Skip to content

[mlir][MemRef][~NFC] Move getStridesAndOffset() onto layouts #138011

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.

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
Show file tree
Hide file tree
Changes from all commits
Commits
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
6 changes: 6 additions & 0 deletions mlir/include/mlir/IR/BuiltinAttributeInterfaces.h
Original file line number Diff line number Diff line change
Expand Up @@ -270,6 +270,12 @@ LogicalResult
verifyAffineMapAsLayout(AffineMap m, ArrayRef<int64_t> shape,
function_ref<InFlightDiagnostic()> emitError);

// Return the strides and offsets that can be inferred from the given affine
// layout map given the map and a memref shape.
LogicalResult getAffineMapStridesAndOffset(AffineMap map,
ArrayRef<int64_t> shape,
SmallVectorImpl<int64_t> &strides,
int64_t &offset);
} // namespace detail

} // namespace mlir
Expand Down
17 changes: 17 additions & 0 deletions mlir/include/mlir/IR/BuiltinAttributeInterfaces.td
Original file line number Diff line number Diff line change
Expand Up @@ -509,6 +509,23 @@ def MemRefLayoutAttrInterface : AttrInterface<"MemRefLayoutAttrInterface"> {
return ::mlir::detail::verifyAffineMapAsLayout($_attr.getAffineMap(),
shape, emitError);
}]
>,

InterfaceMethod<
[{Return the strides (using ShapedType::kDynamic for the dynamic case)
that this layout corresponds to into `strides` and `offset` if such exist
and can be determined from a combination of the layout and the given
`shape`. If these strides cannot be inferred, return failure().
The values of `strides` and `offset` are undefined on failure.}],
"::llvm::LogicalResult", "getStridesAndOffset",
(ins "::llvm::ArrayRef<int64_t>":$shape,
"::llvm::SmallVectorImpl<int64_t>&":$strides,
"int64_t&":$offset),
[{}],
[{
return ::mlir::detail::getAffineMapStridesAndOffset(
$_attr.getAffineMap(), shape, strides, offset);
}]
>
];
}
Expand Down
2 changes: 1 addition & 1 deletion mlir/include/mlir/IR/BuiltinAttributes.td
Original file line number Diff line number Diff line change
Expand Up @@ -1003,7 +1003,7 @@ def Builtin_SparseElementsAttr : Builtin_Attr<

def StridedLayoutAttr : Builtin_Attr<"StridedLayout", "strided_layout",
[DeclareAttrInterfaceMethods<MemRefLayoutAttrInterface,
["verifyLayout"]>]> {
["verifyLayout", "getStridesAndOffset"]>]> {
let summary = "An Attribute representing a strided layout of a shaped type";
let description = [{
Syntax:
Expand Down
135 changes: 135 additions & 0 deletions mlir/lib/IR/BuiltinAttributeInterfaces.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -83,3 +83,138 @@ LogicalResult mlir::detail::verifyAffineMapAsLayout(

return success();
}

// Fallback cases for terminal dim/sym/cst that are not part of a binary op (
// i.e. single term). Accumulate the AffineExpr into the existing one.
static void extractStridesFromTerm(AffineExpr e,
AffineExpr multiplicativeFactor,
MutableArrayRef<AffineExpr> strides,
AffineExpr &offset) {
if (auto dim = dyn_cast<AffineDimExpr>(e))
strides[dim.getPosition()] =
strides[dim.getPosition()] + multiplicativeFactor;
else
offset = offset + e * multiplicativeFactor;
}

/// Takes a single AffineExpr `e` and populates the `strides` array with the
/// strides expressions for each dim position.
/// The convention is that the strides for dimensions d0, .. dn appear in
/// order to make indexing intuitive into the result.
static LogicalResult extractStrides(AffineExpr e,
AffineExpr multiplicativeFactor,
MutableArrayRef<AffineExpr> strides,
AffineExpr &offset) {
auto bin = dyn_cast<AffineBinaryOpExpr>(e);
if (!bin) {
extractStridesFromTerm(e, multiplicativeFactor, strides, offset);
return success();
}

if (bin.getKind() == AffineExprKind::CeilDiv ||
bin.getKind() == AffineExprKind::FloorDiv ||
bin.getKind() == AffineExprKind::Mod)
return failure();

if (bin.getKind() == AffineExprKind::Mul) {
auto dim = dyn_cast<AffineDimExpr>(bin.getLHS());
if (dim) {
strides[dim.getPosition()] =
strides[dim.getPosition()] + bin.getRHS() * multiplicativeFactor;
return success();
}
// LHS and RHS may both contain complex expressions of dims. Try one path
// and if it fails try the other. This is guaranteed to succeed because
// only one path may have a `dim`, otherwise this is not an AffineExpr in
// the first place.
if (bin.getLHS().isSymbolicOrConstant())
return extractStrides(bin.getRHS(), multiplicativeFactor * bin.getLHS(),
strides, offset);
return extractStrides(bin.getLHS(), multiplicativeFactor * bin.getRHS(),
strides, offset);
}

if (bin.getKind() == AffineExprKind::Add) {
auto res1 =
extractStrides(bin.getLHS(), multiplicativeFactor, strides, offset);
auto res2 =
extractStrides(bin.getRHS(), multiplicativeFactor, strides, offset);
return success(succeeded(res1) && succeeded(res2));
}

llvm_unreachable("unexpected binary operation");
}

/// A stride specification is a list of integer values that are either static
/// or dynamic (encoded with ShapedType::kDynamic). Strides encode
/// the distance in the number of elements between successive entries along a
/// particular dimension.
///
/// For example, `memref<42x16xf32, (64 * d0 + d1)>` specifies a view into a
/// non-contiguous memory region of `42` by `16` `f32` elements in which the
/// distance between two consecutive elements along the outer dimension is `1`
/// and the distance between two consecutive elements along the inner dimension
/// is `64`.
///
/// The convention is that the strides for dimensions d0, .. dn appear in
/// order to make indexing intuitive into the result.
static LogicalResult getStridesAndOffset(AffineMap m, ArrayRef<int64_t> shape,
SmallVectorImpl<AffineExpr> &strides,
AffineExpr &offset) {
if (m.getNumResults() != 1 && !m.isIdentity())
return failure();

auto zero = getAffineConstantExpr(0, m.getContext());
auto one = getAffineConstantExpr(1, m.getContext());
offset = zero;
strides.assign(shape.size(), zero);

// Canonical case for empty map.
if (m.isIdentity()) {
// 0-D corner case, offset is already 0.
if (shape.empty())
return success();
auto stridedExpr = makeCanonicalStridedLayoutExpr(shape, m.getContext());
if (succeeded(extractStrides(stridedExpr, one, strides, offset)))
return success();
assert(false && "unexpected failure: extract strides in canonical layout");
}

// Non-canonical case requires more work.
auto stridedExpr =
simplifyAffineExpr(m.getResult(0), m.getNumDims(), m.getNumSymbols());
if (failed(extractStrides(stridedExpr, one, strides, offset))) {
offset = AffineExpr();
strides.clear();
return failure();
}

// Simplify results to allow folding to constants and simple checks.
unsigned numDims = m.getNumDims();
unsigned numSymbols = m.getNumSymbols();
offset = simplifyAffineExpr(offset, numDims, numSymbols);
for (auto &stride : strides)
stride = simplifyAffineExpr(stride, numDims, numSymbols);

return success();
}

LogicalResult mlir::detail::getAffineMapStridesAndOffset(
AffineMap map, ArrayRef<int64_t> shape, SmallVectorImpl<int64_t> &strides,
int64_t &offset) {
AffineExpr offsetExpr;
SmallVector<AffineExpr, 4> strideExprs;
if (failed(::getStridesAndOffset(map, shape, strideExprs, offsetExpr)))
return failure();
if (auto cst = llvm::dyn_cast<AffineConstantExpr>(offsetExpr))
offset = cst.getValue();
else
offset = ShapedType::kDynamic;
for (auto e : strideExprs) {
if (auto c = llvm::dyn_cast<AffineConstantExpr>(e))
strides.push_back(c.getValue());
else
strides.push_back(ShapedType::kDynamic);
}
return success();
}
9 changes: 9 additions & 0 deletions mlir/lib/IR/BuiltinAttributes.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -258,6 +258,15 @@ LogicalResult StridedLayoutAttr::verifyLayout(
return success();
}

LogicalResult
StridedLayoutAttr::getStridesAndOffset(ArrayRef<int64_t>,
SmallVectorImpl<int64_t> &strides,
int64_t &offset) const {
llvm::append_range(strides, getStrides());
offset = getOffset();
return success();
}

//===----------------------------------------------------------------------===//
// StringAttr
//===----------------------------------------------------------------------===//
Expand Down
143 changes: 1 addition & 142 deletions mlir/lib/IR/BuiltinTypes.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -715,150 +715,9 @@ MemRefType MemRefType::canonicalizeStridedLayout() {
return MemRefType::Builder(*this).setLayout({});
}

// Fallback cases for terminal dim/sym/cst that are not part of a binary op (
// i.e. single term). Accumulate the AffineExpr into the existing one.
static void extractStridesFromTerm(AffineExpr e,
AffineExpr multiplicativeFactor,
MutableArrayRef<AffineExpr> strides,
AffineExpr &offset) {
if (auto dim = dyn_cast<AffineDimExpr>(e))
strides[dim.getPosition()] =
strides[dim.getPosition()] + multiplicativeFactor;
else
offset = offset + e * multiplicativeFactor;
}

/// Takes a single AffineExpr `e` and populates the `strides` array with the
/// strides expressions for each dim position.
/// The convention is that the strides for dimensions d0, .. dn appear in
/// order to make indexing intuitive into the result.
static LogicalResult extractStrides(AffineExpr e,
AffineExpr multiplicativeFactor,
MutableArrayRef<AffineExpr> strides,
AffineExpr &offset) {
auto bin = dyn_cast<AffineBinaryOpExpr>(e);
if (!bin) {
extractStridesFromTerm(e, multiplicativeFactor, strides, offset);
return success();
}

if (bin.getKind() == AffineExprKind::CeilDiv ||
bin.getKind() == AffineExprKind::FloorDiv ||
bin.getKind() == AffineExprKind::Mod)
return failure();

if (bin.getKind() == AffineExprKind::Mul) {
auto dim = dyn_cast<AffineDimExpr>(bin.getLHS());
if (dim) {
strides[dim.getPosition()] =
strides[dim.getPosition()] + bin.getRHS() * multiplicativeFactor;
return success();
}
// LHS and RHS may both contain complex expressions of dims. Try one path
// and if it fails try the other. This is guaranteed to succeed because
// only one path may have a `dim`, otherwise this is not an AffineExpr in
// the first place.
if (bin.getLHS().isSymbolicOrConstant())
return extractStrides(bin.getRHS(), multiplicativeFactor * bin.getLHS(),
strides, offset);
return extractStrides(bin.getLHS(), multiplicativeFactor * bin.getRHS(),
strides, offset);
}

if (bin.getKind() == AffineExprKind::Add) {
auto res1 =
extractStrides(bin.getLHS(), multiplicativeFactor, strides, offset);
auto res2 =
extractStrides(bin.getRHS(), multiplicativeFactor, strides, offset);
return success(succeeded(res1) && succeeded(res2));
}

llvm_unreachable("unexpected binary operation");
}

/// A stride specification is a list of integer values that are either static
/// or dynamic (encoded with ShapedType::kDynamic). Strides encode
/// the distance in the number of elements between successive entries along a
/// particular dimension.
///
/// For example, `memref<42x16xf32, (64 * d0 + d1)>` specifies a view into a
/// non-contiguous memory region of `42` by `16` `f32` elements in which the
/// distance between two consecutive elements along the outer dimension is `1`
/// and the distance between two consecutive elements along the inner dimension
/// is `64`.
///
/// The convention is that the strides for dimensions d0, .. dn appear in
/// order to make indexing intuitive into the result.
static LogicalResult getStridesAndOffset(MemRefType t,
SmallVectorImpl<AffineExpr> &strides,
AffineExpr &offset) {
AffineMap m = t.getLayout().getAffineMap();

if (m.getNumResults() != 1 && !m.isIdentity())
return failure();

auto zero = getAffineConstantExpr(0, t.getContext());
auto one = getAffineConstantExpr(1, t.getContext());
offset = zero;
strides.assign(t.getRank(), zero);

// Canonical case for empty map.
if (m.isIdentity()) {
// 0-D corner case, offset is already 0.
if (t.getRank() == 0)
return success();
auto stridedExpr =
makeCanonicalStridedLayoutExpr(t.getShape(), t.getContext());
if (succeeded(extractStrides(stridedExpr, one, strides, offset)))
return success();
assert(false && "unexpected failure: extract strides in canonical layout");
}

// Non-canonical case requires more work.
auto stridedExpr =
simplifyAffineExpr(m.getResult(0), m.getNumDims(), m.getNumSymbols());
if (failed(extractStrides(stridedExpr, one, strides, offset))) {
offset = AffineExpr();
strides.clear();
return failure();
}

// Simplify results to allow folding to constants and simple checks.
unsigned numDims = m.getNumDims();
unsigned numSymbols = m.getNumSymbols();
offset = simplifyAffineExpr(offset, numDims, numSymbols);
for (auto &stride : strides)
stride = simplifyAffineExpr(stride, numDims, numSymbols);

return success();
}

LogicalResult MemRefType::getStridesAndOffset(SmallVectorImpl<int64_t> &strides,
int64_t &offset) {
// Happy path: the type uses the strided layout directly.
if (auto strided = llvm::dyn_cast<StridedLayoutAttr>(getLayout())) {
llvm::append_range(strides, strided.getStrides());
offset = strided.getOffset();
return success();
}

// Otherwise, defer to the affine fallback as layouts are supposed to be
// convertible to affine maps.
AffineExpr offsetExpr;
SmallVector<AffineExpr, 4> strideExprs;
if (failed(::getStridesAndOffset(*this, strideExprs, offsetExpr)))
return failure();
if (auto cst = llvm::dyn_cast<AffineConstantExpr>(offsetExpr))
offset = cst.getValue();
else
offset = ShapedType::kDynamic;
for (auto e : strideExprs) {
if (auto c = llvm::dyn_cast<AffineConstantExpr>(e))
strides.push_back(c.getValue());
else
strides.push_back(ShapedType::kDynamic);
}
return success();
return getLayout().getStridesAndOffset(getShape(), strides, offset);
}

std::pair<SmallVector<int64_t>, int64_t> MemRefType::getStridesAndOffset() {
Expand Down