outputRanges: use isl::map::get_range_simple_fixed_box_hull

The computation is fairly low-level.  Let isl take care of it.
TensorReferenceGroup::promotion no longer needs to fiddle
with the output space because the space is preserved
by isl::map::get_range_simple_fixed_box_hull.

Author: Sven Verdoolaege

tc/core/polyhedral/memory_promotion.cc

@@ -31,71 +31,12 @@ namespace polyhedral {
 using detail::ScheduleTree;
 
 namespace {
-std::pair<isl::val, isl::aff> outputRange(
-    isl::basic_set wrappedAccess,
-    isl::constraint cstr) {
-  auto emptyRange =
-      std::make_pair(isl::val::nan(wrappedAccess.get_ctx()), isl::aff());
-  int pos = cstr.dim(isl::dim_type::set) - 1;
-  if (!cstr.is_lower_bound(isl::dim_type::set, pos)) {
-    return emptyRange;
-  }
-  if (cstr.involves_dims(isl::dim_type::div, 0, cstr.dim(isl::dim_type::div))) {
-    return emptyRange;
-  }
-
-  auto lowerBound = cstr.get_bound(isl::dim_type::set, pos).ceil();
-  auto aff = lowerBound.neg().add_coefficient_si(isl::dim_type::in, pos, 1);
-  lowerBound = lowerBound.drop_dims(isl::dim_type::in, pos, 1);
-
-  auto range = wrappedAccess.max_val(aff);
-  if (range.is_int()) {
-    return std::make_pair(range + 1, lowerBound);
-  }
-  return emptyRange;
-}
-
-std::pair<isl::val, isl::aff> outputRangeSingle(isl::map access) {
-  CHECK_EQ(access.dim(isl::dim_type::out), 1u)
-      << "expected 1-dim output, call outputRanges instead";
-  access = access.detect_equalities();
-  auto wrappedAccess = access.wrap().flatten().compute_divs().simple_hull();
+ScopedFootprint outputRanges(isl::map access) {
+  ScopedFootprint footprint;
 
   // TODO: also compute strides
 
-  isl::val minRange;
-  isl::aff lowerBoundWithMinRange;
-  for (auto cstr : wrappedAccess.get_constraint_list()) {
-    auto range = outputRange(wrappedAccess, cstr);
-    if (range.first.is_nan()) {
-      continue;
-    }
-    if (minRange.is_null() || range.first < minRange) {
-      minRange = range.first;
-      lowerBoundWithMinRange = range.second;
-    }
-  }
-  if (minRange.is_null()) {
-    return std::make_pair(
-        isl::val::nan(access.get_ctx()), lowerBoundWithMinRange);
-  }
-
-  return std::make_pair(minRange, lowerBoundWithMinRange);
-}
-
-ScopedFootprint outputRanges(isl::map access) {
-  int nSubscripts = access.dim(isl::dim_type::out);
-  ScopedFootprint footprint;
-  for (int i = 0; i < nSubscripts; ++i) {
-    auto singleDim =
-        access.project_out(isl::dim_type::out, 0, i)
-            .project_out(isl::dim_type::out, 1, nSubscripts - i - 1);
-    auto range = outputRangeSingle(singleDim);
-    if (range.first.is_nan()) {
-      return {};
-    }
-    footprint.emplace_back(range.second, range.first);
-  }
+  footprint.box = access.get_range_simple_fixed_box_hull();
   return footprint;
 }
 
@@ -126,7 +67,7 @@ std::unique_ptr<TensorReferenceGroup> TensorReferenceGroup::makeSingleton(
   group->references.push_back(std::move(ref));
   group->approximation = outputRanges(scopedAccess);
 
-  if (group->approximation.size() != scopedAccess.dim(isl::dim_type::out)) {
+  if (!group->approximation.box.is_valid()) {
     std::stringstream ss;
     ss << "could not compute rectangular overapproximation of: "
        << scopedAccess;
@@ -137,32 +78,25 @@ std::unique_ptr<TensorReferenceGroup> TensorReferenceGroup::makeSingleton(
 }
 
 isl::set ScopedFootprint::footprint(isl::set domain) const {
-  auto space = add_range(domain.get_space(), size());
+  auto space = box.get_space();
   auto accessed = isl::map::universe(space).intersect_domain(domain);
   auto lspace = isl::local_space(accessed.get_space().range());
 
-  for (size_t i = 0; i < size(); ++i) {
-    auto dim = at(i);
+  for (size_t i = 0; i < dim(); ++i) {
+    auto dimLowerBound = lowerBound(i);
     auto rhs = isl::aff(lspace, isl::dim_type::set, i);
-    isl::map partial = (isl::aff_map(dim.lowerBound) <= rhs) &
-        (isl::aff_map(dim.lowerBound + dim.size) > rhs);
+    isl::map partial = (isl::aff_map(dimLowerBound) <= rhs) &
+        (isl::aff_map(dimLowerBound + size(i)) > rhs);
     accessed = accessed & partial;
   }
   return accessed.range();
 }
 
 isl::multi_aff ScopedFootprint::lowerBounds() const {
-  if (size() == 0) {
+  if (dim() == 0) {
     throw promotion::PromotionNYI("promotion for scalars");
   }
-  auto space = add_range(at(0).lowerBound.get_space().domain(), size());
-  auto ma = isl::multi_aff::zero(space);
-
-  int i = 0;
-  for (const auto& a : *this) {
-    ma = ma.set_aff(i++, a.lowerBound);
-  }
-  return ma;
+  return box.get_offset();
 }
 
 bool TensorReferenceGroup::isReadOnly() const {
@@ -193,9 +127,9 @@ isl::set TensorReferenceGroup::promotedFootprint() const {
 
 std::vector<size_t> TensorReferenceGroup::approximationSizes() const {
   std::vector<size_t> result;
-  result.reserve(approximation.size());
-  for (const auto& dim : approximation) {
-    result.push_back(dim.size.get_num_si());
+  result.reserve(approximation.dim());
+  for (const auto& size : approximation.box.get_size().get_val_list()) {
+    result.push_back(size.get_num_si());
   }
   return result;
 }
@@ -382,9 +316,7 @@ isl::multi_aff TensorReferenceGroup::promotion() const {
   // lower bounds space is S -> P; which we transform into [S -> O] -> P
   auto lowerBounds = approximation.lowerBounds().pullback(
       isl::multi_aff::domain_map(accessSpace));
-  auto promotion = isl::multi_aff::range_map(accessSpace)
-                       .reset_tuple_id(isl::dim_type::out) -
-      lowerBounds;
+  auto promotion = isl::multi_aff::range_map(accessSpace) - lowerBounds;
   return promotion;
 }
 

tc/core/polyhedral/memory_promotion.h

@@ -30,28 +30,25 @@ namespace polyhedral {
 
 enum class AccessType : short { Read, Write };
 
-// A single dimension of the ScopedFootprint.
-// The scope is defined by a specific position in a schedule tree (const
-// ScheduleTree*), the user is responsible for maintaining the correspondance
-// between schedule tree positions and footprints.
-// Overapproximates one dimension by its lower bound, affine function of
+// Rectangular overapproximation of a tensor elements accessed through a single
+// reference.
+// Each dimension is overapproximated by a lower bound, an affine function of
 // parameters and schedule dimensions visible around the scope, and by a
 // constant size.
-struct ScopedFootprintDim {
- public:
-  ScopedFootprintDim(isl::aff lb, isl::val s) : lowerBound(lb), size(s) {}
-
- public:
-  isl::aff lowerBound;
-  isl::val size;
-};
-
-// Rectangular overapproximation of a tensor elements accessed through a single
-// reference.  Each dimension is described independently.
 // The scope is defined by a specific position in a schedule tree (const
 // ScheduleTree*), the user is responsible for maintaining the correspondance
 // between schedule tree positions and footprints.
-struct ScopedFootprint : std::vector<ScopedFootprintDim> {
+struct ScopedFootprint {
+  size_t dim() const {
+    return box.get_size().size();
+  }
+  isl::val size(size_t pos) const {
+    return box.get_size().get_val(pos);
+  }
+  isl::aff lowerBound(size_t pos) const {
+    return box.get_offset().get_aff(pos);
+  }
+  isl::fixed_box box;
   isl::set footprint(isl::set domain) const;
   isl::multi_aff lowerBounds() const;
 };
@@ -159,10 +156,11 @@ class TensorReferenceGroup {
 };
 
 inline std::ostream& operator<<(std::ostream& os, const ScopedFootprint& fp) {
-  os << "{\n";
-  for (const auto& f : fp) {
-    os << f.lowerBound << " of size " << f.size << "\n";
+  if (!fp.box) {
+    return os;
   }
+  os << "{\n";
+  os << fp.box.get_offset() << " of size " << fp.box.get_size() << "\n";
   os << "}";
   return os;
 }

test/test_cuda_mapper_memory_promotion.cc

@@ -281,14 +281,14 @@ def fun(float(N, M) A, float(N, M) B) -> (C) {
       }
 
       auto ref = oneGroup->references[0].get();
-      ASSERT_EQ(oneGroup->approximation.size(), 2u)
+      ASSERT_EQ(oneGroup->approximation.dim(), 2u)
           << "Could not compute footprint for" << ref->scopedAccess;
 
       EXPECT_EQ(
-          oneGroup->approximation[0].size,
+          oneGroup->approximation.size(0),
           isl::val(ctx, std::min(tile1, problemSize1)));
       EXPECT_EQ(
-          oneGroup->approximation[1].size,
+          oneGroup->approximation.size(1),
           isl::val(ctx, std::min(tile2, problemSize2)));
       auto footprint =
           oneGroup->approximateFootprint().intersect_params(blockZero);
@@ -346,12 +346,12 @@ def fun(float(N, M) A) -> (B, C) {
 
     const auto& groupsB = groups.at(idB);
     ASSERT_EQ(groupsB.size(), 1u) << "expected B refereces to be grouped";
-    ASSERT_EQ(groupsB[0]->approximation.size(), 2u) << "B should be a 2D array";
+    ASSERT_EQ(groupsB[0]->approximation.dim(), 2u) << "B should be a 2D array";
     EXPECT_EQ(
-        groupsB[0]->approximation[0].size,
+        groupsB[0]->approximation.size(0),
         isl::val(ctx, std::min(tile1, problemSize1)));
     EXPECT_EQ(
-        groupsB[0]->approximation[1].size,
+        groupsB[0]->approximation.size(1),
         isl::val(ctx, std::min(tile2, problemSize2)));
 
     auto t = scop.scheduleRoot()->child(childPos);