[llvm-branch-commits] [mlir] [MLIR][ArmSVE] Add initial lowering of vector.contract to SVE `*MMLA` instructions (PR #135636)

[Momchil Velikov via llvm-branch-commits]

================
@@ -0,0 +1,304 @@
+//===- LowerContractionToSMMLAPattern.cpp - Contract to SMMLA ---*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements lowering patterns from vector.contract to
+// SVE I8MM operations.
+//
+//===---
+
+#include "mlir/Dialect/Arith/IR/Arith.h"
+#include "mlir/Dialect/ArmSVE/IR/ArmSVEDialect.h"
+#include "mlir/Dialect/ArmSVE/Transforms/Transforms.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/LLVMIR/LLVMDialect.h"
+#include "mlir/Dialect/Utils/IndexingUtils.h"
+#include "mlir/Dialect/Vector/IR/VectorOps.h"
+#include "mlir/IR/AffineMap.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+#include "mlir/Dialect/UB/IR/UBOps.h"
+
+#define DEBUG_TYPE "lower-contract-to-arm-sve-i8mm"
+
+using namespace mlir;
+using namespace mlir::arm_sve;
+
+namespace {
+// Check if the given value is a result of the operation `T` (which must be
+// sign- or zero- extend) from i8 to i32. Return the value before the extension.
+template <typename T>
+inline std::enable_if_t<(std::is_base_of_v<arith::ExtSIOp, T> ||
+                         std::is_base_of_v<arith::ExtUIOp, T>),
+                        std::optional<Value>>
+extractExtOperand(Value v, Type i8Ty, Type i32Ty) {
+  auto extOp = dyn_cast_or_null<T>(v.getDefiningOp());
+  if (!extOp)
+    return {};
+
+  auto inOp = extOp.getIn();
+  auto inTy = dyn_cast<VectorType>(inOp.getType());
+  if (!inTy || inTy.getElementType() != i8Ty)
+    return {};
+
+  auto outTy = dyn_cast<VectorType>(extOp.getType());
+  if (!outTy || outTy.getElementType() != i32Ty)
+    return {};
+
+  return inOp;
+}
+
+// Designate the operation (resp. instruction) used to do sub-tile matrix
+// multiplications.
+enum class MMLA {
+  Signed,      // smmla
+  Unsigned,    // ummla
+  Mixed,       // usmmla
+  MixedSwapped // usmmla with LHS and RHS swapped
+};
+
+// Create the matrix multply and accumulate operation according to `op`.
+Value createMMLA(PatternRewriter &rewriter, MMLA op, Location loc,
+                 mlir::VectorType accType, Value acc, Value lhs, Value rhs) {
+  switch (op) {
+  case MMLA::Signed:
+    return rewriter.create<arm_sve::SmmlaOp>(loc, accType, acc, lhs, rhs);
+  case MMLA::Unsigned:
+    return rewriter.create<arm_sve::UmmlaOp>(loc, accType, acc, lhs, rhs);
+  case MMLA::Mixed:
+    return rewriter.create<arm_sve::UsmmlaOp>(loc, accType, acc, lhs, rhs);
+  case MMLA::MixedSwapped:
+    // The accumulator comes transposed and the result will be transposed
+    // later, so all we have to do here is swap the operands.
+    return rewriter.create<arm_sve::UsmmlaOp>(loc, accType, acc, rhs, lhs);
+  }
+}
+
+class LowerContractionToSVEI8MMPattern
+    : public OpRewritePattern<vector::ContractionOp> {
+public:
+  using OpRewritePattern::OpRewritePattern;
+  LogicalResult matchAndRewrite(vector::ContractionOp op,
+                                PatternRewriter &rewriter) const override {
+
+    Location loc = op.getLoc();
+    mlir::VectorType lhsType = op.getLhsType();
+    mlir::VectorType rhsType = op.getRhsType();
+
+    // For now handle LHS<Mx8> and RHS<8x[N]> - these are the types we
+    // eventually expect from MMT4D. M and N dimensions must be even and at
+    // least 2.
+    if (!lhsType.hasRank() || lhsType.getRank() != 2 || !rhsType.hasRank() ||
----------------
momchil-velikov wrote:

Done

https://github.com/llvm/llvm-project/pull/135636