VectorToXeGPU.cpp

Go to the documentation of this file.

//===- VectorToXeGPU.cpp - Convert vector to XeGPU dialect ------*- 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 of vector operations to XeGPU dialect ops.
//
//===----------------------------------------------------------------------===//
 
#include "mlir/Conversion/VectorToXeGPU/VectorToXeGPU.h"
#include "mlir/Conversion/VectorToGPU/VectorToGPU.h"
 
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/MemRef/IR/MemRef.h"
#include "mlir/Dialect/UB/IR/UBOps.h"
#include "mlir/Dialect/Utils/IndexingUtils.h"
#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h"
#include "mlir/Dialect/XeGPU/IR/XeGPU.h"
#include "mlir/Dialect/XeGPU/Utils/XeGPUUtils.h"
#include "mlir/Interfaces/SideEffectInterfaces.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
#include "llvm/ADT/TypeSwitch.h"
 
#include <algorithm>
#include <optional>
 
namespace mlir {
#define GEN_PASS_DEF_CONVERTVECTORTOXEGPU
#include "mlir/Conversion/Passes.h.inc"
} // namespace mlir
 
using namespace mlir;
 
namespace {
 
// Return true if value represents a zero constant.
static bool isZeroConstant(Value val) {
  auto constant = val.getDefiningOp<arith::ConstantOp>();
  if (!constant)
    return false;
 
  return TypeSwitch<Attribute, bool>(constant.getValue())
      .Case([](FloatAttr floatAttr) { return floatAttr.getValue().isZero(); })
      .Case([](IntegerAttr intAttr) { return intAttr.getValue().isZero(); })
      .Default(false);
}
 
// Return true if the transfer padding value is compatible with the implicit
// padding of an nd block load. LoadNdOp fills out-of-bounds elements with zero,
// so a zero constant matches its semantics exactly. A poison padding means the
// out-of-bounds elements are "don't care", so any implicit padding (including
// zero) is also acceptable.
static bool isZeroOrPoisonPadding(Value val) {
  return isZeroConstant(val) || val.getDefiningOp<ub::PoisonOp>();
}
 
// Return true if the permutation map keeps every dimension in place except the
// innermost two, which are swapped, e.g.:
//   (d0, d1) -> (d1, d0)
//   (d0, d1, d2) -> (d2, d1)
//   (d0, d1, d2, d3) -> (d0, d1, d3, d2)
// This is the only non-identity permutation an nd block load can realize (by
// loading the untransposed block and applying a trailing vector.transpose).
static bool isInnermostTwoDimsTransposed(AffineMap map) {
  unsigned numResults = map.getNumResults();
  if (numResults < 2)
    return false;
  MLIRContext *ctx = map.getContext();
  unsigned numInputs = map.getNumInputs();
  // All but the innermost two results must match the minor-identity map.
  for (unsigned i = 0; i + 2 < numResults; ++i)
    if (map.getResult(i) != getAffineDimExpr(numInputs - numResults + i, ctx))
      return false;
  // The innermost two results must be the last two input dims, swapped.
  return map.getResult(numResults - 2) ==
             getAffineDimExpr(numInputs - 1, ctx) &&
         map.getResult(numResults - 1) == getAffineDimExpr(numInputs - 2, ctx);
}
 
static LogicalResult transferPreconditions(PatternRewriter &rewriter,
                                           VectorTransferOpInterface xferOp) {
  if (xferOp.getMask())
    return rewriter.notifyMatchFailure(xferOp,
                                       "Masked transfer is not supported");
 
  auto srcTy = dyn_cast<MemRefType>(xferOp.getShapedType());
  if (!srcTy)
    return rewriter.notifyMatchFailure(xferOp, "Expects memref source");
 
  // Validate further transfer op semantics.
  SmallVector<int64_t> strides;
  int64_t offset;
  if (failed(srcTy.getStridesAndOffset(strides, offset)))
    return rewriter.notifyMatchFailure(xferOp,
                                       "The memref strides cannot be inferred");
  if (strides.empty())
    return rewriter.notifyMatchFailure(xferOp, "0D memref is not supported");
  if (strides.back() != 1)
    return rewriter.notifyMatchFailure(
        xferOp, "Buffer must be contiguous in the innermost dimension");
 
  VectorType vecTy = xferOp.getVectorType();
  unsigned vecRank = vecTy.getRank();
  if (vecRank == 0)
    return rewriter.notifyMatchFailure(xferOp, "0D vectors are not supported");
  if (xferOp.hasOutOfBoundsDim() && vecRank < 2)
    return rewriter.notifyMatchFailure(
        xferOp, "Boundary check is available only for block instructions.");
 
  AffineMap map = xferOp.getPermutationMap();
  if (!map.isProjectedPermutation(/*allowZeroInResults=*/false))
    return rewriter.notifyMatchFailure(xferOp, "Unsupported permutation map");
  unsigned numInputDims = map.getNumInputs();
  for (AffineExpr expr : map.getResults().take_back(vecRank)) {
    auto dim = dyn_cast<AffineDimExpr>(expr);
    if (dim.getPosition() < (numInputDims - vecRank))
      return rewriter.notifyMatchFailure(
          xferOp, "Only the innermost dimensions can be accessed");
  }
 
  return success();
}
 
// Adjusts the strides of a memref according to a given permutation map for
// vector operations.
//
// This function updates the innermost strides in the `strides` array to
// reflect the permutation specified by `permMap`. The permutation is computed
// using the inverse and broadcasting-aware version of the permutation map,
// and is applied to the relevant strides. This ensures that memory accesses
// are consistent with the logical permutation of vector elements.
//
// Example:
//   Suppose we have a memref of rank 4 with strides `[s0, s1, s2, s3]`.
//   If the permutation map swaps the last two dimensions (e.g., [0, 1] -> [1,
//   0]), then after calling this function, the last two strides will be
//   swapped:
//     Original strides: [s0, s1, s2, s3]
//     After permutation: [s0, s1, s3, s2]
//
static void adjustStridesForPermutation(AffineMap permMap,
                                        SmallVectorImpl<Value> &strides) {
 
  AffineMap invMap = inverseAndBroadcastProjectedPermutation(permMap);
  SmallVector<unsigned> perms;
  invMap.isPermutationOfMinorIdentityWithBroadcasting(perms);
  SmallVector<int64_t> perms64(perms.begin(), perms.end());
  strides = applyPermutation(strides, perms64);
}
 
// Computes memory strides and a memref offset for vector transfer operations,
// handling both static and dynamic memrefs while applying permutation
// transformations for XeGPU lowering.
template <
    typename OpType,
    typename = std::enable_if_t<llvm::is_one_of<
        std::decay_t<OpType>, vector::TransferReadOp, vector::TransferWriteOp,
        vector::GatherOp, vector::ScatterOp>::value>>
static std::pair<SmallVector<Value>, Value>
computeMemrefMeta(OpType xferOp, PatternRewriter &rewriter) {
  SmallVector<Value> strides;
  Value baseMemref = xferOp.getBase();
  MemRefType memrefType = dyn_cast<MemRefType>(baseMemref.getType());
 
  Location loc = xferOp.getLoc();
  Value offsetVal = nullptr;
  if (memrefType.hasStaticShape()) {
    int64_t offset;
    SmallVector<int64_t> intStrides;
    if (failed(memrefType.getStridesAndOffset(intStrides, offset)))
      return {{}, offsetVal};
    bool hasDynamicStrides = llvm::any_of(intStrides, [](int64_t strideVal) {
      return ShapedType::isDynamic(strideVal);
    });
 
    if (!hasDynamicStrides)
      for (int64_t s : intStrides)
        strides.push_back(arith::ConstantIndexOp::create(rewriter, loc, s));
 
    if (!ShapedType::isDynamic(offset))
      offsetVal = arith::ConstantIndexOp::create(rewriter, loc, offset);
  }
 
  if (strides.empty() || !offsetVal) {
    // For dynamic shape memref, use memref.extract_strided_metadata to get
    // stride values
    unsigned rank = memrefType.getRank();
    Type indexType = rewriter.getIndexType();
 
    // Result types: [base_memref, offset, stride0, stride1, ..., strideN-1,
    // size0, size1, ..., sizeN-1]
    SmallVector<Type> resultTypes;
    resultTypes.push_back(MemRefType::get(
        {}, memrefType.getElementType())); // base memref (unranked)
    resultTypes.push_back(indexType);      // offset
 
    for (unsigned i = 0; i < rank; ++i)
      resultTypes.push_back(indexType); // strides
 
    for (unsigned i = 0; i < rank; ++i)
      resultTypes.push_back(indexType); // sizes
 
    auto meta = memref::ExtractStridedMetadataOp::create(
        rewriter, loc, resultTypes, baseMemref);
 
    if (strides.empty())
      strides.append(meta.getStrides().begin(), meta.getStrides().end());
 
    if (!offsetVal)
      offsetVal = meta.getOffset();
  }
 
  // Strides are returned in original memref order; permutation is applied in
  // computeOffsets only where offsets are indexed in vector order.
  return {strides, offsetVal};
}
 
// This function compute the vectors of localOffsets for scattered load/stores.
// It is used in the lowering of vector.transfer_read/write to
// load_gather/store_scatter Example:
//   %0 = vector.transfer_read %expand_shape[%block_id_y, %c0, %c0, %c0, %c0],
//               %cst {in_bounds = [true, true, true, true]}>} :
//               memref<8x4x2x6x32xbf16>, vector<4x2x6x32xbf16>
//
//   %6 = vector.step: vector<4xindex>
//   %7 = vector.step: vector<2xindex>
//   %8 = vector.step: vector<6xindex>
//   %9 = vector.step: vector<32xindex>
//   %10 = arith.mul %6, 384
//   %11 = arith.mul %7, 192
//   %12 = arith.mul %8, 32
//   %13 = arith.mul %9, 1
//   %14 = vector.shape_cast %10: vector<4xindex> -> vector<4x1x1x1xbf16>
//   %15 = vector.shape_cast %11: vector<2xindex> -> vector<1x2x1x1xbf16>
//   %16 = vector.shape_cast %12: vector<6xindex> -> vector<1x1x6x1xbf16>
//   %17 = vector.shape_cast %13: vector<32xindex> -> vector<1x1x1x32xbf16>
//   %18 = vector.broadcast %14: vector<4x1x1x1xbf16> -> vector<4x2x6x32xindex>
//   %19 = vector.broadcast %15: vector<1x2x1x1xbf16> -> vector<4x2x6x32xindex>
//   %20 = vector.broadcast %16: vector<1x1x6x1xbf16> -> vector<4x2x6x32xindex>
//   %21 = vector.broadcast %17: vector<1x1x1x32xbf16> -> vector<4x2x6x32xindex>
//   %22 = arith.add %18, %19
//   %23 = arith.add %20, %21
//   %local_offsets = arith.add %22, %23
//   %orig_offset = %block_id_y * 4x2x6x32 // consider using affine map
//   %offsets =  memref_offset + orig_offset + local_offsets
static Value computeOffsets(VectorTransferOpInterface xferOp,
                            PatternRewriter &rewriter, ArrayRef<Value> strides,
                            Value baseOffset) {
  Location loc = xferOp.getLoc();
  VectorType vectorType = xferOp.getVectorType();
  SmallVector<Value> indices(xferOp.getIndices().begin(),
                             xferOp.getIndices().end());
  ArrayRef<int64_t> vectorShape = vectorType.getShape();
 
  // Create vector.step operations for each dimension
  SmallVector<Value> stepVectors;
  llvm::map_to_vector(vectorShape, [&](int64_t dim) {
    auto stepType = VectorType::get({dim}, rewriter.getIndexType());
    auto stepOp = vector::StepOp::create(rewriter, loc, stepType);
    stepVectors.push_back(stepOp);
    return stepOp;
  });
 
  // Local offsets are indexed in vector order, so permute strides; the base
  // offset below uses the original memref-order strides.
  SmallVector<Value> permutedStrides(strides.begin(), strides.end());
  adjustStridesForPermutation(xferOp.getPermutationMap(), permutedStrides);
 
  // Multiply step vectors by corresponding strides
  size_t memrefRank = permutedStrides.size();
  size_t vectorRank = vectorShape.size();
  SmallVector<Value> strideMultiplied;
  for (size_t i = 0; i < vectorRank; ++i) {
    size_t memrefDim = memrefRank - vectorRank + i;
    Value strideValue = permutedStrides[memrefDim];
    auto mulType = dyn_cast<VectorType>(stepVectors[i].getType());
    auto bcastOp =
        vector::BroadcastOp::create(rewriter, loc, mulType, strideValue);
    auto mulOp = arith::MulIOp::create(rewriter, loc, stepVectors[i], bcastOp);
    strideMultiplied.push_back(mulOp);
  }
 
  // Shape cast each multiplied vector to add singleton dimensions
  SmallVector<Value> shapeCasted;
  for (size_t i = 0; i < vectorRank; ++i) {
    SmallVector<int64_t> newShape(vectorRank, 1);
    newShape[i] = vectorShape[i];
    auto newType = VectorType::get(newShape, rewriter.getIndexType());
    auto castOp = vector::ShapeCastOp::create(rewriter, loc, newType,
                                              strideMultiplied[i]);
    shapeCasted.push_back(castOp);
  }
 
  // Broadcast each shape-casted vector to full vector shape
  SmallVector<Value> broadcasted;
  auto fullIndexVectorType =
      VectorType::get(vectorShape, rewriter.getIndexType());
  for (Value shapeCastVal : shapeCasted) {
    auto broadcastOp = vector::BroadcastOp::create(
        rewriter, loc, fullIndexVectorType, shapeCastVal);
    broadcasted.push_back(broadcastOp);
  }
 
  // Add all broadcasted vectors together to compute local offsets
  Value localOffsets = broadcasted[0];
  for (size_t i = 1; i < broadcasted.size(); ++i)
    localOffsets =
        arith::AddIOp::create(rewriter, loc, localOffsets, broadcasted[i]);
 
  // Compute base offset from transfer read indices
  for (size_t i = 0; i < indices.size(); ++i) {
    Value strideVal = strides[i];
    Value offsetContrib =
        arith::MulIOp::create(rewriter, loc, indices[i], strideVal);
    baseOffset =
        arith::AddIOp::create(rewriter, loc, baseOffset, offsetContrib);
  }
  // Broadcast base offset to match vector shape
  Value bcastBase = vector::BroadcastOp::create(
      rewriter, loc, fullIndexVectorType, baseOffset);
  localOffsets = arith::AddIOp::create(rewriter, loc, bcastBase, localOffsets);
  return localOffsets;
}
 
// Compute the element-wise offsets for vector.gather or vector.scatter ops.
//
// This function linearizes the base offsets of the gather/scatter operation
// and combines them with the per-element indices to produce a final vector of
// memory offsets.
template <
    typename OpType,
    typename = std::enable_if_t<llvm::is_one_of<
        std::decay_t<OpType>, vector::GatherOp, vector::ScatterOp>::value>>
static Value computeOffsets(PatternRewriter &rewriter, OpType gatScatOp,
                            ArrayRef<Value> strides, Value baseOffset) {
  Location loc = gatScatOp.getLoc();
  SmallVector<Value> offsets = gatScatOp.getOffsets();
  for (size_t i = 0; i < offsets.size(); ++i) {
    Value offsetContrib =
        arith::MulIOp::create(rewriter, loc, offsets[i], strides[i]);
    baseOffset =
        arith::AddIOp::create(rewriter, loc, baseOffset, offsetContrib);
  }
  Value indices = gatScatOp.getIndices();
  VectorType vecType = cast<VectorType>(indices.getType());
 
  Value strideVector =
      vector::BroadcastOp::create(rewriter, loc, vecType, strides.back())
          .getResult();
  Value stridedIndices =
      arith::MulIOp::create(rewriter, loc, strideVector, indices).getResult();
 
  Value baseVector =
      vector::BroadcastOp::create(
          rewriter, loc,
          VectorType::get(vecType.getShape(), rewriter.getIndexType()),
          baseOffset)
          .getResult();
  return arith::AddIOp::create(rewriter, loc, baseVector, stridedIndices)
      .getResult();
}
 
// Collapses shapes of a nD memref to the target rank while applying offsets for
// the collapsed dimensions. Returns the new memref value and the remaining
// offsets for the last targetRank dimensions. For example:
//   input: %memref = memref<2x4x8x32xf32>, offsets=[%i0, %i1, %i2, %i3],
//   output: %memref[%i0, %i1, 0, 0] -> memref<8x32xf32>, offsets: [%i2, %i3]
static std::pair<Value, SmallVector<OpFoldResult>>
convertMemrefAndOffsetsToTargetRank(PatternRewriter &rewriter, Location loc,
                                    Value memref,
                                    SmallVector<OpFoldResult> offsets,
                                    int64_t targetRank) {
  auto memrefType = cast<MemRefType>(memref.getType());
  unsigned rank = memrefType.getRank();
 
  if (rank <= targetRank)
    return {memref, offsets};
 
  int64_t numCombinedDims = rank - targetRank;
  SmallVector<OpFoldResult> subviewOffsets;
  SmallVector<OpFoldResult> subviewSizes;
  SmallVector<OpFoldResult> subviewStrides;
 
  // For the combined dimensions: use the provided offsets, size=1, stride=1
  for (unsigned i = 0; i < numCombinedDims; ++i) {
    subviewOffsets.push_back(offsets[i]);
    subviewSizes.push_back(rewriter.getI64IntegerAttr(1));
    subviewStrides.push_back(rewriter.getI64IntegerAttr(1));
  }
 
  // For the last targetRank dimensions: offset=0, use full size, stride=1
  SmallVector<int64_t> resultShape;
  auto originalShape = memrefType.getShape();
  auto meta = memref::ExtractStridedMetadataOp::create(rewriter, loc, memref);
  for (unsigned i = numCombinedDims; i < rank; ++i) {
    subviewOffsets.push_back(rewriter.getI64IntegerAttr(0));
    if (ShapedType::isDynamic(originalShape[i])) {
      subviewSizes.push_back(meta.getSizes()[i]);
      resultShape.push_back(ShapedType::kDynamic);
    } else {
      subviewSizes.push_back(rewriter.getI64IntegerAttr(originalShape[i]));
      resultShape.push_back(originalShape[i]);
    }
    subviewStrides.push_back(rewriter.getI64IntegerAttr(1));
  }
 
  auto resultType = memref::SubViewOp::inferRankReducedResultType(
      resultShape, memrefType, subviewOffsets, subviewSizes, subviewStrides);
  auto subviewOp =
      memref::SubViewOp::create(rewriter, loc, resultType, memref,
                                subviewOffsets, subviewSizes, subviewStrides);
 
  // Return the remaining offsets for the last targetRank dimensions
  SmallVector<OpFoldResult> newOffsets(offsets.begin() + numCombinedDims,
                                       offsets.end());
  return {subviewOp.getResult(), newOffsets};
}
 
template <
    typename OpType,
    typename = std::enable_if_t<llvm::is_one_of<
        std::decay_t<OpType>, vector::TransferReadOp, vector::TransferWriteOp,
        vector::GatherOp, vector::ScatterOp>::value>>
// Convert memref to i64 base pointer
static Value memrefToIndexPtr(OpType xferOp, PatternRewriter &rewriter) {
  Location loc = xferOp.getLoc();
  auto indexPtr = memref::ExtractAlignedPointerAsIndexOp::create(
                      rewriter, loc, xferOp.getBase())
                      .getResult();
  return arith::IndexCastOp::create(rewriter, loc, rewriter.getI64Type(),
                                    indexPtr)
      .getResult();
}
 
static LogicalResult lowerToScatteredLoadOp(vector::TransferReadOp readOp,
                                            PatternRewriter &rewriter) {
 
  Location loc = readOp.getLoc();
  VectorType vectorType = readOp.getVectorType();
  ArrayRef<int64_t> vectorShape = vectorType.getShape();
  auto memrefType = dyn_cast<MemRefType>(readOp.getShapedType());
  if (!memrefType)
    return rewriter.notifyMatchFailure(readOp, "Expected memref source");
 
  auto meta = computeMemrefMeta(readOp, rewriter);
  if (meta.first.empty())
    return rewriter.notifyMatchFailure(readOp, "Failed to compute strides");
 
  Value localOffsets =
      computeOffsets(readOp, rewriter, meta.first, meta.second);
 
  Value flatMemref = memrefToIndexPtr(readOp, rewriter);
 
  Value mask = vector::ConstantMaskOp::create(
      rewriter, loc, VectorType::get(vectorShape, rewriter.getI1Type()),
      vectorShape);
  auto gatherOp = xegpu::LoadGatherOp::create(
      rewriter, loc, vectorType, flatMemref, localOffsets, mask,
      /*chunk_size=*/IntegerAttr{},
      /*l1_hint=*/xegpu::CachePolicyAttr{},
      /*l2_hint=*/xegpu::CachePolicyAttr{},
      /*l3_hint=*/xegpu::CachePolicyAttr{},
      /*layout=*/nullptr, /*contiguity=*/nullptr);
 
  rewriter.replaceOp(readOp, gatherOp.getResult());
  return success();
}
 
static LogicalResult lowerToScatteredStoreOp(vector::TransferWriteOp writeOp,
                                             PatternRewriter &rewriter) {
 
  Location loc = writeOp.getLoc();
  VectorType vectorType = writeOp.getVectorType();
  ArrayRef<int64_t> vectorShape = vectorType.getShape();
 
  auto memrefType = dyn_cast<MemRefType>(writeOp.getShapedType());
  if (!memrefType)
    return rewriter.notifyMatchFailure(writeOp, "Expected memref source");
 
  auto meta = computeMemrefMeta(writeOp, rewriter);
  if (meta.first.empty())
    return rewriter.notifyMatchFailure(writeOp, "Failed to compute strides");
 
  Value localOffsets =
      computeOffsets(writeOp, rewriter, meta.first, meta.second);
 
  Value flatMemref = memrefToIndexPtr(writeOp, rewriter);
 
  Value mask = vector::ConstantMaskOp::create(
      rewriter, loc, VectorType::get(vectorShape, rewriter.getI1Type()),
      vectorShape);
  xegpu::StoreScatterOp::create(rewriter, loc, writeOp.getVector(), flatMemref,
                                localOffsets, mask,
                                /*chunk_size=*/IntegerAttr{},
                                /*l1_hint=*/xegpu::CachePolicyAttr{},
                                /*l2_hint=*/xegpu::CachePolicyAttr{},
                                /*l3_hint=*/xegpu::CachePolicyAttr{},
                                /*layout=*/nullptr, /*contiguity=*/nullptr);
  rewriter.eraseOp(writeOp);
  return success();
}
 
struct TransferReadLowering : public OpRewritePattern<vector::TransferReadOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::TransferReadOp readOp,
                                PatternRewriter &rewriter) const override {
    Location loc = readOp.getLoc();
 
    if (failed(transferPreconditions(rewriter, readOp)))
      return failure();
    auto readMemTy = cast<MemRefType>(readOp.getShapedType());
    VectorType loadedVecTy = readOp.getVectorType();
    bool isOutOfBounds = readOp.hasOutOfBoundsDim();
    // Check if the memref has address space 3 (shared local memory)
    bool isSharedMemory = xegpu::XeGPUDialect::isSharedMemory(readMemTy);
    // Handle the SLM case.
    if (isSharedMemory) {
      // load_matrix supports 1D and 2D loads from SLM.
      if (loadedVecTy.getRank() != 1 && loadedVecTy.getRank() != 2)
        return rewriter.notifyMatchFailure(
            readOp, "Only 1D and 2D vector loads are supported for SLM");
      AffineMap readMap = readOp.getPermutationMap();
      if (!readMap.isMinorIdentity())
        return rewriter.notifyMatchFailure(
            readOp,
            "Non identity transposition is not supported for SLM loads.");
      // Out of bounds case is not supported for SLM loads.
      if (isOutOfBounds)
        return rewriter.notifyMatchFailure(
            readOp, "Out-of-bounds access is not supported for SLM loads");
 
      // Create mem_desc for SLM
      auto memDescType =
          xegpu::MemDescType::get(rewriter.getContext(), readMemTy.getShape(),
                                  readMemTy.getElementType(),
                                  /*mem_layout=*/nullptr);
      auto createMemDescOp = xegpu::CreateMemDescOp::create(
          rewriter, loc, memDescType, readOp.getBase());
      // Convert indices to OpFoldResult for LoadMatrixOp
      SmallVector<OpFoldResult> indices =
          getAsOpFoldResult(readOp.getIndices());
      auto loadMatrixOp = xegpu::LoadMatrixOp::create(
          rewriter, loc, loadedVecTy, createMemDescOp.getResult(), indices,
          /*layout=*/nullptr);
 
      rewriter.replaceOp(readOp, loadMatrixOp.getResult());
      return success();
    }
 
    // TODO: This check needs to be replaced with proper uArch capability check.
    auto chip = xegpu::getChipStr(readOp);
    bool hasBlockLoadSupport =
        (chip == "pvc" || chip == "bmg" || chip == "cri");
 
    // An nd block load can realize a minor-identity map directly, or an
    // innermost-two-dims transpose via a trailing vector.transpose. Any other
    // permutation (e.g. a mid-vector transpose of a high-dim load) is left to
    // the scattered path, which permutes strides explicitly.
    AffineMap readMap = readOp.getPermutationMap();
    bool isTransposeLoad = isInnermostTwoDimsTransposed(readMap);
 
    // Prefer an nd block load. It requires HW block-load support, a vector of
    // rank >= 2 backed by a scalar-element memref, and a map the block load can
    // realize. 1D vectors use the scattered xegpu.load path instead, which has
    // a richer interface (e.g. layout capabilities). Out-of-bounds reads are
    // allowed as long as the padding matches load_nd's implicit zero padding.
    bool canLowerToLoadNd =
        hasBlockLoadSupport && loadedVecTy.getRank() > 1 &&
        (readMap.isMinorIdentity() || isTransposeLoad) &&
        readMemTy.getElementType().isIntOrFloat() &&
        (!isOutOfBounds || isZeroOrPoisonPadding(readOp.getPadding()));
 
    if (canLowerToLoadNd) {
      auto elementType = loadedVecTy.getElementType();
 
      SmallVector<int64_t> descShape(loadedVecTy.getShape());
      if (isTransposeLoad) {
        // If load is transposed, simply swap the last two dimensions of the
        // loaded vector type to get the descriptor shape.
        size_t rank = descShape.size();
        assert(rank >= 2 && "Transpose requires at least 2 dimensions");
        std::swap(descShape[rank - 1], descShape[rank - 2]);
        loadedVecTy = VectorType::get(descShape, elementType);
      }
      auto descType = xegpu::TensorDescType::get(
          descShape, elementType, /*array_length=*/1,
          /*boundary_check=*/isOutOfBounds, xegpu::MemorySpace::Global);
      auto [src, indices] = convertMemrefAndOffsetsToTargetRank(
          rewriter, loc, readOp.getBase(),
          getAsOpFoldResult(readOp.getIndices()), loadedVecTy.getRank());
      // By default, no specific caching policy is assigned.
      xegpu::CachePolicyAttr hint = nullptr;
      xegpu::CreateNdDescOp ndDesc = xegpu::CreateNdDescOp::create(
          rewriter, loc, descType, dyn_cast<TypedValue<MemRefType>>(src));
 
      Operation *loadedOp =
          xegpu::LoadNdOp::create(rewriter, loc, loadedVecTy, ndDesc, indices,
                                  /*packed=*/nullptr, /*transpose=*/nullptr,
                                  /*l1_hint=*/hint,
                                  /*l2_hint=*/hint, /*l3_hint=*/hint,
                                  /*layout=*/nullptr);
      if (isTransposeLoad) {
        // Undo the innermost-two-dims swap with a trailing vector.transpose:
        // keep the leading dimensions in place and interchange only the last
        // two.
        int64_t rank = loadedVecTy.getRank();
        SmallVector<int64_t> perm(llvm::to_vector(llvm::seq<int64_t>(0, rank)));
        std::swap(perm[rank - 1], perm[rank - 2]);
        loadedOp = vector::TransposeOp::create(rewriter, loc,
                                               loadedOp->getResult(0), perm);
      }
      rewriter.replaceOp(readOp, loadedOp);
      return success();
    }
 
    // Fall back to a scattered load. It supports arbitrary permutations and any
    // rank, but cannot express out-of-bounds accesses.
    // TODO: add support for OutOfBound access.
    if (isOutOfBounds)
      return failure();
    return lowerToScatteredLoadOp(readOp, rewriter);
  }
};
 
struct TransferWriteLowering
    : public OpRewritePattern<vector::TransferWriteOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::TransferWriteOp writeOp,
                                PatternRewriter &rewriter) const override {
    Location loc = writeOp.getLoc();
 
    if (failed(transferPreconditions(rewriter, writeOp)))
      return failure();
    // Perform common data transfer checks.
    VectorType vecTy = writeOp.getVectorType();
    auto writeMemTy = cast<MemRefType>(writeOp.getShapedType());
    // Check if the memref has address space 3 (shared local memory)
    bool isSharedMemory = xegpu::XeGPUDialect::isSharedMemory(writeMemTy);
 
    // For shared local memory (address space 3), use create_mem_desc +
    // store_matrix
    if (isSharedMemory) {
      // store_matrix supports 1D and 2D stores to SLM.
      if (vecTy.getRank() != 1 && vecTy.getRank() != 2)
        return rewriter.notifyMatchFailure(
            writeOp, "Only 1D and 2D vector stores are supported for SLM");
      // Create mem_desc for SLM
      auto memDescType =
          xegpu::MemDescType::get(rewriter.getContext(), writeMemTy.getShape(),
                                  writeMemTy.getElementType(),
                                  /*mem_layout=*/nullptr);
 
      auto createMemDescOp = xegpu::CreateMemDescOp::create(
          rewriter, loc, memDescType, writeOp.getBase());
 
      // Convert indices to OpFoldResult for StoreMatrixOp
      SmallVector<OpFoldResult> indices =
          getAsOpFoldResult(writeOp.getIndices());
 
      xegpu::StoreMatrixOp::create(rewriter, loc, writeOp.getVector(),
                                   createMemDescOp.getResult(), indices,
                                   /*layout=*/nullptr);
 
      rewriter.eraseOp(writeOp);
      return success();
    }
 
    // TODO: This check needs to be replaced with proper uArch capability check.
    auto chip = xegpu::getChipStr(writeOp);
    bool hasBlockStoreSupport =
        (chip == "pvc" || chip == "bmg" || chip == "cri");
 
    // Prefer an nd block store. It requires HW block-store support, a vector of
    // rank >= 2 backed by a scalar-element memref, and a minor-identity map
    // (block stores have no transpose support). 1D vectors use the scattered
    // xegpu.store path instead, which has a richer interface. Out-of-bounds
    // writes are handled by the descriptor's boundary check.
    AffineMap map = writeOp.getPermutationMap();
    bool canLowerToStoreNd = hasBlockStoreSupport && vecTy.getRank() > 1 &&
                             map.isMinorIdentity() &&
                             writeMemTy.getElementType().isIntOrFloat();
 
    if (canLowerToStoreNd) {
      auto [src, indices] = convertMemrefAndOffsetsToTargetRank(
          rewriter, loc, writeOp.getBase(),
          getAsOpFoldResult(writeOp.getIndices()), vecTy.getRank());
 
      auto descType = xegpu::TensorDescType::get(
          vecTy.getShape(), vecTy.getElementType(),
          /*array_length=*/1, /*boundary_check=*/writeOp.hasOutOfBoundsDim(),
          xegpu::MemorySpace::Global);
      // By default, no specific caching policy is assigned.
      xegpu::CachePolicyAttr hint = nullptr;
      xegpu::CreateNdDescOp ndDesc = xegpu::CreateNdDescOp::create(
          rewriter, loc, descType, dyn_cast<TypedValue<MemRefType>>(src));
 
      auto storeOp = xegpu::StoreNdOp::create(
          rewriter, loc, writeOp.getVector(), ndDesc, indices,
          /*l1_hint=*/hint,
          /*l2_hint=*/hint, /*l3_hint=*/hint,
          /*layout=*/nullptr);
      rewriter.replaceOp(writeOp, storeOp);
      return success();
    }
 
    // Fall back to a scattered store. It supports arbitrary permutations and
    // any rank, but cannot express out-of-bounds accesses.
    // TODO: add support for OutOfBound access.
    if (writeOp.hasOutOfBoundsDim())
      return failure();
    return lowerToScatteredStoreOp(writeOp, rewriter);
  }
};
 
struct GatherLowering : public OpRewritePattern<vector::GatherOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::GatherOp gatherOp,
                                PatternRewriter &rewriter) const override {
    auto srcTy = dyn_cast<MemRefType>(gatherOp.getBase().getType());
    if (!srcTy)
      return rewriter.notifyMatchFailure(gatherOp, "Expects memref source");
 
    Location loc = gatherOp.getLoc();
    VectorType vectorType = gatherOp.getVectorType();
 
    auto meta = computeMemrefMeta(gatherOp, rewriter);
    if (meta.first.empty())
      return rewriter.notifyMatchFailure(gatherOp, "Failed to compute strides");
 
    Value localOffsets =
        computeOffsets(rewriter, gatherOp, meta.first, meta.second);
    Value flatMemref = memrefToIndexPtr(gatherOp, rewriter);
 
    auto xeGatherOp = xegpu::LoadGatherOp::create(
        rewriter, loc, vectorType, flatMemref, localOffsets, gatherOp.getMask(),
        /*chunk_size=*/IntegerAttr{},
        /*l1_hint=*/xegpu::CachePolicyAttr{},
        /*l2_hint=*/xegpu::CachePolicyAttr{},
        /*l3_hint=*/xegpu::CachePolicyAttr{},
        /*layout=*/nullptr, /*contiguity=*/nullptr);
 
    auto selectOp =
        arith::SelectOp::create(rewriter, loc, gatherOp.getMask(),
                                xeGatherOp.getResult(), gatherOp.getPassThru());
    rewriter.replaceOp(gatherOp, selectOp.getResult());
    return success();
  }
};
 
struct ScatterLowering : public OpRewritePattern<vector::ScatterOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::ScatterOp scatterOp,
                                PatternRewriter &rewriter) const override {
    auto srcTy = dyn_cast<MemRefType>(scatterOp.getBase().getType());
    if (!srcTy)
      return rewriter.notifyMatchFailure(scatterOp, "Expects memref source");
 
    Location loc = scatterOp.getLoc();
    auto meta = computeMemrefMeta(scatterOp, rewriter);
    if (meta.first.empty())
      return rewriter.notifyMatchFailure(scatterOp,
                                         "Failed to compute strides");
 
    Value localOffsets =
        computeOffsets(rewriter, scatterOp, meta.first, meta.second);
    Value flatMemref = memrefToIndexPtr(scatterOp, rewriter);
 
    xegpu::StoreScatterOp::create(rewriter, loc, scatterOp.getValueToStore(),
                                  flatMemref, localOffsets, scatterOp.getMask(),
                                  /*chunk_size=*/IntegerAttr{},
                                  /*l1_hint=*/xegpu::CachePolicyAttr{},
                                  /*l2_hint=*/xegpu::CachePolicyAttr{},
                                  /*l3_hint=*/xegpu::CachePolicyAttr{},
                                  /*layout=*/nullptr,
                                  /*contiguity=*/nullptr);
    rewriter.eraseOp(scatterOp);
    return success();
  }
};
 
struct LoadLowering : public OpRewritePattern<vector::LoadOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::LoadOp loadOp,
                                PatternRewriter &rewriter) const override {
    Location loc = loadOp.getLoc();
 
    VectorType vecTy = loadOp.getResult().getType();
    MemRefType memTy = loadOp.getBase().getType();
    // The plain vector.load lowering only supports 1D/2D block loads.
    if (vecTy.getRank() != 1 && vecTy.getRank() != 2)
      return rewriter.notifyMatchFailure(loadOp, "Expects 1D or 2D vector");
    if (!memTy.getElementType().isIntOrFloat())
      return rewriter.notifyMatchFailure(
          loadOp, "Unsupported memref element type: expected integer or float");
 
    // Boundary check is available only for block instructions.
    bool boundaryCheck = vecTy.getRank() > 1;
    // By default, no specific caching policy is assigned.
    xegpu::CachePolicyAttr hint = nullptr;
 
    auto [src, indices] = convertMemrefAndOffsetsToTargetRank(
        rewriter, loc, loadOp.getBase(), getAsOpFoldResult(loadOp.getIndices()),
        vecTy.getRank());
 
    auto descType = xegpu::TensorDescType::get(
        vecTy.getShape(), vecTy.getElementType(), /*array_length=*/1,
        boundaryCheck, xegpu::MemorySpace::Global);
 
    xegpu::CreateNdDescOp ndDesc = xegpu::CreateNdDescOp::create(
        rewriter, loc, descType, dyn_cast<TypedValue<MemRefType>>(src));
    auto loadNdOp =
        xegpu::LoadNdOp::create(rewriter, loc, vecTy, ndDesc, indices,
                                /*packed=*/nullptr, /*transpose=*/nullptr,
                                /*l1_hint=*/hint,
                                /*l2_hint=*/hint, /*l3_hint=*/hint,
                                /*layout=*/nullptr);
    rewriter.replaceOp(loadOp, loadNdOp);
 
    return success();
  }
};
 
struct StoreLowering : public OpRewritePattern<vector::StoreOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::StoreOp storeOp,
                                PatternRewriter &rewriter) const override {
    Location loc = storeOp.getLoc();
 
    TypedValue<VectorType> vector = storeOp.getValueToStore();
    VectorType vecTy = vector.getType();
    MemRefType memTy = storeOp.getBase().getType();
    // The plain vector.store lowering only supports 1D/2D block stores.
    if (vecTy.getRank() != 1 && vecTy.getRank() != 2)
      return rewriter.notifyMatchFailure(storeOp, "Expects 1D or 2D vector");
    if (!memTy.getElementType().isIntOrFloat())
      return rewriter.notifyMatchFailure(
          storeOp,
          "Unsupported memref element type: expected integer or float");
 
    // Boundary check is available only for block instructions.
    bool boundaryCheck = vecTy.getRank() > 1;
 
    auto [src, indices] = convertMemrefAndOffsetsToTargetRank(
        rewriter, loc, storeOp.getBase(),
        getAsOpFoldResult(storeOp.getIndices()), vecTy.getRank());
 
    auto descType = xegpu::TensorDescType::get(
        vecTy.getShape(), vecTy.getElementType(),
        /*array_length=*/1, boundaryCheck, xegpu::MemorySpace::Global);
 
    // By default, no specific caching policy is assigned.
    xegpu::CachePolicyAttr hint = nullptr;
    xegpu::CreateNdDescOp ndDesc = xegpu::CreateNdDescOp::create(
        rewriter, loc, descType, dyn_cast<TypedValue<MemRefType>>(src));
 
    auto storeNdOp =
        xegpu::StoreNdOp::create(rewriter, loc, vector, ndDesc, indices,
                                 /*l1_hint=*/hint,
                                 /*l2_hint=*/hint, /*l3_hint=*/hint,
                                 /*layout=*/nullptr);
 
    rewriter.replaceOp(storeOp, storeNdOp);
 
    return success();
  }
};
 
// If `indexingMaps` describe a (batched) row-major matmul
//   lhs[b..., m, k], rhs[b..., k, n], acc[b..., m, n]
// return the number of leading batch dims (0 for a plain 2D matmul);
// otherwise return std::nullopt.
static std::optional<int64_t>
getRowMajorMatmulBatchRank(ArrayAttr indexingMaps) {
  if (indexingMaps.size() != 3)
    return std::nullopt;
 
  AffineMap mapA = cast<AffineMapAttr>(indexingMaps[0]).getValue();
  AffineMap mapB = cast<AffineMapAttr>(indexingMaps[1]).getValue();
  AffineMap mapC = cast<AffineMapAttr>(indexingMaps[2]).getValue();
 
  // The result map exposes the batch dims followed by the core (m, n) dims.
  if (mapC.getNumResults() < 2)
    return std::nullopt;
  int64_t batchRank = mapC.getNumResults() - 2;
 
  // A single `k` reduction gives batchRank + 3 iteration dims; each operand
  // map exposes batchRank + 2 dims (batch dims + 2 core dims).
  unsigned numDims = static_cast<unsigned>(batchRank) + 3;
  unsigned numOperandResults = static_cast<unsigned>(batchRank) + 2;
  if (mapA.getNumInputs() != numDims || mapB.getNumInputs() != numDims ||
      mapC.getNumInputs() != numDims)
    return std::nullopt;
  if (mapA.getNumResults() != numOperandResults ||
      mapB.getNumResults() != numOperandResults)
    return std::nullopt;
 
  // Reconstruct the canonical maps from the batch/m/n dims of the result and
  // the k dim of lhs, then compare against the actual maps.
  MLIRContext *context = indexingMaps.getContext();
  ArrayRef<AffineExpr> batchDims = mapC.getResults().take_front(batchRank);
  AffineExpr m = mapC.getResult(batchRank);
  AffineExpr n = mapC.getResult(batchRank + 1);
  AffineExpr k = mapA.getResult(batchRank + 1);
 
  SmallVector<AffineExpr> aDims = llvm::to_vector(batchDims);
  aDims.push_back(m);
  aDims.push_back(k);
  SmallVector<AffineExpr> bDims = llvm::to_vector(batchDims);
  bDims.push_back(k);
  bDims.push_back(n);
  SmallVector<AffineExpr> cDims = llvm::to_vector(batchDims);
  cDims.push_back(m);
  cDims.push_back(n);
 
  auto expected = ArrayAttr::get(
      context,
      {AffineMapAttr::get(AffineMap::get(numDims, 0, aDims, context)),
       AffineMapAttr::get(AffineMap::get(numDims, 0, bDims, context)),
       AffineMapAttr::get(AffineMap::get(numDims, 0, cDims, context))});
  if (indexingMaps != expected)
    return std::nullopt;
  return batchRank;
}
 
struct ContractionLowering : public OpRewritePattern<vector::ContractionOp> {
  using Base::Base;
 
  LogicalResult matchAndRewrite(vector::ContractionOp contractOp,
                                PatternRewriter &rewriter) const override {
    Location loc = contractOp.getLoc();
 
    if (contractOp.getKind() != vector::CombiningKind::ADD)
      return rewriter.notifyMatchFailure(contractOp,
                                         "Expects add combining kind");
 
    TypedValue<VectorType> lhs = contractOp.getLhs();
    TypedValue<VectorType> rhs = contractOp.getRhs();
    TypedValue<Type> acc = contractOp.getAcc();
    VectorType accType = dyn_cast<VectorType>(acc.getType());
    if (!accType)
      return rewriter.notifyMatchFailure(contractOp, "Expects vector acc");
 
    std::optional<int64_t> batchRank =
        getRowMajorMatmulBatchRank(contractOp.getIndexingMapsAttr());
    if (!batchRank)
      return rewriter.notifyMatchFailure(
          contractOp,
          "Expects a (batched) row-major matmul: leading dims must "
          "be batch dims shared by lhs, rhs, and acc; innermost two "
          "dims must be (M, K), (K, N), and (M, N)");
 
    // xegpu.dpas operands are limited to 2 batch + 2 core dims.
    if (*batchRank > 2)
      return rewriter.notifyMatchFailure(contractOp,
                                         "Expects operands of rank 4 or less");
 
    auto dpasOp = xegpu::DpasOp::create(rewriter, loc,
                                        TypeRange{contractOp.getResultType()},
                                        ValueRange{lhs, rhs, acc});
    rewriter.replaceOp(contractOp, dpasOp);
 
    return success();
  }
};
 
// Returns `memrefTy` with its memory space replaced by `newMemSpace`.
static MemRefType withMemorySpace(MemRefType memrefTy, Attribute newMemSpace) {
  return MemRefType::get(memrefTy.getShape(), memrefTy.getElementType(),
                         memrefTy.getLayout(), newMemSpace);
}
 
// Rewrite every `memref.alloca` not already in shared local memory (SLM) to
// be in SLM (address space 3), and propagate the new memory space through
// memref-producing aliasing users (e.g. memref.cast, memref.subview,
// memref.expand_shape, ...). Consumers that take a memref operand but
// produce a non-memref result (e.g. vector.transfer_read, vector.load) are
// left untouched: their operand type simply reflects the new memory space.
//
// This makes `xegpu.load_matrix`/`xegpu.store_matrix` lowering work end-to-end
// for IR coming from bufferization, which by default assigns memory space 0/1
// to allocations.
static void promoteAllocasToSLM(Operation *root) {
  MLIRContext *ctx = root->getContext();
  Attribute slmAttr = IntegerAttr::get(IntegerType::get(ctx, 64), 3);
 
  // A user is treated as a memref-producing alias (e.g. memref.cast,
  // memref.subview, memref.expand_shape, ...) if it is side-effect free and
  // produces at least one memref result. This excludes ops like memref.copy
  // that have memory effects.
  auto isMemrefResultOp = [](Operation *op) {
    if (!isMemoryEffectFree(op))
      return false;
    return llvm::any_of(op->getResultTypes(),
                        [](Type t) { return isa<MemRefType>(t); });
  };
 
  // Update `v`'s type to have SLM memory space, then walk forward through
  // memref-producing users and update their result types accordingly.
  std::function<void(Value)> propagate = [&](Value v) {
    auto memrefTy = dyn_cast<MemRefType>(v.getType());
    if (!memrefTy || xegpu::XeGPUDialect::isSharedMemory(memrefTy))
      return;
    v.setType(withMemorySpace(memrefTy, slmAttr));
    for (Operation *user : v.getUsers()) {
      if (!isMemrefResultOp(user))
        continue;
      for (Value result : user->getResults())
        propagate(result);
    }
  };
 
  SmallVector<memref::AllocaOp> allocas;
  root->walk([&](memref::AllocaOp op) {
    auto memrefTy = dyn_cast<MemRefType>(op.getResult().getType());
    if (!memrefTy || xegpu::XeGPUDialect::isSharedMemory(memrefTy))
      return;
    allocas.push_back(op);
  });
 
  for (memref::AllocaOp alloca : allocas) {
    OpBuilder builder(alloca);
    auto memrefTy = cast<MemRefType>(alloca.getResult().getType());
    auto newTy = withMemorySpace(memrefTy, slmAttr);
    auto newOp = memref::AllocaOp::create(
        builder, alloca.getLoc(), newTy, alloca.getDynamicSizes(),
        alloca.getSymbolOperands(), alloca.getAlignmentAttr());
    alloca.getResult().replaceAllUsesWith(newOp.getResult());
    alloca.erase();
    // Propagate the new memory space through memref-producing consumers.
    for (Operation *user : newOp.getResult().getUsers()) {
      if (!isMemrefResultOp(user))
        continue;
      for (Value result : user->getResults())
        propagate(result);
    }
  }
}
 
struct ConvertVectorToXeGPUPass
    : public impl::ConvertVectorToXeGPUBase<ConvertVectorToXeGPUPass> {
  void runOnOperation() override {
    // Promote local allocations to SLM (address space 3) so that
    // load_matrix/store_matrix lowerings have well-typed memref operands.
    promoteAllocasToSLM(getOperation());
 
    RewritePatternSet patterns(&getContext());
    populateVectorToXeGPUConversionPatterns(patterns);
    populatePrepareVectorToMMAPatterns(patterns);
    if (failed(applyPatternsGreedily(getOperation(), std::move(patterns))))
      return signalPassFailure();
  }
};
 
} // namespace
 
void mlir::populateVectorToXeGPUConversionPatterns(
    RewritePatternSet &patterns) {
  patterns
      .add<TransferReadLowering, TransferWriteLowering, LoadLowering,
           ScatterLowering, GatherLowering, StoreLowering, ContractionLowering>(
          patterns.getContext());
}