Merge pull request #20188 from pschuh/3

Break TensorGroup into InputTensorGroup and OutputTensorGroup.

Author: pschuh

include/swift/AST/KnownProtocols.def

@@ -73,7 +73,8 @@ PROTOCOL(Numeric)
 PROTOCOL(FloatingPoint)
 PROTOCOL(ParameterGroup)
 PROTOCOL(Parameterized)
-PROTOCOL(TensorGroup)
+PROTOCOL(InputTensorGroup)
+PROTOCOL(OutputTensorGroup)
 PROTOCOL(TensorProtocol)
 PROTOCOL(TensorSendableReceivable)
 PROTOCOL(VectorNumeric)

lib/AST/ASTContext.cpp

@@ -972,7 +972,8 @@ ProtocolDecl *ASTContext::getProtocol(KnownProtocolKind kind) const {
   case KnownProtocolKind::AccelerableByTensorFlow:
   case KnownProtocolKind::ParameterGroup:
   case KnownProtocolKind::Parameterized:
-  case KnownProtocolKind::TensorGroup:
+  case KnownProtocolKind::InputTensorGroup:
+  case KnownProtocolKind::OutputTensorGroup:
   case KnownProtocolKind::TensorSendableReceivable:
   case KnownProtocolKind::TensorProtocol:
     M = getLoadedModule(Id_TensorFlow);

lib/IRGen/GenMeta.cpp

@@ -4055,7 +4055,8 @@ SpecialProtocol irgen::getSpecialProtocolID(ProtocolDecl *P) {
   case KnownProtocolKind::Numeric:
   case KnownProtocolKind::ParameterGroup:
   case KnownProtocolKind::Parameterized:
-  case KnownProtocolKind::TensorGroup:
+  case KnownProtocolKind::InputTensorGroup:
+  case KnownProtocolKind::OutputTensorGroup:
   case KnownProtocolKind::TensorProtocol:
   case KnownProtocolKind::TensorSendableReceivable:
   case KnownProtocolKind::VectorNumeric:

lib/IRGen/IRGenSIL.cpp

@@ -1979,9 +1979,12 @@ void IRGenSILFunction::visitGraphOperationInst(GraphOperationInst *i) {
   // TODO: As an optimization, do this lookup once per CurSILFn
   auto tfModule = astCtx.getLoadedModule(astCtx.Id_TensorFlow);
   assert(tfModule && "could not find TensorFlow module");
-  auto tensorGroupProto =
-      astCtx.getProtocol(KnownProtocolKind::TensorGroup);
-  assert(tensorGroupProto && "could not find TensorGroup protocol");
+  auto inputTensorGroupProto =
+      astCtx.getProtocol(KnownProtocolKind::InputTensorGroup);
+  auto outputTensorGroupProto =
+      astCtx.getProtocol(KnownProtocolKind::OutputTensorGroup);
+  assert(inputTensorGroupProto && "could not find InputTensorGroup protocol");
+  assert(outputTensorGroupProto && "could not find OutputTensorGroup protocol");
 
   if (!llvm::TFDynamicCompilation) {
     // If we are not in dynamic compilation mode, then deabstraction may not
@@ -2051,25 +2054,18 @@ void IRGenSILFunction::visitGraphOperationInst(GraphOperationInst *i) {
   // returns an Int32 value for the number of inputs that it has added. There
   // are a few different cases that can be unpacked:
   // - if `opInput` is a TensorFlow value, then we just add its handle;
-  // - if `opInput` is an archetype conforming to TensorGroup, then we
+  // - if `opInput` is an archetype conforming to InputTensorGroup, then we
   //   ask the conformance for the handles and add those;
   // This function crashes if it receives an unhandled case. Earlier
   // typechecking should ensure that inputs match the cases that this function
   // handles.
-  //
-  // TODO: We should also handle the following cases:
-  // - if `opInput` is an array of TensorFlow values, then we add all the
-  //   elements' handles;
-  // - if `opInput` is an array of archetypes conforming to TensorGroup,
-  //   then we ask the conformance for all the elements' handles, flatten the
-  //   results together, and add those;
   auto unpackAndAddInput = [&](SILValue opInput) -> llvm::Value* {
     LLVM_DEBUG(llvm::dbgs()
                << "   Adding input of type " << opInput->getType() << ".\n");
 
     // If this is a known TensorFlow value, add it directly.
     // TODO: We could also handle concrete structs of known TensorFlow values
-    // here, to avoid falling through to the slower TensorGroup case.
+    // here, to avoid falling through to the slower InputTensorGroup case.
     if (tf::isTensorFlowValue(opInput->getType())) {
       auto *tensorHandleValue = getLoweredSingletonExplosion(opInput);
       auto *opAddInputFromTensorHandleFn =
@@ -2082,13 +2078,13 @@ void IRGenSILFunction::visitGraphOperationInst(GraphOperationInst *i) {
       return llvm::ConstantInt::get(IGM.Int32Ty, 1);
     }
 
-    // Otherwise, this must conform to TensorGroup so we can add it using
+    // Otherwise, this must conform to InputTensorGroup so we can add it using
     // TFC_OpAddInputFromTensorGroup.
 
     auto canType = opInput->getType().getASTType()->getCanonicalType();
-    auto conformance = tfModule->lookupConformance(canType,
-                                                   tensorGroupProto);
-    assert(conformance && "input type does not conform to TensorGroup");
+    auto conformance =
+        tfModule->lookupConformance(canType, inputTensorGroupProto);
+    assert(conformance && "input type does not conform to InputTensorGroup");
     auto *typeMetadata = emitTypeMetadataRef(canType);
     auto *wtable = emitWitnessTableRef(*this, canType, *conformance);
 
@@ -2228,8 +2224,8 @@ void IRGenSILFunction::visitGraphOperationInst(GraphOperationInst *i) {
       outParameterCanType =
           silValue->getType().getASTType()->getCanonicalType();
       auto conformance = tfModule->lookupConformance(outParameterCanType,
-                                                     tensorGroupProto);
-      assert(conformance && "out type does not conform to TensorGroup");
+                                                     outputTensorGroupProto);
+      assert(conformance && "out type does not conform to OutputTensorGroup");
       outParameterTypeMetadata = emitTypeMetadataRef(outParameterCanType);
       outParameterTensorGroupWitnessTable =
           emitWitnessTableRef(*this, outParameterCanType, *conformance);
@@ -2888,13 +2884,13 @@ void IRGenSILFunction::visitGraphOperationInst(GraphOperationInst *i) {
       // TensorGroup for the number of outputs that it needs.
 
       assert(hasOpaqueTensorGroupResults &&
-             "found an unexpected opaque TensorGroup result");
+             "found an unexpected opaque OutputTensorGroup result");
 
       // Emit the type metadata and witness table.
       auto canType = silResult->getType().getASTType()->getCanonicalType();
       auto conformance = tfModule->lookupConformance(canType,
-                                                     tensorGroupProto);
-      assert(conformance && "out type does not conform to TensorGroup");
+                                                     outputTensorGroupProto);
+      assert(conformance && "out type does not conform to OutputTensorGroup");
       auto *typeMetadata = emitTypeMetadataRef(canType);
       directResultTypeMetadatas.push_back(typeMetadata);
       auto *witnessTable =

lib/SILOptimizer/Mandatory/TFDeabstraction.cpp

@@ -2198,9 +2198,9 @@ static bool unpackTensorAggregates(
 
   auto tfModule = ctx.getLoadedModule(ctx.Id_TensorFlow);
   assert(tfModule && "could not find TensorFlow module");
-  auto tensorGroupProto =
-      ctx.getProtocol(KnownProtocolKind::TensorGroup);
-  assert(tensorGroupProto && "could not find TensorGroup protocol");
+  auto inputTensorGroupProto =
+      ctx.getProtocol(KnownProtocolKind::InputTensorGroup);
+  assert(inputTensorGroupProto && "could not find TensorGroup protocol");
 
   std::function<bool(SILValue)> recurse;
   recurse = [&](SILValue aggregate) -> bool {
@@ -2209,6 +2209,12 @@ static bool unpackTensorAggregates(
       inputList.push_back(aggregate);
       return false;
     }
+    if (acceptTensorGroupConformingLeaves &&
+        tfModule->lookupConformance(aggregateTy.getASTType(),
+                                    inputTensorGroupProto)) {
+      inputList.push_back(aggregate);
+      return false;
+    }
     if (auto tupleTy = aggregateTy.getAs<TupleType>()) {
       for (auto i : range(tupleTy->getNumElements())) {
         auto eltIdx = std::pair<SILValue, unsigned>(aggregate, i);
@@ -2237,12 +2243,6 @@ static bool unpackTensorAggregates(
       }
       return false;
     }
-    if (acceptTensorGroupConformingLeaves &&
-        tfModule->lookupConformance(aggregateTy.getASTType(),
-                                    tensorGroupProto)) {
-      inputList.push_back(aggregate);
-      return false;
-    }
     return true;
   };
 

stdlib/public/TensorFlow/CompilerRuntime.swift

@@ -1177,12 +1177,12 @@ func _TFCOpAddInputFromTensorHandle(_ op: CTFEOp,
 /// Adds `t` as an input or inputs to `op`. Returns the number of inputs added.
 @usableFromInline
 @_silgen_name("_swift_tfc_OpAddInputFromTensorGroup")
-func _TFCOpAddInputFromTensorGroup<T : TensorGroup>(
+func _TFCOpAddInputFromTensorGroup<T : InputTensorGroup>(
     _ op: CTFEOp, _ t: T, _ status: CTFStatus
 ) -> Int32 {
-  let count = Int(T._tensorHandleCount)
+  let count = t._inputTensorHandleCount
   let buffer =
-      UnsafeMutableBufferPointer<CTensorHandle>.allocate(capacity: count)
+      UnsafeMutableBufferPointer<CTensorHandle>.allocate(capacity: Int(count))
   defer { buffer.deallocate() }
   t._unpackTensorHandles(into: buffer.baseAddress)
   for handle in buffer {
@@ -1191,14 +1191,14 @@ func _TFCOpAddInputFromTensorGroup<T : TensorGroup>(
       return 0
     }
   }
-  return T._tensorHandleCount
+  return count
 }
 
 /// Initializes a TensorGroup value, taking ownership of all the tensor
 /// handles in `tensorHandles`.
 @usableFromInline
 @_silgen_name("_swift_tfc_InitTensorGroup")
-func _TFCInitTensorGroup<T : TensorGroup>(
+func _TFCInitTensorGroup<T : OutputTensorGroup>(
     _ tensorHandles: UnsafeMutablePointer<CTensorHandle>
 ) -> T {
   return T(_owning: tensorHandles)
@@ -1223,10 +1223,10 @@ func _TFCDeallocateCHandleBuffer(
 
 /// Returns the number of CTensorHandles in a TensorGroup of type T.
 @_silgen_name("_swift_tfc_GetTensorGroupCHandleCount")
-public func _TFCGetTensorGroupCHandleCount<T : TensorGroup>(
+public func _TFCGetTensorGroupCHandleCount<T : OutputTensorGroup>(
     _ type: T.Type
 ) -> Int32 {
-  return T._tensorHandleCount
+  return T._outputTensorHandleCount
 }
 
 @inlinable

stdlib/public/TensorFlow/TensorGroup.swift

@@ -16,51 +16,72 @@
 
 import CTensorFlow
 
-/// A protocol for types that can be used as tensor operation inputs and
-/// outputs. When a TensorGroup is used as an input, it gets passed to the
-/// tensor operation as an input list whose elements are the tensor fields of
-/// the type. When a TensorGroup is used as an output, it gets initialized
-/// with its tensor fields set to the tensor operation's tensor outputs.
+/// A protocol for types that can be used as tensor operation inputs. When a
+/// TensorGroup is used as an input, it gets passed to the tensor operation as
+/// an input list whose elements are the tensor fields of the type.
 ///
-/// TODO: Add a derived conformance to TensorGroup so that users don't have
-/// to write the conformance themselves.
-public protocol TensorGroup {
-  /// The types of the tensor stored properties in this type.
-  static var _typeList: [TensorDataType] { get }
-
+/// This protocol is divided from OutputTensorGroup in order for the number of
+/// tensors to be determined at runtime. For example, Array<Tensor<Float>> may
+/// have an unknown number of elements compile time.
+public protocol InputTensorGroup {
   /// Writes the tensor handles to `address`, which must be allocated
   /// with enough capacity to hold `_tensorHandleCount` handles. The tensor
   /// handles written to `address` are borrowed: this container still
   /// owns them.
   func _unpackTensorHandles(into address: UnsafeMutablePointer<CTensorHandle>?)
 
+  var _inputTensorHandleCount : Int32 { get }
+}
+
+/// A protocol for types that can be used as tensor operation outputs. When a
+/// TensorGroup is used as an output, it gets initialized with its tensor fields
+/// set to the tensor operation's tensor outputs.
+/// The number of tensors must be known at compile time.
+public protocol OutputTensorGroup {
+  /// The types of the tensor stored properties in this type.
+  static var _outputTypeList: [TensorDataType] { get }
+
   /// Initializes a value of this type, taking ownership of the
   /// `_tensorHandleCount` tensors that are at `tensorHandles`.
   init(_owning tensorHandles: UnsafePointer<CTensorHandle>?)
 }
 
-public extension TensorGroup {
+public extension OutputTensorGroup {
   /// The number of tensor fields in this type.
-  static var _tensorHandleCount: Int32 {
-    return Int32(_typeList.count)
+  static var _outputTensorHandleCount: Int32 {
+    return Int32(_outputTypeList.count)
   }
 
   /// An array of `nil`s with size equal to `_tensorHandleCount`. The `nil`
   /// represents unknown shape.
   static var _unknownShapeList: [TensorShape?] {
-    return Array(repeating: nil, count: Int(_tensorHandleCount))
+    return Array(repeating: nil, count: Int(_outputTensorHandleCount))
   }
 }
 
+/// A protocol for types that can be used as tensor operation inputs and
+/// outputs. When a TensorGroup is used as an input, it gets passed to the
+/// tensor operation as an input list whose elements are the tensor fields of
+/// the type. When a TensorGroup is used as an output, it gets initialized
+/// with its tensor fields set to the tensor operation's tensor outputs.
+///
+/// TODO: Add a derived conformance to TensorGroup so that users don't have
+/// to write the conformance themselves.
+public protocol TensorGroup : InputTensorGroup & OutputTensorGroup {}
+
 //===----------------------------------------------------------------------===//
 // Conform standard TensorFlow types to TensorGroup
 //===----------------------------------------------------------------------===//
 
 extension TensorHandle : TensorGroup {
-  public static var _typeList: [TensorDataType] {
+  public static var _outputTypeList: [TensorDataType] {
     return [Scalar.tensorFlowDataType]
   }
 
+  public var _inputTensorHandleCount : Int32 {
+    get { return Int32(TensorHandle._outputTypeList.count) }
+  }
+
   public func _unpackTensorHandles(
       into address: UnsafeMutablePointer<CTensorHandle>?) {
     address!.initialize(to: _cTensorHandle)
@@ -72,10 +93,14 @@ extension TensorHandle : TensorGroup {
 }
 
 extension ResourceHandle : TensorGroup {
-  public static var _typeList: [TensorDataType] {
+  public static var _outputTypeList: [TensorDataType] {
     return [TensorDataType(TF_RESOURCE)]
   }
 
+  public var _inputTensorHandleCount : Int32 {
+    get { return Int32(ResourceHandle._outputTypeList.count) }
+  }
+
   public func _unpackTensorHandles(
       into address: UnsafeMutablePointer<CTensorHandle>?) {
     address!.initialize(to: _cTensorHandle)
@@ -87,10 +112,14 @@ extension ResourceHandle : TensorGroup {
 }
 
 extension VariantHandle : TensorGroup {
-  public static var _typeList: [TensorDataType] {
+  public static var _outputTypeList: [TensorDataType] {
     return [TensorDataType(TF_VARIANT)]
   }
 
+  public var _inputTensorHandleCount : Int32 {
+    get { return Int32(VariantHandle._outputTypeList.count) }
+  }
+
   public func _unpackTensorHandles(
       into address: UnsafeMutablePointer<CTensorHandle>?) {
     address!.initialize(to: _cTensorHandle)
@@ -102,10 +131,14 @@ extension VariantHandle : TensorGroup {
 }
 
 extension Tensor : TensorGroup {
-  public static var _typeList: [TensorDataType] {
+  public static var _outputTypeList: [TensorDataType] {
     return [Scalar.tensorFlowDataType]
   }
 
+  public var _inputTensorHandleCount : Int32 {
+    get { return Int32(Tensor._outputTypeList.count) }
+  }
+
   public func _unpackTensorHandles(
       into address: UnsafeMutablePointer<CTensorHandle>?) {
     address!.initialize(to: handle._cTensorHandle)
@@ -117,10 +150,14 @@ extension Tensor : TensorGroup {
 }
 
 extension TensorElementLiteral : TensorGroup {
-  public static var _typeList: [TensorDataType] {
+  public static var _outputTypeList: [TensorDataType] {
     return [Scalar.tensorFlowDataType]
   }
 
+  public var _inputTensorHandleCount : Int32 {
+    get { return Int32(TensorElementLiteral._outputTypeList.count) }
+  }
+
   public func _unpackTensorHandles(
       into address: UnsafeMutablePointer<CTensorHandle>?) {
     address!.initialize(to: handle._cTensorHandle)
@@ -130,3 +167,18 @@ extension TensorElementLiteral : TensorGroup {
     self.init(handle: TensorHandle(_owning: tensorHandles!.pointee))
   }
 }
+
+extension Array : InputTensorGroup where Element : InputTensorGroup {
+  public func _unpackTensorHandles(into address: UnsafeMutablePointer<CTensorHandle>?) {
+    var ptr = address
+    for elem in self {
+      elem._unpackTensorHandles(into: ptr)
+      ptr = ptr!.advanced(by: Int(elem._inputTensorHandleCount))
+    }
+  }
+  public var _inputTensorHandleCount : Int32 { get {
+    var count: Int32 = 0
+    for elem in self { count += elem._inputTensorHandleCount }
+    return count
+  } }
+}