[Relay] GradientCell Relay Pass (apache#5039)

* save

* gradient.rly

* fix

* NOT WORKING: gradient cell pass

* test gradient pass

* fixed basic call ops

* more tests

* fix bug

* transform calls to one ones_like zero zero_like

* maintenance stuff

* fix linting

* linting

* linting

* throw default

* remove unrelated changes

* import gradent.rly in pass

* comment

* linting

* remove changes to test files

* move gradient_cell.cc to transforms

* revert change

* update files with new commits

* type

* wrapper function to main outermost function type

* fix linting

* fix unsigned and signed int comparison

* review

* GetConstructor definition in module and change op comparison

* update node instantiations

* increase code readability

Co-authored-by: Marisa Kirisame <[email protected]>

include/tvm/ir/module.h

@@ -162,6 +162,14 @@ class IRModuleNode : public Object {
    */
   TVM_DLL Array<GlobalTypeVar> GetGlobalTypeVars() const;
 
+  /*!
+   * \brief Find constructor of ADT using name
+   * \param adt name of the ADT the constructor belongs to
+   * \param cons name of the constructor
+   * \returns Constructor of ADT, error if not found
+   */
+  TVM_DLL Constructor GetConstructor(const std::string& adt, const std::string& cons) const;
+
   /*!
    * \brief Look up a global function by its variable.
    * \param var The global var to lookup.

include/tvm/relay/transform.h

@@ -77,6 +77,20 @@ TVM_DLL Pass CreateFunctionPass(const runtime::TypedPackedFunc<
  */
 TVM_DLL Pass DeadCodeElimination(bool inline_once = false);
 
+/*!
+* \brief Convert all expressions of TensorType into GradCell,
+* an algebraic data type defined in gradient.rly.
+*
+* This will delay or decrease memory usage. All calls to
+* ones, ones_like, zeros, zeros_like will not immediately instantiate a tensor in memory,
+* rather only instantiate if needed. It also defines + and * operation
+* between GradCell types which can increase performance when using
+* zero-filled or one-filled tensors, which is the case in reverse mode ad.
+*
+* \return the pass
+*/
+TVM_DLL Pass LazyGradientInit();
+
 /*!
  * \brief Fold constant expressions.
  *

python/tvm/relay/std/gradient.rly

@@ -0,0 +1,55 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied.  See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+v0.0.4
+
+/*
+ * Store the Gradient Value of a Tensor of type T.
+ * Note that Gradient of T is stored inside a Ref(GradCell[T]) instead of GradCell[T].
+ */
+type GradCell[T] {
+  Raw(T),
+  One(fn() -> T),
+  Zero(fn() -> T)
+}
+
+def @FromGradCell[T](%g: GradCell[T]) -> T {
+  match (%g) {
+    Raw(%x) => %x,
+    One(%x) => %x(),
+    Zero(%x) => %x()
+  }
+}
+
+def @MultiplyGradCell[T](%multiply: fn(T, T) -> T, %l: GradCell[T], %r: GradCell[T]) -> GradCell[T] {
+  match((%l, %r)) {
+    (Zero(_), _) => %l,
+    (_, Zero(_)) => %r,
+    (One(_), _) => %r,
+    (_, One(_)) => %l,
+    _ => Raw(%multiply(@FromGradCell(%l), @FromGradCell(%r)))
+  }
+}
+
+def @AddGradCell[T](%add: fn(T, T) -> T, %l: GradCell[T], %r: GradCell[T]) -> GradCell[T] {
+  match ((%l, %r)) {
+    (Zero(_), _) => %r,
+    (_, Zero(_)) => %l,
+    _ => Raw(%add(@FromGradCell(%l), @FromGradCell(%r)))
+  }
+}

python/tvm/relay/transform/transform.py

@@ -219,6 +219,19 @@ def DeadCodeElimination(inline_once=False):
     """
     return _ffi_api.DeadCodeElimination(inline_once)
 
+def LazyGradientInit():
+    """Reduces memory usage of gradient tensors
+
+    Parameters
+    ----------
+
+    Returns
+    -------
+    ret: tvm.relay.Pass
+        A pass which delays and/or reduces memory allocation,
+        by lazily allocating 0 or one filled tensors.
+    """
+    return _ffi_api.LazyGradientInit()
 
 def FoldConstant():
     """Fold the constant expressions in a Relay program.

src/ir/module.cc

@@ -96,6 +96,18 @@ GlobalTypeVar IRModuleNode::GetGlobalTypeVar(const std::string& name) const {
   return (*it).second;
 }
 
+Constructor IRModuleNode::GetConstructor(const std::string& adt, const std::string& cons) const {
+  TypeData typeDef = this->LookupTypeDef(adt);
+  for (Constructor c : typeDef->constructors) {
+    if (cons.compare(c->name_hint) == 0) {
+      return c;
+    }
+  }
+
+  LOG(FATAL) << adt << " does not contain constructor " << cons;
+  throw std::runtime_error("Constructor Not Found.");
+}
+
 tvm::Array<GlobalTypeVar> IRModuleNode::GetGlobalTypeVars() const {
   std::vector<GlobalTypeVar> global_type_vars;
   for (const auto& pair : global_type_var_map_) {