Automated sync from github.com/tensorflow/tensorflow

tensorflow/lite/builtin_ops.h

@@ -186,6 +186,7 @@ typedef enum {
   kTfLiteBuiltinAtan2 = 156,
   kTfLiteBuiltinUnsortedSegmentMin = 157,
   kTfLiteBuiltinSign = 158,
+  kTfLiteBuiltinBitcast = 159,
 } TfLiteBuiltinOperator;
 
 #ifdef __cplusplus

tensorflow/lite/core/api/flatbuffer_conversions.cc

@@ -885,6 +885,7 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
     case BuiltinOperator_UNSORTED_SEGMENT_SUM:
     case BuiltinOperator_ATAN2:
     case BuiltinOperator_SIGN:
+    case BuiltinOperator_BITCAST:
     case BuiltinOperator_WHERE:
       return kTfLiteOk;
     case BuiltinOperator_PLACEHOLDER_FOR_GREATER_OP_CODES:

tensorflow/lite/kernels/internal/reference/mul.h

@@ -1,4 +1,4 @@
-/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2023 The TensorFlow Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -56,7 +56,7 @@ inline void Mul(const ArithmeticParams& params,
   const int flat_size =
       MatchingExtendedShapeFlatSize(input1_shape, input2_shape, output_shape);
   for (int i = 0; i < flat_size; ++i) {
-    output_data[i] = ActivationFunctionWithMinMax(
+    output_data[i] = ActivationFunctionWithMinMax<T>(
         input1_data[i] * input2_data[i], output_activation_min,
         output_activation_max);
   }
@@ -128,14 +128,18 @@ inline void BroadcastMul4DSlow(const ArithmeticParams& params,
   }
 }
 
-template <typename T>
-void BroadcastMul4DSlow(const ArithmeticParams& params,
-                        const RuntimeShape& unextended_input1_shape,
-                        const T* input1_data,
-                        const RuntimeShape& unextended_input2_shape,
-                        const T* input2_data,
-                        const RuntimeShape& unextended_output_shape,
-                        T* output_data) {
+template <typename T,
+          // For unquantized mul on small integers, explictly set to true.
+          bool enable_for_short_integers = false>
+inline typename std::enable_if<
+    !is_small_integer<T>::value || enable_for_short_integers, void>::type
+BroadcastMul4DSlow(const ArithmeticParams& params,
+                   const RuntimeShape& unextended_input1_shape,
+                   const T* input1_data,
+                   const RuntimeShape& unextended_input2_shape,
+                   const T* input2_data,
+                   const RuntimeShape& unextended_output_shape,
+                   T* output_data) {
   T output_activation_min;
   T output_activation_max;
   GetActivationParams(params, &output_activation_min, &output_activation_max);
@@ -167,7 +171,7 @@ void BroadcastMul4DSlow(const ArithmeticParams& params,
       for (int x = 0; x < output_shape.Dims(2); ++x) {
         for (int c = 0; c < output_shape.Dims(3); ++c) {
           output_data[Offset(output_shape, b, y, x, c)] =
-              ActivationFunctionWithMinMax(
+              ActivationFunctionWithMinMax<T>(
                   input1_data[SubscriptToIndex(desc1, b, y, x, c)] *
                       input2_data[SubscriptToIndex(desc2, b, y, x, c)],
                   output_activation_min, output_activation_max);

tensorflow/lite/kernels/internal/types.h

@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2023 The TensorFlow Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -1025,6 +1025,12 @@ inline void SetActivationParams(int32_t min, int32_t max, P* params) {
   params->quantized_activation_max = max;
 }
 
+template <typename P>
+inline void SetActivationParams(uint32_t min, uint32_t max, P* params) {
+  params->quantized_activation_min = min;
+  params->quantized_activation_max = max;
+}
+
 template <typename P>
 inline void SetActivationParams(int16_t min, int16_t max, P* params) {
   params->int16_activation_min = min;
@@ -1043,6 +1049,12 @@ inline void GetActivationParams(const P& params, int32_t* min, int32_t* max) {
   *max = params.quantized_activation_max;
 }
 
+template <typename P>
+inline void GetActivationParams(const P& params, uint32_t* min, uint32_t* max) {
+  *min = params.quantized_activation_min;
+  *max = params.quantized_activation_max;
+}
+
 template <typename P>
 inline void GetActivationParams(const P& params, int16_t* min, int16_t* max) {
   *min = params.int16_activation_min;

tensorflow/lite/python/schema_py_generated.py

@@ -937,6 +937,70 @@ def Pack(self, builder):
 from flatbuffers.compat import import_numpy
 np = import_numpy()
 
+class BitcastOptions(object):
+    __slots__ = ['_tab']
+
+    @classmethod
+    def GetRootAs(cls, buf, offset=0):
+        n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+        x = BitcastOptions()
+        x.Init(buf, n + offset)
+        return x
+
+    @classmethod
+    def GetRootAsBitcastOptions(cls, buf, offset=0):
+        """This method is deprecated. Please switch to GetRootAs."""
+        return cls.GetRootAs(buf, offset)
+    @classmethod
+    def BitcastOptionsBufferHasIdentifier(cls, buf, offset, size_prefixed=False):
+        return flatbuffers.util.BufferHasIdentifier(buf, offset, b"\x54\x46\x4C\x33", size_prefixed=size_prefixed)
+
+    # BitcastOptions
+    def Init(self, buf, pos):
+        self._tab = flatbuffers.table.Table(buf, pos)
+
+def BitcastOptionsStart(builder): builder.StartObject(0)
+def Start(builder):
+    return BitcastOptionsStart(builder)
+def BitcastOptionsEnd(builder): return builder.EndObject()
+def End(builder):
+    return BitcastOptionsEnd(builder)
+
+class BitcastOptionsT(object):
+
+    # BitcastOptionsT
+    def __init__(self):
+        pass
+
+    @classmethod
+    def InitFromBuf(cls, buf, pos):
+        bitcastOptions = BitcastOptions()
+        bitcastOptions.Init(buf, pos)
+        return cls.InitFromObj(bitcastOptions)
+
+    @classmethod
+    def InitFromObj(cls, bitcastOptions):
+        x = BitcastOptionsT()
+        x._UnPack(bitcastOptions)
+        return x
+
+    # BitcastOptionsT
+    def _UnPack(self, bitcastOptions):
+        if bitcastOptions is None:
+            return
+
+    # BitcastOptionsT
+    def Pack(self, builder):
+        BitcastOptionsStart(builder)
+        bitcastOptions = BitcastOptionsEnd(builder)
+        return bitcastOptions
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tflite
+
+from flatbuffers.compat import import_numpy
+np = import_numpy()
+
 class BroadcastToOptions(object):
     __slots__ = ['_tab']
 
@@ -1394,6 +1458,7 @@ class BuiltinOperator(object):
     ATAN2 = 156
     UNSORTED_SEGMENT_MIN = 157
     SIGN = 158
+    BITCAST = 159
 # automatically generated by the FlatBuffers compiler, do not modify
 
 # namespace: tflite
@@ -1523,6 +1588,7 @@ class BuiltinOptions(object):
     UnsortedSegmentSumOptions = 121
     ATan2Options = 122
     SignOptions = 123
+    BitcastOptions = 124
 
 def BuiltinOptionsCreator(unionType, table):
     from flatbuffers.table import Table
@@ -1774,6 +1840,8 @@ def BuiltinOptionsCreator(unionType, table):
         return ATan2OptionsT.InitFromBuf(table.Bytes, table.Pos)
     if unionType == BuiltinOptions().SignOptions:
         return SignOptionsT.InitFromBuf(table.Bytes, table.Pos)
+    if unionType == BuiltinOptions().BitcastOptions:
+        return BitcastOptionsT.InitFromBuf(table.Bytes, table.Pos)
     return None
 # automatically generated by the FlatBuffers compiler, do not modify
 
@@ -7454,7 +7522,7 @@ def __init__(self):
         self.inputs = None  # type: List[int]
         self.outputs = None  # type: List[int]
         self.builtinOptionsType = 0  # type: int
-        self.builtinOptions = None  # type: Union[None, Conv2DOptionsT, DepthwiseConv2DOptionsT, ConcatEmbeddingsOptionsT, LSHProjectionOptionsT, Pool2DOptionsT, SVDFOptionsT, RNNOptionsT, FullyConnectedOptionsT, SoftmaxOptionsT, ConcatenationOptionsT, AddOptionsT, L2NormOptionsT, LocalResponseNormalizationOptionsT, LSTMOptionsT, ResizeBilinearOptionsT, CallOptionsT, ReshapeOptionsT, SkipGramOptionsT, SpaceToDepthOptionsT, EmbeddingLookupSparseOptionsT, MulOptionsT, PadOptionsT, GatherOptionsT, BatchToSpaceNDOptionsT, SpaceToBatchNDOptionsT, TransposeOptionsT, ReducerOptionsT, SubOptionsT, DivOptionsT, SqueezeOptionsT, SequenceRNNOptionsT, StridedSliceOptionsT, ExpOptionsT, TopKV2OptionsT, SplitOptionsT, LogSoftmaxOptionsT, CastOptionsT, DequantizeOptionsT, MaximumMinimumOptionsT, ArgMaxOptionsT, LessOptionsT, NegOptionsT, PadV2OptionsT, GreaterOptionsT, GreaterEqualOptionsT, LessEqualOptionsT, SelectOptionsT, SliceOptionsT, TransposeConvOptionsT, SparseToDenseOptionsT, TileOptionsT, ExpandDimsOptionsT, EqualOptionsT, NotEqualOptionsT, ShapeOptionsT, PowOptionsT, ArgMinOptionsT, FakeQuantOptionsT, PackOptionsT, LogicalOrOptionsT, OneHotOptionsT, LogicalAndOptionsT, LogicalNotOptionsT, UnpackOptionsT, FloorDivOptionsT, SquareOptionsT, ZerosLikeOptionsT, FillOptionsT, BidirectionalSequenceLSTMOptionsT, BidirectionalSequenceRNNOptionsT, UnidirectionalSequenceLSTMOptionsT, FloorModOptionsT, RangeOptionsT, ResizeNearestNeighborOptionsT, LeakyReluOptionsT, SquaredDifferenceOptionsT, MirrorPadOptionsT, AbsOptionsT, SplitVOptionsT, UniqueOptionsT, ReverseV2OptionsT, AddNOptionsT, GatherNdOptionsT, CosOptionsT, WhereOptionsT, RankOptionsT, ReverseSequenceOptionsT, MatrixDiagOptionsT, QuantizeOptionsT, MatrixSetDiagOptionsT, HardSwishOptionsT, IfOptionsT, WhileOptionsT, DepthToSpaceOptionsT, NonMaxSuppressionV4OptionsT, NonMaxSuppressionV5OptionsT, ScatterNdOptionsT, SelectV2OptionsT, DensifyOptionsT, SegmentSumOptionsT, BatchMatMulOptionsT, CumsumOptionsT, CallOnceOptionsT, BroadcastToOptionsT, Rfft2dOptionsT, Conv3DOptionsT, HashtableOptionsT, HashtableFindOptionsT, HashtableImportOptionsT, HashtableSizeOptionsT, VarHandleOptionsT, ReadVariableOptionsT, AssignVariableOptionsT, RandomOptionsT, BucketizeOptionsT, GeluOptionsT, DynamicUpdateSliceOptionsT, UnsortedSegmentProdOptionsT, UnsortedSegmentMaxOptionsT, UnsortedSegmentMinOptionsT, UnsortedSegmentSumOptionsT, ATan2OptionsT, SignOptionsT]
+        self.builtinOptions = None  # type: Union[None, Conv2DOptionsT, DepthwiseConv2DOptionsT, ConcatEmbeddingsOptionsT, LSHProjectionOptionsT, Pool2DOptionsT, SVDFOptionsT, RNNOptionsT, FullyConnectedOptionsT, SoftmaxOptionsT, ConcatenationOptionsT, AddOptionsT, L2NormOptionsT, LocalResponseNormalizationOptionsT, LSTMOptionsT, ResizeBilinearOptionsT, CallOptionsT, ReshapeOptionsT, SkipGramOptionsT, SpaceToDepthOptionsT, EmbeddingLookupSparseOptionsT, MulOptionsT, PadOptionsT, GatherOptionsT, BatchToSpaceNDOptionsT, SpaceToBatchNDOptionsT, TransposeOptionsT, ReducerOptionsT, SubOptionsT, DivOptionsT, SqueezeOptionsT, SequenceRNNOptionsT, StridedSliceOptionsT, ExpOptionsT, TopKV2OptionsT, SplitOptionsT, LogSoftmaxOptionsT, CastOptionsT, DequantizeOptionsT, MaximumMinimumOptionsT, ArgMaxOptionsT, LessOptionsT, NegOptionsT, PadV2OptionsT, GreaterOptionsT, GreaterEqualOptionsT, LessEqualOptionsT, SelectOptionsT, SliceOptionsT, TransposeConvOptionsT, SparseToDenseOptionsT, TileOptionsT, ExpandDimsOptionsT, EqualOptionsT, NotEqualOptionsT, ShapeOptionsT, PowOptionsT, ArgMinOptionsT, FakeQuantOptionsT, PackOptionsT, LogicalOrOptionsT, OneHotOptionsT, LogicalAndOptionsT, LogicalNotOptionsT, UnpackOptionsT, FloorDivOptionsT, SquareOptionsT, ZerosLikeOptionsT, FillOptionsT, BidirectionalSequenceLSTMOptionsT, BidirectionalSequenceRNNOptionsT, UnidirectionalSequenceLSTMOptionsT, FloorModOptionsT, RangeOptionsT, ResizeNearestNeighborOptionsT, LeakyReluOptionsT, SquaredDifferenceOptionsT, MirrorPadOptionsT, AbsOptionsT, SplitVOptionsT, UniqueOptionsT, ReverseV2OptionsT, AddNOptionsT, GatherNdOptionsT, CosOptionsT, WhereOptionsT, RankOptionsT, ReverseSequenceOptionsT, MatrixDiagOptionsT, QuantizeOptionsT, MatrixSetDiagOptionsT, HardSwishOptionsT, IfOptionsT, WhileOptionsT, DepthToSpaceOptionsT, NonMaxSuppressionV4OptionsT, NonMaxSuppressionV5OptionsT, ScatterNdOptionsT, SelectV2OptionsT, DensifyOptionsT, SegmentSumOptionsT, BatchMatMulOptionsT, CumsumOptionsT, CallOnceOptionsT, BroadcastToOptionsT, Rfft2dOptionsT, Conv3DOptionsT, HashtableOptionsT, HashtableFindOptionsT, HashtableImportOptionsT, HashtableSizeOptionsT, VarHandleOptionsT, ReadVariableOptionsT, AssignVariableOptionsT, RandomOptionsT, BucketizeOptionsT, GeluOptionsT, DynamicUpdateSliceOptionsT, UnsortedSegmentProdOptionsT, UnsortedSegmentMaxOptionsT, UnsortedSegmentMinOptionsT, UnsortedSegmentSumOptionsT, ATan2OptionsT, SignOptionsT, BitcastOptionsT]
         self.customOptions = None  # type: List[int]
         self.customOptionsFormat = 0  # type: int
         self.mutatingVariableInputs = None  # type: List[bool]

tensorflow/lite/schema/schema.fbs

@@ -413,7 +413,8 @@ enum BuiltinOperator : int32 {
   UNSORTED_SEGMENT_SUM = 155,
   ATAN2 = 156,
   UNSORTED_SEGMENT_MIN = 157,
-  SIGN = 158
+  SIGN = 158,
+  BITCAST = 159
 }
 // LINT.ThenChange(nnapi_linter/linter.proto)
 
@@ -541,7 +542,8 @@ union BuiltinOptions {
   UnsortedSegmentMinOptions,
   UnsortedSegmentSumOptions,
   ATan2Options,
-  SignOptions
+  SignOptions,
+  BitcastOptions
 }
 
 // LINT.IfChange
@@ -1178,6 +1180,9 @@ table UnsortedSegmentMinOptions{
 table SignOptions {
 }
 
+table BitcastOptions {
+}
+
 
 // An OperatorCode can be an enum value (BuiltinOperator) if the operator is a
 // builtin, or a string if the operator is custom.