Skip to content
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

[Dispatch] Fold collapse(expand) unit dims #19357

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
+256 −29
Open

[Dispatch] Fold collapse(expand) unit dims #19357

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
f7bebfa
[Dispatch] Fold expand/collapse unit dims
IanWood1 Dec 2, 2024
3880283
Fix two problems from sharktank tests
IanWood1 Dec 4, 2024
3849574
Small cleanup
IanWood1 Dec 5, 2024
65a399d
Merge branch 'main' into fold_reshape_unit_dims
IanWood1 Dec 12, 2024
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
142 changes: 142 additions & 0 deletions compiler/src/iree/compiler/DispatchCreation/FoldUnitExtentDims.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,7 @@
#include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Utils/ReshapeOpsUtils.h"
#include "mlir/Pass/Pass.h"
#include "mlir/Transforms/GreedyPatternRewriteDriver.h"

Expand All @@ -32,6 +33,146 @@ namespace mlir::iree_compiler::DispatchCreation {
#define GEN_PASS_DEF_FOLDUNITEXTENTDIMSPASS
#include "iree/compiler/DispatchCreation/Passes.h.inc"

namespace {

/// Simplify collapse_shape(expand_shape) by removing unneeded unit dimensions
/// that get expanded and subsequently collapsed.
struct DropUnitDimsFromCollapseOfExpand
: OpRewritePattern<tensor::CollapseShapeOp> {
using OpRewritePattern::OpRewritePattern;

LogicalResult matchAndRewrite(tensor::CollapseShapeOp collapseOp,
PatternRewriter &rewriter) const override {
auto expandOp = collapseOp.getSrc().getDefiningOp<tensor::ExpandShapeOp>();
if (!expandOp) {
return failure();
}

const auto collapseReassoc = collapseOp.getReassociationIndices();
ArrayRef<int64_t> interShape = expandOp.getType().getShape();
ArrayRef<int64_t> outShape = collapseOp.getType().getShape();
SmallVector<int64_t> interToOutMap(expandOp.getType().getRank());
llvm::SmallDenseSet<int64_t> toDrop;
for (const auto &[outDim, indicies] : llvm::enumerate(collapseReassoc)) {
for (auto [innerIdx, inDim] : llvm::enumerate(indicies)) {
// Can't drop this dim if it isnt statically 1 or if it isn't being
// combined with any other dimensions.
if (indicies.size() == 1 || interShape[inDim] != 1) {
continue;
}

// If we are collapsing multiple unit dims together, at least 1 must be
// kept (prefer the first).
if (outShape[outDim] == 1 && innerIdx != 0) {
continue;
}
toDrop.insert(inDim);
}
}

const auto expandReassoc = expandOp.getReassociationIndices();
for (const auto &[inDim, indicies] : llvm::enumerate(expandReassoc)) {
// Can't drop unit dim if it isn't from an expansion.
if (indicies.size() == 1) {
toDrop.erase(indicies[0]);
}
}

if (toDrop.empty()) {
return rewriter.notifyMatchFailure(collapseOp,
"Didn't find any unit dims to drop");
}

SmallVector<int64_t> newInterShape;
newInterShape.reserve(interShape.size() - toDrop.size());
for (auto [idx, length] : llvm::enumerate(interShape)) {
if (!toDrop.contains(idx)) {
newInterShape.push_back(length);
}
}

/// Returns true if new `ReassociationIndices` were appended to `reassoc`.
auto appendDroppedReassocation =
[&toDrop](SmallVector<ReassociationIndices, 4> &reassoc, int64_t start,
int64_t count, int64_t origStart) {
reassoc.emplace_back();
auto &indicies = reassoc.back();
indicies.reserve(count);
int64_t dim = start;
for (int64_t idx : llvm::seq<int64_t>(origStart, origStart + count)) {
if (!toDrop.contains(idx)) {
indicies.push_back(dim++);
}
}

// All indicies have been dropped.
if (indicies.empty()) {
reassoc.pop_back();
return false;
}
return true;
};

auto dropOutputOfr = [&toDrop](const SmallVector<OpFoldResult> &sizes) {
return llvm::map_to_vector(
llvm::make_filter_range(
llvm::enumerate(sizes),
[&toDrop](auto pair) { return !toDrop.contains(pair.index()); }),
[](auto pair) -> OpFoldResult { return pair.value(); });
};

auto isIdentityReassociation = [](ArrayRef<ReassociationIndices> reassoc) {
return llvm::all_of(reassoc,
[](auto &indices) { return indices.size() == 1; });
};

SmallVector<ReassociationIndices, 4> newCollapseReassoc;
int64_t collapsedDim = 0;
for (auto dim : llvm::seq<int64_t>(0, outShape.size())) {
bool changed = appendDroppedReassocation(newCollapseReassoc, collapsedDim,
collapseReassoc[dim].size(),
collapseReassoc[dim].front());
if (changed) {
collapsedDim += newCollapseReassoc.back().size();
}
}

SmallVector<ReassociationIndices, 4> newExpandReassoc;
ArrayRef<int64_t> srcShape = expandOp.getSrcType().getShape();
int64_t expandedDim = 0;
for (auto dim : llvm::seq<int64_t>(0, srcShape.size())) {
bool changed = appendDroppedReassocation(newExpandReassoc, expandedDim,
expandReassoc[dim].size(),
expandReassoc[dim].front());
if (changed) {
expandedDim += newExpandReassoc.back().size();
}
}

auto outputSizes = getMixedValues(expandOp.getStaticOutputShape(),
expandOp.getOutputShape(), rewriter);
Value newExpanded = expandOp.getSrc();
if (!isIdentityReassociation(newExpandReassoc)) {
newExpanded = rewriter.create<tensor::ExpandShapeOp>(
expandOp.getLoc(),
RankedTensorType::get(newInterShape,
expandOp.getType().getElementType()),
expandOp.getSrc(), newExpandReassoc, dropOutputOfr(outputSizes));
}

Value newCollapsed = newExpanded;
if (!isIdentityReassociation(newCollapseReassoc)) {
newCollapsed = rewriter.create<tensor::CollapseShapeOp>(
collapseOp.getLoc(), collapseOp.getType(), newExpanded,
newCollapseReassoc);
}
rewriter.replaceOp(collapseOp, newCollapsed);
return success();
}
};

} // namespace

//===----------------------------------------------------------------------===//
// Pass helpers
//===----------------------------------------------------------------------===//
Expand Down Expand Up @@ -155,6 +296,7 @@ void FoldUnitExtentDimsPass::runOnOperation() {
};
linalg::populateFoldUnitExtentDimsPatterns(foldUnitDimsPatterns, options);
linalg::populateMoveInitOperandsToInputPattern(foldUnitDimsPatterns);
foldUnitDimsPatterns.insert<DropUnitDimsFromCollapseOfExpand>(context);
if (failed(applyPatternsAndFoldGreedily(moduleOp,
std::move(foldUnitDimsPatterns)))) {
return signalPassFailure();
Expand Down
143 changes: 114 additions & 29 deletions compiler/src/iree/compiler/DispatchCreation/test/fold_unit_dims.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,12 +15,13 @@ util.func public @no_fold_unit_dims_in_dispatches(%arg0 : tensor<1x1x10xf32>) ->
}
util.return %1 : tensor<1x1x10xf32>
}
// CHECK: util.func public @no_fold_unit_dims_in_dispatches(%[[ARG0:.+]]: tensor<1x1x10xf32>)
// CHECK: %[[DISPATCH:.+]] = flow.dispatch.region
// CHECK: %[[GENERIC:.+]] = linalg.generic
// CHECK-SAME: ins(%[[ARG0]] : tensor<1x1x10xf32>)
// CHECK: flow.return %[[GENERIC]]
// CHECK: util.return %[[DISPATCH]]
// CHECK-LABEL: util.func public @no_fold_unit_dims_in_dispatches
// CHECK-SAME: (%[[ARG0:.+]]: tensor<1x1x10xf32>)
// CHECK: %[[DISPATCH:.+]] = flow.dispatch.region
// CHECK: %[[GENERIC:.+]] = linalg.generic
// CHECK-SAME: ins(%[[ARG0]] : tensor<1x1x10xf32>)
// CHECK: flow.return %[[GENERIC]]
// CHECK: util.return %[[DISPATCH]]


// -----
Expand All @@ -46,15 +47,15 @@ module @fold_unit_dims {
}
}

// CHECK: module @fold_unit_dims
// CHECK: util.global private mutable @[[GLOBAL:.+]] {inlining_policy = #util.inline.never} = #util.uninitialized : tensor<32x64xf32>
// CHECK: util.global private mutable @[[UNIT_GLOBAL:.+]] = #util.uninitialized : tensor<f32>
// CHECK: util.func public @fold_global_unit_dims
// CHECK: %[[LOAD0:.+]] = util.global.load @[[GLOBAL]] : tensor<32x64xf32>
// CHECK: %[[LOAD1:.+]] = util.global.load @[[UNIT_GLOBAL]] : tensor<f32>
// CHECK: %[[GENERIC:.+]] = linalg.generic
// CHECK-SAME: ins(%[[LOAD0]], %[[LOAD1]]
// CHECK: util.global.store %[[GENERIC]], @[[GLOBAL]] : tensor<32x64xf32>
// CHECK-LABEL: module @fold_unit_dims
// CHECK: util.global private mutable @[[GLOBAL:.+]] {inlining_policy = #util.inline.never} = #util.uninitialized : tensor<32x64xf32>
// CHECK: util.global private mutable @[[UNIT_GLOBAL:.+]] = #util.uninitialized : tensor<f32>
// CHECK: util.func public @fold_global_unit_dims
// CHECK: %[[LOAD0:.+]] = util.global.load @[[GLOBAL]] : tensor<32x64xf32>
// CHECK: %[[LOAD1:.+]] = util.global.load @[[UNIT_GLOBAL]] : tensor<f32>
// CHECK: %[[GENERIC:.+]] = linalg.generic
// CHECK-SAME: ins(%[[LOAD0]], %[[LOAD1]]
// CHECK: util.global.store %[[GENERIC]], @[[GLOBAL]] : tensor<32x64xf32>
// CHECK: util.return %[[GENERIC]]

// -----
Expand All @@ -68,12 +69,12 @@ module @no_fold_immutable {
}
}

// CHECK: module @no_fold_immutable
// CHECK: util.global private @[[GLOBAL:.+]] : tensor<1x32x1x1x64xf32>
// CHECK: util.func public @no_fold_global_unit_dims
// CHECK: %[[LOAD:.+]] = util.global.load @[[GLOBAL]] : tensor<1x32x1x1x64xf32>
// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[LOAD]]
// CHECK: util.return %[[COLLAPSE]]
// CHECK-LABEL: module @no_fold_immutable
// CHECK: util.global private @[[GLOBAL:.+]] : tensor<1x32x1x1x64xf32>
// CHECK: util.func public @no_fold_global_unit_dims
// CHECK: %[[LOAD:.+]] = util.global.load @[[GLOBAL]] : tensor<1x32x1x1x64xf32>
// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[LOAD]]
// CHECK: util.return %[[COLLAPSE]]

// -----

Expand All @@ -86,11 +87,11 @@ module @no_fold_public {
}
}

// CHECK: module @no_fold_public
// CHECK: util.global public mutable @[[GLOBAL:.+]] : tensor<1x32x1x1x64xf32>
// CHECK: util.func public @no_fold_global_unit_dims
// CHECK: %[[LOAD:.+]] = util.global.load @[[GLOBAL]] : tensor<1x32x1x1x64xf32>
// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[LOAD]]
// CHECK-LABEL: module @no_fold_public
// CHECK: util.global public mutable @[[GLOBAL:.+]] : tensor<1x32x1x1x64xf32>
// CHECK: util.func public @no_fold_global_unit_dims
// CHECK: %[[LOAD:.+]] = util.global.load @[[GLOBAL]] : tensor<1x32x1x1x64xf32>
// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[LOAD]]

// -----

Expand All @@ -102,7 +103,91 @@ module @fold_stream_parameter {
}
}

// CHECK: module @fold_stream_parameter
// CHECK: util.global private mutable @[[GLOBAL:.+]] = #stream.parameter.named<"module"::"global"> : tensor<10xf32>
// CHECK: util.func public @fold_stream_parameter
// CHECK-LABEL: module @fold_stream_parameter
// CHECK: util.global private mutable @[[GLOBAL:.+]] = #stream.parameter.named<"module"::"global"> : tensor<10xf32>
// CHECK: util.func public @fold_stream_parameter
// CHECK: %[[LOAD:.+]] = util.global.load @[[GLOBAL]] : tensor<10xf32>

// -----

util.func @collapse_of_expand_0(%arg0: tensor<?x128xf16>, %arg1: index) -> tensor<4x?x128xf16> {
%expanded = tensor.expand_shape %arg0 [[0, 1, 2], [3, 4]] output_shape [4, %arg1, 1, 1, 128] : tensor<?x128xf16> into tensor<4x?x1x1x128xf16>
%collapsed = tensor.collapse_shape %expanded [[0], [1, 2, 3], [4]] : tensor<4x?x1x1x128xf16> into tensor<4x?x128xf16>
util.return %collapsed : tensor<4x?x128xf16>
}

// CHECK-LABEL: util.func public @collapse_of_expand_0
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x128xf16>, %[[ARG1:.+]]: index
// CHECK: %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]
// CHECK-SAME: tensor<?x128xf16> into tensor<4x?x128xf16>
// CHECK: util.return %[[EXPAND]] : tensor<4x?x128xf16>

// -----

util.func @collapse_of_expand_1(%arg0: tensor<?x128xf16>, %arg1: index) -> tensor<4x?x64xf16> {
%expanded = tensor.expand_shape %arg0 [[0, 1, 2], [3, 4]] output_shape [4, %arg1, 1, 2, 64] : tensor<?x128xf16> into tensor<4x?x1x2x64xf16>
%collapsed = tensor.collapse_shape %expanded [[0], [1, 2, 3], [4]] : tensor<4x?x1x2x64xf16> into tensor<4x?x64xf16>
util.return %collapsed : tensor<4x?x64xf16>
}

// CHECK-LABEL: util.func public @collapse_of_expand_1
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x128xf16>, %[[ARG1:.+]]: index
// CHECK: %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]
// CHECK-SAME: tensor<?x128xf16> into tensor<4x?x2x64xf16>
// CHECK: %[[COLLAPSE:.+]] = tensor.collapse_shape %[[EXPAND]]
// CHECK-SAME: tensor<4x?x2x64xf16> into tensor<4x?x64xf16>
// CHECK: util.return %[[COLLAPSE]] : tensor<4x?x64xf16>

// -----

util.func @collapse_of_expand_2(%arg0: tensor<?x1xf16>, %arg1: index) -> tensor<4x?x1xf16> {
%expanded = tensor.expand_shape %arg0 [[0, 1, 2], [3, 4]] output_shape [4, %arg1, 1, 1, 1] : tensor<?x1xf16> into tensor<4x?x1x1x1xf16>
%collapsed = tensor.collapse_shape %expanded [[0], [1, 2, 3], [4]] : tensor<4x?x1x1x1xf16> into tensor<4x?x1xf16>
util.return %collapsed : tensor<4x?x1xf16>
}

// CHECK-LABEL: util.func public @collapse_of_expand_2
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x1xf16>, %[[ARG1:.+]]: index
// CHECK: %[[EXPAND:.+]] = tensor.expand_shape %[[ARG0]]
// CHECK-SAME: tensor<?x1xf16> into tensor<4x?x1xf16>
// CHECK: util.return %[[EXPAND]] : tensor<4x?x1xf16>

// -----

util.func @collapse_of_expand_3(%arg0: tensor<?x?xf16>, %arg1: index, %arg2: index) -> tensor<?x?xf16> {
%expanded = tensor.expand_shape %arg0 [[0, 1], [2, 3]] output_shape [%arg1, 1, 1, %arg2] : tensor<?x?xf16> into tensor<?x1x1x?xf16>
%collapsed = tensor.collapse_shape %expanded [[0], [1, 2, 3]] : tensor<?x1x1x?xf16> into tensor<?x?xf16>
util.return %collapsed : tensor<?x?xf16>
}

// CHECK-LABEL: util.func public @collapse_of_expand_3
// CHECK-SAME: %[[ARG0:.+]]: tensor<?x?xf16>
// CHECK: util.return %[[ARG0]] : tensor<?x?xf16>

// -----

util.func @collapse_of_expand_4(%arg0: tensor<1x1xf16>, %arg1: index, %arg2: index) -> tensor<1xf16> {
%expanded = tensor.expand_shape %arg0 [[0, 1, 2], [3]] output_shape [%arg1, 1, 1, %arg2] : tensor<1x1xf16> into tensor<1x1x1x1xf16>
%collapsed = tensor.collapse_shape %expanded [[0, 1, 2, 3]] : tensor<1x1x1x1xf16> into tensor<1xf16>
util.return %collapsed : tensor<1xf16>
}

// CHECK-LABEL: util.func public @collapse_of_expand_4
// CHECK-SAME: %[[ARG0:.+]]: tensor<1x1xf16>
// CHECK: %[[COLLAPSED:.+]] = tensor.collapse_shape %[[ARG0]]
// CHECK-SAME: tensor<1x1xf16> into tensor<1xf16>
// CHECK: util.return %[[COLLAPSED]] : tensor<1xf16>

// -----

util.func @collapse_of_expand_5(%arg0: tensor<1x?x4x32xf16>, %arg1: index) -> tensor<?x4x32xf16> {
%expanded = tensor.expand_shape %arg0 [[0], [1], [2], [3, 4]] output_shape [1, %arg1, 4, 1, 32] : tensor<1x?x4x32xf16> into tensor<1x?x4x1x32xf16>
%collapsed = tensor.collapse_shape %expanded [[0, 1], [2, 3], [4]] : tensor<1x?x4x1x32xf16> into tensor<?x4x32xf16>
util.return %collapsed : tensor<?x4x32xf16>
}

// CHECK-LABEL: util.func public @collapse_of_expand_5
// CHECK-SAME: %[[ARG0:.+]]: tensor<1x?x4x32xf16>
// CHECK: %[[COLLAPSED:.+]] = tensor.collapse_shape %[[ARG0]]
// CHECK-SAME: tensor<1x?x4x32xf16> into tensor<?x4x32xf16>
// CHECK: util.return %[[COLLAPSED]] : tensor<?x4x32xf16>
Loading