Skip to content

Commit

Permalink
[mlir][ArmSME] Add support for lowering masked tile_store ops
Browse files Browse the repository at this point in the history 
This patch extends ArmSMEToSCF to support lowering of masked tile_store
ops. Only masks created by 'vector.create_mask' are currently supported.

Example:

  %mask = vector.create_mask %c3, %c2 : vector<[4]x[4]xi1>
  arm_sme.tile_store %tile, %dest[%c0, %c0], %mask : memref<?x?xi32>,
vector<[4]x[4]xi32>

Produces:

  %num_rows = arith.constant 3 : index
  %num_cols = vector.create_mask %c2 : vector<[4]xi1>
  scf.for %slice_idx = %c0 to %num_rows step %c1
    arm_sme.store_tile_slice %tile, %slice_idx, %num_cols, %dest[%slice_idx, %c0]
      : memref<?x?xi32>, vector<[4]xi1>, vector<[4]x[4]xi32>
  • Loading branch information
@c-rhodes
c-rhodes committed Oct 16, 2023
1 parent e5e3e14 commit f5bbd5a
Show file tree
Hide file tree
Showing 2 changed files with 66 additions and 24 deletions.
65 changes: 43 additions & 22 deletions mlir/lib/Conversion/ArmSMEToSCF/ArmSMEToSCF.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -420,38 +420,59 @@ struct TileStoreOpConversion : public OpRewritePattern<arm_sme::TileStoreOp> {
auto tileType = tileStoreOp.getVectorType();
auto tileElementType = tileType.getElementType();

// Create a loop that stores each ZA tile slice from memory.
auto predicateType =
VectorType::get(tileType.getDimSize(1), rewriter.getI1Type(), true);

Value maskCols;
Value upperBound;
auto maskOp = tileStoreOp.getMask();
if (maskOp) {
auto createMaskOp = maskOp.getDefiningOp<vector::CreateMaskOp>();
if (!createMaskOp)
return rewriter.notifyMatchFailure(
tileStoreOp, "unsupported mask op, only 'vector.create_mask' is "
"currently supported");

auto numRows = createMaskOp.getOperands()[0];
auto numCols = createMaskOp.getOperands()[1];

upperBound = numRows;
maskCols =
rewriter.create<vector::CreateMaskOp>(loc, predicateType, numCols);
} else {
// Store all tile slices if no mask.
auto minTileSlices = rewriter.create<arith::ConstantIndexOp>(
loc, arm_sme::getSMETileSliceMinNumElts(tileElementType));
auto vscale =
rewriter.create<vector::VectorScaleOp>(loc, rewriter.getIndexType());
// This describes both the number of ZA tile slices and the number of
// elements in a vector of SVL bits for a given element type (SVL_B,
// SVL_H,
// ..., SVL_Q).
auto numTileSlices =
rewriter.create<arith::MulIOp>(loc, minTileSlices, vscale);

upperBound = numTileSlices;
// Create an 'all true' predicate for the tile slice.
maskCols = rewriter.create<arith::ConstantOp>(
loc, DenseElementsAttr::get(predicateType, true));
}

// Create a loop that stores each (active) active ZA tile slice from memory.
auto step = rewriter.create<arith::ConstantIndexOp>(loc, 1);
auto minTileSlices = rewriter.create<arith::ConstantIndexOp>(
loc, arm_sme::getSMETileSliceMinNumElts(tileElementType));
auto vscale =
rewriter.create<vector::VectorScaleOp>(loc, rewriter.getIndexType());
auto lowerBound = rewriter.create<arith::ConstantIndexOp>(loc, 0);
// This describes both the number of ZA tile slices and the number of
// elements in a vector of SVL bits for a given element type (SVL_B, SVL_H,
// ..., SVL_Q).
auto numTileSlices =
rewriter.create<arith::MulIOp>(loc, minTileSlices, vscale);
auto forOp =
rewriter.create<scf::ForOp>(loc, lowerBound, numTileSlices, step);
auto forOp = rewriter.create<scf::ForOp>(loc, lowerBound, upperBound, step);

rewriter.setInsertionPointToStart(forOp.getBody());

// Create an 'all true' predicate for the tile slice.
auto predicateType =
VectorType::get(tileType.getDimSize(1), rewriter.getI1Type(), true);
auto allTruePredicate = rewriter.create<arith::ConstantOp>(
loc, DenseElementsAttr::get(predicateType, true));

SmallVector<Value> memrefIndices;
auto tileSliceIndex = forOp.getInductionVar();
getMemrefIndices(tileStoreOp.getIndices(),
tileStoreOp.getMemRefType().getRank(), tileSliceIndex,
numTileSlices, memrefIndices, loc, rewriter);
upperBound, memrefIndices, loc, rewriter);
rewriter.replaceOpWithNewOp<arm_sme::StoreTileSliceOp>(
tileStoreOp, tileStoreOp.getValueToStore(), tileSliceIndex,
allTruePredicate, tileStoreOp.getBase(), memrefIndices,
tileStoreOp.getLayout());
tileStoreOp, tileStoreOp.getValueToStore(), tileSliceIndex, maskCols,
tileStoreOp.getBase(), memrefIndices, tileStoreOp.getLayout());

return success();
}
Expand Down
25 changes: 23 additions & 2 deletions mlir/test/Conversion/ArmSMEToSCF/arm-sme-to-scf.mlir
View file
Original file line number Diff line number Diff line change
Expand Up @@ -102,9 +102,9 @@ func.func @arm_sme_tile_load_hor_with_mask_and_nonzero_pad(%src : memref<?x?xi32
// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[C4:.*]] = arith.constant 4 : index
// CHECK-DAG: %[[VSCALE:.*]] = vector.vscale
// CHECK: %[[NUM_TILE_SLICES:.*]] = arith.muli %[[C4]], %[[VSCALE]] : index
// CHECK-DAG: %[[PTRUE_S:.*]] = arith.constant dense<true> : vector<[4]xi1>
// CHECK-DAG: %[[NUM_TILE_SLICES:.*]] = arith.muli %[[C4]], %[[VSCALE]] : index
// CHECK: scf.for %[[TILE_SLICE_INDEX:.*]] = %[[C0]] to %[[NUM_TILE_SLICES]] step %[[C1]] {
// CHECK: %[[PTRUE_S:.*]] = arith.constant dense<true> : vector<[4]xi1>
// CHECK: %[[OFFSET:.*]] = arith.addi %[[C0]], %[[TILE_SLICE_INDEX]] : index
// CHECK: arm_sme.store_tile_slice %[[TILE]], %[[TILE_SLICE_INDEX]], %[[PTRUE_S]], %[[DEST]]{{\[}}%[[OFFSET]], %[[C0]]] : memref<?x?xi32>, vector<[4]xi1>, vector<[4]x[4]xi32>
func.func @arm_sme_tile_store_hor(%tile : vector<[4]x[4]xi32>, %dest : memref<?x?xi32>) {
Expand All @@ -123,6 +123,27 @@ func.func @arm_sme_tile_store_ver(%tile : vector<[4]x[4]xi32>, %dest : memref<?x
return
}

// -----

// CHECK-LABEL: func.func @arm_sme_tile_store_hor_with_mask(
// CHECK-SAME: %[[TILE:.*]]: vector<[4]x[4]xi32>,
// CHECK-SAME: %[[DEST:.*]]: memref<?x?xi32>) {
// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
// CHECK-DAG: %[[C1:.*]] = arith.constant 1 : index
// CHECK-DAG: %[[NUM_ROWS:.*]] = arith.constant 3 : index
// CHECK-DAG: %[[NUM_COLS:.*]] = vector.create_mask %c2 : vector<[4]xi1>
// CHECK-NEXT: scf.for %[[TILE_SLICE_INDEX:.*]] = %[[C0]] to %[[NUM_ROWS]] step %[[C1]] {
// CHECK-NEXT: %[[OFFSET:.*]] = arith.addi %[[C0]], %[[TILE_SLICE_INDEX]] : index
// CHECK-NEXT: arm_sme.store_tile_slice %[[TILE]], %[[TILE_SLICE_INDEX]], %[[NUM_COLS]], %[[DEST]]{{\[}}%[[OFFSET]], %[[C0]]] : memref<?x?xi32>, vector<[4]xi1>, vector<[4]x[4]xi32>
func.func @arm_sme_tile_store_hor_with_mask(%tile : vector<[4]x[4]xi32>, %dest : memref<?x?xi32>) {
%c0 = arith.constant 0 : index
%c2 = arith.constant 2 : index
%c3 = arith.constant 3 : index
%mask = vector.create_mask %c3, %c2 : vector<[4]x[4]xi1>
arm_sme.tile_store %tile, %dest[%c0, %c0], %mask : memref<?x?xi32>, vector<[4]x[4]xi32>
return
}

//===----------------------------------------------------------------------===//
// vector.print
//===----------------------------------------------------------------------===//
Expand Down

0 comments on commit f5bbd5a

Please sign in to comment.