Make Tensor comparator and hash to be aware of same op and index, init checkin of the ir generation

include/tvm/operation.h

@@ -86,4 +86,17 @@ inline Tensor Compute(Array<Expr> shape,
 
 }  // namespace tvm
 
+
+namespace std {
+template <>
+struct hash<::tvm::Tensor> {
+  std::size_t operator()(const ::tvm::Tensor& k) const {
+    if (k.defined() && k->op.defined()) {
+      return k->op.hash();
+    } else{
+      return k.hash();
+    }
+  }
+};
+}  // namespace std
 #endif  // TVM_OPERATION_H_

include/tvm/tensor.h

@@ -47,6 +47,12 @@ class Tensor : public FunctionRef {
    * \return the pointer to the internal node container
    */
   inline const TensorNode* operator->() const;
+  /*!
+   * \brief check if two tensors equals each other.
+   * \param other tensor to be checked.
+   * \return whether the two tensors equals each other.
+   */
+  inline bool operator==(const Tensor& other) const;
   /*! \return The dimension of the tensor */
   inline size_t ndim() const;
   /*!
@@ -201,6 +207,17 @@ inline size_t Tensor::ndim() const {
   return (*this)->shape.size();
 }
 
+inline bool Tensor::operator==(const Tensor& other) const {
+  if (get() == other.get()) return true;
+  if (get() == nullptr || other.get() == nullptr) return false;
+  if ((*this)->op.defined() || other->op.defined()) {
+    return (*this)->op == other->op &&
+        (*this)->value_index == other->value_index;
+  } else {
+    return false;
+  }
+}
+
 // macro to turn every operation of slice to expression
 #define DEFINE_OVERLOAD_SLICE_UNARY_OP(Op)                              \
   inline Expr operator Op (const Tensor::Slice& a) {                    \

python/tvm/_ctypes/_api.py

@@ -29,9 +29,12 @@ kNodeHandle = 4
 def _type_key(handle):
     ret_val = ArgVariant()
     ret_typeid = ctypes.c_int()
+    ret_success = ctypes.c_int()
     check_call(_LIB.TVMNodeGetAttr(
         handle, c_str("type_key"),
-        ctypes.byref(ret_val), ctypes.byref(ret_typeid)))
+        ctypes.byref(ret_val),
+        ctypes.byref(ret_typeid),
+        ctypes.byref(ret_success)))
     return py_str(ret_val.v_str)
 
 NODE_TYPE = {

python/tvm/tensor.py

 from __future__ import absolute_import as _abs
 from ._ctypes._api import NodeBase, SliceBase, register_node, convert
 from . import collections as _collections
+from . import _function_internal
 from . import make as _make
 from . import expr as _expr
 
@@ -38,6 +39,35 @@ class Tensor(NodeBase):
     def __getitem__(self, indices):
         return TensorSlice(self, indices)
 
+    def __hash__(self):
+        return _function_internal._TensorHash(self)
+
+    def __eq__(self, other):
+        if not isinstance(other, Tensor):
+            return False
+        return _function_internal._TensorEqual(self, other)
+
     @property
     def ndim(self):
         return len(self.shape)
+
+
+class Operation(NodeBase):
+    def output(self, index):
+        """Get the index-th output of the operation
+
+        Parameters
+        ----------
+        index : int
+            The index size.
+
+        Returns
+        -------
+        out : Tensor
+            The i-th output.
+        """
+        return _function_internal._OpGetOutput(self, index)
+
+@register_node
+class ComputeOp(Operation):
+    pass

src/c_api/c_api_lang.cc

@@ -149,6 +149,17 @@ TVM_REGISTER_API(_Tensor)
                             args.at(4));
   });
 
+TVM_REGISTER_API(_TensorEqual)
+.set_body([](const ArgStack& args,  RetValue *ret) {
+    *ret = args.at(0).operator Tensor() == args.at(1).operator Tensor();
+  });
+
+TVM_REGISTER_API(_TensorHash)
+.set_body([](const ArgStack& args,  RetValue *ret) {
+    *ret = static_cast<int64_t>(
+        std::hash<Tensor>()(args.at(0).operator Tensor()));
+  });
+
 TVM_REGISTER_API(_ComputeOp)
 .set_body([](const ArgStack& args,  RetValue *ret) {
     *ret = ComputeOpNode::make(args.at(0),
@@ -156,6 +167,12 @@ TVM_REGISTER_API(_ComputeOp)
                                args.at(2));
   });
 
+TVM_REGISTER_API(_OpGetOutput)
+.set_body([](const ArgStack& args,  RetValue *ret) {
+    *ret = args.at(0).operator Operation().output(
+        args.at(1).operator size_t());
+  });
+
 
 TVM_REGISTER_API(_IterVar)
 .set_body([](const ArgStack& args,  RetValue *ret) {

src/pass/schedule_ops.cc

@@ -5,11 +5,225 @@
 #include <tvm/ir.h>
 #include <tvm/ir_mutator.h>
 #include <tvm/ir_pass.h>
+#include <tvm/ir_visitor.h>
 #include "./scope.h"
 
 namespace tvm {
 namespace ir {
 namespace {
+
+/*!
+ * \brief use message passing to calculate the assignment of each Var inside the loop body.
+ * \param s The schedule to be used.
+ * \param dom_map The domain map of each iteration variable's domain
+ * \param p_state The message passing state
+ *     IterVar->The assignment.
+ */
+void PassUpOffset(const Schedule& s,
+                  const std::unordered_map<IterVar, Range>& dom_map,
+                  std::unordered_map<IterVar, Expr>* p_state) {
+  auto& state = *p_state;
+  for (size_t i = s->relations.size(); i != 0; --i) {
+    IterVarRelation rel = s->relations[i - 1];
+    if (rel.as<SplitNode>()) {
+      const SplitNode* s = rel.as<SplitNode>();
+      Expr outer = state.at(s->outer);
+      Expr inner = state.at(s->outer);
+      Expr factor = dom_map.at(s->outer)->extent;
+      Expr offset = inner + outer * factor;
+      Expr outer_min = dom_map.at(s->parent)->min;
+      if (!is_zero(outer_min)) {
+        offset = outer_min + offset;
+      }
+      state[s->parent] = offset;
+    } else if (rel.as<FuseNode>()) {
+      const FuseNode* s = rel.as<FuseNode>();
+      Expr value = state.at(s->fused);
+      Expr factor = dom_map.at(s->outer)->extent;
+      state[s->outer] = value / factor;
+      state[s->inner] = value % factor;
+    } else {
+      LOG(FATAL) << "unknown relation type";
+    }
+  }
+}
+
+/*!
+ * \brief split the expr by addition.
+ * \param expr The expression to be splitted.
+ * \param loop_level The loop level of each Variable
+ * \param result vector of (level, expr)
+ *   The level gives the mimimum loop level this expression need to be computed.
+ *   The Expr gives the expression content.
+ */
+void SplitByAdd(Expr expr,
+                const std::unordered_map<const Variable*, size_t>& loop_level,
+                std::vector<std::pair<size_t, Expr> > *result) {
+  const Add* op = expr.as<Add>();
+  if (op != nullptr) {
+    SplitByAdd(op->a, loop_level, result);
+    SplitByAdd(op->b, loop_level, result);
+  } else {
+    size_t max_level = 0;
+    auto fvisit = [&max_level, &loop_level](const NodeRef& n) {
+      const Variable* op = n.as<Variable>();
+      if (op != nullptr) {
+        auto it = loop_level.find(op);
+        if (it != loop_level.end()) {
+          max_level = std::max(max_level, it->second);
+        }
+      }
+    };
+    PostOrderVisit(expr, fvisit);
+    result->push_back(std::make_pair(max_level, expr));
+  }
+}
+
+/*!
+ * \brief combine the nest stmt, whose body is not defined.
+ * \param nest A list of For and LetStmt, whose body is not defined.
+ * \param body body
+ */
+Stmt CombineNest(std::vector<Stmt>&& nest, Stmt body) {
+  while (!nest.empty()) {
+    Stmt s = std::move(nest.back());
+    nest.pop_back();
+    if (s.as<For>()) {
+      auto n = std::make_shared<For>(*s.as<For>());
+      n->body = body;
+      body = Stmt(n);
+    } else if (s.as<LetStmt>()) {
+      auto n = std::make_shared<LetStmt>(*s.as<LetStmt>());
+      n->body = body;
+      body = Stmt(n);
+    } else if (s.as<AttrStmt>()) {
+      auto n = std::make_shared<AttrStmt>(*s.as<AttrStmt>());
+      n->body = body;
+      body = Stmt(n);
+    } else {
+      LOG(FATAL) << "not supported nest type";
+    }
+  }
+  return body;
+}
+
+/*!
+ * \brief Make the loop nest of the correspondings schedule.
+ * \param sch The schedule.
+ * \param dom_map The domain map.
+ */
+std::vector<Stmt> MakeLoopNest(
+    const Schedule& sch,
+    const std::unordered_map<IterVar, Range>& dom_map) {
+  // optional, use let to define some CSE in dom_map.
+  auto leaf_iter_vars = sch->leaf_iter_vars;
+  std::unordered_map<IterVar, Expr> offset;
+  std::unordered_map<const Variable*, size_t> loop_level;
+
+  // create the loop nest
+  std::vector<Stmt> nest;
+  nest.resize(leaf_iter_vars.size() + 1, Stmt());
+
+  for (size_t i = 0; i < leaf_iter_vars.size(); ++i) {
+    auto iv = leaf_iter_vars[i];
+    // initialize the offset and loop_level
+    offset[iv] = iv->var;
+    loop_level[iv->var.as<Variable>()] = i + 1;
+
+    nest[i] = AttrStmt::make(iv->var, "scope", iv, Stmt());
+    if (iv->thread_tag.length() == 0) {
+      Range dom = dom_map.at(iv);
+      nest[i] = For::make(iv->var, dom->min, dom->extent,
+                          ForType::Serial, DeviceAPI::None, nest[i]);
+    }
+  }
+  // message passing to get offset of root iter vars.
+  PassUpOffset(sch, dom_map, &offset);
+  for (IterVar iv : sch->op->root_iter_vars()) {
+    Expr value = offset.at(iv);
+    if (value.same_as(iv->var)) continue;
+    using Entry = std::pair<size_t, Expr>;
+    std::vector<Entry> splits;
+    SplitByAdd(value, loop_level, &splits);
+
+    Expr offset = 0;
+    for (size_t i = 0; i <= leaf_iter_vars.size(); ++i) {
+      auto iv = leaf_iter_vars[i];
+      for (const auto& kv : splits) {
+        if (kv.first == i) {
+          offset = offset + splits[i].second;
+        }
+      }
+      std::ostringstream os;
+      os << iv->var->name_hint << ".at.l" << i;
+      Var base_offset(os.str());
+      nest[i] = LetStmt::make(base_offset, offset, nest[i]);
+      offset = base_offset;
+    }
+    nest.back() = LetStmt::make(iv->var, offset, nest.back());
+  }
+  return nest;
+}
+
+/*!
+ * \brief Make the loop nest of the correspondings schedule.
+ * \param op The operation.
+ */
+Stmt MakeBody(const Operation& op) {
+  Stmt body;
+  if (op.as<ComputeOpNode>()) {
+    const ComputeOpNode* compute = op.as<ComputeOpNode>();
+    // Note: Tensor's address cannot uniquely
+    Tensor t = op.output(0);
+    Array<Expr> args;
+    for (IterVar iv : compute->axis) {
+      args.push_back(iv->var);
+    }
+    body = Provide::make(t, {compute->body}, args);
+  } else {
+    LOG(FATAL) << "not supported op";
+  }
+  return body;
+}
+
+Stmt MakePipeline(const Schedule& sch, Stmt body) {
+  return body;
+}
+
+// inject the operator's realization on the stmt.
+class InjectRealize : public IRMutator {
+ public:
+  explicit InjectRealize(Schedule sch)
+      : sch_(sch) {}
+
+  Stmt Mutate(Stmt stmt) final {
+    const AttrStmt* op = stmt.as<AttrStmt>();
+    if (op != nullptr) {
+      attr_scope_.Push({op->node, op->type_key}, op->value);
+      stmt = IRMutator::Mutate(stmt);
+      attr_scope_.Pop({op->node, op->type_key});
+    } else {
+      stmt = IRMutator::Mutate(stmt);
+    }
+
+    if (op != nullptr &&
+        op->type_key == "scope" &&
+        op->node == sch_->attach_parent) {
+      return AttrStmt::make(
+          op->node, op->type_key, op->value,
+          MakePipeline(sch_, op->body));
+    } else {
+      return stmt;
+    }
+  }
+
+ private:
+  // the operations to be carried
+  Schedule sch_;
+  Scope<AttrKey, Expr> attr_scope_;
+};
+
+
 }  // namespace
 }  // namespace ir
 }  // namespace tvm

src/schedule/bound.cc

@@ -101,7 +101,11 @@ void PassToOperation(
     const Tensor& tensor,
     const std::vector<IntSet>& dim_bounds,
     std::unordered_map<IterVar, std::vector<IntSet> >* result) {
-
+  // This is a push style operation, given output bound, push to the op IterVar bound.
+  // It cannot handle complicated cases where op bound is coupled with bounds of
+  // all of its outputs, without having a simple communicative union relation.
+  //
+  // Eventually, we need to change the inference to be a Pull style inference
   if (tensor->op.as<ComputeOpNode>()) {
     auto root_iter_vars = tensor->op->root_iter_vars();
     CHECK_EQ(tensor.ndim(), root_iter_vars.size());

src/schedule/int_set.cc

@@ -220,23 +220,25 @@ void PassUp(const SplitNode* s,
     *parent = IntSet::make_range(dom_map.at(s->parent));
     return;
   }
+  Expr factor = dom_map.at(s->outer)->extent;
   CHECK(outer.defined());
   CHECK(inner.defined());
+  CHECK(factor.defined());
   // copy construct
   auto n = std::make_shared<IntSetNode>(*(inner.operator->()));
 
   if (IsNumber(outer)) {
     // shift the base offset
     n->base = Range::make_with_min_extent(
-        AsNumber(outer) * s->factor + inner->base->min,
+        AsNumber(outer) * factor + inner->base->min,
         inner->base->extent);
   } else {
     // default use all domains in the data.
     n->domain.push_back(outer->base);
-    n->stride.push_back(s->factor);
+    n->stride.push_back(factor);
     for (size_t i = 0; i < outer->domain.size(); ++i) {
       n->domain.push_back(outer->domain[i]);
-      n->stride.push_back(outer->stride[i] * s->factor);
+      n->stride.push_back(outer->stride[i] * factor);
     }
   }
   *parent = IntSet(n);

tests/python/test_tensor.py

@@ -11,6 +11,12 @@ def test_tensor():
     print(T.op.body)
     assert(tuple(T.shape) == (m, n, l))
     assert(A.op is None)
+    assert(A == A)
+    assert(T.op.output(0) == T)
+    assert(T.op.output(0).__hash__() == T.__hash__())
+    d = {T.op.output(0) : 1}
+    assert(d[T] == 1)
+
 
 def test_tensor_reduce():
     m = tvm.Var('m')
@@ -21,7 +27,6 @@ def test_tensor_reduce():
     T = tvm.compute((m, n, l), lambda i, j, k: A[i, k] * B[j, k])
     rv = tvm.IterVar((0, A.shape[1]), name="k")
     C = tvm.compute((m, n), lambda i, j: tvm.sum(T(i, j, rv+1), rdom=rv))
-
     print(C.op.body)
 
 if __name__ == "__main__":