| 1 | //===- LowerContractionToSVEI8MMPattern.cpp - Contract to I8MM --*- 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 FEAT_I8MM extension. |
| 11 | // |
| 12 | // TODO: There may be opportunities to unify this with a similar pattern |
| 13 | // for Neon. See: |
| 14 | // https://github.com/llvm/llvm-project/issues/145559 |
| 15 | // LowerContractionToNeonI8MMPattern.cpp |
| 16 | // |
| 17 | //===----------------------------------------------------------------------===// |
| 18 | |
| 19 | #include "mlir/Dialect/Arith/IR/Arith.h" |
| 20 | #include "mlir/Dialect/ArmSVE/IR/ArmSVEDialect.h" |
| 21 | #include "mlir/Dialect/ArmSVE/Transforms/Transforms.h" |
| 22 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
| 23 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
| 24 | #include "mlir/IR/AffineMap.h" |
| 25 | #include "mlir/IR/PatternMatch.h" |
| 26 | |
| 27 | #include "mlir/Dialect/UB/IR/UBOps.h" |
| 28 | |
| 29 | #define DEBUG_TYPE "lower-contract-to-arm-sve-i8mm" |
| 30 | |
| 31 | using namespace mlir; |
| 32 | |
| 33 | namespace { |
| 34 | // Get the operand of a `vector.contract`. This function is intended to abstract |
| 35 | // away from the particular way a value is extended before feeding it into the |
| 36 | // `vector.contract` - via zero-extend or an explicit or implicit sign-extend |
| 37 | // (for implicit sign-extension see `vector.contract` documentation). |
| 38 | // |
| 39 | // The template parameter `Op` indicates the extension operation (explicit or |
| 40 | // implicit) for which we are checking. |
| 41 | // |
| 42 | // Return success only for extensions from `i8` to `i32`. |
| 43 | template <typename Op> |
| 44 | std::optional<Value> getExtOperand(Value v) { |
| 45 | |
| 46 | static_assert(llvm::is_one_of<Op, arith::ExtSIOp, arith::ExtUIOp>::value, |
| 47 | "Must be instantiated with either sign- or zero- extension op" ); |
| 48 | |
| 49 | // If the operand is not defined by an explicit extend operation of the |
| 50 | // accepted operation type allow for an implicit sign-extension. |
| 51 | auto extOp = dyn_cast_or_null<Op>(v.getDefiningOp()); |
| 52 | if (!extOp) { |
| 53 | if constexpr (std::is_same<Op, arith::ExtSIOp>::value) { |
| 54 | auto vTy = cast<VectorType>(Val: v.getType()); |
| 55 | if (!vTy.getElementType().isSignlessInteger(width: 8)) |
| 56 | return {}; |
| 57 | return v; |
| 58 | } |
| 59 | return {}; |
| 60 | } |
| 61 | |
| 62 | // If the operand is defined by an explicit extend operation of the accepted |
| 63 | // operation type, check it's extended from `i8` to `i32`. |
| 64 | auto inOp = extOp.getIn(); |
| 65 | auto inTy = dyn_cast<VectorType>(inOp.getType()); |
| 66 | if (!inTy || !inTy.getElementType().isSignlessInteger(8)) |
| 67 | return {}; |
| 68 | |
| 69 | auto outTy = dyn_cast<VectorType>(extOp.getType()); |
| 70 | if (!outTy || !outTy.getElementType().isSignlessInteger(32)) |
| 71 | return {}; |
| 72 | |
| 73 | return inOp; |
| 74 | } |
| 75 | |
| 76 | // Designate the operation (resp. instruction) used to do sub-tile matrix |
| 77 | // multiplications. |
| 78 | enum class MMLA { |
| 79 | Signed, // smmla |
| 80 | Unsigned, // ummla |
| 81 | Mixed, // usmmla |
| 82 | MixedSwapped // usmmla with LHS and RHS swapped |
| 83 | }; |
| 84 | |
| 85 | // Create the matrix mulitply and accumulate operation according to `op`. |
| 86 | Value createMMLA(PatternRewriter &rewriter, MMLA op, Location loc, |
| 87 | mlir::VectorType accType, Value acc, Value lhs, Value rhs) { |
| 88 | switch (op) { |
| 89 | case MMLA::Signed: |
| 90 | return rewriter.create<arm_sve::SmmlaOp>(location: loc, args&: accType, args&: acc, args&: lhs, args&: rhs); |
| 91 | case MMLA::Unsigned: |
| 92 | return rewriter.create<arm_sve::UmmlaOp>(location: loc, args&: accType, args&: acc, args&: lhs, args&: rhs); |
| 93 | case MMLA::Mixed: |
| 94 | return rewriter.create<arm_sve::UsmmlaOp>(location: loc, args&: accType, args&: acc, args&: lhs, args&: rhs); |
| 95 | case MMLA::MixedSwapped: |
| 96 | // The accumulator comes transposed and the result will be transposed |
| 97 | // later, so all we have to do here is swap the operands. |
| 98 | return rewriter.create<arm_sve::UsmmlaOp>(location: loc, args&: accType, args&: acc, args&: rhs, args&: lhs); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | /// Lower a contraction operation that performs a matrix multiplication |
| 103 | /// of two 8-bit integer matrix tiles with logical dimensions <Mx8> and <8x[N]> |
| 104 | /// for the left-hand side and the right-hand side, respectively, |
| 105 | /// yielding a <Mx[N]> 32-bit integer result. |
| 106 | /// |
| 107 | /// The operands' shapes are such that the operands can be evenly split into |
| 108 | /// sub-tiles with dimensions as expected by the targeted FEAT_I8MM |
| 109 | /// instructions. The intent is that M and N are chosen (by higher level |
| 110 | /// transforms) in such a way as to maximise register usage. The main use case |
| 111 | /// we envision as of now is MMT4D, thus the RHS operand is expected |
| 112 | /// pre-transposed. |
| 113 | /// |
| 114 | /// The matrix multiplication is performed by unrolling the usual tiled matrix |
| 115 | /// multiplication algorithm using sub-tiles with dimensions <2x8> for the LHS, |
| 116 | /// <8x[2]> for the RHS, and <2x[2]> for the result and the input accumulator. |
| 117 | /// |
| 118 | /// One way to illustrate the operation is as follows: |
| 119 | /// |
| 120 | /// RHS<8x[N]>: <8x[2]> <8x[2]> ... <8x[2]> |
| 121 | /// +----------------------------- |
| 122 | /// LHS<Mx8>: <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
| 123 | /// <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
| 124 | /// ... | ... ... ... ... |
| 125 | /// <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
| 126 | /// |
| 127 | /// The RHS operand is unpacked into N/2 values, each representing a sequence of |
| 128 | /// VSCALE number of sub-tiles with dimensions <8x2>. |
| 129 | /// The LHS operand is initially unpacked into M/2 values, each representing a |
| 130 | /// sub-tile with dimensions <2x8>, and then each such sub-tile is replicated |
| 131 | /// VSCALE times. |
| 132 | /// Multiplying thus replicated LHS sub-tile by the corresponding RHS sub-tile |
| 133 | /// correctly computes an entire result sub-tile. |
| 134 | class LowerContractionToSVEI8MMPattern |
| 135 | : public OpRewritePattern<vector::ContractionOp> { |
| 136 | public: |
| 137 | using OpRewritePattern::OpRewritePattern; |
| 138 | LogicalResult matchAndRewrite(vector::ContractionOp op, |
| 139 | PatternRewriter &rewriter) const override { |
| 140 | |
| 141 | Location loc = op.getLoc(); |
| 142 | mlir::VectorType lhsType = op.getLhsType(); |
| 143 | mlir::VectorType rhsType = op.getRhsType(); |
| 144 | |
| 145 | // Check the rank the types so we can safely examine their dimensions. |
| 146 | if (lhsType.getRank() != 2 || rhsType.getRank() != 2) |
| 147 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching operand shape" ); |
| 148 | |
| 149 | auto M = lhsType.getDimSize(idx: 0); |
| 150 | auto N = rhsType.getDimSize(idx: 0); |
| 151 | auto K = rhsType.getDimSize(idx: 1); |
| 152 | |
| 153 | // Check the operands have the expected shape: |
| 154 | // * for LHS: fixed vector MxK |
| 155 | // * for RHS: scalable vector [N]xK |
| 156 | // * K == 8 |
| 157 | // * M and N even and at least 2 |
| 158 | if (lhsType.isScalable() || !rhsType.getScalableDims()[0] || |
| 159 | rhsType.getScalableDims()[1] || lhsType.getDimSize(idx: 1) != K || K != 8 || |
| 160 | M < 2 || M % 2 != 0 || N < 2 || N % 2 != 0 || |
| 161 | !rhsType.getScalableDims()[0]) |
| 162 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching operand shape" ); |
| 163 | |
| 164 | // Check permutation maps. For now only accept |
| 165 | // lhs: (d0, d1, d2) -> (d0, d2) |
| 166 | // rhs: (d0, d1, d2) -> (d1, d2) |
| 167 | // acc: (d0, d1, d2) -> (d0, d1) |
| 168 | // This corresponds to matrix multiplication with transposed RHS. |
| 169 | if (op.getIndexingMapsArray()[0] != |
| 170 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{0u, 2u}, |
| 171 | context: op.getContext()) || |
| 172 | op.getIndexingMapsArray()[1] != |
| 173 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{1u, 2u}, |
| 174 | context: op.getContext()) || |
| 175 | op.getIndexingMapsArray()[2] != |
| 176 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{0u, 1u}, |
| 177 | context: op.getContext())) |
| 178 | return rewriter.notifyMatchFailure(arg&: op, msg: "non-matching permutation maps" ); |
| 179 | |
| 180 | // Check iterator types for matrix multiplication. |
| 181 | auto itTypes = op.getIteratorTypesArray(); |
| 182 | if (itTypes.size() != 3 || itTypes[0] != vector::IteratorType::parallel || |
| 183 | itTypes[1] != vector::IteratorType::parallel || |
| 184 | itTypes[2] != vector::IteratorType::reduction) |
| 185 | return rewriter.notifyMatchFailure( |
| 186 | arg&: op, msg: "iterator types do not correspond to matrix multiplication" ); |
| 187 | |
| 188 | // Check the combining kind is addition. |
| 189 | if (op.getKind() != vector::CombiningKind::ADD) |
| 190 | return rewriter.notifyMatchFailure(arg&: op, |
| 191 | msg: "combining kind is not an addition" ); |
| 192 | |
| 193 | // Check the output is a vector of i32 elements. |
| 194 | auto outTy = dyn_cast<VectorType>(Val: op.getResultType()); |
| 195 | if (!outTy || outTy.getElementType() != rewriter.getI32Type()) |
| 196 | return rewriter.notifyMatchFailure(arg&: op, |
| 197 | msg: "output type is not a vector of i32" ); |
| 198 | |
| 199 | // Check inputs are sign-/zero- extensions from i8 to i32. Get the values |
| 200 | // before the extension. All four signed/unsigned combinations for input |
| 201 | // operands are supported, but they are lowered to different operations. |
| 202 | // Determine which is the appropriate operation to lower to. |
| 203 | MMLA mmlaOp = MMLA::Signed; |
| 204 | auto maybeLhs = getExtOperand<arith::ExtSIOp>(v: op.getLhs()); |
| 205 | if (!maybeLhs) { |
| 206 | mmlaOp = MMLA::Unsigned; |
| 207 | maybeLhs = getExtOperand<arith::ExtUIOp>(v: op.getLhs()); |
| 208 | } |
| 209 | if (!maybeLhs) |
| 210 | return rewriter.notifyMatchFailure( |
| 211 | arg&: op, msg: "LHS is not a sign- or zero- extended i8" ); |
| 212 | |
| 213 | auto maybeRhs = getExtOperand<arith::ExtSIOp>(v: op.getRhs()); |
| 214 | if (maybeRhs) { |
| 215 | if (mmlaOp == MMLA::Unsigned) |
| 216 | mmlaOp = MMLA::Mixed; |
| 217 | } else { |
| 218 | if (mmlaOp == MMLA::Signed) |
| 219 | mmlaOp = MMLA::MixedSwapped; |
| 220 | maybeRhs = getExtOperand<arith::ExtUIOp>(v: op.getRhs()); |
| 221 | } |
| 222 | if (!maybeRhs) |
| 223 | return rewriter.notifyMatchFailure( |
| 224 | arg&: op, msg: "RHS is not a sign- or zero- extended i8" ); |
| 225 | |
| 226 | // One-dimensional vector types for arm_sve.*mmla |
| 227 | auto nxv16i8 = VectorType::get(/*shape=*/16, elementType: rewriter.getI8Type(), |
| 228 | /*scalableDims=*/{true}); |
| 229 | auto nxv4i32 = VectorType::get(/*shape=*/4, elementType: rewriter.getI32Type(), |
| 230 | /*scalableDims=*/{true}); |
| 231 | |
| 232 | // Extract LHS sub-tiles with logicall shape <2x8>. |
| 233 | SmallVector<Value> lhsTile; |
| 234 | for (int64_t i = 0; i < M; i += 2) { |
| 235 | // Extract two consecutive rows of the LHS tile. |
| 236 | auto r0 = rewriter.create<vector::ExtractOp>(location: loc, args&: *maybeLhs, |
| 237 | args: ArrayRef<int64_t>{i}); |
| 238 | auto r1 = rewriter.create<vector::ExtractOp>(location: loc, args&: *maybeLhs, |
| 239 | args: ArrayRef<int64_t>{i + 1}); |
| 240 | // Concatenate to obtain a 16 x i8 flattened sub-tile. |
| 241 | auto t = rewriter.create<vector::ShuffleOp>( |
| 242 | location: loc, args&: r0, args&: r1, |
| 243 | args: llvm::ArrayRef<int64_t>{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, |
| 244 | 14, 15}); |
| 245 | // Turn it into a scalable vector. |
| 246 | auto s = rewriter.create<vector::ScalableInsertOp>( |
| 247 | location: loc, args&: t, args: rewriter.create<ub::PoisonOp>(location: loc, args&: nxv16i8), args: 0); |
| 248 | // Replicate the sub-tile VSCALE times to fill the entire vector. |
| 249 | auto r = rewriter.create<arm_sve::DupQLaneOp>(location: loc, args&: s, args: 0); |
| 250 | lhsTile.push_back(Elt: r); |
| 251 | } |
| 252 | |
| 253 | // "Flatten" the RHS tile from <[N]x8> to <[8*N]>. |
| 254 | auto rhs = rewriter.create<vector::ShapeCastOp>( |
| 255 | location: maybeRhs->getLoc(), |
| 256 | args: VectorType::get(/*shape=*/8 * N, elementType: rewriter.getI8Type(), |
| 257 | /*scalableDims=*/{true}), |
| 258 | args&: *maybeRhs); |
| 259 | |
| 260 | // Extract the RHS sub-tiles with logical shape <8x[2]>. |
| 261 | SmallVector<Value> rhsTile; |
| 262 | for (int64_t j = 0; j < N; j += 2) |
| 263 | rhsTile.push_back( |
| 264 | Elt: rewriter.create<vector::ScalableExtractOp>(location: loc, args&: nxv16i8, args&: rhs, args: j * 8)); |
| 265 | |
| 266 | // Handy types for packing/unpacking of the accumulator tile. |
| 267 | auto accRowTy = VectorType::get(/*shape=*/N, elementType: rewriter.getI32Type(), |
| 268 | /*scalableDims=*/{true}); |
| 269 | auto accRowX2Ty = VectorType::get(/*shape=*/2 * N, elementType: rewriter.getI32Type(), |
| 270 | /*scalableDims=*/{true}); |
| 271 | auto accRow64Ty = VectorType::get(/*shape=*/N / 2, elementType: rewriter.getI64Type(), |
| 272 | /*scalableDims=*/{true}); |
| 273 | auto accRowX264Ty = VectorType::get(/*shape=*/N, elementType: rewriter.getI64Type(), |
| 274 | /*scalableDims=*/{true}); |
| 275 | |
| 276 | // Extract and pack the ACC sub-tiles. |
| 277 | SmallVector<Value> accTile; |
| 278 | for (int64_t i = 0; i < M; i += 2) { |
| 279 | // Extract two consecutive rows of the accumulator tile. |
| 280 | auto r0 = rewriter.create<vector::ExtractOp>(location: loc, args: op.getAcc(), |
| 281 | args: ArrayRef<int64_t>{i}); |
| 282 | auto r1 = rewriter.create<vector::ExtractOp>(location: loc, args: op.getAcc(), |
| 283 | args: ArrayRef<int64_t>{i + 1}); |
| 284 | Value accTileVec; |
| 285 | if (mmlaOp == MMLA::MixedSwapped) { |
| 286 | // We need to swap the positions of the LHS and RHS (since we don't have |
| 287 | // a signed * unsigned operation), but then each individual 2x2 tile of |
| 288 | // the acumulator and (later) the result need to be transposed. |
| 289 | accTileVec = rewriter.create<vector::InterleaveOp>(location: loc, args&: r0, args&: r1); |
| 290 | } else { |
| 291 | // Bitcast them to 64-bit elements, so subsequent |
| 292 | // interleave/deinterleave work on pairs of 32-bit numbers. |
| 293 | auto r0I64 = rewriter.create<vector::BitCastOp>(location: loc, args&: accRow64Ty, args&: r0); |
| 294 | auto r1I64 = rewriter.create<vector::BitCastOp>(location: loc, args&: accRow64Ty, args&: r1); |
| 295 | |
| 296 | // Interleave the rows, effectively flattening each 2x2 tile into 4 |
| 297 | // consecutive elements. |
| 298 | auto intrI64 = rewriter.create<vector::InterleaveOp>(location: loc, args&: r0I64, args&: r1I64); |
| 299 | |
| 300 | // Bitcast back to 32-bit elements. |
| 301 | accTileVec = |
| 302 | rewriter.create<vector::BitCastOp>(location: loc, args&: accRowX2Ty, args&: intrI64); |
| 303 | } |
| 304 | // Extract ACC sub-tiles. |
| 305 | for (int64_t j = 0; j < N; j += 2) |
| 306 | accTile.push_back(Elt: rewriter.create<vector::ScalableExtractOp>( |
| 307 | location: loc, args&: nxv4i32, args&: accTileVec, args: j * 2)); |
| 308 | } |
| 309 | |
| 310 | // Emit sub-tile matrix multiplications. |
| 311 | SmallVector<Value> outTile; |
| 312 | for (int64_t i = 0; i < M / 2; ++i) |
| 313 | for (int64_t j = 0; j < N / 2; ++j) { |
| 314 | Value mmla = createMMLA(rewriter, op: mmlaOp, loc, accType: nxv4i32, |
| 315 | acc: accTile[i * N / 2 + j], lhs: lhsTile[i], rhs: rhsTile[j]); |
| 316 | outTile.push_back(Elt: mmla); |
| 317 | } |
| 318 | |
| 319 | // Unpack the OUT sub-tiles and insert into the result. |
| 320 | Value result = rewriter.create<ub::PoisonOp>(location: loc, args: op.getResultType()); |
| 321 | for (int64_t i = 0; i < M / 2; ++i) { |
| 322 | // Collect a number of sub-tiles in a row. |
| 323 | Value row = rewriter.create<ub::PoisonOp>(location: loc, args&: accRowX2Ty); |
| 324 | for (int64_t j = 0; j < N / 2; ++j) |
| 325 | row = rewriter.create<vector::ScalableInsertOp>( |
| 326 | location: loc, args&: outTile[i * N / 2 + j], args&: row, args: j * 4); |
| 327 | |
| 328 | // Unpack the row to obtain two rows of the output. If we have the out |
| 329 | // sub-tiles transposed we obtain two consecutive output rows by |
| 330 | // separating even and odd elements, i.e. a simple deinterleave. |
| 331 | // Otherwise, the interleave is by pairs. |
| 332 | Value out0, out1; |
| 333 | if (mmlaOp == MMLA::MixedSwapped) { |
| 334 | auto tmp = rewriter.create<vector::DeinterleaveOp>(location: loc, args&: row); |
| 335 | out0 = tmp.getRes1(); |
| 336 | out1 = tmp.getRes2(); |
| 337 | } else { |
| 338 | // Deinterleave by pairs. |
| 339 | auto row64 = rewriter.create<vector::BitCastOp>(location: loc, args&: accRowX264Ty, args&: row); |
| 340 | auto deintr64 = rewriter.create<vector::DeinterleaveOp>(location: loc, args&: row64); |
| 341 | |
| 342 | // Bitcast back into 32-bit elements and insert into the result. |
| 343 | out0 = rewriter.create<vector::BitCastOp>(location: loc, args&: accRowTy, |
| 344 | args: deintr64.getRes1()); |
| 345 | out1 = rewriter.create<vector::BitCastOp>(location: loc, args&: accRowTy, |
| 346 | args: deintr64.getRes2()); |
| 347 | } |
| 348 | result = rewriter.create<vector::InsertOp>(location: loc, args&: out0, args&: result, args: i * 2); |
| 349 | result = rewriter.create<vector::InsertOp>(location: loc, args&: out1, args&: result, args: i * 2 + 1); |
| 350 | } |
| 351 | |
| 352 | rewriter.replaceOp(op, newValues: result); |
| 353 | return success(); |
| 354 | } |
| 355 | }; |
| 356 | |
| 357 | } // namespace |
| 358 | |
| 359 | void mlir::populateLowerContractionToSVEI8MMPatternPatterns( |
| 360 | RewritePatternSet &patterns) { |
| 361 | MLIRContext *context = patterns.getContext(); |
| 362 | patterns.add<LowerContractionToSVEI8MMPattern>(arg&: context, /*benefit=*/args: 2); |
| 363 | } |
| 364 | |