| 1 | //===- LowerContractToSVEPatterns.cpp - Contract to I8MM/BF16 ---*- 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 | // This file implements lowering patterns from vector.contract to operations |
| 10 | // that map to instructions from the SVE ISA, including matrix-multiplication |
| 11 | // extensions: FEAT_I8MM and FEAT_BF16. |
| 12 | // |
| 13 | // TODO: There may be opportunities to unify this with a similar pattern |
| 14 | // for Neon. See: |
| 15 | // https://github.com/llvm/llvm-project/issues/145559 |
| 16 | // LowerContractToNeonPatterns.cpp |
| 17 | // |
| 18 | //===----------------------------------------------------------------------===// |
| 19 | |
| 20 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 21 | #include "mlir/Dialect/ArmSVE/IR/ArmSVEDialect.h" |
| 22 | #include "mlir/Dialect/ArmSVE/Transforms/Transforms.h" |
| 23 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 24 | #include "mlir/Dialect/LLVMIR/LLVMVectorIntrinsicOps.h" |
| 25 | #include "mlir/Dialect/UB/IR/UBOps.h" |
| 26 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 27 | #include "mlir/IR/AffineMap.h" |
| 28 | #include "mlir/IR/PatternMatch.h" |
| 29 | |
| 30 | #include <cassert> |
| 31 | #include <numeric> |
| 32 | |
| 33 | #define DEBUG_TYPE "lower-contract-to-arm-sve-i8mm" |
| 34 | |
| 35 | using namespace mlir; |
| 36 | using namespace mlir::arm_sve; |
| 37 | |
| 38 | namespace { |
| 39 | // Get the operand of a `vector.contract`. This function is intended to abstract |
| 40 | // away from the particular way a value is extended before feeding it into the |
| 41 | // `vector.contract` - via zero-extend or an explicit or implicit sign-extend |
| 42 | // (for implicit sign-extension see `vector.contract` documentation). |
| 43 | // |
| 44 | // The template parameter `Op` indicates the extension operation (explicit or |
| 45 | // implicit) for which we are checking. |
| 46 | // |
| 47 | // Return success only for extensions from `i8` to `i32`. |
| 48 | template <typename Op> |
| 49 | std::optional<Value> getExtOperand(Value v) { |
| 50 | |
| 51 | static_assert(llvm::is_one_of<Op, arith::ExtSIOp, arith::ExtUIOp>::value, |
| 52 | "Must be instantiated with either sign- or zero- extension op" ); |
| 53 | |
| 54 | // If the operand is not defined by an explicit extend operation of the |
| 55 | // accepted operation type allow for an implicit sign-extension. |
| 56 | auto extOp = v.getDefiningOp<Op>(); |
| 57 | if (!extOp) { |
| 58 | if constexpr (std::is_same<Op, arith::ExtSIOp>::value) { |
| 59 | auto vTy = cast<VectorType>(Val: v.getType()); |
| 60 | if (!vTy.getElementType().isSignlessInteger(width: 8)) |
| 61 | return {}; |
| 62 | return v; |
| 63 | } |
| 64 | return {}; |
| 65 | } |
| 66 | |
| 67 | // If the operand is defined by an explicit extend operation of the accepted |
| 68 | // operation type, check it's extended from `i8` to `i32`. |
| 69 | auto inOp = extOp.getIn(); |
| 70 | auto inTy = dyn_cast<VectorType>(inOp.getType()); |
| 71 | if (!inTy || !inTy.getElementType().isSignlessInteger(8)) |
| 72 | return {}; |
| 73 | |
| 74 | auto outTy = dyn_cast<VectorType>(extOp.getType()); |
| 75 | if (!outTy || !outTy.getElementType().isSignlessInteger(32)) |
| 76 | return {}; |
| 77 | |
| 78 | return inOp; |
| 79 | } |
| 80 | |
| 81 | /// This class encapsulates the algorithm and parametrisation (in terms of types |
| 82 | /// and dimensions) of lowering a `vector.contract` to "primitive" matrix |
| 83 | /// multiplication operations of the SVE dialect (here "primitive" would mean |
| 84 | /// corresponding to a single target instruction). |
| 85 | /// |
| 86 | /// Supported are lowering to FEAT_I8MM `smmla`, `ummla`, and `usmmla`, and to |
| 87 | /// FEAT_BF16 `bfmmla`. All the transformations are very similar to each other |
| 88 | /// for concreteness the description below is given for `smmla`. |
| 89 | /// |
| 90 | /// The lowering triggers for a contraction operation that performs a matrix |
| 91 | /// multiply of two 8-bit integer matrix tiles with logical dimensions |
| 92 | /// <Mx8> and <8x[N]> for the left-hand side (LHS) and the right-hand side |
| 93 | /// (RHS), respectively, added to a 32-bit integer accumulator operand (ACC) |
| 94 | /// with dimensions <Mx[N]>, yielding a <Mx[N]> 32-bit integer result (OUT). |
| 95 | /// |
| 96 | /// The operands' shapes are such that the operands can be evenly split into |
| 97 | /// sub-tiles with dimensions as expected by the targeted FEAT_I8MM |
| 98 | /// instructions. The intent is that M and N are chosen (by higher level |
| 99 | /// transforms) in such a way as to maximise register usage. The main use case |
| 100 | /// we envision as of now is MMT4D, thus the RHS operand is expected |
| 101 | /// pre-transposed. |
| 102 | /// |
| 103 | /// The matrix multiplication is performed by unrolling the usual tiled matrix |
| 104 | /// multiplication algorithm using sub-tiles with dimensions <2x8> for the |
| 105 | /// LHS, <8x[2]> for the RHS, and <2x[2]> for the result and the input |
| 106 | /// accumulator. |
| 107 | /// |
| 108 | /// One way to illustrate the operation is as follows: |
| 109 | /// |
| 110 | /// RHS<8x[N]>: <8x[2]> <8x[2]> ... <8x[2]> |
| 111 | /// +----------------------------- |
| 112 | /// LHS<Mx8>: <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
| 113 | /// <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
| 114 | /// ... | ... ... ... ... |
| 115 | /// <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
| 116 | /// |
| 117 | /// The RHS operand is unpacked into N/2 values, each representing a sequence |
| 118 | /// of VSCALE number of sub-tiles with dimensions <8x2>. |
| 119 | /// The LHS operand is initially unpacked into M/2 values, each representing a |
| 120 | /// sub-tile with dimensions <2x8>, and then each such sub-tile is replicated |
| 121 | /// VSCALE times. Multiplying thus replicated LHS sub-tile by the corresponding |
| 122 | /// RHS sub-tile correctly computes an entire result sub-tile. |
| 123 | /// The 2x2 sub-tiles of the ACC and OUT have rows that are not adjacent |
| 124 | /// (in memory or when imposing a row-major layout on the 2D vector value). |
| 125 | /// Reading the ACC is implemented as reading two consecutive rows and |
| 126 | /// interleaving the by pairs to obtain a vector having length twice the length |
| 127 | /// of an ACC row. This vector now is a sequence of one-dimensional tiles with |
| 128 | /// the exact layout needed by the `smmla`/`bfmmla`/etc instructions, which |
| 129 | /// tiles are extracted one by one. For illustration, if we have an 2x4 ACC tile |
| 130 | /// a0 a1 b0 b1 |
| 131 | /// a2 a3 b2 b3 |
| 132 | /// we read the two rows as separate values and then interleave by pairs |
| 133 | /// to obtain |
| 134 | /// a0 a1 a2 a3 b0 b1 b2 b3 |
| 135 | /// from which we extract `a0 a1 a2 a3` and `b0 b1 b2 b3`. |
| 136 | /// |
| 137 | /// Writing the OUT tile is done by the reverse of the above procedure, |
| 138 | /// concatenate two "flattened" sub-tiles into |
| 139 | /// c0 c1 c2 c3 d0 d1 d2 d3 |
| 140 | /// deinterleave by pairs to obtain as separate values |
| 141 | /// c0 c1 d0 d1 |
| 142 | /// c2 c3 d2 d3 |
| 143 | /// which are then inserted into the final result. |
| 144 | /// |
| 145 | /// Multiplication of a signed LHS by an unsigned LHS is performed by |
| 146 | /// swapping the order of the operands and emitting an `usmmla` (since there |
| 147 | /// isn't an `summla` instruction). Therefore each ACC sub-tile needs |
| 148 | /// to be transposed before the addition and the sum, an OUT sub-tile, |
| 149 | /// needs to be transposed before insertion into the final result. |
| 150 | /// This is done very elegantly by a modification of the above to |
| 151 | /// interleave/deinterleave not by pairs, but by individual elements, e.g. |
| 152 | /// after ordinary interleave we obtain |
| 153 | /// a0 a2 a1 a3 b0 b2 b1 b3 |
| 154 | /// which is exactly the desired layout of having each individual 2x2 tile |
| 155 | /// transposed. |
| 156 | /// |
| 157 | /// All of the above readily applies to FEAT_BF16 `bfmmla` with the |
| 158 | /// difference that the shapes of the LHS, RHS are <Mx4>, <4x[M]>, and |
| 159 | /// respectively, that is the "K" dimension is fixed to 4, instead of 8 (like |
| 160 | /// for the integer case). |
| 161 | class VectorContractRewriter { |
| 162 | protected: |
| 163 | // Designate the operation (resp. instruction) used to do sub-tile matrix |
| 164 | // multiplications. |
| 165 | enum class MMLA { |
| 166 | Nop, |
| 167 | SignedInt, // smmla |
| 168 | UnsignedInt, // ummla |
| 169 | MixedInt, // usmmla |
| 170 | Bfloat // bfmmla |
| 171 | }; |
| 172 | |
| 173 | // Lower-level operation to be emitted. |
| 174 | MMLA mmlaOp = MMLA::Nop; |
| 175 | |
| 176 | // Indicate if the operands for the ArmSVE dialect operation need to be |
| 177 | // swapped. Currently this is needed in order to emulate an "summla" |
| 178 | // operation. |
| 179 | bool swapOperands = false; |
| 180 | |
| 181 | // The operand tiles. These are not necessarily the operends of |
| 182 | // `vector.contract`, for example they could be operands to `arith.extsi` |
| 183 | // that is in turn fed into `vector.contract`. |
| 184 | Value lhs; |
| 185 | Value rhs; |
| 186 | Value acc; |
| 187 | |
| 188 | // Conventional names for matrix dimensions. |
| 189 | int64_t m = 0; |
| 190 | int64_t n = 0; |
| 191 | int64_t k = 0; |
| 192 | |
| 193 | // Create the matrix mulitply and accumulate operation according to |
| 194 | // `mmlaOp`. |
| 195 | Value createMMLA(PatternRewriter &rewriter, Location loc, Value acc, |
| 196 | Value lhs, Value rhs); |
| 197 | |
| 198 | // Check general preconditions for applying the transformation, common to the |
| 199 | // integer and the bfloat16 case. |
| 200 | LogicalResult match(vector::ContractionOp op, PatternRewriter &rewriter); |
| 201 | |
| 202 | public: |
| 203 | VectorContractRewriter() = default; |
| 204 | |
| 205 | // Do the actuall rewrite. This member function is shared by both integer and |
| 206 | // bfloat16 rewrites. |
| 207 | Value lower(vector::ContractionOp op, PatternRewriter &rewriter); |
| 208 | }; |
| 209 | |
| 210 | Value VectorContractRewriter::createMMLA(PatternRewriter &rewriter, |
| 211 | Location loc, Value acc, Value lhs, |
| 212 | Value rhs) { |
| 213 | |
| 214 | Type resTy = acc.getType(); |
| 215 | if (swapOperands) |
| 216 | std::swap(a&: lhs, b&: rhs); |
| 217 | |
| 218 | switch (mmlaOp) { |
| 219 | case MMLA::SignedInt: |
| 220 | return arm_sve::SmmlaOp::create(builder&: rewriter, location: loc, dst: resTy, acc, src1: lhs, src2: rhs); |
| 221 | case MMLA::UnsignedInt: |
| 222 | return arm_sve::UmmlaOp::create(builder&: rewriter, location: loc, dst: resTy, acc, src1: lhs, src2: rhs); |
| 223 | case MMLA::MixedInt: |
| 224 | return arm_sve::UsmmlaOp::create(builder&: rewriter, location: loc, dst: resTy, acc, src1: lhs, src2: rhs); |
| 225 | case MMLA::Bfloat: |
| 226 | return arm_sve::BfmmlaOp::create(builder&: rewriter, location: loc, res: resTy, acc, src1: lhs, src2: rhs); |
| 227 | default: |
| 228 | llvm_unreachable("Uninitialized operation kind" ); |
| 229 | } |
| 230 | } |
| 231 | |
| 232 | LogicalResult VectorContractRewriter::match(vector::ContractionOp op, |
| 233 | PatternRewriter &rewriter) { |
| 234 | // Check iterator types for matrix multiplication. |
| 235 | auto itTypes = op.getIteratorTypesArray(); |
| 236 | if (itTypes.size() != 3 || itTypes[0] != vector::IteratorType::parallel || |
| 237 | itTypes[1] != vector::IteratorType::parallel || |
| 238 | itTypes[2] != vector::IteratorType::reduction) |
| 239 | return rewriter.notifyMatchFailure( |
| 240 | arg&: op, msg: "iterator types do not correspond to matrix multiplication" ); |
| 241 | |
| 242 | // Check permutation maps. For now only accept |
| 243 | // lhs: (d0, d1, d2) -> (d0, d2) |
| 244 | // rhs: (d0, d1, d2) -> (d1, d2) |
| 245 | // acc: (d0, d1, d2) -> (d0, d1) |
| 246 | // This corresponds to matrix multiplication with transposed RHS. |
| 247 | if (op.getIndexingMapsArray()[0] != |
| 248 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{0u, 2u}, |
| 249 | context: op.getContext()) || |
| 250 | op.getIndexingMapsArray()[1] != |
| 251 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{1u, 2u}, |
| 252 | context: op.getContext()) || |
| 253 | op.getIndexingMapsArray()[2] != AffineMap::getMultiDimMapWithTargets( |
| 254 | numDims: 3, targets: ArrayRef{0u, 1u}, context: op.getContext())) |
| 255 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching permutation maps" ); |
| 256 | |
| 257 | // Check the combining kind is addition. |
| 258 | if (op.getKind() != vector::CombiningKind::ADD) |
| 259 | return rewriter.notifyMatchFailure(arg&: op, msg: "combining kind is not an addition" ); |
| 260 | |
| 261 | return success(); |
| 262 | } |
| 263 | |
| 264 | Value VectorContractRewriter::lower(vector::ContractionOp op, |
| 265 | PatternRewriter &rewriter) { |
| 266 | |
| 267 | // Initialize some helper types. |
| 268 | Type operandEltType = cast<VectorType>(Val: lhs.getType()).getElementType(); |
| 269 | Type resultEltType = cast<VectorType>(Val: op.getResultType()).getElementType(); |
| 270 | |
| 271 | const int64_t numOperandSubTileElts = |
| 272 | 128 / operandEltType.getIntOrFloatBitWidth(); |
| 273 | |
| 274 | assert(resultEltType.getIntOrFloatBitWidth() == 32 && |
| 275 | "Only implemented for i32 or f32 output" ); |
| 276 | const int64_t numResultSubTileElts = 4; |
| 277 | |
| 278 | // Single-dimensional vector types for the operands of the ArmSVE dialect |
| 279 | // op. |
| 280 | auto flatLhsType = |
| 281 | VectorType::get(/*shape=*/numOperandSubTileElts, elementType: operandEltType, |
| 282 | /*scalableDims=*/{true}); |
| 283 | auto flatRhsType = |
| 284 | VectorType::get(/*shape=*/numOperandSubTileElts, elementType: operandEltType, |
| 285 | /*scalableDims=*/{true}); |
| 286 | auto flatAccType = |
| 287 | VectorType::get(/*shape=*/numResultSubTileElts, elementType: resultEltType, |
| 288 | /*scalableDims=*/{true}); |
| 289 | |
| 290 | // Single-dimension vector type for the entire RHS tile. |
| 291 | |
| 292 | auto flatRhsTileType = VectorType::get(/*shape=*/k * n, elementType: operandEltType, |
| 293 | /*scalableDims=*/{true}); |
| 294 | |
| 295 | // Vector type having the same number of elements as a row in the |
| 296 | // accumulator/output tile and the same element type. |
| 297 | auto accRowTy = VectorType::get(/*shape=*/n, elementType: resultEltType, |
| 298 | /*scalableDims=*/{true}); |
| 299 | |
| 300 | // Vector type having twice the number of elements as a row in the |
| 301 | // accumulator/output tile the same element type. |
| 302 | auto accRowX2Ty = VectorType::get(/*shape=*/2 * n, elementType: resultEltType, |
| 303 | /*scalableDims=*/{true}); |
| 304 | // Vector type having half the number of elements as a row in the |
| 305 | // accumulator/output tile and an integer element type with twice the bit |
| 306 | // width. |
| 307 | auto accRow64Ty = VectorType::get(/*shape=*/n / 2, elementType: rewriter.getI64Type(), |
| 308 | /*scalableDims=*/{true}); |
| 309 | // Vector type having the same the number of elements as a row in the |
| 310 | // accumulator/output tile and an integer element type with twice the bit |
| 311 | // width. |
| 312 | auto accRowX264Ty = VectorType::get(/*shape=*/n, elementType: rewriter.getI64Type(), |
| 313 | /*scalableDims=*/{true}); |
| 314 | |
| 315 | Location loc = op.getLoc(); |
| 316 | |
| 317 | // Extract LHS sub-tiles with logical shape <2xK>. |
| 318 | SmallVector<Value> lhsTile; |
| 319 | for (int64_t i = 0; i < m; i += 2) { |
| 320 | // Extract two consecutive rows of the LHS tile. |
| 321 | auto r0 = |
| 322 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: lhs, position: ArrayRef<int64_t>{i}); |
| 323 | auto r1 = |
| 324 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: lhs, position: ArrayRef<int64_t>{i + 1}); |
| 325 | // Concatenate to obtain a 2 x K x <input-type> flattened sub-tile. |
| 326 | SmallVector<int64_t> shuffleIdx(2 * k); |
| 327 | std::iota(first: shuffleIdx.begin(), last: shuffleIdx.end(), value: 0); |
| 328 | auto t = vector::ShuffleOp::create(builder&: rewriter, location: loc, v1: r0, v2: r1, mask: shuffleIdx); |
| 329 | // Turn it into a scalable vector. |
| 330 | auto s = vector::ScalableInsertOp::create( |
| 331 | builder&: rewriter, location: loc, valueToStore: t, dest: ub::PoisonOp::create(builder&: rewriter, location: loc, result: flatLhsType), pos: 0); |
| 332 | // Replicate the sub-tile VSCALE times to fill the entire vector. |
| 333 | auto r = arm_sve::DupQLaneOp::create(builder&: rewriter, location: loc, src: s, lane: 0); |
| 334 | lhsTile.push_back(Elt: r); |
| 335 | } |
| 336 | |
| 337 | // "Flatten" the RHS tile from <[N]xK> to <[N*K]>. |
| 338 | auto rhs = vector::ShapeCastOp::create(builder&: rewriter, location: this->rhs.getLoc(), |
| 339 | result: flatRhsTileType, source: this->rhs); |
| 340 | |
| 341 | // Extract the RHS sub-tiles with logical shape <Kx[2]>. |
| 342 | SmallVector<Value> rhsTile; |
| 343 | for (int64_t j = 0; j < n; j += 2) |
| 344 | rhsTile.push_back(Elt: vector::ScalableExtractOp::create( |
| 345 | builder&: rewriter, location: loc, result: flatRhsType, source: rhs, pos: j * k)); |
| 346 | |
| 347 | // Extract and pack the ACC sub-tiles. |
| 348 | SmallVector<Value> accTile; |
| 349 | for (int64_t i = 0; i < m; i += 2) { |
| 350 | // Extract two consecutive rows of the accumulator tile. |
| 351 | auto r0 = vector::ExtractOp::create(builder&: rewriter, location: loc, source: op.getAcc(), |
| 352 | position: ArrayRef<int64_t>{i}); |
| 353 | auto r1 = vector::ExtractOp::create(builder&: rewriter, location: loc, source: op.getAcc(), |
| 354 | position: ArrayRef<int64_t>{i + 1}); |
| 355 | Value accTileVec; |
| 356 | if (swapOperands) { |
| 357 | // We are performing the operation with swapped LHS and RHS we need to |
| 358 | // transpose each individual 2x2 tile of the accumulator and (later) the |
| 359 | // final result. |
| 360 | accTileVec = vector::InterleaveOp::create(builder&: rewriter, location: loc, lhs: r0, rhs: r1); |
| 361 | } else { |
| 362 | // Bitcast accumulator rows to double-width integer elements, so |
| 363 | // subsequent interleave/deinterleave work on pairs of elements. |
| 364 | auto r0I64 = vector::BitCastOp::create(builder&: rewriter, location: loc, result: accRow64Ty, source: r0); |
| 365 | auto r1I64 = vector::BitCastOp::create(builder&: rewriter, location: loc, result: accRow64Ty, source: r1); |
| 366 | |
| 367 | // Interleave the rows, effectively flattening each 2x2 tile into 4 |
| 368 | // consecutive elements. |
| 369 | auto intrI64 = vector::InterleaveOp::create(builder&: rewriter, location: loc, lhs: r0I64, rhs: r1I64); |
| 370 | |
| 371 | // Bitcast back to original element type. |
| 372 | accTileVec = |
| 373 | vector::BitCastOp::create(builder&: rewriter, location: loc, result: accRowX2Ty, source: intrI64); |
| 374 | } |
| 375 | // Extract ACC sub-tiles. |
| 376 | for (int64_t j = 0; j < n; j += 2) |
| 377 | accTile.push_back(Elt: vector::ScalableExtractOp::create( |
| 378 | builder&: rewriter, location: loc, result: flatAccType, source: accTileVec, pos: j * 2)); |
| 379 | } |
| 380 | |
| 381 | // Emit sub-tile matrix multiplications. |
| 382 | SmallVector<Value> outTile; |
| 383 | for (int64_t i = 0; i < m / 2; ++i) |
| 384 | for (int64_t j = 0; j < n / 2; ++j) { |
| 385 | Value mmla = createMMLA(rewriter, loc, acc: accTile[i * n / 2 + j], lhs: lhsTile[i], |
| 386 | rhs: rhsTile[j]); |
| 387 | outTile.push_back(Elt: mmla); |
| 388 | } |
| 389 | |
| 390 | // Unpack the OUT sub-tiles and insert into the result. |
| 391 | Value result = ub::PoisonOp::create(builder&: rewriter, location: loc, result: op.getResultType()); |
| 392 | for (int64_t i = 0; i < m / 2; ++i) { |
| 393 | // Collect a number of sub-tiles in a row. |
| 394 | Value row = ub::PoisonOp::create(builder&: rewriter, location: loc, result: accRowX2Ty); |
| 395 | for (int64_t j = 0; j < n / 2; ++j) |
| 396 | row = vector::ScalableInsertOp::create( |
| 397 | builder&: rewriter, location: loc, valueToStore: outTile[i * n / 2 + j], dest: row, pos: j * 4); |
| 398 | |
| 399 | // Unpack the row to obtain two rows of the output. If we have the out |
| 400 | // sub-tiles transposed we obtain two consecutive output rows by |
| 401 | // separating even and odd elements, i.e. a simple deinterleave. |
| 402 | // Otherwise, the interleave is by pairs. |
| 403 | Value out0, out1; |
| 404 | if (swapOperands) { |
| 405 | auto tmp = vector::DeinterleaveOp::create(builder&: rewriter, location: loc, source: row); |
| 406 | out0 = tmp.getResultEven(); |
| 407 | out1 = tmp.getResultOdd(); |
| 408 | } else { |
| 409 | // Deinterleave by pairs. |
| 410 | auto row64 = vector::BitCastOp::create(builder&: rewriter, location: loc, result: accRowX264Ty, source: row); |
| 411 | auto deintr64 = vector::DeinterleaveOp::create(builder&: rewriter, location: loc, source: row64); |
| 412 | |
| 413 | // Bitcast back into original element type and insert into the result. |
| 414 | out0 = vector::BitCastOp::create(builder&: rewriter, location: loc, result: accRowTy, |
| 415 | source: deintr64.getResultEven()); |
| 416 | out1 = vector::BitCastOp::create(builder&: rewriter, location: loc, result: accRowTy, |
| 417 | source: deintr64.getResultOdd()); |
| 418 | } |
| 419 | result = vector::InsertOp::create(builder&: rewriter, location: loc, valueToStore: out0, dest: result, position: i * 2); |
| 420 | result = vector::InsertOp::create(builder&: rewriter, location: loc, valueToStore: out1, dest: result, position: i * 2 + 1); |
| 421 | } |
| 422 | |
| 423 | return result; |
| 424 | } |
| 425 | |
| 426 | class VectorContractRewriterI8MM : public VectorContractRewriter { |
| 427 | public: |
| 428 | // Check the specific preconditions for the integer case. Initialise |
| 429 | // parametrisation types and dimensions. |
| 430 | LogicalResult matchAndInit(vector::ContractionOp op, |
| 431 | PatternRewriter &rewriter) { |
| 432 | if (failed(Result: match(op, rewriter))) |
| 433 | return failure(); |
| 434 | |
| 435 | VectorType lhsType = op.getLhsType(); |
| 436 | VectorType rhsType = op.getRhsType(); |
| 437 | |
| 438 | m = lhsType.getDimSize(idx: 0); |
| 439 | n = rhsType.getDimSize(idx: 0); |
| 440 | k = rhsType.getDimSize(idx: 1); |
| 441 | |
| 442 | // Check the operands have the expected shape: |
| 443 | // * for LHS: fixed vector MxK |
| 444 | // * for RHS: scalable vector [N]xK |
| 445 | // * K == 8 |
| 446 | // * M and N even and at least 2 |
| 447 | if (lhsType.isScalable() || !rhsType.getScalableDims()[0] || |
| 448 | rhsType.getScalableDims()[1] || lhsType.getDimSize(idx: 1) != k || k != 8 || |
| 449 | m < 2 || m % 2 != 0 || n < 2 || n % 2 != 0) |
| 450 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching operand shape" ); |
| 451 | |
| 452 | // Check the output is a vector of i32 elements. |
| 453 | auto outTy = dyn_cast<VectorType>(Val: op.getResultType()); |
| 454 | if (!outTy || outTy.getElementType() != rewriter.getI32Type()) |
| 455 | return rewriter.notifyMatchFailure(arg&: op, |
| 456 | msg: "output type is not a vector of i32" ); |
| 457 | |
| 458 | // Check inputs are sign-/zero- extensions from i8 to i32. Get the values |
| 459 | // before the extension. All four signed/unsigned combinations for input |
| 460 | // operands are supported, but they are lowered to different operations. |
| 461 | // Determine which is the appropriate operation to lower to. |
| 462 | mmlaOp = MMLA::SignedInt; |
| 463 | swapOperands = false; |
| 464 | auto maybeLhs = getExtOperand<arith::ExtSIOp>(v: op.getLhs()); |
| 465 | if (!maybeLhs) { |
| 466 | mmlaOp = MMLA::UnsignedInt; |
| 467 | maybeLhs = getExtOperand<arith::ExtUIOp>(v: op.getLhs()); |
| 468 | } |
| 469 | if (!maybeLhs) |
| 470 | return rewriter.notifyMatchFailure( |
| 471 | arg&: op, msg: "LHS is not a sign- or zero- extended i8" ); |
| 472 | |
| 473 | auto maybeRhs = getExtOperand<arith::ExtSIOp>(v: op.getRhs()); |
| 474 | if (maybeRhs) { |
| 475 | if (mmlaOp == MMLA::UnsignedInt) |
| 476 | mmlaOp = MMLA::MixedInt; |
| 477 | } else { |
| 478 | if (mmlaOp == MMLA::SignedInt) { |
| 479 | mmlaOp = MMLA::MixedInt; |
| 480 | swapOperands = true; |
| 481 | } |
| 482 | maybeRhs = getExtOperand<arith::ExtUIOp>(v: op.getRhs()); |
| 483 | } |
| 484 | if (!maybeRhs) |
| 485 | return rewriter.notifyMatchFailure( |
| 486 | arg&: op, msg: "RHS is not a sign- or zero- extended i8" ); |
| 487 | |
| 488 | // Initialise algorithm parameters. |
| 489 | lhs = *maybeLhs; |
| 490 | rhs = *maybeRhs; |
| 491 | acc = op.getAcc(); |
| 492 | |
| 493 | return success(); |
| 494 | } |
| 495 | }; |
| 496 | |
| 497 | class VectorContractRewriterBfloat : public VectorContractRewriter { |
| 498 | public: |
| 499 | // Check the specific preconditions for the bfloat16 case. Initialise |
| 500 | // parametrisation types and dimensions. |
| 501 | LogicalResult matchAndInit(vector::ContractionOp op, |
| 502 | PatternRewriter &rewriter) { |
| 503 | if (failed(Result: match(op, rewriter))) |
| 504 | return failure(); |
| 505 | |
| 506 | VectorType lhsType = op.getLhsType(); |
| 507 | VectorType rhsType = op.getRhsType(); |
| 508 | |
| 509 | m = lhsType.getDimSize(idx: 0); |
| 510 | n = rhsType.getDimSize(idx: 0); |
| 511 | k = rhsType.getDimSize(idx: 1); |
| 512 | |
| 513 | // Check the operands have the expected shape: |
| 514 | // * for LHS: fixed vector MxK |
| 515 | // * for RHS: scalable vector [N]xK |
| 516 | // * K == 4 |
| 517 | // * M and N even and at least 2 |
| 518 | if (lhsType.isScalable() || !rhsType.getScalableDims()[0] || |
| 519 | rhsType.getScalableDims()[1] || lhsType.getDimSize(idx: 1) != k || k != 4 || |
| 520 | m < 2 || m % 2 != 0 || n < 2 || n % 2 != 0) |
| 521 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching operand shape" ); |
| 522 | |
| 523 | // Check the output is a vector of Float32 elements. |
| 524 | auto outTy = dyn_cast<VectorType>(Val: op.getResultType()); |
| 525 | if (!outTy || outTy.getElementType() != rewriter.getF32Type()) |
| 526 | return rewriter.notifyMatchFailure(arg&: op, |
| 527 | msg: "output type is not a vector of f32" ); |
| 528 | |
| 529 | // Check the inputs are vectors of BFloat16 elements. |
| 530 | if (lhsType.getElementType() != rewriter.getBF16Type()) |
| 531 | return rewriter.notifyMatchFailure(arg&: op, |
| 532 | msg: "input type is not a vector of bf16" ); |
| 533 | |
| 534 | // Initialise algorithm parameters. |
| 535 | mmlaOp = MMLA::Bfloat; |
| 536 | swapOperands = false; |
| 537 | lhs = op.getLhs(); |
| 538 | rhs = op.getRhs(); |
| 539 | acc = op.getAcc(); |
| 540 | |
| 541 | return success(); |
| 542 | } |
| 543 | }; |
| 544 | |
| 545 | class LowerContractionToSVEI8MMPattern |
| 546 | : public OpRewritePattern<vector::ContractionOp> { |
| 547 | public: |
| 548 | using Base::Base; |
| 549 | LogicalResult matchAndRewrite(vector::ContractionOp op, |
| 550 | PatternRewriter &rewriter) const override { |
| 551 | |
| 552 | // Match i8xi8 -> i32 matrix multiply and accumulate. |
| 553 | VectorContractRewriterI8MM vcr; |
| 554 | if (failed(Result: vcr.matchAndInit(op, rewriter))) |
| 555 | return failure(); |
| 556 | |
| 557 | Value result = vcr.lower(op, rewriter); |
| 558 | rewriter.replaceOp(op, newValues: result); |
| 559 | |
| 560 | return success(); |
| 561 | } |
| 562 | }; |
| 563 | |
| 564 | class LowerContractionToSVEBFMMLAPattern |
| 565 | : public OpRewritePattern<vector::ContractionOp> { |
| 566 | public: |
| 567 | using Base::Base; |
| 568 | LogicalResult matchAndRewrite(vector::ContractionOp op, |
| 569 | PatternRewriter &rewriter) const override { |
| 570 | |
| 571 | // Match bf16xbf16 -> f32 matrix multiply and accumulate. |
| 572 | VectorContractRewriterBfloat vcr; |
| 573 | if (failed(Result: vcr.matchAndInit(op, rewriter))) |
| 574 | return failure(); |
| 575 | |
| 576 | Value result = vcr.lower(op, rewriter); |
| 577 | rewriter.replaceOp(op, newValues: result); |
| 578 | |
| 579 | return success(); |
| 580 | } |
| 581 | }; |
| 582 | |
| 583 | /// A helper for LowerContractionToSVEIndexedFMLAPatterns to check the |
| 584 | /// preconditions. Check the pattern for an overview of the conditions to |
| 585 | /// check. |
| 586 | static LogicalResult matchIndexed(vector::ContractionOp op, |
| 587 | PatternRewriter &rewriter) { |
| 588 | // Check iterator types for matrix multiplication. |
| 589 | // TODO: Extract to a helper hook. |
| 590 | auto itTypes = op.getIteratorTypesArray(); |
| 591 | if (itTypes.size() != 3 || itTypes[0] != vector::IteratorType::parallel || |
| 592 | itTypes[1] != vector::IteratorType::parallel || |
| 593 | itTypes[2] != vector::IteratorType::reduction) |
| 594 | return rewriter.notifyMatchFailure( |
| 595 | arg&: op, msg: "iterator types do not correspond to matrix multiplication" ); |
| 596 | |
| 597 | // Verify that the permutation maps represent a plain matmul: |
| 598 | // lhs: (d0, d1, d2) -> (d0, d2) |
| 599 | // rhs: (d0, d1, d2) -> (d1, d2) |
| 600 | // acc: (d0, d1, d2) -> (d0, d1) |
| 601 | // TODO: Extract to a helper hook. |
| 602 | if (op.getIndexingMapsArray()[0] != |
| 603 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{0u, 2u}, |
| 604 | context: op.getContext()) || |
| 605 | op.getIndexingMapsArray()[1] != |
| 606 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{2u, 1u}, |
| 607 | context: op.getContext()) || |
| 608 | op.getIndexingMapsArray()[2] != AffineMap::getMultiDimMapWithTargets( |
| 609 | numDims: 3, targets: ArrayRef{0u, 1u}, context: op.getContext())) |
| 610 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching permutation maps" ); |
| 611 | |
| 612 | // Check the combining kind is addition. |
| 613 | // TODO: Extract to a helper hook. |
| 614 | if (op.getKind() != vector::CombiningKind::ADD) |
| 615 | return rewriter.notifyMatchFailure(arg&: op, msg: "combining kind is not an addition" ); |
| 616 | |
| 617 | // Check the operands have the supported shape: |
| 618 | // * LHS: <1 x 4> |
| 619 | // * RHS: <4 x [8]> |
| 620 | // * ACC: <1 x [8]> |
| 621 | VectorType lhsType = op.getLhsType(); |
| 622 | VectorType rhsType = op.getRhsType(); |
| 623 | auto m = lhsType.getDimSize(idx: 0); |
| 624 | auto n = rhsType.getDimSize(idx: 1); |
| 625 | auto k = rhsType.getDimSize(idx: 0); |
| 626 | |
| 627 | if (rhsType.getScalableDims()[0] || !rhsType.getScalableDims()[1] || |
| 628 | lhsType.isScalable() || lhsType.getDimSize(idx: 1) != k || k != 4 || m != 1 || |
| 629 | n != 8) |
| 630 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching operand shape" ); |
| 631 | |
| 632 | // Check that output element type is Float32. |
| 633 | auto outTy = dyn_cast<VectorType>(Val: op.getResultType()); |
| 634 | if (!outTy || outTy.getElementType() != rewriter.getF32Type()) |
| 635 | return rewriter.notifyMatchFailure(arg&: op, |
| 636 | msg: "output type is not a vector of f32" ); |
| 637 | |
| 638 | // Check that the input element type is Float32. |
| 639 | if (lhsType.getElementType() != rewriter.getF32Type()) |
| 640 | return rewriter.notifyMatchFailure(arg&: op, |
| 641 | msg: "input type is not a vector of bf16" ); |
| 642 | |
| 643 | return success(); |
| 644 | } |
| 645 | |
| 646 | /// Lowers a matching contraction operation to SVE's "indexed" FMLAs (FP32) |
| 647 | /// |
| 648 | /// The input contraction must represent a plain matrix-multiplication with the |
| 649 | /// following dimensions: |
| 650 | /// |
| 651 | /// vector.contract #matmul_trait %lhs, %rhs, %acc |
| 652 | /// : vector<1x4xf32>, vector<4x[8]xf32> into vector<1x[8]xf32> |
| 653 | /// |
| 654 | /// It is lowered to a sequence of indexed FMLAs, one per reduction dimension |
| 655 | /// (pseudo-code illustration): |
| 656 | /// |
| 657 | /// %acc = fmla %acc, %rhs, %lhs[0] |
| 658 | /// %acc = fmla %acc, %rhs, %lhs[1] |
| 659 | /// %acc = fmla %acc, %rhs, %lhs[2] |
| 660 | /// %acc = fmla %acc, %rhs, %lhs[3] |
| 661 | /// |
| 662 | /// Note that the relative order of `%rhs` and `%lhs` is reversed to match the |
| 663 | /// SVE Op semantics. |
| 664 | /// |
| 665 | /// The actual output is a bit more complex - input vectors |
| 666 | /// (e.g. vector<[8]xf32>) are broken down into valid SVE vectors (e.g. |
| 667 | /// vector<[4]xf32>) that the FMLA Op expects. |
| 668 | class LowerContractionToSVEIndexedFMLAPatterns |
| 669 | : public OpRewritePattern<vector::ContractionOp> { |
| 670 | using Base::Base; |
| 671 | |
| 672 | public: |
| 673 | LogicalResult matchAndRewrite(vector::ContractionOp op, |
| 674 | PatternRewriter &rewriter) const override { |
| 675 | // Check the specific preconditions. |
| 676 | if (failed(Result: matchIndexed(op, rewriter))) |
| 677 | return failure(); |
| 678 | |
| 679 | auto loc = op.getLoc(); |
| 680 | auto acc = op.getAcc(); |
| 681 | |
| 682 | // Legal type for indexed FMLA: |
| 683 | // * vector<[4]xf32> rather than the input vector<[8]xf32> |
| 684 | mlir::VectorType legalType = |
| 685 | VectorType::get(shape: {4}, elementType: rewriter.getF32Type(), scalableDims: {true}); |
| 686 | |
| 687 | // Prepare LHS for indexed FMLA (2D -> 1D): |
| 688 | // * vector<1x4xf32> -> vector<[4]xf32> |
| 689 | auto lhs1D = vector::ExtractOp::create(builder&: rewriter, location: loc, source: op.getLhs(), position: 0); |
| 690 | auto repeatedTy = |
| 691 | VectorType::get(shape: {4}, elementType: rewriter.getF32Type(), /*scalableDims=*/{true}); |
| 692 | Value lhsScalable = |
| 693 | LLVM::vector_repeat::create(builder&: rewriter, location: loc, res: repeatedTy, vec: lhs1D) |
| 694 | .getResult(); |
| 695 | |
| 696 | // Prepare ACC for indexed FMLA (split into legal vectors): |
| 697 | // vector<1x[8]xf32> -> [vector<[4]xf32>, vector<[4]xf32>] |
| 698 | auto accRow = vector::ExtractOp::create(builder&: rewriter, location: loc, source: acc, position: 0); |
| 699 | auto acc0 = |
| 700 | vector::ScalableExtractOp::create(builder&: rewriter, location: loc, result: legalType, source: accRow, pos: 0) |
| 701 | .getResult(); |
| 702 | auto acc1 = |
| 703 | vector::ScalableExtractOp::create(builder&: rewriter, location: loc, result: legalType, source: accRow, pos: 4) |
| 704 | .getResult(); |
| 705 | |
| 706 | // Loop over the reduction lanes |
| 707 | for (size_t laneIdx = 0; laneIdx < 4; ++laneIdx) { |
| 708 | auto lane = arith::ConstantIntOp::create(builder&: rewriter, location: loc, value: laneIdx, width: 32); |
| 709 | |
| 710 | // Prepare RHS for indexed FMLA (select one row and split): |
| 711 | // vector<4x[8]xf32> -> [vector<[4]xf32>, vector<[4]xf32>] |
| 712 | auto rhsRow = |
| 713 | vector::ExtractOp::create(builder&: rewriter, location: loc, source: op.getRhs(), position: laneIdx); |
| 714 | auto rhs0 = vector::ScalableExtractOp::create(builder&: rewriter, location: loc, result: legalType, |
| 715 | source: rhsRow, pos: 0); |
| 716 | auto rhs1 = vector::ScalableExtractOp::create(builder&: rewriter, location: loc, result: legalType, |
| 717 | source: rhsRow, pos: 4); |
| 718 | |
| 719 | // FMLAs - note that the LHS and RHS operands from vector.contract are |
| 720 | // passed in the reverse order: |
| 721 | // %acc = fmla %acc, %rhs, %lhs[0] |
| 722 | acc0 = arm_sve::FmlaOp::create(builder&: rewriter, location: loc, dst: legalType, acc: acc0, src1: rhs0, |
| 723 | src2: lhsScalable, lane) |
| 724 | .getResult(); |
| 725 | acc1 = arm_sve::FmlaOp::create(builder&: rewriter, location: loc, dst: legalType, acc: acc1, src1: rhs1, |
| 726 | src2: lhsScalable, lane) |
| 727 | .getResult(); |
| 728 | } |
| 729 | |
| 730 | // Store the result back |
| 731 | acc = vector::ScalableInsertOp::create(builder&: rewriter, location: loc, valueToStore: acc0, dest: accRow, pos: 0); |
| 732 | acc = vector::ScalableInsertOp::create(builder&: rewriter, location: loc, valueToStore: acc1, dest: acc, pos: 4); |
| 733 | |
| 734 | rewriter.replaceOp(op, newOp: vector::ShapeCastOp::create( |
| 735 | builder&: rewriter, location: loc, result: op.getResultType(), source: acc)); |
| 736 | |
| 737 | return success(); |
| 738 | } |
| 739 | }; |
| 740 | } // namespace |
| 741 | |
| 742 | void mlir::arm_sve::populateLowerContractionToSVEI8MMPatterns( |
| 743 | RewritePatternSet &patterns) { |
| 744 | MLIRContext *context = patterns.getContext(); |
| 745 | patterns.add<LowerContractionToSVEI8MMPattern>(arg&: context, /*benefit=*/args: 2); |
| 746 | } |
| 747 | |
| 748 | void mlir::arm_sve::populateLowerContractionToSVEBFMMLAPatterns( |
| 749 | RewritePatternSet &patterns) { |
| 750 | MLIRContext *context = patterns.getContext(); |
| 751 | patterns.add<LowerContractionToSVEBFMMLAPattern>(arg&: context, /*benefit=*/args: 2); |
| 752 | } |
| 753 | |
| 754 | void mlir::arm_sve::populateLowerContractionToSVEIndexedFMLAPatterns( |
| 755 | RewritePatternSet &patterns) { |
| 756 | MLIRContext *context = patterns.getContext(); |
| 757 | patterns.add<LowerContractionToSVEIndexedFMLAPatterns>(arg&: context, |
| 758 | /*benefit=*/args: 2); |
| 759 | } |
| 760 | |