1 /* Copyright 2022 The TensorFlow Authors. All Rights Reserved.
2
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15
16 #include "tensorflow/compiler/mlir/lite/python/graphdef_to_tfl_flatbuffer.h"
17
18 #include <ostream>
19 #include <string>
20 #include <utility>
21
22 #include "llvm/ADT/None.h"
23 #include "llvm/Support/ToolOutputFile.h"
24 #include "mlir/Dialect/Func/IR/FuncOps.h" // from @llvm-project
25 #include "mlir/IR/BuiltinOps.h" // from @llvm-project
26 #include "mlir/IR/MLIRContext.h" // from @llvm-project
27 #include "mlir/Pass/Pass.h" // from @llvm-project
28 #include "mlir/Support/FileUtilities.h" // from @llvm-project
29 #include "mlir/Transforms/ViewOpGraph.h" // from @llvm-project
30 #include "tensorflow/compiler/mlir/lite/common/tfl_pass_config.h"
31 #include "tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.h"
32 #include "tensorflow/compiler/mlir/lite/tf_tfl_passes.h"
33 #include "tensorflow/compiler/mlir/lite/tf_to_tfl_flatbuffer.h"
34 #include "tensorflow/compiler/mlir/lite/transforms/passes.h"
35 #include "tensorflow/compiler/mlir/tensorflow/translate/import_model.h"
36 #include "tensorflow/compiler/mlir/tensorflow/translate/mlir_roundtrip_flags.h"
37 #include "tensorflow/core/framework/graph.pb.h"
38 #include "tensorflow/core/framework/types.pb.h"
39 #include "tensorflow/core/lib/core/errors.h"
40 #include "tensorflow/core/platform/status.h"
41 #include "tensorflow/core/protobuf/graph_debug_info.pb.h"
42 #include "tensorflow/lite/toco/model_flags.pb.h"
43 #include "tensorflow/lite/toco/toco_flags.pb.h"
44 #include "tensorflow/lite/toco/types.pb.h"
45 #include "tensorflow/stream_executor/lib/statusor.h"
46
47 namespace tensorflow {
ConvertGraphDefToTFLiteFlatBuffer(const toco::ModelFlags & model_flags,const toco::TocoFlags & toco_flags,const GraphDebugInfo & debug_info,const GraphDef & input,string * result)48 Status ConvertGraphDefToTFLiteFlatBuffer(const toco::ModelFlags& model_flags,
49 const toco::TocoFlags& toco_flags,
50 const GraphDebugInfo& debug_info,
51 const GraphDef& input,
52 string* result) {
53 using ::tflite::optimize::ReducedPrecisionSupport;
54 mlir::MLIRContext context;
55 GraphImportConfig specs;
56 mlir::quant::QuantizationSpecs quant_specs;
57
58 // Parse input arrays.
59 std::vector<string> node_names;
60 std::vector<string> node_dtypes;
61 std::vector<llvm::Optional<std::vector<int>>> node_shapes;
62 std::vector<llvm::Optional<double>> node_mins;
63 std::vector<llvm::Optional<double>> node_maxs;
64
65 // Populate quantization specs.
66 TF_RETURN_IF_ERROR(internal::PopulateQuantizationSpecs(
67 model_flags, toco_flags, &quant_specs, &node_names, &node_dtypes,
68 &node_shapes, &node_mins, &node_maxs));
69
70 TF_RETURN_IF_ERROR(tensorflow::ParseInputArrayInfo(
71 node_names, node_dtypes, node_shapes, &specs.inputs));
72
73 // Parse output arrays.
74 std::vector<string> output_arrays(model_flags.output_arrays().begin(),
75 model_flags.output_arrays().end());
76 TF_RETURN_IF_ERROR(
77 tensorflow::ParseOutputArrayInfo(output_arrays, &specs.outputs));
78
79 // Parse control output arrays.
80 std::vector<string> control_output_arrays(
81 model_flags.control_output_arrays().begin(),
82 model_flags.control_output_arrays().end());
83 TF_RETURN_IF_ERROR(tensorflow::ParseOutputArrayInfo(control_output_arrays,
84 &specs.control_outputs));
85
86 specs.prune_unused_nodes = true;
87 specs.convert_legacy_fed_inputs = true;
88 specs.graph_as_function = false;
89 specs.upgrade_legacy = true;
90 specs.unconditionally_use_set_output_shapes = true;
91 internal::WarningUnusedFlags(model_flags, toco_flags);
92
93 // Register all custom ops, including user-specified custom ops.
94 TF_RETURN_IF_ERROR(internal::RegisterAllCustomOps(toco_flags));
95
96 TF_ASSIGN_OR_RETURN(
97 auto module, ConvertGraphdefToMlir(input, debug_info, specs, &context));
98
99 mlir::TFL::PassConfig pass_config(quant_specs);
100 bool emit_builtin_tflite_ops = !toco_flags.force_select_tf_ops();
101 pass_config.emit_builtin_tflite_ops = emit_builtin_tflite_ops;
102 pass_config.unfold_batch_matmul = toco_flags.unfold_batchmatmul();
103 pass_config.lower_tensor_list_ops = toco_flags.lower_tensor_list_ops();
104 // Disable the unfolding of the 16x16 TF::BatchMatMulOp to avoid the
105 // conversion to an unsupported 16x16 TFL::FullyConnectedOp.
106 if (toco_flags.inference_type() == toco::IODataType::QUANTIZED_INT16) {
107 pass_config.unfold_batch_matmul = false;
108 }
109 pass_config.unfold_large_splat_constant =
110 toco_flags.unfold_large_splat_constant();
111 pass_config.enable_dynamic_update_slice =
112 toco_flags.enable_dynamic_update_slice();
113 pass_config.preserve_assert_op = toco_flags.preserve_assert_op();
114 pass_config.guarantee_all_funcs_one_use =
115 toco_flags.guarantee_all_funcs_one_use();
116
117 return internal::ConvertMLIRToTFLiteFlatBuffer(
118 model_flags, toco_flags, std::move(module), pass_config,
119 /*saved_model_tags=*/{}, result,
120 /*session=*/llvm::None);
121 }
122
123 } // namespace tensorflow
124