Shape and type deduction (apache#7)

* Shape and type deduction.

* Fix header.

* Add call attrs to the deduce signature.

* Address comments.

* Add DiagnosticContext to IRBuilder and inference signature.

* Fix nits.

Author: YuchenJin

include/tvm/relax/ir_builder.h

@@ -82,7 +82,7 @@ class IRBuilderNode : public Object {
   /*!
    * \brief Generate an output for the current dataflow block or function.
    * \param output The output variable of the block/function.
-   * \return The variable being binded to \p ouput.
+   * \return The variable being binded to \p output.
    */
   Var EmitOutput(const Expr& output);
   /*!
@@ -107,13 +107,15 @@ class IRBuilderNode : public Object {
 
  private:
   /*! \brief The state of the function currently being built. */
-  RelaxFunction func;
+  RelaxFunction func_;
   /*! \brief A flag tracking if currently inside a dataflow block or not. */
-  bool is_dataflow = false;
+  bool is_dataflow_ = false;
   /*! \brief A global variable counter for naming global variables. */
-  int global_var_counter = 0;
+  int global_var_counter_ = 0;
   /*! \brief A dataflow variable counter for naming dataflow variables. */
-  int dataflow_var_counter = 0;
+  int dataflow_var_counter_ = 0;
+  /*! \brief A diagnostic context for reporting errors. */
+  DiagnosticContext diag_ctx_ = DiagnosticContext::Default(IRModule({}, {}));
 };
 
 class IRBuilder : public ObjectRef {

include/tvm/relax/op_attr_types.h

@@ -0,0 +1,59 @@
+/*
+ * 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/relax/op_attr_types.h
+ * \brief Data structures that can appear in operator attributes.
+ */
+#ifndef TVM_RELAX_OP_ATTR_TYPES_H_
+#define TVM_RELAX_OP_ATTR_TYPES_H_
+
+#include <tvm/relay/expr.h>
+#include <tvm/relay/type.h>
+#include <tvm/te/schedule.h>
+#include <tvm/te/tensor.h>
+
+#include <string>
+
+namespace tvm {
+namespace relax {
+
+using relay::Call;
+
+/*!
+ * \brief Infer the output shape for operators. This function will
+ * be invoked to fill the \p shape_ field of expressions.
+ * \param call The call node.
+ * \param diag_ctx The diagnostic context for reporting errors.
+ * \return The inferred output shape expression.
+ */
+using FInferShape = runtime::TypedPackedFunc<Optional<RelayExpr>(const Call& call, DiagnosticContext diag_ctx)>;
+
+/*!
+ * \brief Infer the output type for operators. This function will
+ * be invoked to fill the \p checked_type_ field of expressions.
+ * \param call The call node.
+ * \param diag_ctx The diagnostic context for reporting errors.
+ * \return The inferred output type.
+ */
+using FInferType = runtime::TypedPackedFunc<Type(const Call& call, DiagnosticContext diag_ctx)>;
+
+}  // namespace relax
+}  // namespace tvm
+#endif  // TVM_RELAX_OP_ATTR_TYPES_H_

python/tvm/ir/expr.py

@@ -45,7 +45,7 @@ def checked_type(self):
         checked_type : tvm.relay.Type
             The checked type.
         """
-        ret = self._checked_type_
+        ret = self.checked_type_
         if ret is None:
             raise ValueError("The type checker has not populated" " the checked_type for this node")
         return ret

python/tvm/relax/expr.py

@@ -66,6 +66,8 @@ def __init__(
         type_annotation: Optional[Type] = None,
         span: Span = None,
     ) -> None:
+        if shape_annotation is not None:
+            shape_annotation = make_shape(shape_annotation)
         self.__init_handle_by_constructor__(
             _ffi_api.Var, name_hint, shape_annotation, type_annotation, span
         )
@@ -86,6 +88,8 @@ def __init__(
         type_annotation: Optional[Type] = None,
         span: Span = None,
     ) -> None:
+        if shape_annotation is not None:
+            shape_annotation = make_shape(shape_annotation)
         self.__init_handle_by_constructor__(
             _ffi_api.DataflowVar, name_hint, shape_annotation, type_annotation, span
         )

python/tvm/relax/op/tensor.py

@@ -1,3 +1,19 @@
+# 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
+"""Basic tensor operations."""
 from . import _ffi_api
 from ..expr import Expr
 

src/relax/ir_builder.cc

@@ -22,6 +22,7 @@
  */
 
 #include <tvm/relax/ir_builder.h>
+#include <tvm/relax/op_attr_types.h>
 #include <tvm/relay/op.h>
 
 namespace tvm {
@@ -38,59 +39,84 @@ IRBuilder IRBuilderNode::Create() {
 
 void IRBuilderNode::FillFuncNameParam(const Array<Var>& params, const std::string& func_name) {
   if (!func_name.empty()) {
-    this->func.func_name = GlobalVar(func_name);
+    this->func_.func_name = GlobalVar(func_name);
   }
-  
-  this->func.params = params;
+
+  this->func_.params = params;
 }
 
 void IRBuilderNode::BuildFunction() {
-  SeqExpr seq = SeqExpr(this->func.binding_blocks, this->func.ret);
-  this->func.func = Function(this->func.func_name, this->func.params, seq, {});
-  this->global_var_counter = 0;
+  SeqExpr seq = SeqExpr(this->func_.binding_blocks, this->func_.ret);
+  this->func_.func = Function(this->func_.func_name, this->func_.params, seq, {});
+  this->global_var_counter_ = 0;
 }
 
 void IRBuilderNode::BuildBlock() {
-  if (!this->func.bindings.empty()) {
-    if (is_dataflow) {
-      this->func.binding_blocks.emplace_back(DataflowBlock(this->func.bindings));
+  if (!this->func_.bindings.empty()) {
+    if (is_dataflow_) {
+      this->func_.binding_blocks.emplace_back(DataflowBlock(this->func_.bindings));
     } else {
-      this->func.binding_blocks.emplace_back(BindingBlock(this->func.bindings));
+      this->func_.binding_blocks.emplace_back(BindingBlock(this->func_.bindings));
     }
-    this->func.bindings.clear();
+    this->func_.bindings.clear();
   }
-  this->dataflow_var_counter = 0;
-  this->is_dataflow = !this->is_dataflow;
+  this->dataflow_var_counter_ = 0;
+  this->is_dataflow_ = !this->is_dataflow_;
+}
+
+Optional<RelayExpr> InferShape(const Call& call, DiagnosticContext diag_ctx) {
+  auto op_map = Op::GetAttrMap<FInferShape>("FInferShape");
+  Op op = Downcast<Op>(call->op);
+  return op_map[op](call, diag_ctx);
+}
+
+Type InferType(const Call& call, DiagnosticContext diag_ctx) {
+  auto op_map = Op::GetAttrMap<FInferType>("FInferType");
+  Op op = Downcast<Op>(call->op);
+  return op_map[op](call, diag_ctx);
 }
 
 Var IRBuilderNode::Emit(const Call& call) {
   Var var;
-  if (is_dataflow) {
-    var = DataflowVar(Id("lv" + std::to_string(dataflow_var_counter++)), NullOpt, NullOpt);
+  if (is_dataflow_) {
+    var = DataflowVar(Id("lv" + std::to_string(dataflow_var_counter_++)), NullOpt, NullOpt);
   } else {
-    var = Var(Id("gv" + std::to_string(global_var_counter++)), NullOpt, NullOpt);
+    var = Var(Id("gv" + std::to_string(global_var_counter_++)), NullOpt, NullOpt);
+  }
+
+  // Shape inference
+  auto inferred_shape = InferShape(call, this->diag_ctx_);
+  if (inferred_shape.defined()) {
+    if (auto* shape_expr = inferred_shape.value().as<ShapeExprNode>()) {
+      call->shape_ = GetRef<Expr>(shape_expr);
+      var->shape_ = call->shape_;
+    }
   }
+  // Type inference
+  auto inferred_type = InferType(call, this->diag_ctx_);
+  call->checked_type_ = inferred_type;
+  var->checked_type_ = inferred_type;
 
-  this->func.bindings.emplace_back(VarBinding(var, call));
+  this->func_.bindings.emplace_back(VarBinding(var, call));
   return var;
 }
 
 Var IRBuilderNode::EmitOutput(const Expr& output) {
   Var ret;
-  if (is_dataflow) {
-    ret = Var(Id("gv" + std::to_string(global_var_counter++)), NullOpt, NullOpt);
+  if (is_dataflow_) {
+    ret = Var(Id("gv" + std::to_string(global_var_counter_++)), NullOpt, NullOpt);
     ret->shape_ = output->shape_;
     ret->checked_type_ = output->checked_type_;
-    this->func.bindings.emplace_back(VarBinding(ret, output));
+    this->func_.bindings.emplace_back(VarBinding(ret, output));
   } else {
-    this->func.ret = output;
+    this->func_.ret = output;
   }
   return ret;
 }
 
-Function IRBuilderNode::Get() { return this->func.func; }
+Function IRBuilderNode::Get() { return this->func_.func; }
 
-std::vector<BindingBlock> IRBuilderNode::GetBlocks() { return this->func.binding_blocks; }
+std::vector<BindingBlock> IRBuilderNode::GetBlocks() { return this->func_.binding_blocks; }
 
 class FunctionScope::Internal {
  public:
@@ -121,20 +147,16 @@ DataflowScope::DataflowScope(IRBuilder ib) {
   data_ = std::move(n);
 }
 
-void DataflowScope::EnterWithScope() {
-  this->get()->ir_builder->BuildBlock();
-}
+void DataflowScope::EnterWithScope() { this->get()->ir_builder->BuildBlock(); }
 
-void DataflowScope::ExitWithScope() {
-  this->get()->ir_builder->BuildBlock();
-}
+void DataflowScope::ExitWithScope() { this->get()->ir_builder->BuildBlock(); }
 
 TVM_REGISTER_GLOBAL("relax.IRBuilderCreate").set_body_typed(IRBuilderNode::Create);
 
 TVM_REGISTER_GLOBAL("relax.IRBuilderFillFuncNameParam")
-.set_body_typed([](IRBuilder builder, const Array<Var>& params, const std::string& func_name) {
-  return builder->FillFuncNameParam(params, func_name);
-});
+    .set_body_typed([](IRBuilder builder, const Array<Var>& params, const std::string& func_name) {
+      return builder->FillFuncNameParam(params, func_name);
+    });
 
 TVM_REGISTER_GLOBAL("relax.IRBuilderBuildFunction").set_body_typed([](IRBuilder builder) {
   return builder->BuildFunction();
@@ -145,9 +167,9 @@ TVM_REGISTER_GLOBAL("relax.IRBuilderEmit").set_body_typed([](IRBuilder builder,
 });
 
 TVM_REGISTER_GLOBAL("relax.IRBuilderEmitOutput")
-.set_body_typed([](IRBuilder builder, const Expr& output) {
-  return builder->EmitOutput(output);
-});
+    .set_body_typed([](IRBuilder builder, const Expr& output) {
+      return builder->EmitOutput(output);
+    });
 
 TVM_REGISTER_GLOBAL("relax.IRBuilderGet").set_body_typed([](IRBuilder builder) {
   return builder->Get();

src/relax/op/op_common.h

@@ -25,9 +25,9 @@
 #ifndef TVM_RELAX_OP_OP_COMMON_H_
 #define TVM_RELAX_OP_OP_COMMON_H_
 
+#include <tvm/relax/op_attr_types.h>
 #include <tvm/relay/expr.h>
 #include <tvm/relay/op.h>
-#include <tvm/relay/op_attr_types.h>
 
 namespace tvm {
 namespace relax {
@@ -42,15 +42,17 @@ namespace relax {
  *
  * \param OpName the name of registry.
  */
-#define RELAX_REGISTER_BINARY_OP(OpName)                                          \
+#define RELAX_REGISTER_BINARY_BROADCAST_OP(OpName)                                          \
   TVM_REGISTER_GLOBAL("relax.op." OpName).set_body_typed([](Expr lhs, Expr rhs) { \
-    static const Op& op = Op::Get(OpName);                                        \
+    static const Op& op = Op::Get("relax." OpName);                               \
     return Call(op, {lhs, rhs}, Attrs(), {});                                     \
   });                                                                             \
   RELAY_REGISTER_OP("relax." OpName)                                              \
       .set_num_inputs(2)                                                          \
       .add_argument("lhs", "Tensor", "The left hand side tensor.")                \
-      .add_argument("rhs", "Tensor", "The right hand side tensor.")
+      .add_argument("rhs", "Tensor", "The right hand side tensor.")               \
+      .set_attr<FInferShape>("FInferShape", InferShapeBinaryBroadcast)            \
+      .set_attr<FInferType>("FInferType", InferTypeBinaryBroadcast)
 
 }  // namespace relax
 }  // namespace tvm

src/relax/op/tensor/binary.cc

@@ -22,32 +22,18 @@
  * \brief binary broadcast operators.
  */
 
-#include <tvm/arith/analyzer.h>
-#include <tvm/relax/expr.h>
-#include <tvm/relax/type.h>
-#include <tvm/tir/op.h>
-#include <tvm/topi/broadcast.h>
+#include "binary.h"
 
 #include "../op_common.h"
 
 namespace tvm {
 namespace relax {
 
-using Expr = tvm::RelayExpr;
-using relay::Call;
-
-#define RELAX_BINARY_COMPUTE(FTOPI)                       \
-  [](const Attrs& attrs, const Array<te::Tensor>& inputs, \
-     const Type& out_type) -> Array<te::Tensor> {         \
-    ICHECK_EQ(inputs.size(), 2U);                         \
-    return {FTOPI(inputs[0], inputs[1])};                 \
-  }
-
-RELAX_REGISTER_BINARY_OP("add")
+RELAX_REGISTER_BINARY_BROADCAST_OP("add")
     .describe("Elementwise add with broadcasting")
     .set_support_level(1);
 
-RELAX_REGISTER_BINARY_OP("multiply")
+RELAX_REGISTER_BINARY_BROADCAST_OP("multiply")
     .describe("Elementwise multiply with broadcasting")
     .set_support_level(1);