1 | //===- LowerContractionToSMMLAPattern.cpp - Contract to SMMLA ---*- 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 | //===----------------------------------------------------------------------===// |
13 | |
14 | #include "mlir/Dialect/Arith/IR/Arith.h" |
15 | #include "mlir/Dialect/ArmSVE/IR/ArmSVEDialect.h" |
16 | #include "mlir/Dialect/ArmSVE/Transforms/Transforms.h" |
17 | #include "mlir/Dialect/Func/IR/FuncOps.h" |
18 | #include "mlir/Dialect/LLVMIR/LLVMDialect.h" |
19 | #include "mlir/Dialect/Utils/IndexingUtils.h" |
20 | #include "mlir/Dialect/Vector/IR/VectorOps.h" |
21 | #include "mlir/IR/AffineMap.h" |
22 | #include "mlir/IR/PatternMatch.h" |
23 | #include "mlir/Transforms/GreedyPatternRewriteDriver.h" |
24 | |
25 | #include "mlir/Dialect/UB/IR/UBOps.h" |
26 | |
27 | #define DEBUG_TYPE "lower-contract-to-arm-sve-i8mm" |
28 | |
29 | using namespace mlir; |
30 | |
31 | namespace { |
32 | // Get the operand of a `vector.contract`. This function is intended to abstract |
33 | // away from the particular way a value is extended before feeding it into the |
34 | // `vector.contract` - via zero-extend or an explicit or implicit sign-extend |
35 | // (for implicit sign-extension see `vector.contract` documentation). |
36 | // |
37 | // The template parameter `Op` indicates the extension operation (explicit or |
38 | // implicit) for which we are checking. |
39 | // |
40 | // Return success only for extensions from `i8` to `i32`. |
41 | template <typename Op> |
42 | std::optional<Value> getExtOperand(Value v, Type i8Ty, Type i32Ty) { |
43 | |
44 | static_assert(llvm::is_one_of<Op, arith::ExtSIOp, arith::ExtUIOp>::value, |
45 | "Must be instantiated with either sign- or zero- extension op" ); |
46 | |
47 | // If the operand is not defined by an explicit extend operation of the |
48 | // accepted operation type allow for an implicit sign-extension. |
49 | auto extOp = dyn_cast_or_null<Op>(v.getDefiningOp()); |
50 | if (!extOp) { |
51 | if constexpr (std::is_same<Op, arith::ExtSIOp>::value) { |
52 | auto vTy = cast<VectorType>(v.getType()); |
53 | if (vTy.getElementType() != i8Ty) |
54 | return {}; |
55 | return v; |
56 | } |
57 | return {}; |
58 | } |
59 | |
60 | // If the operand is defined by an explicit extend operation of the accepted |
61 | // operation type, check it's extended from `i8` to `i32`. |
62 | auto inOp = extOp.getIn(); |
63 | auto inTy = dyn_cast<VectorType>(inOp.getType()); |
64 | if (!inTy || inTy.getElementType() != i8Ty) |
65 | return {}; |
66 | |
67 | auto outTy = dyn_cast<VectorType>(extOp.getType()); |
68 | if (!outTy || outTy.getElementType() != i32Ty) |
69 | return {}; |
70 | |
71 | return inOp; |
72 | } |
73 | |
74 | // Designate the operation (resp. instruction) used to do sub-tile matrix |
75 | // multiplications. |
76 | enum class MMLA { |
77 | Signed, // smmla |
78 | Unsigned, // ummla |
79 | Mixed, // usmmla |
80 | MixedSwapped // usmmla with LHS and RHS swapped |
81 | }; |
82 | |
83 | // Create the matrix mulitply and accumulate operation according to `op`. |
84 | Value createMMLA(PatternRewriter &rewriter, MMLA op, Location loc, |
85 | mlir::VectorType accType, Value acc, Value lhs, Value rhs) { |
86 | switch (op) { |
87 | case MMLA::Signed: |
88 | return rewriter.create<arm_sve::SmmlaOp>(loc, accType, acc, lhs, rhs); |
89 | case MMLA::Unsigned: |
90 | return rewriter.create<arm_sve::UmmlaOp>(loc, accType, acc, lhs, rhs); |
91 | case MMLA::Mixed: |
92 | return rewriter.create<arm_sve::UsmmlaOp>(loc, accType, acc, lhs, rhs); |
93 | case MMLA::MixedSwapped: |
94 | // The accumulator comes transposed and the result will be transposed |
95 | // later, so all we have to do here is swap the operands. |
96 | return rewriter.create<arm_sve::UsmmlaOp>(loc, accType, acc, rhs, lhs); |
97 | } |
98 | } |
99 | |
100 | /// Lower a contraction operation that performs a matrix multiplication |
101 | /// of two 8-bit integer matrix tiles with logical dimensions <Mx8> and <8x[N]> |
102 | /// for the left-hand side and the right-hand side, respectively, |
103 | /// yielding a <Mx[N]> 32-bit integer result. |
104 | /// |
105 | /// The operands' shapes are such that the operands can be evenly split into |
106 | /// sub-tiles with dimensions as expected by the targeted FEAT_I8MM |
107 | /// instructions. The intent is that M and N are chosen (by higher level |
108 | /// transforms) in such a way as to maximise register usage. The main use case |
109 | /// we envision as of now is MMT4D, thus the RHS operand is expected |
110 | /// pre-transposed. |
111 | /// |
112 | /// The matrix multiplication is performed by unrolling the usual tiled matrix |
113 | /// multiplication algorithm using sub-tiles with dimensions <2x8> for the LHS, |
114 | /// <8x[2]> for the RHS, and <2x[2]> for the result and the input accumulator. |
115 | /// |
116 | /// One way to illustrate the operation is as follows: |
117 | /// |
118 | /// RHS<8x[N]>: <8x[2]> <8x[2]> ... <8x[2]> |
119 | /// +----------------------------- |
120 | /// LHS<Mx8>: <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
121 | /// <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
122 | /// ... | ... ... ... ... |
123 | /// <2x8> | <2x[2]> <2x[2]> ... <2x[2]> |
124 | /// |
125 | /// The RHS operand is unpacked into N/2 values, each representing a sequence of |
126 | /// VSCALE number of sub-tiles with dimensions <8x2>. |
127 | /// The LHS operand is initially unpacked into M/2 values, each representing a |
128 | /// sub-tile with dimensions <2x8>, and then each such sub-tile is replicated |
129 | /// VSCALE times. |
130 | /// Multiplying thus replicated LHS sub-tile by the corresponding RHS sub-tile |
131 | /// correctly computes an entire result sub-tile. |
132 | class LowerContractionToSVEI8MMPattern |
133 | : public OpRewritePattern<vector::ContractionOp> { |
134 | public: |
135 | using OpRewritePattern::OpRewritePattern; |
136 | LogicalResult matchAndRewrite(vector::ContractionOp op, |
137 | PatternRewriter &rewriter) const override { |
138 | |
139 | Location loc = op.getLoc(); |
140 | mlir::VectorType lhsType = op.getLhsType(); |
141 | mlir::VectorType rhsType = op.getRhsType(); |
142 | |
143 | // Check the rank the types so we can safely examine their dimensions. |
144 | if (lhsType.getRank() != 2 || rhsType.getRank() != 2) |
145 | return rewriter.notifyMatchFailure(op, "non-matching operand shape" ); |
146 | |
147 | auto M = lhsType.getDimSize(0); |
148 | auto N = rhsType.getDimSize(0); |
149 | auto K = rhsType.getDimSize(1); |
150 | |
151 | // Check the operands have the expected shape: |
152 | // * for LHS: fixed vector MxK |
153 | // * for RHS: scalable vector [N]xK |
154 | // * K == 8 |
155 | // * M and N even and at least 2 |
156 | if (lhsType.isScalable() || !rhsType.getScalableDims()[0] || |
157 | rhsType.getScalableDims()[1] || lhsType.getDimSize(1) != K || K != 8 || |
158 | M < 2 || M % 2 != 0 || N < 2 || N % 2 != 0 || |
159 | !rhsType.getScalableDims()[0]) |
160 | return rewriter.notifyMatchFailure(op, "non-matching operand shape" ); |
161 | |
162 | // Check permutation maps. For now only accept |
163 | // lhs: (d0, d1, d2) -> (d0, d2) |
164 | // rhs: (d0, d1, d2) -> (d1, d2) |
165 | // acc: (d0, d1, d2) -> (d0, d1) |
166 | // This corresponds to matrix multiplication with transposed RHS. |
167 | if (op.getIndexingMapsArray()[0] != |
168 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{0u, 2u}, |
169 | context: op.getContext()) || |
170 | op.getIndexingMapsArray()[1] != |
171 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{1u, 2u}, |
172 | context: op.getContext()) || |
173 | op.getIndexingMapsArray()[2] != |
174 | AffineMap::getMultiDimMapWithTargets(numDims: 3, targets: ArrayRef{0u, 1u}, |
175 | context: op.getContext())) |
176 | return rewriter.notifyMatchFailure(op, "non-matching permutation maps" ); |
177 | |
178 | // Check iterator types for matrix multiplication. |
179 | auto itTypes = op.getIteratorTypesArray(); |
180 | if (itTypes.size() != 3 || itTypes[0] != vector::IteratorType::parallel || |
181 | itTypes[1] != vector::IteratorType::parallel || |
182 | itTypes[2] != vector::IteratorType::reduction) |
183 | return rewriter.notifyMatchFailure( |
184 | op, "iterator types do not correspond to matrix multiplication" ); |
185 | |
186 | // Check the combining kind is addition. |
187 | if (op.getKind() != vector::CombiningKind::ADD) |
188 | return rewriter.notifyMatchFailure(op, |
189 | "combining kind is not an addition" ); |
190 | |
191 | // Check the output is a vector of i32 elements. |
192 | auto outTy = dyn_cast<VectorType>(op.getResultType()); |
193 | if (!outTy || outTy.getElementType() != rewriter.getI32Type()) |
194 | return rewriter.notifyMatchFailure(op, |
195 | "output type is not a vector of i32" ); |
196 | |
197 | // Check inputs are sign-/zero- extensions from i8 to i32. Get the values |
198 | // before the extension. All four signed/unsigned combinations for input |
199 | // operands are supported, but they are lowered to different operations. |
200 | // Determine which is the appropriate operation to lower to. |
201 | MMLA mmlaOp = MMLA::Signed; |
202 | auto maybeLhs = getExtOperand<arith::ExtSIOp>( |
203 | op.getLhs(), rewriter.getI8Type(), rewriter.getI32Type()); |
204 | if (!maybeLhs) { |
205 | mmlaOp = MMLA::Unsigned; |
206 | maybeLhs = getExtOperand<arith::ExtUIOp>( |
207 | op.getLhs(), rewriter.getI8Type(), rewriter.getI32Type()); |
208 | } |
209 | if (!maybeLhs) |
210 | return rewriter.notifyMatchFailure( |
211 | op, "LHS is not a sign- or zero- extended i8" ); |
212 | |
213 | auto maybeRhs = getExtOperand<arith::ExtSIOp>( |
214 | op.getRhs(), rewriter.getI8Type(), rewriter.getI32Type()); |
215 | if (maybeRhs) { |
216 | if (mmlaOp == MMLA::Unsigned) |
217 | mmlaOp = MMLA::Mixed; |
218 | } else { |
219 | if (mmlaOp == MMLA::Signed) |
220 | mmlaOp = MMLA::MixedSwapped; |
221 | maybeRhs = getExtOperand<arith::ExtUIOp>( |
222 | op.getRhs(), rewriter.getI8Type(), rewriter.getI32Type()); |
223 | } |
224 | if (!maybeRhs) |
225 | return rewriter.notifyMatchFailure( |
226 | op, "RHS is not a sign- or zero- extended i8" ); |
227 | |
228 | // One-dimensional vector types for arm_sve.*mmla |
229 | auto nxv16i8 = VectorType::get(/*shape=*/16, rewriter.getI8Type(), |
230 | /*scalableDims=*/{true}); |
231 | auto nxv4i32 = VectorType::get(/*shape=*/4, rewriter.getI32Type(), |
232 | /*scalableDims=*/{true}); |
233 | |
234 | // Extract LHS sub-tiles with logicall shape <2x8>. |
235 | SmallVector<Value> lhsTile; |
236 | for (int64_t i = 0; i < M; i += 2) { |
237 | // Extract two consecutive rows of the LHS tile. |
238 | auto r0 = rewriter.create<vector::ExtractOp>(loc, *maybeLhs, |
239 | ArrayRef<int64_t>{i}); |
240 | auto r1 = rewriter.create<vector::ExtractOp>(loc, *maybeLhs, |
241 | ArrayRef<int64_t>{i + 1}); |
242 | // Concatenate to obtain a 16 x i8 flattened sub-tile. |
243 | auto t = rewriter.create<vector::ShuffleOp>( |
244 | loc, r0, r1, |
245 | llvm::ArrayRef<int64_t>{0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, |
246 | 14, 15}); |
247 | // Turn it into a scalable vector. |
248 | auto s = rewriter.create<vector::ScalableInsertOp>( |
249 | loc, t, rewriter.create<ub::PoisonOp>(loc, nxv16i8), 0); |
250 | // Replicate the sub-tile VSCALE times to fill the entire vector. |
251 | auto r = rewriter.create<arm_sve::DupQLaneOp>(loc, s, 0); |
252 | lhsTile.push_back(Elt: r); |
253 | } |
254 | |
255 | // "Flatten" the RHS tile from <[N]x8> to <[8*N]>. |
256 | auto rhs = rewriter.create<vector::ShapeCastOp>( |
257 | maybeRhs->getLoc(), |
258 | VectorType::get(/*shape=*/8 * N, rewriter.getI8Type(), |
259 | /*scalableDims=*/{true}), |
260 | *maybeRhs); |
261 | |
262 | // Extract the RHS sub-tiles with logical shape <8x[2]>. |
263 | SmallVector<Value> rhsTile; |
264 | for (int64_t j = 0; j < N; j += 2) |
265 | rhsTile.push_back( |
266 | rewriter.create<vector::ScalableExtractOp>(loc, nxv16i8, rhs, j * 8)); |
267 | |
268 | // Handy types for packing/unpacking of the accumulator tile. |
269 | auto accRowTy = VectorType::get(/*shape=*/N, rewriter.getI32Type(), |
270 | /*scalableDims=*/{true}); |
271 | auto accRowX2Ty = VectorType::get(/*shape=*/2 * N, rewriter.getI32Type(), |
272 | /*scalableDims=*/{true}); |
273 | auto accRow64Ty = VectorType::get(/*shape=*/N / 2, rewriter.getI64Type(), |
274 | /*scalableDims=*/{true}); |
275 | auto accRowX264Ty = VectorType::get(/*shape=*/N, rewriter.getI64Type(), |
276 | /*scalableDims=*/{true}); |
277 | |
278 | // Extract and pack the ACC sub-tiles. |
279 | SmallVector<Value> accTile; |
280 | for (int64_t i = 0; i < M; i += 2) { |
281 | // Extract two consecutive rows of the accumulator tile. |
282 | auto r0 = rewriter.create<vector::ExtractOp>(loc, op.getAcc(), |
283 | ArrayRef<int64_t>{i}); |
284 | auto r1 = rewriter.create<vector::ExtractOp>(loc, op.getAcc(), |
285 | ArrayRef<int64_t>{i + 1}); |
286 | Value accTileVec; |
287 | if (mmlaOp == MMLA::MixedSwapped) { |
288 | // We need to swap the positions of the LHS and RHS (since we don't have |
289 | // a signed * unsigned operation), but then each individual 2x2 tile of |
290 | // the acumulator and (later) the result need to be transposed. |
291 | accTileVec = rewriter.create<vector::InterleaveOp>(loc, r0, r1); |
292 | } else { |
293 | // Bitcast them to 64-bit elements, so subsequent |
294 | // interleave/deinterleave work on pairs of 32-bit numbers. |
295 | auto r0I64 = rewriter.create<vector::BitCastOp>(loc, accRow64Ty, r0); |
296 | auto r1I64 = rewriter.create<vector::BitCastOp>(loc, accRow64Ty, r1); |
297 | |
298 | // Interleave the rows, effectively flattening each 2x2 tile into 4 |
299 | // consecutive elements. |
300 | auto intrI64 = rewriter.create<vector::InterleaveOp>(loc, r0I64, r1I64); |
301 | |
302 | // Bitcast back to 32-bit elements. |
303 | accTileVec = |
304 | rewriter.create<vector::BitCastOp>(loc, accRowX2Ty, intrI64); |
305 | } |
306 | // Extract ACC sub-tiles. |
307 | for (int64_t j = 0; j < N; j += 2) |
308 | accTile.push_back(rewriter.create<vector::ScalableExtractOp>( |
309 | loc, nxv4i32, accTileVec, j * 2)); |
310 | } |
311 | |
312 | // Emit sub-tile matrix multiplications. |
313 | SmallVector<Value> outTile; |
314 | for (int64_t i = 0; i < M / 2; ++i) |
315 | for (int64_t j = 0; j < N / 2; ++j) { |
316 | Value mmla = createMMLA(rewriter, mmlaOp, loc, nxv4i32, |
317 | accTile[i * N / 2 + j], lhsTile[i], rhsTile[j]); |
318 | outTile.push_back(Elt: mmla); |
319 | } |
320 | |
321 | // Unpack the OUT sub-tiles and insert into the result. |
322 | Value result = rewriter.create<ub::PoisonOp>(loc, op.getResultType()); |
323 | for (int64_t i = 0; i < M / 2; ++i) { |
324 | // Collect a number of sub-tiles in a row. |
325 | Value row = rewriter.create<ub::PoisonOp>(loc, accRowX2Ty); |
326 | for (int64_t j = 0; j < N / 2; ++j) |
327 | row = rewriter.create<vector::ScalableInsertOp>( |
328 | loc, outTile[i * N / 2 + j], row, j * 4); |
329 | |
330 | // Unpack the row to obtain two rows of the output. If we have the out |
331 | // sub-tiles transposed we obtain two consecutive output rows by |
332 | // separating even and odd elements, i.e. a simple deinterleave. |
333 | // Otherwise, the interleave is by pairs. |
334 | Value out0, out1; |
335 | if (mmlaOp == MMLA::MixedSwapped) { |
336 | auto tmp = rewriter.create<vector::DeinterleaveOp>(loc, row); |
337 | out0 = tmp.getRes1(); |
338 | out1 = tmp.getRes2(); |
339 | } else { |
340 | // Deinterleave by pairs. |
341 | auto row64 = rewriter.create<vector::BitCastOp>(loc, accRowX264Ty, row); |
342 | auto deintr64 = rewriter.create<vector::DeinterleaveOp>(loc, row64); |
343 | |
344 | // Bitcast back into 32-bit elements and insert into the result. |
345 | out0 = rewriter.create<vector::BitCastOp>(loc, accRowTy, |
346 | deintr64.getRes1()); |
347 | out1 = rewriter.create<vector::BitCastOp>(loc, accRowTy, |
348 | deintr64.getRes2()); |
349 | } |
350 | result = rewriter.create<vector::InsertOp>(loc, out0, result, i * 2); |
351 | result = rewriter.create<vector::InsertOp>(loc, out1, result, i * 2 + 1); |
352 | } |
353 | |
354 | rewriter.replaceOp(op, result); |
355 | return success(); |
356 | } |
357 | }; |
358 | |
359 | } // namespace |
360 | |
361 | void mlir::populateLowerContractionToSVEI8MMPatternPatterns( |
362 | RewritePatternSet &patterns) { |
363 | MLIRContext *context = patterns.getContext(); |
364 | patterns.add<LowerContractionToSVEI8MMPattern>(arg&: context, /*benefit=*/args: 2); |
365 | } |
366 | |