maxjhandsome
/
llvm-project
forked from OSchip/llvm-project
1
0
Fork 0
Code Issues Pull Requests Packages Releases Wiki Activity

[MLIR][Presburger] Support computing volumes via hyperrectangular overapproximation 

Add support for computing an overapproximation of the number of integer points
in a polyhedron. The returned result is actually the number of integer points
one gets by computing the "rational shadow" obtained by projecting out the
local IDs, finding the minimal axis-parallel hyperrectangular approximation
of the shadow, and returning the number of integer points in that. This does
not currently support symbols.

Reviewed By: Groverkss

Differential Revision: https://reviews.llvm.org/D119228
This commit is contained in:
Arjun P 2022-02-08 21:06:24 +05:30
parent ae9414d562
commit 1096fcff7d
7 changed files with 231 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
mlir/include/mlir/Analysis/Presburger/IntegerPolyhedron.h
View File

@ -277,6 +277,11 @@ public:
/// otherwise.
Optional<SmallVector<int64_t, 8>> findIntegerSample() const;
/// Compute an overapproximation of the number of integer points in the
/// polyhedron. Symbol ids are currently not supported. If the computed
/// overapproximation is infinite, an empty optional is returned.
Optional<uint64_t> computeVolume() const;
/// Returns true if the given point satisfies the constraints, or false
/// otherwise.
///

9
mlir/include/mlir/Analysis/Presburger/PresburgerSet.h
View File

@ -99,6 +99,15 @@ public:
/// any of the Polys in the union are unbounded.
bool findIntegerSample(SmallVectorImpl<int64_t> &sample);
/// Compute an overapproximation of the number of integer points in the
/// polyhedron. Symbol ids are currently not supported. If the computed
/// overapproximation is infinite, an empty optional is returned.
///
/// This currently just sums up the overapproximations of the volumes of the
/// disjuncts, so the approximation might be far from the true volume in the
/// case when there is a lot of overlap between disjuncts.
Optional<uint64_t> computeVolume() const;
/// Simplifies the representation of a PresburgerSet.
///
/// In particular, removes all Polys which are subsets of other Polys in the

54
mlir/lib/Analysis/Presburger/IntegerPolyhedron.cpp
View File

@ -1065,6 +1065,60 @@ void IntegerPolyhedron::removeRedundantConstraints() {
equalities.resizeVertically(pos);
}
Optional<uint64_t> IntegerPolyhedron::computeVolume() const {
assert(getNumSymbolIds() == 0 && "Symbols are not yet supported!");
Simplex simplex(*this);
// If the polytope is rationally empty, there are certainly no integer
// points.
if (simplex.isEmpty())
return 0;
// Just find the maximum and minimum integer value of each non-local id
// separately, thus finding the number of integer values each such id can
// take. Multiplying these together gives a valid overapproximation of the
// number of integer points in the polyhedron. The result this gives is
// equivalent to projecting (rationally) the polyhedron onto its non-local ids
// and returning the number of integer points in a minimal axis-parallel
// hyperrectangular overapproximation of that.
//
// We also handle the special case where one dimension is unbounded and
// another dimension can take no integer values. In this case, the volume is
// zero.
//
// If there is no such empty dimension, if any dimension is unbounded we
// just return the result as unbounded.
uint64_t count = 1;
SmallVector<int64_t, 8> dim(getNumIds() + 1);
bool hasUnboundedId = false;
for (unsigned i = 0, e = getNumDimAndSymbolIds(); i < e; ++i) {
dim[i] = 1;
Optional<int64_t> min, max;
std::tie(min, max) = simplex.computeIntegerBounds(dim);
dim[i] = 0;
// One of the dimensions is unbounded. Note this fact. We will return
// unbounded if none of the other dimensions makes the volume zero.
if (!min || !max) {
hasUnboundedId = true;
continue;
}
// In this case there are no valid integer points and the volume is
// definitely zero.
if (*min > *max)
return 0;
count *= (*max - *min + 1);
}
if (count == 0)
return 0;
if (hasUnboundedId)
return {};
return count;
}
void IntegerPolyhedron::eliminateRedundantLocalId(unsigned posA,
unsigned posB) {
assert(posA < getNumLocalIds() && "Invalid local id position");

14
mlir/lib/Analysis/Presburger/PresburgerSet.cpp
View File

@ -385,6 +385,20 @@ bool PresburgerSet::findIntegerSample(SmallVectorImpl<int64_t> &sample) {
return false;
}
Optional<uint64_t> PresburgerSet::computeVolume() const {
assert(getNumSymbolIds() == 0 && "Symbols are not yet supported!");
// The sum of the volumes of the disjuncts is a valid overapproximation of the
// volume of their union, even if they overlap.
uint64_t result = 0;
for (const IntegerPolyhedron &poly : integerPolyhedrons) {
Optional<uint64_t> volume = poly.computeVolume();
if (!volume)
return {};
result += *volume;
}
return result;
}
PresburgerSet PresburgerSet::coalesce() const {
PresburgerSet newSet =
PresburgerSet::getEmptySet(getNumDimIds(), getNumSymbolIds());

62
mlir/unittests/Analysis/Presburger/IntegerPolyhedronTest.cpp
View File

@ -1118,4 +1118,66 @@ TEST(IntegerPolyhedronTest, getRationalLexMin) {
expectNoRationalLexMin(parsePoly("(x) : (2*x >= 0, -x - 1 >= 0)", &context));
}
static void
expectComputedVolumeIsValidOverapprox(const IntegerPolyhedron &poly,
Optional<uint64_t> trueVolume,
Optional<uint64_t> resultBound) {
expectComputedVolumeIsValidOverapprox(poly.computeVolume(), trueVolume,
resultBound);
}
TEST(IntegerPolyhedronTest, computeVolume) {
MLIRContext context;
// 0 <= x <= 3 + 1/3, -5.5 <= y <= 2 + 3/5, 3 <= z <= 1.75.
// i.e. 0 <= x <= 3, -5 <= y <= 2, 3 <= z <= 3 + 1/4.
// So volume is 4 * 8 * 1 = 32.
expectComputedVolumeIsValidOverapprox(
parsePoly("(x, y, z) : (x >= 0, -3*x + 10 >= 0, 2*y + 11 >= 0,"
"-5*y + 13 >= 0, z - 3 >= 0, -4*z + 13 >= 0)",
&context),
/*trueVolume=*/32ull, /*resultBound=*/32ull);
// Same as above but y has bounds 2 + 1/5 <= y <= 2 + 3/5. So the volume is
// zero.
expectComputedVolumeIsValidOverapprox(
parsePoly("(x, y, z) : (x >= 0, -3*x + 10 >= 0, 5*y - 11 >= 0,"
"-5*y + 13 >= 0, z - 3 >= 0, -4*z + 13 >= 0)",
&context),
/*trueVolume=*/0ull, /*resultBound=*/0ull);
// Now x is unbounded below but y still has no integer values.
expectComputedVolumeIsValidOverapprox(
parsePoly("(x, y, z) : (-3*x + 10 >= 0, 5*y - 11 >= 0,"
"-5*y + 13 >= 0, z - 3 >= 0, -4*z + 13 >= 0)",
&context),
/*trueVolume=*/0ull, /*resultBound=*/0ull);
// A diamond shape, 0 <= x + y <= 10, 0 <= x - y <= 10,
// with vertices at (0, 0), (5, 5), (5, 5), (10, 0).
// x and y can take 11 possible values so result computed is 11*11 = 121.
expectComputedVolumeIsValidOverapprox(
parsePoly("(x, y) : (x + y >= 0, -x - y + 10 >= 0, x - y >= 0,"
"-x + y + 10 >= 0)",
&context),
/*trueVolume=*/61ull, /*resultBound=*/121ull);
// Effectively the same diamond as above; constrain the variables to be even
// and double the constant terms of the constraints. The algorithm can't
// eliminate locals exactly, so the result is an overapproximation by
// computing that x and y can take 21 possible values so result is 21*21 =
// 441.
expectComputedVolumeIsValidOverapprox(
parsePoly("(x, y) : (x + y >= 0, -x - y + 20 >= 0, x - y >= 0,"
" -x + y + 20 >= 0, x - 2*(x floordiv 2) == 0,"
"y - 2*(y floordiv 2) == 0)",
&context),
/*trueVolume=*/61ull, /*resultBound=*/441ull);
// Unbounded polytope.
expectComputedVolumeIsValidOverapprox(
parsePoly("(x, y) : (2*x - y >= 0, y - 3*x >= 0)", &context),
/*trueVolume=*/{}, /*resultBound=*/{});
}
} // namespace mlir

60
mlir/unittests/Analysis/Presburger/PresburgerSetTest.cpp
View File

@ -696,4 +696,64 @@ TEST(SetTest, coalesceContainedEqComplex) {
expectCoalesce(1, set);
}
static void
expectComputedVolumeIsValidOverapprox(const PresburgerSet &set,
Optional<uint64_t> trueVolume,
Optional<uint64_t> resultBound) {
expectComputedVolumeIsValidOverapprox(set.computeVolume(), trueVolume,
resultBound);
}
TEST(SetTest, computeVolume) {
MLIRContext context;
// Diamond with vertices at (0, 0), (5, 5), (5, 5), (10, 0).
PresburgerSet diamond(
parsePoly("(x, y) : (x + y >= 0, -x - y + 10 >= 0, x - y >= 0, -x + y + "
"10 >= 0)",
&context));
expectComputedVolumeIsValidOverapprox(diamond,
/*trueVolume=*/61ull,
/*resultBound=*/121ull);
// Diamond with vertices at (-5, 0), (0, -5), (0, 5), (5, 0).
PresburgerSet shiftedDiamond(parsePoly(
"(x, y) : (x + y + 5 >= 0, -x - y + 5 >= 0, x - y + 5 >= 0, -x + y + "
"5 >= 0)",
&context));
expectComputedVolumeIsValidOverapprox(shiftedDiamond,
/*trueVolume=*/61ull,
/*resultBound=*/121ull);
// Diamond with vertices at (-5, 0), (5, -10), (5, 10), (15, 0).
PresburgerSet biggerDiamond(parsePoly(
"(x, y) : (x + y + 5 >= 0, -x - y + 15 >= 0, x - y + 5 >= 0, -x + y + "
"15 >= 0)",
&context));
expectComputedVolumeIsValidOverapprox(biggerDiamond,
/*trueVolume=*/221ull,
/*resultBound=*/441ull);
// There is some overlap between diamond and shiftedDiamond.
expectComputedVolumeIsValidOverapprox(diamond.unionSet(shiftedDiamond),
/*trueVolume=*/104ull,
/*resultBound=*/242ull);
// biggerDiamond subsumes both the small ones.
expectComputedVolumeIsValidOverapprox(
diamond.unionSet(shiftedDiamond).unionSet(biggerDiamond),
/*trueVolume=*/221ull,
/*resultBound=*/683ull);
// Unbounded polytope.
PresburgerSet unbounded(
parsePoly("(x, y) : (2*x - y >= 0, y - 3*x >= 0)", &context));
expectComputedVolumeIsValidOverapprox(unbounded, /*trueVolume=*/{},
/*resultBound=*/{});
// Union of unbounded with bounded is unbounded.
expectComputedVolumeIsValidOverapprox(unbounded.unionSet(diamond),
/*trueVolume=*/{},
/*resultBound=*/{});
}
} // namespace mlir

27
mlir/unittests/Analysis/Presburger/Utils.h
View File

@ -14,6 +14,8 @@
#define MLIR_UNITTESTS_ANALYSIS_PRESBURGER_UTILS_H
#include "../../Dialect/Affine/Analysis/AffineStructuresParser.h"
#include "mlir/Analysis/Presburger/IntegerPolyhedron.h"
#include "mlir/Analysis/Presburger/PresburgerSet.h"
#include "mlir/IR/MLIRContext.h"
#include <gtest/gtest.h>
@ -27,6 +29,31 @@ static IntegerPolyhedron parsePoly(StringRef str, MLIRContext *context) {
EXPECT_TRUE(succeeded(poly));
return *poly;
}
/// lhs and rhs represent non-negative integers or positive infinity. The
/// infinity case corresponds to when the Optional is empty.
static bool infinityOrUInt64LE(Optional<uint64_t> lhs, Optional<uint64_t> rhs) {
// No constraint.
if (!rhs)
return true;
// Finite rhs provided so lhs has to be finite too.
if (!lhs)
return false;
return *lhs <= *rhs;
}
/// Expect that the computed volume is a valid overapproximation of
/// the true volume `trueVolume`, while also being at least as good an
/// approximation as `resultBound`.
static void
expectComputedVolumeIsValidOverapprox(Optional<uint64_t> computedVolume,
Optional<uint64_t> trueVolume,
Optional<uint64_t> resultBound) {
assert(infinityOrUInt64LE(trueVolume, resultBound) &&
"can't expect result to be less than the true volume");
EXPECT_TRUE(infinityOrUInt64LE(trueVolume, computedVolume));
EXPECT_TRUE(infinityOrUInt64LE(computedVolume, resultBound));
}
} // namespace mlir
#endif // MLIR_UNITTESTS_ANALYSIS_PRESBURGER_UTILS_H