[Relay][AutoTVM] Relay op strategy (apache#4644)

* relay op strategy

fix lint

bitpack strategy

bitserial_dense (#6)

* update strategy

* address comments

fix a few topi test

Dense strategy (#5)

* dense

* add biforst; remove comments

* address comment

Refactor x86 conv2d_NCHWc (#4)

* Refactor x86 conv2d

* Add x86 depthwise_conv2d_NCHWc

* Add back topi x86 conv2d_nchw

* Merge x86 conv2d_nchw and conv2d_NCHWc

* Minor fix for x86 conv2d

fix more strategy

Add x86 conv2d_NCHWc_int8 strategy (#8)

* Add x86 conv2d_NCHWc_int8 strategy

* Remove contrib_conv2d_nchwc_int8

* Fix generic conv2d_NCHWc for int8

* Fix topi arm_cpu conv2d_NCHWc_int8

update x86 conv2d

enable specify relay ops to be tuned for autotvm

add cuda conv2d strategy

add conv2d strategy for rocm

add conv2d strategy for hls

add conv2d strategy for arm cpu

add conv2d strategy for mali

add conv2d strategy for bifrost

add conv2d strategy for intel graphics

clean up and fix lint

remove template keys from autotvm

remove 2 in the func name

address comments

fix

* fix bugs

* lint

* address comments

* add name to op implement

* Modify topi tests (#9)

* Add pooling, reorg, softmax and vision

* Add lrn

* fix topi test

* fix more topi test

* lint

* address comments

* x

* fix more tests & bugs

* Modify more tests (#10)

* Modify tests for bitserial_conv2d, bitserial_dense, bitserial_conv2d_rasp and bnn

* Minor fix

* More minor fix

* fix more test

* try to update vta using strategy

* fix cpptest

* x

* fix rebase err

* Fix two tests (#11)

* change autotvm log format

* lint

* minor fix

* try fix vta test

* fix rebase err

* tweak

* tmp hack for vta pass

* fix tutorial

* fix

* fix more tutorials

* fix vta tutorial

* minor

* address comments

* fix

* address comments

* fix cpptest

* fix docs

* change data structure name and api

* address comments

* lint

* fix rebase err

* updates

* fix winograd test

* fix doc

* rebase

* upgrade tophub version number

* fix bug

* re-enable vta tsim test after tophub is upgraded

* fix vta test to use the correct args so the config can be found in tophub

Co-authored-by: Yao Wang <kevinthesunwy@gmail.com>

Author: 2 people

include/tvm/relay/op_attr_types.h

@@ -29,6 +29,7 @@
 #include <tvm/relay/type.h>
 #include <tvm/relay/expr.h>
 #include <tvm/target/target.h>
+#include <tvm/target/generic_func.h>
 #include <tvm/tir/data_layout.h>
 #include <string>
 
@@ -105,9 +106,8 @@ using TShapeDataDependant = bool;
  */
 using FTVMCompute = runtime::TypedPackedFunc<
   Array<te::Tensor>(const Attrs& attrs,
-                     const Array<te::Tensor>& inputs,
-                     const Type& out_type,
-                     const Target& target)>;
+                    const Array<te::Tensor>& inputs,
+                    const Type& out_type)>;
 
 /*!
  * \brief Build the computation schedule for
@@ -120,8 +120,18 @@ using FTVMCompute = runtime::TypedPackedFunc<
  */
 using FTVMSchedule = runtime::TypedPackedFunc<
   te::Schedule(const Attrs& attrs,
-                const Array<te::Tensor>& outs,
-                const Target& target)>;
+               const Array<te::Tensor>& outs,
+               const Target& target)>;
+
+/*!
+ * \brief Generate the strategy of operators. This function is a generic
+ * function and can be re-defined for different targets.
+ *
+ * The function signature of generic function is:
+ *   OpStrategy(const Attrs& attrs, const Array<Tensor>& inputs,
+ *              const Type& out_type, const Target& target)
+ */
+using FTVMStrategy = GenericFunc;
 
 /*!
  * \brief Alternate the layout of operators or replace the
@@ -136,7 +146,8 @@ using FTVMSchedule = runtime::TypedPackedFunc<
 using FTVMAlterOpLayout = runtime::TypedPackedFunc<
   Expr(const Attrs& attrs,
        const Array<Expr>& args,
-       const Array<te::Tensor>& tinfos)>;
+       const Array<te::Tensor>& tinfos,
+       const Type& out_type)>;
 
 /*!
  * \brief Convert the layout of operators or replace the
@@ -191,9 +202,7 @@ using FForwardRewrite = runtime::TypedPackedFunc<
  * \brief Gradient for a specific op.
  *
  * \param orig_call the original Expr.
- *
  * \param output_grad the gradient of the Expr.
- *
  * \return the gradient for each parameters.
  */
 using FPrimalGradient = runtime::TypedPackedFunc<tvm::Array<Expr>(const Expr& orig_call,
@@ -207,13 +216,13 @@ enum AnyCodegenStrategy {
   kVariableDimensions
 };
 
-/* \brief A runtime representation of shape. */
+/*! \brief A runtime representation of shape. */
 using Shape = Array<IndexExpr>;
 
 using FShapeFunc = runtime::TypedPackedFunc<
   Array<te::Tensor>(const Attrs& attrs,
-                     const Array<te::Tensor>& inputs,
-                     const Array<IndexExpr>& out_ndims)>;
+                    const Array<te::Tensor>& inputs,
+                    const Array<IndexExpr>& out_ndims)>;
 
 }  // namespace relay
 }  // namespace tvm

include/tvm/relay/op_strategy.h

@@ -0,0 +1,164 @@
+/*
+ * 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.
+ */
+
+/*!
+ * \file tvm/relay/op_strategy.h
+ * \brief The Relay operator Strategy and related data structure.
+ */
+
+#ifndef TVM_RELAY_OP_STRATEGY_H_
+#define TVM_RELAY_OP_STRATEGY_H_
+
+#include <tvm/te/tensor.h>
+#include <tvm/te/schedule.h>
+#include <tvm/relay/expr.h>
+#include <tvm/relay/op_attr_types.h>
+#include <tvm/target/target.h>
+#include <string>
+
+namespace tvm {
+namespace relay {
+
+/*!
+ * \brief Operator implementation that includes compute and schedule function.
+ */
+class OpImplementationNode : public Object {
+ public:
+  /*! \brief Compute function */
+  FTVMCompute fcompute;
+  /*! \brief Schedule function */
+  FTVMSchedule fschedule;
+  /*! \brief Name of the implementation */
+  std::string name;
+  /*! \brief Priority level */
+  int plevel;
+
+  void VisitAttrs(tvm::AttrVisitor* v) {
+    v->Visit("name", &name);
+    v->Visit("plevel", &plevel);
+  }
+
+  static constexpr const char* _type_key = "relay.OpImplementation";
+  TVM_DECLARE_FINAL_OBJECT_INFO(OpImplementationNode, Object);
+};
+
+/*!
+ * \brief Operator implementation class.
+ */
+class OpImplementation : public ObjectRef {
+ public:
+  /*!
+   * \brief Invoke the operator compute function.
+   * \param attrs The attribute of the primitive
+   * \param inputs The input tensors.
+   * \param out_type The output type information.
+   * \return The output compute description of the operator.
+   */
+  TVM_DLL Array<te::Tensor> Compute(const Attrs& attrs,
+                                    const Array<te::Tensor>& inputs,
+                                    const Type& out_type);
+  /*!
+   * \brief Build the computation schedule.
+   * \param attrs The attribute of the node.
+   * \param outs The output tensors.
+   * \param target The build target.
+   * \return The computation schedule.
+   */
+  TVM_DLL te::Schedule Schedule(const Attrs& attrs,
+                                const Array<te::Tensor>& outs,
+                                const Target& target);
+
+  TVM_DEFINE_OBJECT_REF_METHODS(OpImplementation, ObjectRef, OpImplementationNode);
+};
+
+/*!
+ * \brief Specialized implementations for operators under certain conditions.
+ */
+class OpSpecializationNode : public Object {
+ public:
+  /*! \brief List of implementations. */
+  Array<OpImplementation> implementations;
+  /*! \brief Condition to enable the specialization.
+   *    Could be undefined to represent generic case. */
+  te::SpecializedCondition condition;
+
+  void VisitAttrs(tvm::AttrVisitor* v) {
+    v->Visit("condition", &condition);
+    v->Visit("implementations", &implementations);
+  }
+
+  static constexpr const char* _type_key = "relay.OpSpecialization";
+  TVM_DECLARE_FINAL_OBJECT_INFO(OpSpecializationNode, ExprNode);
+};
+
+/*!
+ * \brief Operator specialization class.
+ */
+class OpSpecialization : public ObjectRef {
+ public:
+  /*!
+   * \brief Add an implementation.
+   * \param fcompute Compute function
+   * \param fschedule Schedule function
+   * \param name Name of the implementation
+   * \param plevel Priority level of the implementation
+   */
+  TVM_DLL void AddImplementation(FTVMCompute fcompute, FTVMSchedule fschedule,
+                                 std::string name, int plevel);
+
+  TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(OpSpecialization, ObjectRef, OpSpecializationNode);
+};
+
+/*!
+ * \brief Operator strategy to choose implementation.
+ */
+class OpStrategyNode : public Object {
+ public:
+  /*! \brief List of operator specializations. */
+  Array<OpSpecialization> specializations;
+
+  void VisitAttrs(tvm::AttrVisitor* v) {
+    v->Visit("specializations", &specializations);
+  }
+
+  static constexpr const char* _type_key = "relay.OpStrategy";
+  TVM_DECLARE_FINAL_OBJECT_INFO(OpStrategyNode, ExprNode);
+};
+
+/*!
+ * \brief Operator strategy class.
+ */
+class OpStrategy : public ObjectRef {
+ public:
+  /*!
+   * \brief Add an implementation.
+   * \param fcompute Compute function
+   * \param fschedule Schedule function
+   * \param name Name of the implementation
+   * \param plevel Priority level of the implementation
+   */
+  TVM_DLL void AddImplementation(FTVMCompute fcompute, FTVMSchedule fschedule,
+                                 std::string name, int plevel);
+
+  TVM_DEFINE_MUTABLE_OBJECT_REF_METHODS(OpStrategy, ObjectRef, OpStrategyNode);
+};
+
+}  // namespace relay
+}  // namespace tvm
+#endif  // TVM_RELAY_OP_STRATEGY_H_

include/tvm/te/schedule.h

@@ -28,6 +28,7 @@
 #include <tvm/tir/expr.h>
 #include <tvm/te/tensor.h>
 #include <tvm/te/tensor_intrin.h>
+#include <tvm/support/with.h>
 
 #include <string>
 #include <unordered_map>
@@ -742,6 +743,53 @@ class SingletonNode : public IterVarRelationNode {
   TVM_DECLARE_FINAL_OBJECT_INFO(SingletonNode, IterVarRelationNode);
 };
 
+/*! \brief Container for specialization conditions. */
+class SpecializedConditionNode : public Object {
+ public:
+  /*!
+   * \brief List of conditions in conjunctive joint form (CNF).
+   *   Each condition should be a simple expression, e.g., n > 16, m % 8 == 0, etc.,
+   *   where n, m are tvm::Var that represents a dimension in the tensor shape.
+   */
+  Array<PrimExpr> clauses;
+
+  void VisitAttrs(AttrVisitor* v) {
+    v->Visit("clauses", &clauses);
+  }
+
+  static constexpr const char* _type_key = "SpecializedCondition";
+  TVM_DECLARE_FINAL_OBJECT_INFO(SpecializedConditionNode, Object);
+};
+
+/*!
+ * \brief Specialized condition to enable op specialization
+ */
+class SpecializedCondition : public ObjectRef {
+ public:
+  /*!
+   * \brief construct from conditions
+   * \param conditions The clauses in the specialized condition.
+   */
+  TVM_DLL SpecializedCondition(Array<PrimExpr> conditions);  // NOLINT(*)
+
+  /*!
+   * \brief Get the current specialized condition.
+   * \return the current specialized condition.
+   */
+  TVM_DLL static SpecializedCondition Current();
+
+  TVM_DEFINE_OBJECT_REF_METHODS(SpecializedCondition, ObjectRef, SpecializedConditionNode);
+  class Internal;
+
+ private:
+  // enable with syntax.
+  friend class Internal;
+  friend class With<SpecializedCondition>;
+  /*! \brief Push a new specialized condition onto the thread local stack. */
+  TVM_DLL void EnterWithScope();
+  /*! \brief Pop a specialized condition off the thread local context stack. */
+  TVM_DLL void ExitWithScope();
+};
 
 // implementations
 inline const StageNode* Stage::operator->() const {
@@ -765,6 +813,7 @@ inline const IterVarRelationNode* IterVarRelation::operator->() const {
 inline const IterVarAttrNode* IterVarAttr::operator->() const {
   return static_cast<const IterVarAttrNode*>(get());
 }
+
 }  // namespace te
 }  // namespace tvm
 #endif  // TVM_TE_SCHEDULE_H_

python/tvm/autotvm/__init__.py

@@ -41,8 +41,8 @@
 from .measure import measure_option, MeasureInput, MeasureResult, MeasureErrorNo, \
     LocalBuilder, LocalRunner, RPCRunner
 from .tuner import callback
-from .task import template, get_config, create, ConfigSpace, ConfigEntity, \
-    register_topi_compute, register_topi_schedule, \
+from .task import get_config, create, ConfigSpace, ConfigEntity, \
+    register_topi_compute, register_topi_schedule, register_customized_task, \
     DispatchContext, FallbackContext, ApplyHistoryBest as apply_history_best, \
     ApplyGraphBest as apply_graph_best
 from .env import GLOBAL_SCOPE

python/tvm/autotvm/database.py

@@ -125,7 +125,7 @@ def load(self, inp, get_all=False):
         current = self.get(measure_str_key(inp))
         if current is not None:
             records = [decode(x) for x in current.split(RedisDatabase.MAGIC_SPLIT)]
-            results = [rec[1] for rec in records]
+            results = [rec[1] for rec in records if rec is not None]
             if get_all:
                 return results
             return max(results, key=lambda result: result.timestamp)
@@ -167,9 +167,12 @@ def filter(self, func):
             current = self.get(key)
             try:
                 records = [decode(x) for x in current.split(RedisDatabase.MAGIC_SPLIT)]
+                records = [rec for rec in records if rec is not None]
             except TypeError: # got a badly formatted/old format record
                 continue
 
+            if not records:
+                continue
             inps, results = zip(*records)
             inp = inps[0]
             if not func(inp, results):

python/tvm/autotvm/feature.py

@@ -153,7 +153,10 @@ def get_flatten_name(fea):
         from .record import decode
         # flatten line to feature
         line = fea
-        inp, _ = decode(line)
+        ret = decode(line)
+        if ret is None:
+            raise ValueError("Unsupported AutoTVM log format")
+        inp, _ = ret
         target = _target.create(inp.target)
         with target:
             s, args = inp.template.instantiate(inp.config)