Lce benchmark and interpreter flags

larq_compute_engine/tflite/benchmark/BUILD

@@ -29,8 +29,23 @@ tf_cc_binary(
         ],
     }),
     deps = [
+        "//larq_compute_engine/tflite/benchmark:lce_benchmark_tflite_model_lib",
         "//larq_compute_engine/tflite/kernels:lce_op_kernels",
         "@org_tensorflow//tensorflow/lite/tools:logging",
+    ],
+)
+
+cc_library(
+    name = "lce_benchmark_tflite_model_lib",
+    srcs = ["lce_benchmark_tflite_model.cc"],
+    hdrs = ["lce_benchmark_tflite_model.h"],
+    copts = tflite_copts() + select({
+        "@org_tensorflow//tensorflow:ios": [
+            "-xobjective-c++",
+        ],
+        "//conditions:default": [],
+    }),
+    deps = [
         "@org_tensorflow//tensorflow/lite/tools/benchmark:benchmark_tflite_model_lib",
     ],
 )

larq_compute_engine/tflite/benchmark/lce_benchmark_main.cc

@@ -15,23 +15,29 @@ limitations under the License.
 ==============================================================================*/
 
 #include <iostream>
+#include <string>
 
 #include "absl/base/attributes.h"
+#include "larq_compute_engine/tflite/benchmark/lce_benchmark_tflite_model.h"
 #include "larq_compute_engine/tflite/kernels/lce_ops_register.h"
-#include "tensorflow/lite/tools/benchmark/benchmark_tflite_model.h"
 #include "tensorflow/lite/tools/logging.h"
 
+bool use_reference_bconv = false;
+bool use_indirect_bgemm = false;
+
 void ABSL_ATTRIBUTE_WEAK
 RegisterSelectedOps(::tflite::MutableOpResolver* resolver) {
-  compute_engine::tflite::RegisterLCECustomOps(resolver);
+  compute_engine::tflite::RegisterLCECustomOps(resolver, use_reference_bconv,
+                                               use_indirect_bgemm);
 }
 
 namespace tflite {
 namespace benchmark {
 
 int Main(int argc, char** argv) {
   TFLITE_LOG(INFO) << "STARTING!";
-  BenchmarkTfLiteModel benchmark;
+  LceBenchmarkTfLiteModel benchmark(LceBenchmarkTfLiteModel::DefaultParams(),
+                                    use_reference_bconv, use_indirect_bgemm);
   if (benchmark.Run(argc, argv) != kTfLiteOk) {
     TFLITE_LOG(ERROR) << "Benchmarking failed.";
     return EXIT_FAILURE;

larq_compute_engine/tflite/benchmark/lce_benchmark_tflite_model.cc

@@ -0,0 +1,74 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+Modifications copyright (C) 2022 Larq Contributors.
+
+Licensed 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.
+==============================================================================*/
+
+#include "larq_compute_engine/tflite/benchmark/lce_benchmark_tflite_model.h"
+
+#include "tensorflow/lite/tools/logging.h"
+
+namespace tflite {
+namespace benchmark {
+
+BenchmarkParams LceBenchmarkTfLiteModel::DefaultParams() {
+  BenchmarkParams default_params = BenchmarkTfLiteModel::DefaultParams();
+  default_params.AddParam("use_reference_bconv",
+                          BenchmarkParam::Create<bool>(false));
+  default_params.AddParam("use_indirect_bgemm",
+                          BenchmarkParam::Create<bool>(false));
+
+  return default_params;
+}
+
+LceBenchmarkTfLiteModel::LceBenchmarkTfLiteModel(BenchmarkParams params,
+                                                 bool& use_reference_bconv,
+                                                 bool& use_indirect_bgemm)
+    : BenchmarkTfLiteModel(std::move(params)),
+      use_reference_bconv(use_reference_bconv),
+      use_indirect_bgemm(use_indirect_bgemm) {}
+
+std::vector<Flag> LceBenchmarkTfLiteModel::GetFlags() {
+  std::vector<Flag> flags = BenchmarkTfLiteModel::GetFlags();
+  std::vector<Flag> lce_flags = {
+      CreateFlag<bool>(
+          "use_reference_bconv", &params_,
+          "When true, uses the reference implementation of LceBconv2d."),
+      CreateFlag<bool>("use_indirect_bgemm", &params_,
+                       "When true, uses the optimized indirect BGEMM kernel of"
+                       "LceBconv2d.")};
+
+  flags.insert(flags.end(), lce_flags.begin(), lce_flags.end());
+
+  return flags;
+}
+
+void LceBenchmarkTfLiteModel::LogParams() {
+  BenchmarkTfLiteModel::LogParams();
+  const bool verbose = params_.Get<bool>("verbose");
+  LOG_BENCHMARK_PARAM(bool, "use_reference_bconv", "Use reference Bconv",
+                      verbose);
+  LOG_BENCHMARK_PARAM(bool, "use_indirect_bgemm", "Use indirect BGEMM",
+                      verbose);
+}
+
+TfLiteStatus LceBenchmarkTfLiteModel::Run(int argc, char** argv) {
+  TF_LITE_ENSURE_STATUS(ParseFlags(argc, argv));
+  use_reference_bconv = params_.Get<bool>("use_reference_bconv");
+  use_indirect_bgemm = params_.Get<bool>("use_indirect_bgemm");
+
+  return BenchmarkTfLiteModel::Run();
+}
+
+}  // namespace benchmark
+}  // namespace tflite

larq_compute_engine/tflite/benchmark/lce_benchmark_tflite_model.h

@@ -0,0 +1,47 @@
+/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+Modifications copyright (C) 2022 Larq Contributors.
+
+Licensed 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.
+==============================================================================*/
+
+#ifndef COMPUTE_ENGINE_TFLITE_BENCHMARK_LCE_BENCHMARK_TFLITE_MODEL_H_
+#define COMPUTE_ENGINE_TFLITE_BENCHMARK_LCE_BENCHMARK_TFLITE_MODEL_H_
+
+#include "tensorflow/lite/tools/benchmark/benchmark_tflite_model.h"
+
+namespace tflite {
+namespace benchmark {
+
+// Benchmarks a TFLite model by running tflite interpreter.
+class LceBenchmarkTfLiteModel : public BenchmarkTfLiteModel {
+ public:
+  explicit LceBenchmarkTfLiteModel(BenchmarkParams params,
+                                   bool& use_reference_bconv,
+                                   bool& use_indirect_bgemm);
+
+  std::vector<Flag> GetFlags() override;
+  void LogParams() override;
+  static BenchmarkParams DefaultParams();
+
+  using BenchmarkTfLiteModel::Run;
+  TfLiteStatus Run(int argc, char** argv);
+
+ private:
+  bool& use_reference_bconv;
+  bool& use_indirect_bgemm;
+};
+
+}  // namespace benchmark
+}  // namespace tflite
+
+#endif  // COMPUTE_ENGINE_TFLITE_BENCHMARK_LCE_BENCHMARK_TFLITE_MODEL_H_

larq_compute_engine/tflite/build_make/Makefile

@@ -117,6 +117,7 @@ LCE_EXAMPLE_SRCS := \
 	examples/lce_minimal.cc
 
 LCE_BENCHMARK_SRCS := \
+    larq_compute_engine/tflite/benchmark/lce_benchmark_tflite_model.cc \
 	larq_compute_engine/tflite/benchmark/lce_benchmark_main.cc
 
 # These target-specific makefiles should modify or replace options like

larq_compute_engine/tflite/kernels/lce_ops_register.h

@@ -14,12 +14,14 @@ TfLiteRegistration* Register_QUANTIZE();
 TfLiteRegistration* Register_DEQUANTIZE();
 TfLiteRegistration* Register_BCONV_2D();
 TfLiteRegistration* Register_BCONV_2D_REF();
+TfLiteRegistration* Register_BCONV_2D_OPT_INDIRECT_BGEMM();
 TfLiteRegistration* Register_BMAXPOOL_2D();
 
 // By calling this function on TF lite mutable op resolver, all LCE custom ops
 // will be registerd to the op resolver.
 inline void RegisterLCECustomOps(::tflite::MutableOpResolver* resolver,
-                                 const bool use_reference_bconv = false) {
+                                 const bool use_reference_bconv = false,
+                                 const bool use_indirect_bgemm = false) {
   resolver->AddCustom("LceQuantize",
                       compute_engine::tflite::Register_QUANTIZE());
   resolver->AddCustom("LceDequantize",
@@ -28,8 +30,14 @@ inline void RegisterLCECustomOps(::tflite::MutableOpResolver* resolver,
     resolver->AddCustom("LceBconv2d",
                         compute_engine::tflite::Register_BCONV_2D_REF());
   } else {
-    resolver->AddCustom("LceBconv2d",
-                        compute_engine::tflite::Register_BCONV_2D());
+    if (use_indirect_bgemm) {
+      resolver->AddCustom(
+          "LceBconv2d",
+          compute_engine::tflite::Register_BCONV_2D_OPT_INDIRECT_BGEMM());
+    } else {
+      resolver->AddCustom("LceBconv2d",
+                          compute_engine::tflite::Register_BCONV_2D());
+    }
   }
   resolver->AddCustom("LceBMaxPool2d",
                       compute_engine::tflite::Register_BMAXPOOL_2D());

larq_compute_engine/tflite/python/interpreter.py

@@ -23,6 +23,8 @@ class Interpreter(InterpreterBase):
         flatbuffer_model: A serialized Larq Compute Engine model in the flatbuffer format.
         num_threads: The number of threads used by the interpreter.
         use_reference_bconv: When True, uses the reference implementation of LceBconv2d.
+        use_indirect_bgemm: When True, uses the optimized indirect BGEMM kernel of LceBconv2d.
+        use_xnnpack: When True, uses the XNNPack delegate of TFLite.
 
     # Attributes
         input_types: Returns a list of input types.
@@ -40,11 +42,17 @@ def __init__(
         flatbuffer_model: bytes,
         num_threads: int = 1,
         use_reference_bconv: bool = False,
+        use_indirect_bgemm: bool = False,
+        use_xnnpack: bool = False,
     ):
         from larq_compute_engine.tflite.python import interpreter_wrapper_lite
 
         super().__init__(
             interpreter_wrapper_lite.LiteInterpreter(
-                flatbuffer_model, num_threads, use_reference_bconv
+                flatbuffer_model,
+                num_threads,
+                use_reference_bconv,
+                use_indirect_bgemm,
+                use_xnnpack,
             )
         )

larq_compute_engine/tflite/python/interpreter_wrapper_lite.cc

@@ -12,7 +12,9 @@ class LiteInterpreterWrapper
  public:
   LiteInterpreterWrapper(const pybind11::bytes& flatbuffer,
                          const int num_threads = 1,
-                         const bool use_reference_bconv = false);
+                         const bool use_reference_bconv = false,
+                         const bool use_indirect_bgemm = false,
+                         const bool use_xnnpack = false);
   ~LiteInterpreterWrapper(){};
 
  private:
@@ -25,7 +27,8 @@ class LiteInterpreterWrapper
 
 LiteInterpreterWrapper::LiteInterpreterWrapper(
     const pybind11::bytes& flatbuffer, const int num_threads,
-    const bool use_reference_bconv) {
+    const bool use_reference_bconv, const bool use_indirect_bgemm,
+    const bool use_xnnpack) {
   // Make a copy of the flatbuffer because it can get deallocated after the
   // constructor is done
   flatbuffer_ = static_cast<std::string>(flatbuffer);
@@ -37,9 +40,14 @@ LiteInterpreterWrapper::LiteInterpreterWrapper(
   }
 
   // Build the interpreter
-  resolver_ = std::make_unique<tflite::ops::builtin::BuiltinOpResolver>();
-  compute_engine::tflite::RegisterLCECustomOps(resolver_.get(),
-                                               use_reference_bconv);
+  if (use_xnnpack) {
+    resolver_ = std::make_unique<tflite::ops::builtin::BuiltinOpResolver>();
+  } else {
+    resolver_ = std::make_unique<
+        tflite::ops::builtin::BuiltinOpResolverWithoutDefaultDelegates>();
+  }
+  compute_engine::tflite::RegisterLCECustomOps(
+      resolver_.get(), use_reference_bconv, use_indirect_bgemm);
 
   tflite::InterpreterBuilder builder(*model_, *resolver_);
   builder(&interpreter_, num_threads);
@@ -51,7 +59,8 @@ LiteInterpreterWrapper::LiteInterpreterWrapper(
 
 PYBIND11_MODULE(interpreter_wrapper_lite, m) {
   pybind11::class_<LiteInterpreterWrapper>(m, "LiteInterpreter")
-      .def(pybind11::init<const pybind11::bytes&, const int, const bool>())
+      .def(pybind11::init<const pybind11::bytes&, const int, const bool,
+                          const bool, const bool>())
       .def_property("input_types", &LiteInterpreterWrapper::get_input_types,
                     nullptr)
       .def_property("output_types", &LiteInterpreterWrapper::get_output_types,