8000 Revert "[Dispatch Creation] Disable broken reshape fusion by collapsi… by IanWood1 · Pull Request #20653 · iree-org/iree · GitHub
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

Revert "[Dispatch Creation] Disable broken reshape fusion by collapsi… #20653

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 1 commit into from
Apr 28, 2025
Merged

Revert "[Dispatch Creation] Disable broken reshape fusion by collapsi… #20653

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
9 changes: 2 additions & 7 deletions compiler/src/iree/compiler/DispatchCreation/BubbleUpExpandShapes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -131,13 +131,8 @@ void BubbleUpExpandShapesPass::runOnOperation() {
return false;
}

// Do not push down collapse shape across consumer if it is a bit-extend
// op. The bit-extend ops get cloned into producer dispatches, and the
// `collapse_shape` op going past dequant, prevents this clong.
if (IREE::LinalgExt::isBitExtendOp(consumer)) {
return false;
}

// If producer generic op is elementwise op, bubble up the expand shape
// past this operation.
if (auto producerGenericOp = dyn_cast<linalg::GenericOp>(producer)) {
// If producer generic op is elementwise op, bubble up the expand
// shape past this operation.
Expand Down
52 changes: 40 additions & 12 deletions compiler/src/iree/compiler/DispatchCreation/SinkReshapes.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/MemRef/Transforms/Transforms.h"
#include "mlir/Dialect/Tensor/Transforms/Transforms.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"

Expand Down Expand Up @@ -54,17 +55,24 @@ static bool isFusableUsingTileAndFuse(Operation *producer,
IREE::Encoding::UnsetEncodingOp>(producer);
}

/// Control function to check if an `tensor.expand_shape` (which is producer of
/// `opOperand`) should be pushed past the `genericOp` (which is the consumer of
/// `opOperand`).
static bool shouldSinkExpandShapeOp(OpOperand *opOperand) {
auto reshapeOp =
dyn_cast<tensor::ExpandShapeOp>(opOperand->get().getDefiningOp());
if (!reshapeOp) {
/// Control function to check if a `tensor.collapse_shape` (which is the
/// consumer of `opOperand`) should be bubbled through the `genericOp`.
static bool shouldBubbleCollapseShapeOp(tensor::CollapseShapeOp collapseOp,
OpOperand *opOperand) {
auto *producer = opOperand->get().getDefiningOp();
if (!producer) {
return false;
}
return IREE::Flow::isClonableIntoDispatchOp(opOperand->get().getDefiningOp());
}

/// Control function to check if a `tensor.expand_shape` (which is producer of
/// `opOperand`) should be pushed past the `genericOp` (which is the consumer of
/// `opOperand`).
static bool shouldSinkExpandShapeOp(tensor::ExpandShapeOp expandOp,
OpOperand *opOperand) {
Operation *consumer = opOperand->getOwner();
if (!IREE::Flow::isNonNullAndOutsideDispatch({reshapeOp, consumer})) {
if (!IREE::Flow::isNonNullAndOutsideDispatch({expandOp, consumer})) {
return false;
}
auto consumerGenericOp = dyn_cast<linalg::GenericOp>(consumer);
Expand All @@ -84,12 +92,11 @@ static bool shouldSinkExpandShapeOp(OpOperand *opOperand) {
}

// First check that the expand_shape producer and consumer can be fused.
Operation *reshapeProducer = reshapeOp.getSrc().getDefiningOp();
Operation *reshapeProducer = expandOp.getSrc().getDefiningOp();
if (!reshapeProducer) {
return false;
}
if (!isFusableUsingTileAndFuse(reshapeOp.getSrc().getDefiningOp(),
consumer)) {
if (!isFusableUsingTileAndFuse(expandOp.getSrc().getDefiningOp(), consumer)) {
return false;
}

Expand Down Expand Up @@ -165,10 +172,31 @@ void SinkReshapesPass::runOnOperation() {
MLIRContext *context = &getContext();

RewritePatternSet sinkReshapePatterns(context);

auto collapsingControlFn = [](OpOperand *opOperand) {
auto collapseOp =
llvm::dyn_cast_or_null<tensor::CollapseShapeOp>(opOperand->getOwner());
if (collapseOp) {
return shouldBubbleCollapseShapeOp(collapseOp, opOperand);
}

auto expandOp =
dyn_cast<tensor::ExpandShapeOp>(opOperand->get().getDefiningOp());
if (expandOp) {
return shouldSinkExpandShapeOp(expandOp, opOperand);
}
llvm_unreachable("reshape is neither a collapse or expand op");
};
linalg::populateFoldReshapeOpsByCollapsingPatterns(sinkReshapePatterns,
shouldSinkExpandShapeOp);
collapsingControlFn);
// Add patterns to fold `tensor.empty` and reshape ops.
tensor::populateFoldTensorEmptyPatterns(sinkReshapePatterns);
memref::populateResolveRankedShapedTypeResultDimsPatterns(
sinkReshapePatterns);
tensor::ExpandShapeOp::getCanonicalizationPatterns(sinkReshapePatterns,
context);
tensor::CollapseShapeOp::getCanonicalizationPatterns(sinkReshapePatterns,
context);
if (failed(applyPatternsGreedily(getOperation(),
std::move(sinkReshapePatterns)))) {
getOperation()->emitOpError("failed to sink reshape ops");
Expand Down
35 changes: 35 additions & 0 deletions compiler/src/iree/compiler/DispatchCreation/test/sink_reshapes.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -236,3 +236,38 @@ func.func @fuse_softmax_with_truncate(%arg0 : tensor<4x64x?xf32>) -> tensor<4x64
// CHECK: %[[TRUNC:.+]] = linalg.generic {{.*}} ins(%[[SOFTMAX]]
// CHECK: %[[EXPAND:.+]] = tensor.expand_shape %[[TRUNC]]
// CHECK: return %[[EXPAND]]

// -----

func.func @bubble_across_bit_extend(%arg0: tensor<2x64x32xf16>, %arg1 : tensor<2xf32>) -> tensor<2xf32> {
%empty = tensor.empty() : tensor<2x64x32xf32>
%0 = linalg.generic {
indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
iterator_types = ["parallel", "parallel", "parallel"]}
ins(%arg0 : tensor<2x64x32xf16>) outs(%empty : tensor<2x64x32xf32>) {
^bb0(%b0 : f16, %b1 : f32):
%0 = arith.extf %b0 : f16 to f32
linalg.yield %0 : f32
} -> tensor<2x64x32xf32>
%collapse = tensor.collapse_shape %0 [[0], [1, 2]] : tensor<2x64x32xf32> into tensor<2x2048xf32>
%1 = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
affine_map<(d0, d1) -> (d0)>],
iterator_types = ["parallel", "reduction"]}
ins(%collapse : tensor<2x2048xf32>) outs(%arg1 : tensor<2xf32>) {
^bb0(%b0 : f32, %b1 : f32):
%2 = arith.addf %b0, %b1 : f32
linalg.yield %2 : f32
} -> tensor<2xf32>
func.return %1 : tensor<2xf32>
}
// CHECK-LABEL: func @bubble_across_bit_extend
// CHECK-SAME: %[[ARG0:.+]]: tensor<2x64x32xf16>
// CHECK-SAME: %[[ARG1:.+]]: tensor<2xf32>
// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[ARG0]]
// CHECK: %[[GEN0:.+]] = linalg.generic
// CHECK-SAME: ins(%[[COLLAPSE]] :
// CHECK: %[[GEN1:.+]] = linalg.generic
// CHECK-SAME: ins(%[[GEN0]] :
// CHECK: return %[[GEN1]]
Loading
0