498 | xegpu::DistributeLayoutAttr srcLayoutForCvt; |
| 499 | if (auto resSlice = dyn_cast_if_present<xegpu::SliceAttr>(Val&: resLayout)) |
| 500 | srcLayoutForCvt = resSlice.getParent(); |
| 501 | if (!srcLayoutForCvt) |
| 502 | srcLayoutForCvt = |
| 503 | xegpu::getDistributeLayoutAttr(value: reductionOp.getSource()); |
| 504 | if (srcLayoutForCvt) { |
| 505 | // The natural layout of the post-decomposition reduction result |
| 506 | // is a nested SliceAttr: REDUCE_1 (reduces `intraLaneDim` from |
| 507 | // the source) yields `slice<src, [intraLaneDim]>`; REDUCE_2 |
| 508 | // then reduces `adjCrossLaneDim` from that intermediate, giving |
| 509 | // `slice<slice<src, [intraLaneDim]>, [adjCrossLaneDim]>`. |
| 510 | MLIRContext *ctx = reductionOp.getContext(); |
| 511 | int64_t adjCrossLaneDim = |
| 512 | crossLaneDim > intraLaneDim ? crossLaneDim - 1 : crossLaneDim; |
| 513 | auto intermediateLayout = xegpu::SliceAttr::get( |
| 514 | context: ctx, parent: srcLayoutForCvt, dims: DenseI64ArrayAttr::get(context: ctx, content: {intraLaneDim})); |
| 515 | postDecompLayout = xegpu::SliceAttr::get( |
| 516 | context: ctx, parent: intermediateLayout, |
| 517 | dims: DenseI64ArrayAttr::get(context: ctx, content: {adjCrossLaneDim})); |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | SmallVector<int64_t> accShape(sourceVecType.getShape()); |
| 522 | accShape.erase(CI: accShape.begin() + intraLaneDim); |
| 523 | Type eTy = sourceVecType.getElementType(); |
| 524 | Value constNeutralVal = xegpu::createReductionNeutralValue( |
| 525 | builder&: rewriter, loc, type: VectorType::get(shape: accShape, elementType: eTy), kind: reductionOp.getKind(), |
| 526 | fastmath: reductionOp.getFastmathAttr()); |
| 527 | |
| 528 | Value intraLaneReduced = vector::MultiDimReductionOp::create( |
| 529 | builder&: rewriter, location: loc, kind: reductionOp.getKind(), source: reductionOp.getSource(), |
| 530 | acc: constNeutralVal, reduction_dims: ArrayRef<int64_t>(intraLaneDim), |
| 531 | fastmath: reductionOp.getFastmath()); |
| 532 | |
| 533 | // Adjust crossLaneDim after the first reduction. |
| 534 | if (crossLaneDim > intraLaneDim) |
| 535 | crossLaneDim -= 1; |
| 536 | Value crossLaneReduced = vector::MultiDimReductionOp::create( |
| 537 | builder&: rewriter, location: loc, kind: reductionOp.getKind(), source: intraLaneReduced, acc, |
| 538 | reduction_dims: ArrayRef<int64_t>(crossLaneDim), fastmath: reductionOp.getFastmath()); |
| 539 | assert(crossLaneReduced.getType() == reductionOp.getResult().getType() && |
| 540 | "Type mismatch" ); |
| 541 | |
| 542 | Value replacement = crossLaneReduced; |
| 543 | if (resLayout && postDecompLayout) { |
| 544 | // Bridge from the natural post-decomposition layout to the |
| 545 | // original reduction's result layout. This preserves the contract |
| 546 | // any consumer (convert_layout, anchor op, or otherwise) was |
| 547 | // written against, so the rewrite is correct independent of |
| 548 | // whether layout propagation runs afterwards. |
| 549 | auto bridgeOp = xegpu::ConvertLayoutOp::create( |
| 550 | builder&: rewriter, location: loc, result: crossLaneReduced.getType(), source: crossLaneReduced, |
| 551 | input_layout: postDecompLayout, target_layout: resLayout); |
| 552 | replacement = bridgeOp.getResult(); |
| 553 | } |
| 554 | |
| 555 | rewriter.replaceOp(op: reductionOp, newValues: replacement); |
| 556 | return success(); |
| 557 | } |
| 558 | |
| 559 | private: |
| 560 | std::pair<int64_t, int64_t> |
| 561 | getReductionDimOrder(ArrayRef<int64_t> reductionDims, |
| 562 | xegpu::DistributeLayoutAttr layout) const { |
| 563 | assert(layout.isForSubgroup() && "Must know the lane layout" ); |
| 564 | assert(reductionDims.size() == 2 && "Expected 2D reduction" ); |
| 565 | int64_t intra = -1, cross = -1; |
| 566 | xegpu::LayoutAttr layoutAttr = dyn_cast<xegpu::LayoutAttr>(Val&: layout); |
| 567 | if (auto layoutSliceAttr = dyn_cast<xegpu::SliceAttr>(Val&: layout)) |
| 568 | layoutAttr = |
| 569 | dyn_cast<xegpu::LayoutAttr>(Val: layoutSliceAttr.flatten().getParent()); |
| 570 | assert(layoutAttr); |
| 571 | SmallVector<int64_t> laneLayout = layoutAttr.getEffectiveLaneLayoutAsInt(); |
| 572 | |
| 573 | assert(laneLayout.size() && "Expected a non-empty layout" ); |
| 574 | // try to pick a dim that does not communicate |
| 575 | for (auto dim : reductionDims) { |
| 576 | if (laneLayout[dim] == 1) |
| 577 | intra = dim; |
| 578 | else |
| 579 | cross = dim; |
| 580 | } |
| 581 | return {intra, cross}; |
| 582 | } |
| 583 | }; |
| 584 | |
| 585 | } // namespace |
| 586 | |
| 587 | void xegpu::populateXeGPUPeepHoleOptimizerPatterns( |
| 588 | RewritePatternSet &patterns) { |
| 589 | patterns.add<XeGPUCreateNdDescOpPattern, XeGPULoadNdDescOpPattern, |
| 590 | VectorExtractOpPattern, MultiRed2dOpPattern>( |
| 591 | arg: patterns.getContext()); |
| 592 | } |
| 593 | |
| 594 | namespace { |
| 595 | |
| 596 | struct XeGPUPeepHoleOptimizerPass final |
| 597 | : public xegpu::impl::XeGPUPeepHoleOptimizerBase< |
| 598 | XeGPUPeepHoleOptimizerPass> { |
| 599 | void runOnOperation() override { |
| 600 | MLIRContext &context = getContext(); |
| 601 | TypeConverter converter; |
| 602 | RewritePatternSet patterns(&context); |
| 603 | ConversionTarget target(context); |
| 604 | |
| 605 | // This pass is only meant for PVC, BMG or CRI targets. If unsupported |
| 606 | // target is found, exit early. |
| 607 | bool isTargetSupported = false; |
| 608 | getOperation()->walk(callback: [&](gpu::GPUFuncOp funcOp) { |
| 609 | auto chipStr = xegpu::getChipStr(op: funcOp); |
| 610 | if (chipStr && (chipStr.value() == "pvc" || chipStr.value() == "bmg" || |
| 611 | chipStr.value() == "cri" )) |
| 612 | isTargetSupported = true; |
| 613 | }); |
| 614 | |
| 615 | if (!isTargetSupported) { |
| 616 | DBGS() << "XeGPUPeepHoleOptimizerPass only supports PVC, BMG targets." |
| 617 | << "\n" ; |
| 618 | return; |
| 619 | } |
| 620 | |
| 621 | // Run array length optimization patterns first so that subsequent transpose |
| 622 | // peephole patterns operate on the array-length-optimized tensor descs. |
| 623 | { |
| 624 | RewritePatternSet arrayLenPatterns(&context); |
| 625 | xegpu::populateXeGPUArrayLengthOptimizationPatterns(patterns&: arrayLenPatterns); |
| 626 | if (failed(Result: applyPatternsGreedily(op: getOperation(), |
| 627 | patterns: std::move(arrayLenPatterns)))) { |
| 628 | DBGS() << "Array length optimization patterns failed.\n" ; |
| 629 | return signalPassFailure(); |
| 630 | } |
| 631 | } |
| 632 | |
| 633 | // CreateNdDescOp and LoadNdOp with optimizable tensor desc types must be |
| 634 | // converted. |
| 635 | target.addDynamicallyLegalOp<xegpu::CreateNdDescOp>( |
| 636 | callback: [&](xegpu::CreateNdDescOp createNdOp) { |
| 637 | return !canBeOptimizedForTranspose(tdescType: createNdOp.getType()); |
| 638 | }); |
| 639 | target.addDynamicallyLegalOp<xegpu::LoadNdOp>( |
| 640 | callback: [&](xegpu::LoadNdOp loadNdOp) { |
| 641 | return !canBeOptimizedForTranspose(tdescType: loadNdOp.getTensorDescType()); |
| 642 | }); |
| 643 | // Vector ExtractOps can have optimizable layouts if they extract from |
| 644 | // LoadNdOps with array length greater than 1. These ExtractOps must be |
| 645 | // converted. |
| 646 | target.addDynamicallyLegalOp<vector::ExtractOp>( |
| 647 | callback: [&](vector::ExtractOp ) { |
| 648 | auto layout = xegpu::getTemporaryLayout( |
| 649 | operandOrResult: dyn_cast<OpResult>(Val: extractOp.getResult())); |
| 650 | if (!layout) |
| 651 | return true; |
| 652 | auto laneLayout = layout.getEffectiveLaneLayoutAsInt(); |
| 653 | auto laneData = layout.getEffectiveLaneDataAsInt(); |
| 654 | return !canBeOptimizedForTranspose(laneLayout, laneData); |
| 655 | }); |
| 656 | |
| 657 | target.addDynamicallyLegalOp<vector::MultiDimReductionOp>( |
| 658 | callback: [=](Operation *op) -> bool { |
| 659 | auto layout = xegpu::getDistributeLayoutAttr(value: op->getResult(idx: 0)); |
| 660 | if (!layout || !layout.isForSubgroup()) |
| 661 | return true; |
| 662 | if (auto reductionOp = dyn_cast<vector::MultiDimReductionOp>(Val: op)) |
| 663 | return reductionOp.getReductionDims().size() != 2; |
| 664 | return true; |
| 665 | }); |
| 666 | |
| 667 | converter.addConversion(callback: [](Type type) { return type; }); |
| 668 | |
| 669 | target.addLegalDialect<arith::ArithDialect, memref::MemRefDialect, |
| 670 | vector::VectorDialect>(); |
| 671 | // xegpu.convert_layout is left untouched by this pass; mark it legal |
| 672 | // so in-place updates don't trigger re-legalization failures. |
| 673 | target.addLegalOp<xegpu::ConvertLayoutOp>(); |
| 674 | scf::populateSCFStructuralTypeConversionsAndLegality(typeConverter: converter, patterns, |
| 675 | target); |
| 676 | xegpu::populateXeGPUPeepHoleOptimizerPatterns(patterns); |
| 677 | if (failed(Result: applyPartialConversion(op: getOperation(), target, |
| 678 | patterns: std::move(patterns)))) { |
| 679 | DBGS() << "Optimize block loads pass failed.\n" ; |
| 680 | return signalPassFailure(); |
| 681 | } |
| 682 | |
| 683 | // Apply folding for cleaning up IR. |
| 684 | MLIRContext *ctx = &getContext(); |
| 685 | RewritePatternSet emptyPatterns(ctx); |
| 686 | (void)applyPatternsGreedily(op: getOperation(), patterns: std::move(emptyPatterns)); |
| 687 | |
| 688 | xegpu::removeTemporaryLayoutAttrs(op: getOperation()); |
| 689 | } |
| 690 | }; |
| 691 | |
| 692 | } // namespace |
| 693 | |