Skip to content

[mlir][vector] Notify the rewriter when sinking out of warp ops #71964

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 3 commits into from
Nov 10, 2023
Merged

[mlir][vector] Notify the rewriter when sinking out of warp ops #71964

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
48a0036
[mlir][vector] NFC: Notify the rewriter when sinking out of warp ops
qedawkins Nov 10, 2023
3604a33
Address comments and fix shape cast
qedawkins Nov 10, 2023
69701ab
Add comment explaining why this is needed
qedawkins Nov 10, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
48 changes: 46 additions & 2 deletions mlir/lib/Dialect/Vector/Transforms/VectorDistribute.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -645,6 +645,11 @@ struct WarpOpElementwise : public OpRewritePattern<WarpExecuteOnLane0Op> {
});
if (!yieldOperand)
return failure();

// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startRootUpdate(warpOp);

Operation *elementWise = yieldOperand->get().getDefiningOp();
unsigned operandIndex = yieldOperand->getOperandNumber();
Value distributedVal = warpOp.getResult(operandIndex);
Expand Down Expand Up @@ -683,6 +688,7 @@ struct WarpOpElementwise : public OpRewritePattern<WarpExecuteOnLane0Op> {
{newWarpOp.getResult(operandIndex).getType()});
rewriter.replaceAllUsesWith(newWarpOp.getResult(operandIndex),
newOp->getResult(0));
rewriter.finalizeRootUpdate(warpOp);
return success();
}
};
Expand Down Expand Up @@ -713,6 +719,9 @@ struct WarpOpConstant : public OpRewritePattern<WarpExecuteOnLane0Op> {
auto dense = dyn_cast<SplatElementsAttr>(constantOp.getValue());
if (!dense)
return failure();
// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startRootUpdate(warpOp);
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Later, would be nice to wrap it with make_scope_exit(...), to account for all the exits, but that can be a cleanup PR.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

makes sense, can do that later.

unsigned operandIndex = yieldOperand->getOperandNumber();
Attribute scalarAttr = dense.getSplatValue<Attribute>();
auto newAttr = DenseElementsAttr::get(
Expand All @@ -721,6 +730,7 @@ struct WarpOpConstant : public OpRewritePattern<WarpExecuteOnLane0Op> {
rewriter.setInsertionPointAfter(warpOp);
Value distConstant = rewriter.create<arith::ConstantOp>(loc, newAttr);
rewriter.replaceAllUsesWith(warpOp.getResult(operandIndex), distConstant);
rewriter.finalizeRootUpdate(warpOp);
return success();
}
};
Expand Down Expand Up @@ -823,7 +833,9 @@ struct WarpOpTransferRead : public OpRewritePattern<WarpExecuteOnLane0Op> {
OpBuilder::InsertionGuard g(rewriter);
WarpExecuteOnLane0Op newWarpOp = warpOp;
Value newMask = read.getMask();
bool hasMask = false;
if (read.getMask()) {
hasMask = true;
// TODO: Distribution of masked reads with non-trivial permutation maps
// requires the distribution of the mask to elementwise match the
// distribution of the permuted written vector. Currently the details
Expand All @@ -840,6 +852,16 @@ struct WarpOpTransferRead : public OpRewritePattern<WarpExecuteOnLane0Op> {
newRetIndices);
newMask = newWarpOp.getResult(newRetIndices[0]);
distributedVal = newWarpOp.getResult(operandIndex);
} else {
// This pattern does not actually change the warp op directly. Instead it
// just rewrites a new transfer read (when not masked) outside of the warp
// op and replaces the correponding result. There are then follow up
// patterns to erase now dead results of the warp op. This erasure allows
// propagation to continue, but this pattern on its own never actually
// tells the pattern rewriter that the warp op "changed." Notify the
// rewriter here that the warp op is changing. Similar situations are
// noted in following patterns.
rewriter.startRootUpdate(warpOp);
}

rewriter.setInsertionPointAfter(newWarpOp);
Expand All @@ -849,9 +871,12 @@ struct WarpOpTransferRead : public OpRewritePattern<WarpExecuteOnLane0Op> {
SmallVector<Value> delinearizedIds;
if (!delinearizeLaneId(rewriter, read.getLoc(), sequentialType.getShape(),
distributedType.getShape(), newWarpOp.getWarpSize(),
newWarpOp.getLaneid(), delinearizedIds))
newWarpOp.getLaneid(), delinearizedIds)) {
if (!hasMask)
rewriter.cancelRootUpdate(warpOp);
return rewriter.notifyMatchFailure(
read, "cannot delinearize lane ID for distribution");
}
assert(!delinearizedIds.empty() || map.getNumResults() == 0);

for (auto it : llvm::zip_equal(indexMap.getResults(), map.getResults())) {
Expand Down Expand Up @@ -890,10 +915,15 @@ struct WarpOpTransferRead : public OpRewritePattern<WarpExecuteOnLane0Op> {
if (!llvm::all_of(newRead->getOperands(), [&](Value value) {
return (newRead.getMask() && value == newRead.getMask()) ||
newWarpOp.isDefinedOutsideOfRegion(value);
}))
})) {
if (!hasMask)
rewriter.cancelRootUpdate(warpOp);
return failure();
}

rewriter.replaceAllUsesWith(distributedVal, newRead);
if (!hasMask)
rewriter.finalizeRootUpdate(warpOp);
return success();
}
};
Expand Down Expand Up @@ -996,7 +1026,11 @@ struct WarpOpForwardOperand : public OpRewritePattern<WarpExecuteOnLane0Op> {
}
if (!valForwarded)
return failure();
// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startRootUpdate(warpOp);
rewriter.replaceAllUsesWith(warpOp.getResult(resultIndex), valForwarded);
rewriter.finalizeRootUpdate(warpOp);
return success();
}
};
Expand Down Expand Up @@ -1024,6 +1058,9 @@ struct WarpOpBroadcast : public OpRewritePattern<WarpExecuteOnLane0Op> {
if (vector::isBroadcastableTo(broadcastSrcType, destVecType) !=
vector::BroadcastableToResult::Success)
return failure();
// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startRootUpdate(warpOp);
SmallVector<size_t> newRetIndices;
WarpExecuteOnLane0Op newWarpOp = moveRegionToNewWarpOpAndAppendReturns(
rewriter, warpOp, {broadcastSrc}, {broadcastSrcType}, newRetIndices);
Expand All @@ -1032,6 +1069,7 @@ struct WarpOpBroadcast : public OpRewritePattern<WarpExecuteOnLane0Op> {
loc, destVecType, newWarpOp->getResult(newRetIndices[0]));
rewriter.replaceAllUsesWith(newWarpOp->getResult(operandNumber),
broadcasted);
rewriter.finalizeRootUpdate(warpOp);
return success();
}
};
Expand All @@ -1046,6 +1084,7 @@ struct WarpOpShapeCast : public OpRewritePattern<WarpExecuteOnLane0Op> {
warpOp, [](Operation *op) { return isa<vector::ShapeCastOp>(op); });
if (!operand)
return failure();

auto oldCastOp = operand->get().getDefiningOp<vector::ShapeCastOp>();

unsigned int operandNumber = operand->getOperandNumber();
Expand Down Expand Up @@ -1133,6 +1172,10 @@ struct WarpOpCreateMask : public OpRewritePattern<WarpExecuteOnLane0Op> {
mask, "cannot delinearize lane ID for distribution");
assert(!delinearizedIds.empty());

// Notify the rewriter that the warp op is changing (see the comment on
// the WarpOpTransferRead pattern).
rewriter.startRootUpdate(warpOp);

AffineExpr s0, s1;
bindSymbols(rewriter.getContext(), s0, s1);
SmallVector<Value> newOperands;
Expand All @@ -1151,6 +1194,7 @@ struct WarpOpCreateMask : public OpRewritePattern<WarpExecuteOnLane0Op> {
auto newMask =
rewriter.create<vector::CreateMaskOp>(loc, distType, newOperands);
rewriter.replaceAllUsesWith(warpOp.getResult(operandIndex), newMask);
rewriter.finalizeRootUpdate(warpOp);
return success();
}
};
Expand Down
20 changes: 20 additions & 0 deletions mlir/test/Dialect/Vector/vector-warp-distribute.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1351,6 +1351,26 @@ func.func @warp_propagate_masked_transfer_read(%laneid: index, %src: memref<4096

// -----

func.func @warp_propagate_masked_transfer_read_shared_mask(%laneid: index, %src: memref<4096x4096xf32>, %index: index, %index2: index, %mask_ub: index) -> (vector<2xf32>, vector<2xf32>) {
%f0 = arith.constant 0.000000e+00 : f32
%c0 = arith.constant 0 : index
%r:2 = vector.warp_execute_on_lane_0(%laneid)[64] -> (vector<2xf32>, vector<2xf32>) {
%mask = vector.create_mask %mask_ub: vector<128xi1>
%0 = vector.transfer_read %src[%c0, %index], %f0, %mask {in_bounds = [true]} : memref<4096x4096xf32>, vector<128xf32>
%1 = vector.transfer_read %src[%c0, %index2], %f0, %mask {in_bounds = [true]} : memref<4096x4096xf32>, vector<128xf32>
vector.yield %0, %1 : vector<128xf32>, vector<128xf32>
}
return %r#0, %r#1 : vector<2xf32>, vector<2xf32>
}

// CHECK-PROP-LABEL: func.func @warp_propagate_masked_transfer_read_shared_mask
// CHECK-PROP: vector.create_mask %{{.*}} : vector<2xi1>
// CHECK-PROP: vector.transfer_read %{{.*}} : memref<4096x4096xf32>, vector<2xf32>
// CHECK-PROP: vector.create_mask %{{.*}} : vector<2xi1>
// CHECK-PROP: vector.transfer_read %{{.*}} : memref<4096x4096xf32>, vector<2xf32>

// -----

func.func @warp_propagate_unconnected_read_write(%laneid: index, %buffer: memref<128xf32>, %f1: f32) -> (vector<2xf32>, vector<4xf32>) {
%f0 = arith.constant 0.000000e+00 : f32
%c0 = arith.constant 0 : index
Expand Down