[SCHEDULE] Fuse support for 0 rank tensor (#1328)

include/tvm/schedule.h

@@ -130,6 +130,20 @@ class Stage : public NodeRef {
    */
   EXPORT Stage& fuse(IterVar outer, IterVar inner, IterVar* p_target);  // NOLINT(*)
   /*!
+   * \brief Fuse all the axes together into a single axis.
+   *
+   * \param axes All the axes to be fused.
+   * \param p_target The result target domain.
+   *
+   * \note axes can be an empty array,
+   *       in that case, a singleton itervar is created and
+   *       inserted to the outermost loop.
+   *       The fuse of empty array is used to support zero-dimension tensors.
+   *
+   * \return reference to self.
+   */
+  EXPORT Stage& fuse(const Array<IterVar>& axes, IterVar* p_target);  // NOLINT(*)
+  /*!
    * \brief Reorder the iteration
    * \param order The order of iteration variable.
    * \return reference to self.
@@ -151,9 +165,9 @@ class Stage : public NodeRef {
    * \return reference to self.
    */
   EXPORT Stage& tile(IterVar x_parent, IterVar y_parent,   // NOLINT(*)
-              Expr x_factor, Expr y_factor,
-              IterVar* p_x_outer, IterVar* p_y_outer,
-              IterVar* p_x_inner, IterVar* p_y_inner);
+                     Expr x_factor, Expr y_factor,
+                     IterVar* p_x_outer, IterVar* p_y_outer,
+                     IterVar* p_x_inner, IterVar* p_y_inner);
   /*!
    * \brief Vectorize iteration.
    * \param var The axis to be vectorized.
@@ -674,6 +688,25 @@ class RebaseNode : public IterVarRelationNode {
 };
 
 
+/*!
+ * \brief Singleton iterator [0, 1)
+ */
+class SingletonNode : public IterVarRelationNode {
+ public:
+  /*! \brief The singleton iterator */
+  IterVar iter;
+
+  void VisitAttrs(AttrVisitor* v) final {
+    v->Visit("iter", &iter);
+  }
+
+  static IterVarRelation make(IterVar iter);
+
+  static constexpr const char* _type_key = "Singleton";
+  TVM_DECLARE_NODE_TYPE_INFO(SingletonNode, IterVarRelationNode);
+};
+
+
 // implementations
 inline const StageNode* Stage::operator->() const {
   return static_cast<const StageNode*>(node_.get());

python/tvm/schedule.py

@@ -153,6 +153,12 @@ class Fuse(NodeBase):
 
 
 @register_node
+class Singleton(NodeBase):
+    """Singleton axis."""
+    pass
+
+
+@register_node
 class IterVar(NodeBase, _expr.ExprOp):
     """Represent iteration variable.
 
@@ -380,10 +386,7 @@ class Stage(NodeBase):
         fused : IterVar
             The fused variable of iteration.
         """
-        assert len(args) >= 1, "Length of the arguments must be >=1 for fuse."
-        fused = args[0]
-        for i in range(1, len(args)):
-            fused = _api_internal._StageFuse(self, fused, args[i])
+        fused = _api_internal._StageFuse(self, args)
         return fused
 
     def set_scope(self, scope):

src/api/api_lang.cc

@@ -350,7 +350,7 @@ TVM_REGISTER_API("_StageFuse")
 .set_body([](TVMArgs args, TVMRetValue* ret) {
     IterVar fused;
     args[0].operator Stage()
-        .fuse(args[1], args[2], &fused);
+        .fuse(args[1], &fused);
     *ret = fused;
   });
 

src/schedule/message_passing.cc

@@ -82,6 +82,8 @@ void PassDownDomain(const Stage& stage,
       Update(p_state, r->rebased,
              Range::make_by_min_extent(
                  0, state.at(r->parent)->extent));
+    } else if (const SingletonNode* s = rel.as<SingletonNode>()) {
+      Update(p_state, s->iter, Range::make_by_min_extent(0, 1));
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -147,6 +149,7 @@ void PassUpIndex(const Stage& stage,
       } else {
         state[s->parent] = value;
       }
+    } else if (rel.as<SingletonNode>()) {
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -192,6 +195,8 @@ void PassDownIndex(const Stage& stage,
       Expr parent_min = dom_map.at(s->parent)->min;
       CHECK(is_zero(parent_min));
       state[s->rebased] = value;
+    } else if (const SingletonNode* s = rel.as<SingletonNode>()) {
+      state[s->iter] = make_zero(s->iter->var.type());
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -296,6 +301,7 @@ void PassUpDomain(const Stage& stage,
                    state.at(r->rebased),
                    &parent);
       state[r->parent] = parent;
+    } else if (rel.as<SingletonNode>()) {
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -344,6 +350,7 @@ void PassUpBitMaskOr(const Stage& stage,
       } else {
         state[s->parent] |= state[s->rebased];
       }
+    } else if (rel.as<SingletonNode>()) {
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -390,6 +397,8 @@ void PassDownBitMaskOr(const Stage& stage,
       } else {
         state[s->rebased] |= state.at(s->parent);
       }
+    } else if (const SingletonNode* s = rel.as<SingletonNode>()) {
+      state[s->iter] = 0;
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -438,6 +447,8 @@ void PassUpBoundCheck(const Stage& s,
     } else if (rel.as<RebaseNode>()) {
       const RebaseNode* s = rel.as<RebaseNode>();
       state[s->parent] = state.at(s->rebased);
+    } else if (rel.as<SingletonNode>()) {
+      // nop
     } else {
       LOG(FATAL) << "unknown relation type";
     }

src/schedule/schedule_lang.cc

@@ -237,7 +237,6 @@ Stage& Stage::fuse(IterVar outer, IterVar inner, IterVar* p_target) {  // NOLINT
   IterVar fused = IterVarNode::make(
       Range(), Var(fused_name, outer->var.type()), iter_type);
 
-  *p_target = fused;
   ArrayNode* all_vars = self->all_iter_vars.CopyOnWrite();
   ArrayNode* leaf_vars = self->leaf_iter_vars.CopyOnWrite();
 
@@ -255,6 +254,31 @@ Stage& Stage::fuse(IterVar outer, IterVar inner, IterVar* p_target) {  // NOLINT
                         leaf_vars->data.begin() + pos_inner + 1);
   leaf_vars->data.insert(leaf_vars->data.begin() + pos_outer,
                          fused.node_);
+  *p_target = fused;
+  return *this;
+}
+
+Stage& Stage::fuse(const Array<IterVar>& axes, IterVar* p_target) {  // NOLINT(*)
+  if (axes.size() != 0) {
+    IterVar fused = axes[0];
+    for (size_t i = 1; i < axes.size(); ++i) {
+      this->fuse(fused, axes[i], &fused);
+    }
+    *p_target = std::move(fused);
+  } else {
+    StageNode* self = operator->();
+    // special handle fuse empty array.
+    // insert at the outer most loop
+    IterVar singleton = IterVarNode::make(
+        Range::make_by_min_extent(0, 1),
+        Var("singleton", Int(32)), kDataPar);
+    self->relations.push_back(SingletonNode::make(singleton));
+    ArrayNode* all_vars = self->all_iter_vars.CopyOnWrite();
+    ArrayNode* leaf_vars = self->leaf_iter_vars.CopyOnWrite();
+    all_vars->data.push_back(singleton.node_);
+    leaf_vars->data.insert(leaf_vars->data.begin(), singleton.node_);
+    *p_target = singleton;
+  }
   return *this;
 }
 
@@ -732,11 +756,18 @@ IterVarRelation RebaseNode::make(IterVar parent, IterVar rebased) {
   return IterVarRelation(n);
 }
 
+IterVarRelation SingletonNode::make(IterVar iter) {
+  auto n = std::make_shared<SingletonNode>();
+  n->iter = iter;
+  return IterVarRelation(n);
+}
+
 TVM_REGISTER_NODE_TYPE(StageNode);
 TVM_REGISTER_NODE_TYPE(IterVarAttrNode);
 TVM_REGISTER_NODE_TYPE(SplitNode);
 TVM_REGISTER_NODE_TYPE(FuseNode);
 TVM_REGISTER_NODE_TYPE(RebaseNode);
+TVM_REGISTER_NODE_TYPE(SingletonNode);
 TVM_REGISTER_NODE_TYPE(ScheduleNode);
 
 // Printer
@@ -778,6 +809,11 @@ TVM_STATIC_IR_FUNCTOR(IRPrinter, vtable)
     p->print(op->rebased);
     p->stream << ')';
 })
+.set_dispatch<SingletonNode>([](const SingletonNode *op, IRPrinter *p) {
+    p->stream << "singleton(";
+    p->print(op->iter);
+    p->stream << ')';
+})
 .set_dispatch<ScheduleNode>([](const ScheduleNode *op, IRPrinter *p) {
     p->stream << "schedule(" << op << ")";
   });

tests/python/integration/test_ewise.py

@@ -44,10 +44,10 @@ def test_multiple_cache_write():
     n = tvm.convert(1024)
     A0 = tvm.placeholder((n,), name='A0', dtype = "float32")
     A1 = tvm.placeholder((n,), name='A1', dtype = "float32")
-    B0, B1 = tvm.compute((n,), 
-            lambda *i: (A0(*i) + A1(*i), A0(*i) * A1(*i)), 
+    B0, B1 = tvm.compute((n,),
+            lambda *i: (A0(*i) + A1(*i), A0(*i) * A1(*i)),
             name='B')
-    C = tvm.compute((n,), lambda *i: B0(*i) + B1(*i), 
+    C = tvm.compute((n,), lambda *i: B0(*i) + B1(*i),
             name='C')
     s = tvm.create_schedule(C.op)
     # create iter var and assign them tags.
@@ -76,7 +76,7 @@ def test_multiple_cache_write():
         c = tvm.nd.array(np.zeros(n, dtype=C.dtype), ctx)
         func(a0, a1, c)
         np.testing.assert_allclose(
-            c.asnumpy(), a0.asnumpy() + a1.asnumpy() + (a0.asnumpy() * a1.asnumpy()), 
+            c.asnumpy(), a0.asnumpy() + a1.asnumpy() + (a0.asnumpy() * a1.asnumpy()),
             rtol=1e-5)
 
     check_device("cuda", "llvm")
@@ -235,7 +235,6 @@ def try_warp_memory():
         f(a, b)
         np.testing.assert_allclose(
             b.asnumpy(), a.asnumpy() + 3, rtol=1e-6)
-
     check_device("cuda")
 
 

tests/python/unittest/test_lang_schedule.py

@@ -84,6 +84,19 @@ def test_fuse():
     assert any(isinstance(x, tvm.schedule.Fuse) for x in s[T].relations)
     assert tuple(s[T].leaf_iter_vars) == (fused, xi, yi)
 
+
+def test_singleton():
+    A = tvm.placeholder((), name='A')
+    T = tvm.compute((), lambda : A() + 1)
+    s = tvm.create_schedule(T.op)
+    fused = s[T].fuse()
+    assert any(isinstance(x, tvm.schedule.Singleton) for x in s[T].relations)
+    assert tuple(s[T].leaf_iter_vars) == (fused,)
+    dump = pkl.dumps(s)
+    s_loaded = pkl.loads(dump)
+    assert isinstance(s_loaded, tvm.schedule.Schedule)
+
+
 def test_vectorize():
     m = tvm.var('m')
     n = tvm.var('n')
@@ -174,6 +187,7 @@ def test_tensor_intrin():
 
 
 if __name__ == "__main__":
+    test_singleton()
     test_pragma()
     test_tensor_intrin()
     test_rfactor()

topi/tests/python/test_topi_broadcast.py

@@ -94,6 +94,8 @@ def test_broadcast_to():
 
 def test_add():
     verify_broadcast_binary_ele(
+        (), (), topi.add, np.add)
+    verify_broadcast_binary_ele(
         (5, 2, 3), (2, 1), topi.add, np.add)
 
 def test_subtract():
@@ -114,6 +116,8 @@ def test_divide():
     verify_broadcast_binary_ele(
         None, (10,), topi.divide, np.divide, rhs_min=0.0001)
     verify_broadcast_binary_ele(
+        (), None, topi.divide, np.divide, rhs_min=0.0001)
+    verify_broadcast_binary_ele(
         (2, 3, 1, 32), (64, 32), topi.divide, np.divide, rhs_min=0.0001)
 
 def test_maximum_minmum():
@@ -157,10 +161,10 @@ def test_shift():
 
 
 if __name__ == "__main__":
+    test_add()
     test_shift()
     test_cmp()
     test_mod()
-    test_add()
     test_subtract()
     test_multiply()
     test_divide()