1 /* Copyright 2019 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 <utility>
20
21 #include "llvm/ADT/None.h"
22 #include "llvm/Support/ToolOutputFile.h"
23 #include "mlir/IR/BuiltinOps.h" // from @llvm-project
24 #include "mlir/IR/MLIRContext.h" // from @llvm-project
25 #include "mlir/Pass/Pass.h" // from @llvm-project
26 #include "mlir/Support/FileUtilities.h" // from @llvm-project
27 #include "mlir/Transforms/ViewOpGraph.h" // from @llvm-project
28 #include "tensorflow/compiler/mlir/lite/common/tfl_pass_config.h"
29 #include "tensorflow/compiler/mlir/lite/python/tf_tfl_flatbuffer_helpers.h"
30 #include "tensorflow/compiler/mlir/lite/tf_tfl_passes.h"
31 #include "tensorflow/compiler/mlir/lite/tf_to_tfl_flatbuffer.h"
32 #include "tensorflow/compiler/mlir/lite/transforms/passes.h"
33 #include "tensorflow/compiler/mlir/tensorflow/translate/import_model.h"
34 #include "tensorflow/compiler/mlir/tensorflow/translate/mlir_roundtrip_flags.h"
35 #include "tensorflow/core/framework/graph.pb.h"
36 #include "tensorflow/core/framework/types.pb.h"
37 #include "tensorflow/core/lib/core/errors.h"
38 #include "tensorflow/core/platform/status.h"
39 #include "tensorflow/core/protobuf/graph_debug_info.pb.h"
40 #include "tensorflow/lite/toco/model_flags.pb.h"
41 #include "tensorflow/lite/toco/toco_flags.pb.h"
42 #include "tensorflow/lite/toco/types.pb.h"
43 #include "tensorflow/stream_executor/lib/statusor.h"
44
45 namespace tensorflow {
ConvertGraphDefToTFLiteFlatBuffer(const toco::ModelFlags & model_flags,const toco::TocoFlags & toco_flags,const GraphDebugInfo & debug_info,const GraphDef & input,string * result)46 Status ConvertGraphDefToTFLiteFlatBuffer(const toco::ModelFlags& model_flags,
47 const toco::TocoFlags& toco_flags,
48 const GraphDebugInfo& debug_info,
49 const GraphDef& input,
50 string* result) {
51 using ::tflite::optimize::ReducedPrecisionSupport;
52 mlir::MLIRContext context;
53 GraphImportConfig specs;
54 mlir::TFL::QuantizationSpecs quant_specs;
55
56 // Parse input arrays.
57 std::vector<string> node_names;
58 std::vector<string> node_dtypes;
59 std::vector<llvm::Optional<std::vector<int>>> node_shapes;
60 std::vector<llvm::Optional<double>> node_mins;
61 std::vector<llvm::Optional<double>> node_maxs;
62
63 // Populate quantization specs.
64 TF_RETURN_IF_ERROR(internal::PopulateQuantizationSpecs(
65 model_flags, toco_flags, &quant_specs, &node_names, &node_dtypes,
66 &node_shapes, &node_mins, &node_maxs));
67
68 TF_RETURN_IF_ERROR(tensorflow::ParseInputArrayInfo(
69 node_names, node_dtypes, node_shapes, &specs.inputs));
70
71 if (toco_flags.quantize_to_float16() || toco_flags.allow_bfloat16()) {
72 ReducedPrecisionSupport mask = ReducedPrecisionSupport::None;
73 if (toco_flags.quantize_to_float16()) {
74 mask |= ReducedPrecisionSupport::Float16Inference;
75 }
76 if (toco_flags.allow_bfloat16()) {
77 mask |= ReducedPrecisionSupport::Bfloat16Inference;
78 }
79 if (toco_flags.accumulation_type() == toco::IODataType::FLOAT16) {
80 mask |= ReducedPrecisionSupport::Float16Accumulation;
81 } else {
82 mask |= ReducedPrecisionSupport::Float32Accumulation;
83 }
84 quant_specs.support_mask = mask;
85 }
86 // Parse output arrays.
87 std::vector<string> output_arrays(model_flags.output_arrays().begin(),
88 model_flags.output_arrays().end());
89 TF_RETURN_IF_ERROR(
90 tensorflow::ParseOutputArrayInfo(output_arrays, &specs.outputs));
91
92 // Parse control output arrays.
93 std::vector<string> control_output_arrays(
94 model_flags.control_output_arrays().begin(),
95 model_flags.control_output_arrays().end());
96 TF_RETURN_IF_ERROR(tensorflow::ParseOutputArrayInfo(control_output_arrays,
97 &specs.control_outputs));
98
99 specs.prune_unused_nodes = true;
100 specs.convert_legacy_fed_inputs = true;
101 specs.graph_as_function = false;
102 specs.upgrade_legacy = true;
103 internal::WarningUnusedFlags(model_flags, toco_flags);
104
105 // Register all custom ops, including user-specified custom ops.
106 TF_RETURN_IF_ERROR(internal::RegisterAllCustomOps(toco_flags));
107
108 TF_ASSIGN_OR_RETURN(
109 auto module, ConvertGraphdefToMlir(input, debug_info, specs, &context));
110
111 mlir::TFL::PassConfig pass_config(quant_specs);
112 bool emit_builtin_tflite_ops = !toco_flags.force_select_tf_ops();
113 pass_config.emit_builtin_tflite_ops = emit_builtin_tflite_ops;
114 pass_config.unfold_batch_matmul = toco_flags.unfold_batchmatmul();
115 pass_config.lower_tensor_list_ops = toco_flags.lower_tensor_list_ops();
116 // Disable the unfolding of the 16x16 TF::BatchMatMulOp to avoid the
117 // conversion to an unsupported 16x16 TFL::FullyConnectedOp.
118 if (toco_flags.inference_type() == toco::IODataType::QUANTIZED_INT16) {
119 pass_config.unfold_batch_matmul = false;
120 }
121 pass_config.unfold_large_splat_constant =
122 toco_flags.unfold_large_splat_constant();
123
124 return internal::ConvertMLIRToTFLiteFlatBuffer(
125 model_flags, toco_flags, std::move(module), pass_config,
126 /*saved_model_tags=*/{}, result,
127 /*session=*/llvm::None);
128 }
129
130 } // namespace tensorflow
131