[mlir][SCF][NFC] ValueBoundsConstraintSet: Simplify scf.for implementation
#87862
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…mentation (#86239) This commit simplifies the implementation of the `ValueBoundsOpInterface` for `scf.for` based on the newly added `ValueBoundsConstraintSet::compare` API and adds additional documentation. Previously, the interface implementation created a new constraint set just to check if the yielded value and iter_arg are equal. This was inefficient because constraints were added multiple times (to two different constraint sets) for ops that are inside the loop.
|
@llvm/pr-subscribers-mlir Author: Matthias Springer (matthias-springer) ChangesThis commit simplifies the implementation of the Previously, the interface implementation created a new constraint set just to check if the yielded value and iter_arg are equal. This was inefficient because constraints were added multiple times (to two different constraint sets) for ops that are inside the loop. Note: This is a re-upload of #86239. Full diff: https://github.com/llvm/llvm-project/pull/87862.diff 1 Files Affected:
diff --git a/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp b/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp
index 8e9d1021f93e4b..72c5aaa2306783 100644
--- a/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp
@@ -12,7 +12,6 @@
#include "mlir/Interfaces/ValueBoundsOpInterface.h"
using namespace mlir;
-using presburger::BoundType;
namespace mlir {
namespace scf {
@@ -21,7 +20,28 @@ namespace {
struct ForOpInterface
: public ValueBoundsOpInterface::ExternalModel<ForOpInterface, ForOp> {
- /// Populate bounds of values/dimensions for iter_args/OpResults.
+ /// Populate bounds of values/dimensions for iter_args/OpResults. If the
+ /// value/dimension size does not change in an iteration, we can deduce that
+ /// it the same as the initial value/dimension.
+ ///
+ /// Example 1:
+ /// %0 = scf.for ... iter_args(%arg0 = %t) -> tensor<?xf32> {
+ /// ...
+ /// %1 = tensor.insert %f into %arg0[...] : tensor<?xf32>
+ /// scf.yield %1 : tensor<?xf32>
+ /// }
+ /// --> bound(%0)[0] == bound(%t)[0]
+ /// --> bound(%arg0)[0] == bound(%t)[0]
+ ///
+ /// Example 2:
+ /// %0 = scf.for ... iter_args(%arg0 = %t) -> tensor<?xf32> {
+ /// %sz = tensor.dim %arg0 : tensor<?xf32>
+ /// %incr = arith.addi %sz, %c1 : index
+ /// %1 = tensor.empty(%incr) : tensor<?xf32>
+ /// scf.yield %1 : tensor<?xf32>
+ /// }
+ /// --> The yielded tensor dimension size changes with each iteration. Such
+ /// loops are not supported and no constraints are added.
static void populateIterArgBounds(scf::ForOp forOp, Value value,
std::optional<int64_t> dim,
ValueBoundsConstraintSet &cstr) {
@@ -33,59 +53,31 @@ struct ForOpInterface
iterArgIdx = llvm::cast<OpResult>(value).getResultNumber();
}
- // An EQ constraint can be added if the yielded value (dimension size)
- // equals the corresponding block argument (dimension size).
Value yieldedValue = cast<scf::YieldOp>(forOp.getBody()->getTerminator())
.getOperand(iterArgIdx);
Value iterArg = forOp.getRegionIterArg(iterArgIdx);
Value initArg = forOp.getInitArgs()[iterArgIdx];
- auto addEqBound = [&]() {
+ // Populate constraints for the yielded value.
+ cstr.populateConstraints(yieldedValue, dim);
+ // Populate constraints for the iter_arg. This is just to ensure that the
+ // iter_arg is mapped in the constraint set, which is a prerequisite for
+ // `compare`. It may lead to a recursive call to this function in case the
+ // iter_arg was not visited when the constraints for the yielded value were
+ // populated, but no additional work is done.
+ cstr.populateConstraints(iterArg, dim);
+
+ // An EQ constraint can be added if the yielded value (dimension size)
+ // equals the corresponding block argument (dimension size).
+ if (cstr.compare(yieldedValue, dim,
+ ValueBoundsConstraintSet::ComparisonOperator::EQ, iterArg,
+ dim)) {
if (dim.has_value()) {
cstr.bound(value)[*dim] == cstr.getExpr(initArg, dim);
} else {
cstr.bound(value) == initArg;
}
- };
-
- if (yieldedValue == iterArg) {
- addEqBound();
- return;
- }
-
- // Compute EQ bound for yielded value.
- AffineMap bound;
- ValueDimList boundOperands;
- LogicalResult status = ValueBoundsConstraintSet::computeBound(
- bound, boundOperands, BoundType::EQ, yieldedValue, dim,
- [&](Value v, std::optional<int64_t> d, ValueBoundsConstraintSet &cstr) {
- // Stop when reaching a block argument of the loop body.
- if (auto bbArg = llvm::dyn_cast<BlockArgument>(v))
- return bbArg.getOwner()->getParentOp() == forOp;
- // Stop when reaching a value that is defined outside of the loop. It
- // is impossible to reach an iter_arg from there.
- Operation *op = v.getDefiningOp();
- return forOp.getRegion().findAncestorOpInRegion(*op) == nullptr;
- });
- if (failed(status))
- return;
- if (bound.getNumResults() != 1)
- return;
-
- // Check if computed bound equals the corresponding iter_arg.
- Value singleValue = nullptr;
- std::optional<int64_t> singleDim;
- if (auto dimExpr = dyn_cast<AffineDimExpr>(bound.getResult(0))) {
- int64_t idx = dimExpr.getPosition();
- singleValue = boundOperands[idx].first;
- singleDim = boundOperands[idx].second;
- } else if (auto symExpr = dyn_cast<AffineSymbolExpr>(bound.getResult(0))) {
- int64_t idx = symExpr.getPosition() + bound.getNumDims();
- singleValue = boundOperands[idx].first;
- singleDim = boundOperands[idx].second;
}
- if (singleValue == iterArg && singleDim == dim)
- addEqBound();
}
void populateBoundsForIndexValue(Operation *op, Value value,
|
|
@llvm/pr-subscribers-mlir-scf Author: Matthias Springer (matthias-springer) ChangesThis commit simplifies the implementation of the Previously, the interface implementation created a new constraint set just to check if the yielded value and iter_arg are equal. This was inefficient because constraints were added multiple times (to two different constraint sets) for ops that are inside the loop. Note: This is a re-upload of #86239. Full diff: https://github.com/llvm/llvm-project/pull/87862.diff 1 Files Affected:
diff --git a/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp b/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp
index 8e9d1021f93e4b..72c5aaa2306783 100644
--- a/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp
+++ b/mlir/lib/Dialect/SCF/IR/ValueBoundsOpInterfaceImpl.cpp
@@ -12,7 +12,6 @@
#include "mlir/Interfaces/ValueBoundsOpInterface.h"
using namespace mlir;
-using presburger::BoundType;
namespace mlir {
namespace scf {
@@ -21,7 +20,28 @@ namespace {
struct ForOpInterface
: public ValueBoundsOpInterface::ExternalModel<ForOpInterface, ForOp> {
- /// Populate bounds of values/dimensions for iter_args/OpResults.
+ /// Populate bounds of values/dimensions for iter_args/OpResults. If the
+ /// value/dimension size does not change in an iteration, we can deduce that
+ /// it the same as the initial value/dimension.
+ ///
+ /// Example 1:
+ /// %0 = scf.for ... iter_args(%arg0 = %t) -> tensor<?xf32> {
+ /// ...
+ /// %1 = tensor.insert %f into %arg0[...] : tensor<?xf32>
+ /// scf.yield %1 : tensor<?xf32>
+ /// }
+ /// --> bound(%0)[0] == bound(%t)[0]
+ /// --> bound(%arg0)[0] == bound(%t)[0]
+ ///
+ /// Example 2:
+ /// %0 = scf.for ... iter_args(%arg0 = %t) -> tensor<?xf32> {
+ /// %sz = tensor.dim %arg0 : tensor<?xf32>
+ /// %incr = arith.addi %sz, %c1 : index
+ /// %1 = tensor.empty(%incr) : tensor<?xf32>
+ /// scf.yield %1 : tensor<?xf32>
+ /// }
+ /// --> The yielded tensor dimension size changes with each iteration. Such
+ /// loops are not supported and no constraints are added.
static void populateIterArgBounds(scf::ForOp forOp, Value value,
std::optional<int64_t> dim,
ValueBoundsConstraintSet &cstr) {
@@ -33,59 +53,31 @@ struct ForOpInterface
iterArgIdx = llvm::cast<OpResult>(value).getResultNumber();
}
- // An EQ constraint can be added if the yielded value (dimension size)
- // equals the corresponding block argument (dimension size).
Value yieldedValue = cast<scf::YieldOp>(forOp.getBody()->getTerminator())
.getOperand(iterArgIdx);
Value iterArg = forOp.getRegionIterArg(iterArgIdx);
Value initArg = forOp.getInitArgs()[iterArgIdx];
- auto addEqBound = [&]() {
+ // Populate constraints for the yielded value.
+ cstr.populateConstraints(yieldedValue, dim);
+ // Populate constraints for the iter_arg. This is just to ensure that the
+ // iter_arg is mapped in the constraint set, which is a prerequisite for
+ // `compare`. It may lead to a recursive call to this function in case the
+ // iter_arg was not visited when the constraints for the yielded value were
+ // populated, but no additional work is done.
+ cstr.populateConstraints(iterArg, dim);
+
+ // An EQ constraint can be added if the yielded value (dimension size)
+ // equals the corresponding block argument (dimension size).
+ if (cstr.compare(yieldedValue, dim,
+ ValueBoundsConstraintSet::ComparisonOperator::EQ, iterArg,
+ dim)) {
if (dim.has_value()) {
cstr.bound(value)[*dim] == cstr.getExpr(initArg, dim);
} else {
cstr.bound(value) == initArg;
}
- };
-
- if (yieldedValue == iterArg) {
- addEqBound();
- return;
- }
-
- // Compute EQ bound for yielded value.
- AffineMap bound;
- ValueDimList boundOperands;
- LogicalResult status = ValueBoundsConstraintSet::computeBound(
- bound, boundOperands, BoundType::EQ, yieldedValue, dim,
- [&](Value v, std::optional<int64_t> d, ValueBoundsConstraintSet &cstr) {
- // Stop when reaching a block argument of the loop body.
- if (auto bbArg = llvm::dyn_cast<BlockArgument>(v))
- return bbArg.getOwner()->getParentOp() == forOp;
- // Stop when reaching a value that is defined outside of the loop. It
- // is impossible to reach an iter_arg from there.
- Operation *op = v.getDefiningOp();
- return forOp.getRegion().findAncestorOpInRegion(*op) == nullptr;
- });
- if (failed(status))
- return;
- if (bound.getNumResults() != 1)
- return;
-
- // Check if computed bound equals the corresponding iter_arg.
- Value singleValue = nullptr;
- std::optional<int64_t> singleDim;
- if (auto dimExpr = dyn_cast<AffineDimExpr>(bound.getResult(0))) {
- int64_t idx = dimExpr.getPosition();
- singleValue = boundOperands[idx].first;
- singleDim = boundOperands[idx].second;
- } else if (auto symExpr = dyn_cast<AffineSymbolExpr>(bound.getResult(0))) {
- int64_t idx = symExpr.getPosition() + bound.getNumDims();
- singleValue = boundOperands[idx].first;
- singleDim = boundOperands[idx].second;
}
- if (singleValue == iterArg && singleDim == dim)
- addEqBound();
}
void populateBoundsForIndexValue(Operation *op, Value value,
|
matthias-springer
deleted the
users/matthias-springer/value_bounds_scf_for2
branch
yzhang93
added a commit
to iree-org/llvm-project
that referenced
this pull request
Apr 10, 2024
…r` implementation (llvm#87862)" This reverts commit 0ba3e96.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This commit simplifies the implementation of the
ValueBoundsOpInterfaceforscf.forbased on the newly addedValueBoundsConstraintSet::compareAPI and adds additional documentation.Previously, the interface implementation created a new constraint set just to check if the yielded value and iter_arg are equal. This was inefficient because constraints were added multiple times (to two different constraint sets) for ops that are inside the loop.
Note: This is a re-upload of #86239.