Breakpoint, expose the transformed axes for use in TE scheduling.

Lunderberg · Lunderberg · commit b961ad167b22 · 2021-12-13T08:29:55.000-06:00
Final step, exposing the axes generated in .transform_layout for use
in TE scheduling.
diff --git a/include/tvm/te/operation.h b/include/tvm/te/operation.h
@@ -265,6 +265,7 @@ class ComputeOp : public Operation {
                     Array<IterVar> axis, Array<PrimExpr> body);
 
   TVM_DEFINE_OBJECT_REF_METHODS(ComputeOp, Operation, ComputeOpNode);
+  TVM_DEFINE_OBJECT_REF_COW_METHOD(ComputeOpNode);
 };
 
 /*!
diff --git a/include/tvm/te/schedule.h b/include/tvm/te/schedule.h
@@ -272,10 +272,14 @@ class Stage : public ObjectRef {
    * Expressions should be in terms of the variables given in
    * initial_indices.
    *
+   * \param out_iter_vars An optional output location for the updated
+   * loop iteration variables.
+   *
    * \return reference to self
    */
   TVM_DLL Stage& transform_layout(const Array<Var>& initial_indices,
-                                  const Array<PrimExpr>& final_indices);
+                                  const Array<PrimExpr>& final_indices,
+                                  Array<IterVar>* out_iter_vars = nullptr);
   /*! \brief Defines separators between groups of axes.
    *
    * Used to define `BufferNode::axis_separators`, which has
@@ -494,9 +498,27 @@ class StageNode : public Object {
    *  while origin_op remains fixed.
    */
   Operation origin_op;
-  /*! \brief All the nodes in the iter var */
+  /*! \brief All the nodes in the iter var
+   *
+   * Each element of all_iter_vars represents an iteration variable
+   * that may appear within this stage's computation.  Any element
+   * of `all_iter_vars` that is in `leaf_iter_vars` represents a
+   * variable that is directly defined and usable within the stage's
+   * computation.  All other elements of `all_iter_vars` represent
+   * variables whose value must be computed from the variables in
+   * `leaf_iter_vars`.  (e.g. Support index k has been split by
+   * ``ko, ki = s.split(k, factor=4)``.  ko and ki will appear in
+   * `leaf_iter_vars`, while k will not, and must be computed as
+   * `4*ko + ki`.
+   */
   Array<IterVar> all_iter_vars;
-  /*! \brief The current active leaf iter vars in the stage. */
+  /*! \brief The current active leaf iter vars in the stage.
+   *
+   * Each element of leaf_iter_vars will either be replaced with the
+   * bound index (e.g. threadIdx.x), or will be expanded into a loop
+   * over the variable's extent.  `leaf_iter_vars` is a subset of
+   * `all_iter_vars`.
+   */
   Array<IterVar> leaf_iter_vars;
   /*!
    * \brief Specify threads to be launched at the stage.
@@ -809,6 +831,36 @@ class Singleton : public IterVarRelation {
   TVM_DEFINE_OBJECT_REF_METHODS(Singleton, IterVarRelation, SingletonNode);
 };
 
+/*!
+ * \brief Transform iterator according to some arbitrary expression.
+ */
+class TransformNode : public IterVarRelationNode {
+ public:
+  Array<IterVar> original_variables;
+  Array<IterVar> transformed_variables;
+  IndexMap forward_transformation;
+  IndexMap inverse_transformation;
+
+  void VisitAttrs(AttrVisitor* v) {
+    v->Visit("original_variables", &original_variables);
+    v->Visit("transformed_variables", &transformed_variables);
+    v->Visit("forward_transformation", &forward_transformation);
+    v->Visit("inverse_transformation", &inverse_transformation);
+  }
+
+  static constexpr const char* _type_key = "Transform";
+  TVM_DECLARE_FINAL_OBJECT_INFO(TransformNode, IterVarRelationNode);
+};
+
+class Transform : public IterVarRelation {
+ public:
+  TVM_DLL explicit Transform(Array<IterVar> original_variables,
+                             Array<IterVar> transformed_variables, IndexMap forward_transformation,
+                             IndexMap inverse_transformation);
+
+  TVM_DEFINE_OBJECT_REF_METHODS(Transform, IterVarRelation, TransformNode);
+};
+
 /*! \brief Container for specialization conditions. */
 class SpecializedConditionNode : public Object {
  public:
diff --git a/python/tvm/te/schedule.py b/python/tvm/te/schedule.py
@@ -527,9 +527,15 @@ def transform_layout(self, mapping_function: Callable[..., List[tvm.tir.PrimExpr
         """Defines the layout transformation for the current stage's tensor.
 
         The map from initial_indices to final_indices must be an
-        invertible affine transformation.
+        invertible affine transformation.  This method may be called
+        more than once for a given tensor, in which case each
+        transformation is applied sequentially.
 
-        This method may be called more than once for a given tensor, in which case each
+        If the stage is a ComputeOp, then the iteration order of the
+        compute stage is rewritten to be a row-major traversal of the
+        tensor, and the new loop iteration variables are returned.
+        For all other stages, the loop iteration order is unmodified,
+        and the return value is None.
 
         Parameters
         ----------
@@ -543,6 +549,17 @@ def transform_layout(self, mapping_function: Callable[..., List[tvm.tir.PrimExpr
             the current stage's tensor, using the post-transformation
             layout.
 
+        Returns
+        -------
+        new_iter_vars : Optional[List[tvm.tir.IterVar]]
+
+            If the stage is a ComputeOp, then the return will be the
+            updated loop iteration variables over the data array, in
+            the same order as the output values from the
+            `mapping_function`.
+
+            Otherwise, the return value is None.
+
         Examples
         --------
         .. code-block:: python
@@ -557,15 +574,29 @@ def transform_layout(self, mapping_function: Callable[..., List[tvm.tir.PrimExpr
 
         .. code-block:: python
 
-            # ``A`` is a tensor whose compute definition is in format,
-            # and should be transformed such that the last index is
-            # split, with the slower-chan index of the split placed at the
-            # slowest changing dimension.
+            # ``A`` is a tensor whose compute definition is in an
+            # arbitrary format, and should be transformed such that
+            # the last index is split, with the slower-changing index
+            # of the split placed at the slowest changing dimension.
 
             s[A].transform_layout(
                 lambda *indices, i: [i//4, *indices, i%4]
             )
 
+        .. code-block:: python
+
+            # ``B`` is a tensor defined by te.compute to be a copy of
+            # ``A`, and should be transformed such that ``B``'s layout
+            # is a transpose of ``A``'s layout.  The loop iteration
+            # that computes ``B`` will correspond to ``B``'s memory
+            # layout.
+
+            A = te.placeholder([n,m])
+            B = te.compute(A.shape, lambda i,j: A[i,j])
+            s = te.create_schedule(B.op)
+
+            s[B].transform_layout(lambda i,j: [j,i])
+
         """
 
         args = []
@@ -626,9 +657,10 @@ def transform_layout(self, mapping_function: Callable[..., List[tvm.tir.PrimExpr
                     "Instead received {val} of type {type(val)}."
                 )
 
-        _ffi_api.StageTransformLayout(self, initial_indices, final_indices)
+        new_iter_vars = _ffi_api.StageTransformLayout(self, initial_indices, final_indices)
         _ffi_api.StageSetAxisSeparators(self, axis_separators)
 
+        return new_iter_vars or None
 
 
 @tvm._ffi.register_object
diff --git a/src/te/schedule/message_passing.cc b/src/te/schedule/message_passing.cc
@@ -79,6 +79,22 @@ void PassUpThreadBinding(const Stage& stage, std::unordered_map<IterVar, bool>*
     } else if (const RebaseNode* s = rel.as<RebaseNode>()) {
       state[s->parent] = state[s->rebased];
     } else if (rel.as<SingletonNode>()) {
+    } else if (const TransformNode* s = rel.as<TransformNode>()) {
+      // Currently, this marks all original iter vars as deriving from
+      // a thread bind if any of the transformed variables are bound,
+      // even if the inverse expression for that iter var doesn't
+      // depend on the bound variable.
+
+      // TODO(Lunderberg): For each of original variable, check
+      // whether any variable in the inverse expression for it has a
+      // thread binding.
+      bool is_thread_binding = false;
+      for (const auto& iter_var : s->transformed_variables) {
+        is_thread_binding = is_thread_binding || state[iter_var];
+      }
+      for (const auto& iter_var : s->original_variables) {
+        state[iter_var] = is_thread_binding;
+      }
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -157,6 +173,29 @@ void PassDownDomain(const Stage& stage, std::unordered_map<IterVar, Range>* p_st
       Update(p_state, r->rebased, Range::FromMinExtent(0, state.at(r->parent)->extent), actx);
     } else if (const SingletonNode* s = rel.as<SingletonNode>()) {
       Update(p_state, s->iter, Range::FromMinExtent(0, 1), actx);
+    } else if (const TransformNode* s = rel.as<TransformNode>()) {
+      bool missing_originals = false;
+      for (const auto& iter_var : s->original_variables) {
+        if (!state.count(iter_var)) {
+          ICHECK(allow_missing);
+          missing_originals = true;
+        }
+      }
+      if (missing_originals) {
+        continue;
+      }
+
+      Array<Range> original_ranges;
+      for (const auto& iter_var : s->original_variables) {
+        original_ranges.push_back(state[iter_var]);
+      }
+      Array<Range> updated_ranges = s->forward_transformation->MapRanges(original_ranges);
+
+      ICHECK_EQ(updated_ranges.size(), s->transformed_variables.size());
+      for (size_t i = 0; i < updated_ranges.size(); i++) {
+        Update(p_state, s->transformed_variables[i], updated_ranges[i], actx);
+      }
+
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -225,6 +264,39 @@ void PassUpIndex(const Stage& stage, const Map<IterVar, Range>& dom_map,
         state[s->parent] = value;
       }
     } else if (rel.as<SingletonNode>()) {
+    } else if (const TransformNode* s = rel.as<TransformNode>()) {
+      bool missing_transformed = false;
+      for (const auto& iter_var : s->transformed_variables) {
+        if (!state.count(iter_var)) {
+          // for (const auto& kv : state) {
+          //   std::cout << "Looking for " << tvm::PrettyPrint(iter_var) << std::endl;
+          //   std::cout << "State contains " << tvm::PrettyPrint(kv.first) << " -> "
+          //             << tvm::PrettyPrint(kv.second) << std::endl;
+          // }
+          // TODO: Decide whether to have this check, for similarity
+          // with other handlers.  In this case, the indices may
+          // already be in terms of the pre-transformed variables, so
+          // no need to untransform them?
+
+          // ICHECK(allow_missing);
+          missing_transformed = true;
+        }
+      }
+      if (missing_transformed) {
+        continue;
+      }
+
+      Array<PrimExpr> transformed_indices;
+      for (const auto& iter_var : s->transformed_variables) {
+        transformed_indices.push_back(state[iter_var]);
+      }
+      Array<PrimExpr> original_indices = s->inverse_transformation->MapIndices(transformed_indices);
+
+      ICHECK_EQ(original_indices.size(), s->original_variables.size());
+      for (size_t i = 0; i < original_indices.size(); i++) {
+        state[s->original_variables[i]] = original_indices[i];
+      }
+
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -270,6 +342,28 @@ void PassDownIndex(const Stage& stage, const Map<IterVar, Range>& dom_map,
       state[s->rebased] = value;
     } else if (const SingletonNode* s = rel.as<SingletonNode>()) {
       state[s->iter] = make_zero(s->iter->var.dtype());
+    } else if (const TransformNode* s = rel.as<TransformNode>()) {
+      bool missing_originals = false;
+      for (const auto& iter_var : s->original_variables) {
+        if (!state.count(iter_var)) {
+          ICHECK(allow_missing);
+          missing_originals = true;
+        }
+      }
+      if (missing_originals) {
+        continue;
+      }
+
+      Array<PrimExpr> original_indices;
+      for (const auto& iter_var : s->original_variables) {
+        original_indices.push_back(state[iter_var]);
+      }
+      Array<PrimExpr> transformed_indices = s->forward_transformation->MapIndices(original_indices);
+
+      ICHECK_EQ(transformed_indices.size(), s->transformed_variables.size());
+      for (size_t i = 0; i < transformed_indices.size(); i++) {
+        state[s->transformed_variables[i]] = transformed_indices[i];
+      }
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -351,6 +445,26 @@ void PassUpDomain(const RebaseNode* s, const std::unordered_map<IterVar, Range>&
   *parent = arith::EvalSet(s->rebased->var + parent_min, {{s->rebased, rebased}});
 }
 
+Array<IntSet> PassUpDomain(const TransformNode* s,
+                           const std::unordered_map<IterVar, Range>& dom_map,
+                           const Map<IterVar, IntSet>& transformed_domains) {
+  Array<IntSet> output;
+
+  Array<PrimExpr> transformed_indices;
+  for (const auto& iter_var : s->transformed_variables) {
+    transformed_indices.push_back(iter_var->var);
+  }
+
+  Array<PrimExpr> transformed_exprs = s->inverse_transformation->MapIndices(transformed_indices);
+
+  ICHECK_EQ(transformed_exprs.size(), s->original_variables.size());
+  for (size_t i = 0; i < transformed_exprs.size(); i++) {
+    output.push_back(arith::EvalSet(transformed_exprs[i], transformed_domains));
+  }
+
+  return output;
+}
+
 void PassUpDomain(const Stage& stage, const std::unordered_map<IterVar, Range>& dom_map,
                   std::unordered_map<IterVar, IntSet>* p_state) {
   auto& state = *p_state;
@@ -370,6 +484,16 @@ void PassUpDomain(const Stage& stage, const std::unordered_map<IterVar, Range>&
       PassUpDomain(r, dom_map, state.at(r->rebased), &parent);
       state[r->parent] = parent;
     } else if (rel.as<SingletonNode>()) {
+    } else if (const TransformNode* r = rel.as<TransformNode>()) {
+      Map<IterVar, IntSet> transformed_domains;
+      for (const auto& var : r->transformed_variables) {
+        transformed_domains.Set(var, state.at(var));
+      }
+      auto original_ranges = PassUpDomain(r, dom_map, transformed_domains);
+      ICHECK_EQ(original_ranges.size(), r->original_variables.size());
+      for (size_t i = 0; i < original_ranges.size(); i++) {
+        state[r->original_variables[i]] = original_ranges[i];
+      }
     } else {
       LOG(FATAL) << "unknown relation type";
     }
@@ -509,6 +633,22 @@ void PassUpBoundCheck(const Stage& s, const Map<IterVar, Range>& dom_map,
       state[s->parent] = state.at(s->rebased);
     } else if (rel.as<SingletonNode>()) {
       // nop
+    } else if (const TransformNode* s = rel.as<TransformNode>()) {
+      // Currently, this marks all original iter vars as requiring
+      // bounds checks if any of the transformed variables require
+      // bounds checks, even if the inverse expression for that iter
+      // var doesn't depend on the bound variable.
+
+      // TODO(Lunderberg): For each of original variable, check
+      // whether any variable in the inverse expression for it
+      // requires bounds checking.
+      bool needs_bounds_check = false;
+      for (const auto& iter_var : s->transformed_variables) {
+        needs_bounds_check = needs_bounds_check || state[iter_var];
+      }
+      for (const auto& iter_var : s->original_variables) {
+        state[iter_var] = needs_bounds_check;
+      }
     } else {
       LOG(FATAL) << "unknown relation type";
     }
diff --git a/src/te/schedule/schedule_lang.cc b/src/te/schedule/schedule_lang.cc
diff --git a/tests/python/unittest/test_transform_layout.py b/tests/python/unittest/test_transform_layout.py
