[GitHub] [incubator-tvm] comaniac commented on a change in pull request #5078: [DOC] Add doc for Relay op strategy

[GitBox <gi...@apache.org>]

comaniac commented on a change in pull request #5078: [DOC] Add doc for Relay op strategy
URL: https://github.com/apache/incubator-tvm/pull/5078#discussion_r393161533
 
 

 ##########
 File path: docs/dev/relay_op_strategy.rst
 ##########
 @@ -0,0 +1,256 @@
+..  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.
+
+.. _relay-op-strategy:
+
+Relay Operator Strategy
+=======================
+
+In order to lower Relay operators to implementation defined in TOPI library, the
+compute and schedule functions need to be registered to Relay operators.
+However, compute and schedule functions are usually specialized for each target,
+and further, even for the same target, we may have multiple algorithms and
+implementations available. To deal with the complexity, we introduce operator
+strategy to allow developers to define a flexible lowering strategy for each
+operator and target.
+
+
+Operator Strategy Design
+------------------------
+
+The basic element in operator strategy is an ``OpImplementation``. It includes
+the a pair of compute and schedule function, the name of the implementation,
+and a priority level (the usability of priority level will be explained below).
+
+The ``OpStrategy`` includes a list of specializations. Each specialization
+contains a list of ``OpImplementation`` associated with a specialized condition
+(see ``SpecializedCondition`` definition in ``include/tvm/te/schedule.h``).  The
+specialized condition can be null, indicating the implementations are generally
+applicable; otherwise, the implementations should only be used when the
+specialized condition is satisfied. ``OpStrategy`` provides only one API,
+adding an implementation to the strategy:
+
+.. code:: python
+
+    def add_implementation(self, compute, schedule, name="default", plevel=10)
+
+Last, a ``FTVMStrategy`` function is registered to each Relay operator.
+``FTVMStrategy`` is a generic function (see ``include/tvm/target/generic_func.h``),
+that can be overwritten for each target. The function signature is
+
+.. code:: c
+
+    OpStrategy(const Attrs& attrs, const Array<Tensor>& inputs, const Type& out_type, const Target& target)
+
+, that the function returns an ``OpStrategy`` given the op attributes, input
+tensors, output types, and target to compile to,
+
+
+
+Register strategy for a new operator
+------------------------------------
+
+There are three methods to register a strategy function for an operator,
+defined in ``python/tvm/relay/op/op.py``.
+
+First, for operators that have injective, broadcast, or reduction pattern, we
+can call ``register_injective_schedule``, ``register_broadcast_schedule``, and
+``register_reduce_schedule`` repsectively. The schedule function for these
+patterns are already registered by each target and can be applied to these
+operators. We assume the compute function should be same across all targets, and
+``FTVMCompute`` needs to be registered to the op before invoking register
+schedule.
+
+.. code:: python
+
+    register_injective_schedule("my_new_op")
+
+Second, for operators that doesn't have these common patterns mentioned before,
+but also have the same compute function for all targets, we can use
+``register_schedule`` API. But before that, we need to first define the
+``FTVMSchedule`` function as follows:
+
+.. code:: python
+
+    # add to python/tvm/relay/op/strategy/generic.py
+    @generic_func
+    def schedule_my_new_op(attrs, outs, target):
+        ...
+
+    # add to each target file in python/tvm/relay/op/strategy, e.g., x86.py, cuda.py, etc.
+    @schedule_my_new_op.register("cpu")
+    def schedule_my_new_op_cpu(attrs, outs, target):
+        ...
+
+Now that we've created the ``FTVMSchedule`` for this new operator, we can
+register the strategy using ``register_schedule``:
+
+.. code:: python
+
+    register_schedule("my_new_op", strategy.schedule_my_new_op)
+
+Third, for most comprehensive usage of op strategy, we can allow operator to use
+different implementation for both compute and schedule for different targets.
 
 Review comment:
   s/different implementation/different implementations

----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
 
For queries about this service, please contact Infrastructure at:
users@infra.apache.org


With regards,
Apache Git Services