Rolling 20220103

FeatureRequest.md

@@ -11,6 +11,20 @@
 
 ____
 
+
+
+#### #15 Calibration for FishEye image
+
+* https://github.com/jomjol/AI-on-the-edge-device/issues/507
+
+1.  The development of such a correction algorithm with the libraries, that are available for the ESP32 environment.
+2. New module for integration of the flow into the image processing flow.
+3. Extension of the configuration (config.ini) and html-pages
+4. Parameter adjustment and testing for every different fish-eye module
+5. Maintenance for further updates / modules, ...
+
+
+
 #### #14 Backup and restore option for configuration
 
 * https://github.com/jomjol/AI-on-the-edge-device/issues/459

README.md

@@ -52,6 +52,15 @@ In other cases you can contact the developer via email: <img src="https://raw.gi
 
 ------
 
+##### Rolling (2022-01-03)
+
+- Update tflite (new source: https://github.com/espressif/tflite-micro-esp-examples)
+- Update digital neural network (dig-s1-q-20220102.tflite)
+- Increased web-server buffers
+- bug fix: compiler compatibility
+
+
+
 ##### 10.0.2 - Stability Increase (2022-01-01)
 
 - NEW v10.0.2: Corrected JSON error

code/components/jomjol_fileserver_ota/CMakeLists.txt

@@ -2,6 +2,6 @@ FILE(GLOB_RECURSE app_sources ${CMAKE_CURRENT_SOURCE_DIR}/*.*)
 
 idf_component_register(SRCS ${app_sources}
                     INCLUDE_DIRS "." "../../include"
-                    REQUIRES tfmicro esp_http_server app_update esp_http_client nvs_flash jomjol_tfliteclass jomjol_flowcontroll spiffs jomjol_helper jomjol_controlGPIO)
+                    REQUIRES tflite-lib esp_http_server app_update esp_http_client nvs_flash jomjol_tfliteclass jomjol_flowcontroll spiffs jomjol_helper jomjol_controlGPIO)
 
 

code/components/jomjol_fileserver_ota/server_file.cpp

@@ -17,7 +17,14 @@
 #include <sys/param.h>
 #include <sys/unistd.h>
 #include <sys/stat.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 #include <dirent.h>
+#ifdef __cplusplus
+}
+#endif
 
 #include "esp_err.h"
 #include "esp_log.h"

code/components/jomjol_fileserver_ota/server_help.cpp

@@ -5,7 +5,14 @@
 #include <sys/param.h>
 #include <sys/unistd.h>
 #include <sys/stat.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 #include <dirent.h>
+#ifdef __cplusplus
+}
+#endif
 
 #include "esp_err.h"
 #include "esp_log.h"

code/components/jomjol_flowcontroll/ClassFlowControll.cpp

@@ -6,7 +6,15 @@
 #include "freertos/task.h"
 
 #include <sys/stat.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 #include <dirent.h>
+#ifdef __cplusplus
+}
+#endif
+
 #include "ClassLogFile.h"
 #include "time_sntp.h"
 #include "Helper.h"

code/components/jomjol_flowcontroll/ClassFlowImage.cpp

@@ -2,7 +2,15 @@
 #include <string>
 #include <string.h>
 #include <sys/stat.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 #include <dirent.h>
+#ifdef __cplusplus
+}
+#endif
+
 #include "time_sntp.h"
 #include "ClassLogFile.h"
 #include "CImageBasis.h"

code/components/jomjol_helper/CMakeLists.txt

@@ -2,6 +2,6 @@ FILE(GLOB_RECURSE app_sources ${CMAKE_CURRENT_SOURCE_DIR}/*.*)
 
 idf_component_register(SRCS ${app_sources}
                     INCLUDE_DIRS "."
-                    REQUIRES tfmicro jomjol_logfile)
+                    REQUIRES tflite-lib jomjol_logfile)
 
 

code/components/jomjol_helper/Helper.cpp

@@ -6,7 +6,15 @@
 #include "Helper.h"
 #include <sys/types.h>
 #include <sys/stat.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 #include <dirent.h>
+#ifdef __cplusplus
+}
+#endif
+
 #include <string.h>
 #include <esp_log.h>
 

code/components/jomjol_logfile/ClassLogFile.cpp

@@ -3,7 +3,15 @@
 #include <string.h>
 #include <sys/types.h>
 #include <sys/stat.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
 #include <dirent.h>
+#ifdef __cplusplus
+}
+#endif
+
 #include "Helper.h"
 
 static const char *TAG = "log";

code/components/jomjol_mqtt/CMakeLists.txt

@@ -2,6 +2,6 @@ FILE(GLOB_RECURSE app_sources ${CMAKE_CURRENT_SOURCE_DIR}/*.*)
 
 idf_component_register(SRCS ${app_sources}
                     INCLUDE_DIRS "."
-                    REQUIRES tfmicro mqtt jomjol_logfile)
+                    REQUIRES tflite-lib mqtt jomjol_logfile)
 
 

code/components/jomjol_time_sntp/CMakeLists.txt

@@ -2,6 +2,6 @@ FILE(GLOB_RECURSE app_sources ${CMAKE_CURRENT_SOURCE_DIR}/*.*)
 
 idf_component_register(SRCS ${app_sources}
                     INCLUDE_DIRS "."
-                    REQUIRES tfmicro jomjol_logfile)
+                    REQUIRES tflite-lib jomjol_logfile)
 
 

code/components/tflite-lib/CMakeLists.txt

@@ -0,0 +1,50 @@
+cmake_minimum_required(VERSION 3.5)
+
+set(tflite_dir "${CMAKE_CURRENT_SOURCE_DIR}/tensorflow/lite")
+set(tfmicro_dir "${tflite_dir}/micro")
+set(tfmicro_frontend_dir "${tflite_dir}/experimental/microfrontend/lib")
+set(tfmicro_kernels_dir "${tfmicro_dir}/kernels")
+
+file(GLOB srcs_micro
+          "${tfmicro_dir}/*.cc"
+          "${tfmicro_dir}/*.c")
+
+file(GLOB src_micro_frontend
+          "${tfmicro_frontend_dir}/*.c"
+          "${tfmicro_frontend_dir}/*.cc")
+file(GLOB srcs_kernels
+          "${tfmicro_kernels_dir}/*.c"
+          "${tfmicro_kernels_dir}/*.cc")
+
+set(lib_srcs
+          "${srcs_micro}"
+          "${srcs_kernels}"
+          "${src_micro_frontend}"
+          "${tflite_dir}/kernels/kernel_util.cc"
+          "${tflite_dir}/micro/memory_planner/greedy_memory_planner.cc"
+          "${tflite_dir}/micro/memory_planner/linear_memory_planner.cc"
+          "${tflite_dir}/c/common.c"
+          "${tflite_dir}/core/api/error_reporter.cc"
+          "${tflite_dir}/core/api/flatbuffer_conversions.cc"
+          "${tflite_dir}/core/api/op_resolver.cc"
+          "${tflite_dir}/core/api/tensor_utils.cc"
+          "${tflite_dir}/kernels/internal/quantization_util.cc"
+          "${tflite_dir}/schema/schema_utils.cc")
+
+idf_component_register(
+            SRCS "${lib_srcs}"
+            INCLUDE_DIRS "." "third_party/gemmlowp"
+                         "third_party/flatbuffers/include"
+                         "third_party/ruy"
+                         "third_party/kissfft")
+
+# Reduce the level of paranoia to be able to compile TF sources
+target_compile_options(${COMPONENT_LIB} PRIVATE
+  -Wno-maybe-uninitialized
+  -Wno-missing-field-initializers
+  -Wno-type-limits)
+
+target_compile_options(${COMPONENT_LIB} PRIVATE -fno-unwind-tables -ffunction-sections -fdata-sections -fmessage-length=0 -DTF_LITE_STATIC_MEMORY -DTF_LITE_DISABLE_X86_NEON -O3 -Wsign-compare -Wdouble-promotion -Wshadow -Wunused-variable -Wmissing-field-initializers -Wunused-function -Wswitch -Wvla -Wall -Wextra -Wstrict-aliasing -Wno-unused-parameter -DESP -DESP_NN -Wno-nonnull -Wno-nonnull -Wno-nonnull)
+target_compile_options(${COMPONENT_LIB} PRIVATE $<$<COMPILE_LANGUAGE:CXX>: -std=c++11 -fno-rtti -fno-exceptions -fno-threadsafe-statics -fno-unwind-tables -ffunction-sections -fdata-sections -fmessage-length=0 -DTF_LITE_STATIC_MEMORY -DTF_LITE_DISABLE_X86_NEON -O3 -Werror -Wsign-compare -Wdouble-promotion -Wshadow -Wunused-variable -Wmissing-field-initializers -Wunused-function -Wswitch -Wvla -Wall -Wextra -Wstrict-aliasing -Wno-unused-parameter -DESP -DESP_NN -Wno-return-type -Wno-strict-aliasing -std=gnu++14 -Wno-return-type -Wno-strict-aliasing -std=gnu++14 -Wno-return-type -Wno-strict-aliasing -std=gnu++14 >)
+target_compile_options(${COMPONENT_LIB} INTERFACE $<$<IN_LIST:-DTF_LITE_STATIC_MEMORY,$<TARGET_PROPERTY:${COMPONENT_LIB},COMPILE_OPTIONS>>:-DTF_LITE_STATIC_MEMORY>)
+target_link_libraries(${COMPONENT_LIB} PRIVATE -lm)

code/components/tflite-lib/tensorflow/lite/c/builtin_op_data.h

@@ -502,6 +502,22 @@ typedef struct {
   const char* shared_name;
 } TfLiteVarHandleParams;
 
+typedef struct {
+  int seed;
+  int seed2;
+} TfLiteRandomParams;
+
+typedef struct {
+  int num_boundaries;
+  // This points to the memory stored in the model (flatbuffer),
+  // and is not owned.
+  const float* boundaries;
+} TfLiteBucketizeParams;
+
+typedef struct {
+  bool approximate;
+} TfLiteGeluParams;
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif  // __cplusplus

code/components/tflite-lib/tensorflow/lite/c/c_api_types.h

@@ -43,6 +43,9 @@ extern "C" {
 #endif  // _WIN32
 #endif  // SWIG
 
+// Note that new error status values may be added in future in order to
+// indicate more fine-grained internal states, therefore, applications should
+// not rely on status values being members of the enum.
 typedef enum TfLiteStatus {
   kTfLiteOk = 0,
 
@@ -54,7 +57,7 @@ typedef enum TfLiteStatus {
 
   // Generally referring to an error in applying a delegate due to
   // incompatibility between runtime and delegate, e.g., this error is returned
-  // when trying to apply a TfLite delegate onto a model graph that's already
+  // when trying to apply a TF Lite delegate onto a model graph that's already
   // immutable.
   kTfLiteApplicationError = 3,
 
@@ -68,7 +71,12 @@ typedef enum TfLiteStatus {
 
   // Generally referring to data-reading issues in delegate serialization.
   // See tflite::delegates::Serialization.
-  kTfLiteDelegateDataReadError = 5,
+  kTfLiteDelegateDataReadError = 6,
+
+  // Generally referring to issues when the TF Lite model has ops that cannot be
+  // resolved at runtime. This could happen when the specific op is not
+  // registered or built with the TF Lite framework.
+  kTfLiteUnresolvedOps = 7,
 } TfLiteStatus;
 
 // Types supported by tensor

code/components/tflite-lib/tensorflow/lite/c/common.c

@@ -21,9 +21,15 @@ limitations under the License.
 #include <string.h>
 #endif  // TF_LITE_STATIC_MEMORY
 
-int TfLiteIntArrayGetSizeInBytes(int size) {
+size_t TfLiteIntArrayGetSizeInBytes(int size) {
   static TfLiteIntArray dummy;
-  return sizeof(dummy) + sizeof(dummy.data[0]) * size;
+
+  size_t computed_size = sizeof(dummy) + sizeof(dummy.data[0]) * size;
+#if defined(_MSC_VER)
+  // Context for why this is needed is in http://b/189926408#comment21
+  computed_size -= sizeof(dummy.data[0]);
+#endif
+  return computed_size;
 }
 
 int TfLiteIntArrayEqual(const TfLiteIntArray* a, const TfLiteIntArray* b) {
@@ -45,7 +51,7 @@ int TfLiteIntArrayEqualsArray(const TfLiteIntArray* a, int b_size,
 #ifndef TF_LITE_STATIC_MEMORY
 
 TfLiteIntArray* TfLiteIntArrayCreate(int size) {
-  int alloc_size = TfLiteIntArrayGetSizeInBytes(size);
+  size_t alloc_size = TfLiteIntArrayGetSizeInBytes(size);
   if (alloc_size <= 0) return NULL;
   TfLiteIntArray* ret = (TfLiteIntArray*)malloc(alloc_size);
   if (!ret) return ret;
@@ -68,7 +74,13 @@ void TfLiteIntArrayFree(TfLiteIntArray* a) { free(a); }
 
 int TfLiteFloatArrayGetSizeInBytes(int size) {
   static TfLiteFloatArray dummy;
-  return sizeof(dummy) + sizeof(dummy.data[0]) * size;
+
+  int computed_size = sizeof(dummy) + sizeof(dummy.data[0]) * size;
+#if defined(_MSC_VER)
+  // Context for why this is needed is in http://b/189926408#comment21
+  computed_size -= sizeof(dummy.data[0]);
+#endif
+  return computed_size;
 }
 
 #ifndef TF_LITE_STATIC_MEMORY
@@ -176,6 +188,26 @@ void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims,
   tensor->quantization.params = NULL;
 }
 
+TfLiteStatus TfLiteTensorCopy(const TfLiteTensor* src, TfLiteTensor* dst) {
+  if (!src || !dst)
+    return kTfLiteOk;
+  if (src->bytes != dst->bytes)
+    return kTfLiteError;
+  if (src == dst)
+    return kTfLiteOk;
+
+  dst->type = src->type;
+  if (dst->dims)
+    TfLiteIntArrayFree(dst->dims);
+  dst->dims = TfLiteIntArrayCopy(src->dims);
+  memcpy(dst->data.raw, src->data.raw, src->bytes);
+  dst->buffer_handle = src->buffer_handle;
+  dst->data_is_stale = src->data_is_stale;
+  dst->delegate = src->delegate;
+
+  return kTfLiteOk;
+}
+
 void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor) {
   if (tensor->allocation_type != kTfLiteDynamic &&
       tensor->allocation_type != kTfLitePersistentRo) {

code/components/tflite-lib/tensorflow/lite/c/common.h

@@ -80,12 +80,16 @@ typedef struct TfLiteExternalContext {
 // indices
 typedef struct TfLiteIntArray {
   int size;
-// gcc 6.1+ have a bug where flexible members aren't properly handled
+
-// https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
+#if defined(_MSC_VER)
-#if (!defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
+  // Context for why this is needed is in http://b/189926408#comment21
+  int data[1];
+#elif (!defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
        __GNUC_MINOR__ >= 1) ||                                      \
     defined(HEXAGON) ||                                             \
     (defined(__clang__) && __clang_major__ == 7 && __clang_minor__ == 1)
+  // gcc 6.1+ have a bug where flexible members aren't properly handled
+  // https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
   int data[0];
 #else
   int data[];
@@ -94,7 +98,7 @@ typedef struct TfLiteIntArray {
 
 // Given the size (number of elements) in a TfLiteIntArray, calculate its size
 // in bytes.
-int TfLiteIntArrayGetSizeInBytes(int size);
+size_t TfLiteIntArrayGetSizeInBytes(int size);
 
 #ifndef TF_LITE_STATIC_MEMORY
 // Create a array of a given `size` (uninitialized entries).
@@ -121,11 +125,15 @@ void TfLiteIntArrayFree(TfLiteIntArray* a);
 // Fixed size list of floats. Used for per-channel quantization.
 typedef struct TfLiteFloatArray {
   int size;
+#if defined(_MSC_VER)
+  // Context for why this is needed is in http://b/189926408#comment21
+  float data[1];
+#elif (!defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
+       __GNUC_MINOR__ >= 1) ||                                      \
+    defined(HEXAGON) ||                                             \
+    (defined(__clang__) && __clang_major__ == 7 && __clang_minor__ == 1)
   // gcc 6.1+ have a bug where flexible members aren't properly handled
   // https://github.com/google/re2/commit/b94b7cd42e9f02673cd748c1ac1d16db4052514c
-// This also applies to the toolchain used for Qualcomm Hexagon DSPs.
-#if !defined(__clang__) && defined(__GNUC__) && __GNUC__ == 6 && \
-    __GNUC_MINOR__ >= 1
   float data[0];
 #else
   float data[];
@@ -562,6 +570,10 @@ typedef struct TfLiteNode {
   // Outputs to this node expressed as indices into the simulator's tensors.
   TfLiteIntArray* outputs;
 
+  // intermediate tensors to this node expressed as indices into the simulator's
+  // tensors.
+  TfLiteIntArray* intermediates;
+
   // Opaque data provided by the node implementer through `Registration.init`.
   void* user_data;
 
@@ -614,6 +626,16 @@ void TfLiteTensorReset(TfLiteType type, const char* name, TfLiteIntArray* dims,
                        const void* allocation, bool is_variable,
                        TfLiteTensor* tensor);
 
+// Copies the contents of 'src' in 'dst'.
+// Function does nothing if either 'src' or 'dst' is passed as nullptr and
+// return kTfLiteOk.
+// Returns kTfLiteError if 'src' and 'dst' doesn't have matching data size.
+// Note function copies contents, so it won't create new data pointer
+// or change allocation type.
+// All Tensor related properties will be copied from 'src' to 'dst' like
+// quantization, sparsity, ...
+TfLiteStatus TfLiteTensorCopy(const TfLiteTensor* src, TfLiteTensor* dst);
+
 // Resize the allocated data of a (dynamic) tensor. Tensors with allocation
 // types other than kTfLiteDynamic will be ignored.
 void TfLiteTensorRealloc(size_t num_bytes, TfLiteTensor* tensor);
@@ -819,6 +841,9 @@ typedef struct TfLiteContext {
 
 typedef struct TfLiteRegistration {
   // Initializes the op from serialized data.
+  // Called only *once* for the lifetime of the op, so any one-time allocations
+  // should be made here (unless they depend on tensor sizes).
+  //
   // If a built-in op:
   //   `buffer` is the op's params data (TfLiteLSTMParams*).
   //   `length` is zero.
@@ -841,6 +866,7 @@ typedef struct TfLiteRegistration {
   // prepare is called when the inputs this node depends on have been resized.
   // context->ResizeTensor() can be called to request output tensors to be
   // resized.
+  // Can be called multiple times for the lifetime of the op.
   //
   // Returns kTfLiteOk on success.
   TfLiteStatus (*prepare)(TfLiteContext* context, TfLiteNode* node);

code/components/tflite-lib/tensorflow/lite/core/api/flatbuffer_conversions.cc

@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2021 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.
@@ -131,6 +131,17 @@ TfLitePadding ConvertPadding(Padding padding) {
   return kTfLitePaddingUnknown;
 }
 
+// Converts the flatbuffer mirror padding enum to what is used at runtime.
+TfLiteMirrorPaddingMode ConvertMirrorPadding(MirrorPadMode padding) {
+  switch (padding) {
+    case MirrorPadMode_REFLECT:
+      return kTfLiteMirrorPaddingReflect;
+    case MirrorPadMode_SYMMETRIC:
+      return kTfLiteMirrorPaddingSymmetric;
+  }
+  return kTfLiteMirrorPaddingUnknown;
+}
+
 #ifndef TF_LITE_STATIC_MEMORY
 TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
                                ErrorReporter* error_reporter,
@@ -181,6 +192,10 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       return ParseArgMin(op, error_reporter, allocator, builtin_data);
     }
 
+    case BuiltinOperator_ASSIGN_VARIABLE: {
+      return ParseAssignVariable(op, error_reporter, allocator, builtin_data);
+    }
+
     case BuiltinOperator_AVERAGE_POOL_2D: {
       return ParsePool(op, error_reporter, allocator, builtin_data);
     }
@@ -193,6 +208,10 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       return ParseBatchToSpaceNd(op, error_reporter, allocator, builtin_data);
     }
 
+    case BuiltinOperator_CALL_ONCE: {
+      return ParseCallOnce(op, error_reporter, allocator, builtin_data);
+    }
+
     case BuiltinOperator_CEIL: {
       return ParseCeil(op, error_reporter, allocator, builtin_data);
     }
@@ -325,6 +344,10 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       return ParsePool(op, error_reporter, allocator, builtin_data);
     }
 
+    case BuiltinOperator_MIRROR_PAD: {
+      return ParseMirrorPad(op, error_reporter, allocator, builtin_data);
+    }
+
     case BuiltinOperator_MEAN: {
       return ParseReducer(op, error_reporter, allocator, builtin_data);
     }
@@ -369,6 +392,10 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       return ParseQuantize(op, error_reporter, allocator, builtin_data);
     }
 
+    case BuiltinOperator_READ_VARIABLE: {
+      return ParseReadVariable(op, error_reporter, allocator, builtin_data);
+    }
+
     case BuiltinOperator_REDUCE_ANY: {
       return ParseReducer(op, error_reporter, allocator, builtin_data);
     }
@@ -486,6 +513,10 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       return ParseUnpack(op, error_reporter, allocator, builtin_data);
     }
 
+    case BuiltinOperator_VAR_HANDLE: {
+      return ParseVarHandle(op, error_reporter, allocator, builtin_data);
+    }
+
     case BuiltinOperator_ZEROS_LIKE: {
       return ParseZerosLike(op, error_reporter, allocator, builtin_data);
     }
@@ -606,21 +637,8 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       return kTfLiteOk;
     }
     case BuiltinOperator_UNIDIRECTIONAL_SEQUENCE_LSTM: {
-      auto params =
+      return ParseUnidirectionalSequenceLSTM(op, error_reporter, allocator,
-          safe_allocator.Allocate<TfLiteUnidirectionalSequenceLSTMParams>();
+                                             builtin_data);
-      TF_LITE_ENSURE(error_reporter, params != nullptr);
-      if (const auto* seq_lstm_params =
-              op->builtin_options_as_UnidirectionalSequenceLSTMOptions()) {
-        params->activation =
-            ConvertActivation(seq_lstm_params->fused_activation_function());
-        params->cell_clip = seq_lstm_params->cell_clip();
-        params->proj_clip = seq_lstm_params->proj_clip();
-        params->time_major = seq_lstm_params->time_major();
-        params->asymmetric_quantize_inputs =
-            seq_lstm_params->asymmetric_quantize_inputs();
-      }
-      *builtin_data = params.release();
-      return kTfLiteOk;
     }
     case BuiltinOperator_BIDIRECTIONAL_SEQUENCE_LSTM: {
       auto params =
@@ -693,19 +711,6 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       *builtin_data = params.release();
       return kTfLiteOk;
     }
-    case BuiltinOperator_MIRROR_PAD: {
-      auto params = safe_allocator.Allocate<TfLiteMirrorPaddingParams>();
-      TF_LITE_ENSURE(error_reporter, params != nullptr);
-      const auto* mirror_pad_params = op->builtin_options_as_MirrorPadOptions();
-      if (mirror_pad_params != nullptr) {
-        params->mode =
-            mirror_pad_params->mode() == tflite::MirrorPadMode_REFLECT
-                ? TfLiteMirrorPaddingMode::kTfLiteMirrorPaddingReflect
-                : TfLiteMirrorPaddingMode::kTfLiteMirrorPaddingSymmetric;
-      }
-      *builtin_data = params.release();
-      return kTfLiteOk;
-    }
     case BuiltinOperator_UNIQUE: {
       auto params = safe_allocator.Allocate<TfLiteUniqueParams>();
       TF_LITE_ENSURE(error_reporter, params != nullptr);
@@ -750,16 +755,6 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       *builtin_data = params.release();
       return kTfLiteOk;
     }
-    case BuiltinOperator_CALL_ONCE: {
-      auto params = safe_allocator.Allocate<TfLiteCallOnceParams>();
-      TF_LITE_ENSURE(error_reporter, params != nullptr);
-      if (const auto* call_once_params =
-              op->builtin_options_as_CallOnceOptions()) {
-        params->init_subgraph_index = call_once_params->init_subgraph_index();
-      }
-      *builtin_data = params.release();
-      return kTfLiteOk;
-    }
     case BuiltinOperator_CONV_3D:
     case BuiltinOperator_CONV_3D_TRANSPOSE: {
       auto params = safe_allocator.Allocate<TfLiteConv3DParams>();
@@ -793,17 +788,69 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
       *builtin_data = params.release();
       return kTfLiteOk;
     }
-    case BuiltinOperator_VAR_HANDLE: {
+    case BuiltinOperator_MULTINOMIAL: {
-      auto params = safe_allocator.Allocate<TfLiteVarHandleParams>();
+      auto params = safe_allocator.Allocate<TfLiteRandomParams>();
       TF_LITE_ENSURE(error_reporter, params != nullptr);
-      params->container = nullptr;
+      if (const auto* multinomial_params =
-      params->shared_name = nullptr;
+              op->builtin_options_as_RandomOptions()) {
-      if (const auto* var_handle_params =
+        params->seed = multinomial_params->seed();
-              op->builtin_options_as_VarHandleOptions()) {
+        params->seed2 = multinomial_params->seed2();
-        if (var_handle_params->container())
+      }
-          params->container = var_handle_params->container()->c_str();
+      *builtin_data = params.release();
-        if (var_handle_params->shared_name())
+      return kTfLiteOk;
-          params->shared_name = var_handle_params->shared_name()->c_str();
+    }
+    case BuiltinOperator_RANDOM_STANDARD_NORMAL: {
+      auto params = safe_allocator.Allocate<TfLiteRandomParams>();
+      TF_LITE_ENSURE(error_reporter, params != nullptr);
+      if (const auto* random_std_normal_params =
+              op->builtin_options_as_RandomOptions()) {
+        params->seed = random_std_normal_params->seed();
+        params->seed2 = random_std_normal_params->seed2();
+      }
+      *builtin_data = params.release();
+      return kTfLiteOk;
+    }
+    case BuiltinOperator_BUCKETIZE: {
+      auto params = safe_allocator.Allocate<TfLiteBucketizeParams>();
+      TF_LITE_ENSURE(error_reporter, params != nullptr);
+      if (const auto* bucketize_params =
+              op->builtin_options_as_BucketizeOptions()) {
+        const flatbuffers::Vector<float>* boundaries =
+            bucketize_params->boundaries();
+        if (boundaries == nullptr) {
+          TF_LITE_REPORT_ERROR(
+              error_reporter,
+              "boundaries array not provided for operation 'bucketize'.\n");
+          return kTfLiteError;
+        }
+        params->num_boundaries = boundaries->size();
+        if (boundaries->data() == nullptr) {
+          TF_LITE_REPORT_ERROR(error_reporter,
+                               "boundaries.data() returned nullptr for "
+                               "operation 'bucketize'.\n");
+          return kTfLiteError;
+        }
+        params->boundaries = boundaries->data();
+      }
+      *builtin_data = params.release();
+      return kTfLiteOk;
+    }
+    case BuiltinOperator_RANDOM_UNIFORM: {
+      auto params = safe_allocator.Allocate<TfLiteRandomParams>();
+      TF_LITE_ENSURE(error_reporter, params != nullptr);
+      if (const auto* random_uniform_params =
+              op->builtin_options_as_RandomOptions()) {
+        params->seed = random_uniform_params->seed();
+        params->seed2 = random_uniform_params->seed2();
+      }
+      *builtin_data = params.release();
+      return kTfLiteOk;
+    }
+    case BuiltinOperator_GELU: {
+      auto params = safe_allocator.Allocate<TfLiteGeluParams>();
+      TF_LITE_ENSURE(error_reporter, params != nullptr);
+      if (const auto* gelu_params = op->builtin_options_as_GeluOptions()) {
+        params->approximate = gelu_params->approximate();
       }
       *builtin_data = params.release();
       return kTfLiteOk;
@@ -844,8 +891,6 @@ TfLiteStatus ParseOpDataTfLite(const Operator* op, BuiltinOperator op_type,
     case BuiltinOperator_HASHTABLE_FIND:
     case BuiltinOperator_HASHTABLE_IMPORT:
     case BuiltinOperator_HASHTABLE_SIZE:
-    case BuiltinOperator_READ_VARIABLE:
-    case BuiltinOperator_ASSIGN_VARIABLE:
     case BuiltinOperator_BROADCAST_ARGS:
       return kTfLiteOk;
     case BuiltinOperator_PLACEHOLDER_FOR_GREATER_OP_CODES:
@@ -1003,6 +1048,14 @@ TfLiteStatus ParseArgMin(const Operator* op, ErrorReporter* error_reporter,
   return kTfLiteOk;
 }
 
+// We have this parse function instead of directly returning kTfLiteOk from the
+// switch-case in ParseOpData because this function is used as part of the
+// selective registration for the OpResolver implementation in micro.
+TfLiteStatus ParseAssignVariable(const Operator*, ErrorReporter*,
+                                 BuiltinDataAllocator*, void**) {
+  return kTfLiteOk;
+}
+
 // We have this parse function instead of directly returning kTfLiteOk from the
 // switch-case in ParseOpData because this function is used as part of the
 // selective registration for the OpResolver implementation in micro.
@@ -1032,6 +1085,33 @@ TfLiteStatus ParseBatchToSpaceNd(const Operator*, ErrorReporter*,
   return kTfLiteOk;
 }
 
+TfLiteStatus ParseCallOnce(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data) {
+  CheckParsePointerParams(op, error_reporter, allocator, builtin_data);
+
+  SafeBuiltinDataAllocator safe_allocator(allocator);
+  std::unique_ptr<TfLiteCallOnceParams,
+                  SafeBuiltinDataAllocator::BuiltinDataDeleter>
+      params = safe_allocator.Allocate<TfLiteCallOnceParams>();
+  TF_LITE_ENSURE(error_reporter, params != nullptr);
+
+  const CallOnceOptions* schema_params =
+      op->builtin_options_as_CallOnceOptions();
+
+  if (schema_params != nullptr) {
+    params->init_subgraph_index = schema_params->init_subgraph_index();
+
+  } else {
+    // TODO(b/157480169): We should either return kTfLiteError or fill in some
+    // reasonable defaults in the params struct. We are not doing so until we
+    // better undertand the ramifications of changing the legacy behavior.
+  }
+
+  *builtin_data = params.release();
+  return kTfLiteOk;
+}
+
 // We have this parse function instead of directly returning kTfLiteOk from the
 // switch-case in ParseOpData because this function is used as part of the
 // selective registration for the OpResolver implementation in micro.
@@ -1541,6 +1621,32 @@ TfLiteStatus ParseMinimum(const Operator*, ErrorReporter*,
   return kTfLiteOk;
 }
 
+TfLiteStatus ParseMirrorPad(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data) {
+  CheckParsePointerParams(op, error_reporter, allocator, builtin_data);
+
+  SafeBuiltinDataAllocator safe_allocator(allocator);
+  std::unique_ptr<TfLiteMirrorPaddingParams,
+                  SafeBuiltinDataAllocator::BuiltinDataDeleter>
+      params = safe_allocator.Allocate<TfLiteMirrorPaddingParams>();
+  TF_LITE_ENSURE(error_reporter, params != nullptr);
+
+  const MirrorPadOptions* schema_params =
+      op->builtin_options_as_MirrorPadOptions();
+
+  if (schema_params != nullptr) {
+    params->mode = ConvertMirrorPadding(schema_params->mode());
+  } else {
+    // TODO(b/157480169): We should either return kTfLiteError or fill in some
+    // reasonable defaults in the params struct. We are not doing so until we
+    // better undertand the ramifications of changing the legacy behavior.
+  }
+
+  *builtin_data = params.release();
+  return kTfLiteOk;
+}
+
 TfLiteStatus ParseMul(const Operator* op, ErrorReporter* error_reporter,
                       BuiltinDataAllocator* allocator, void** builtin_data) {
   CheckParsePointerParams(op, error_reporter, allocator, builtin_data);
@@ -1676,6 +1782,14 @@ TfLiteStatus ParseQuantize(const Operator*, ErrorReporter*,
   return kTfLiteOk;
 }
 
+// We have this parse function instead of directly returning kTfLiteOk from the
+// switch-case in ParseOpData because this function is used as part of the
+// selective registration for the OpResolver implementation in micro.
+TfLiteStatus ParseReadVariable(const Operator*, ErrorReporter*,
+                               BuiltinDataAllocator*, void**) {
+  return kTfLiteOk;
+}
+
 TfLiteStatus ParseReducer(const Operator* op, ErrorReporter* error_reporter,
                           BuiltinDataAllocator* allocator,
                           void** builtin_data) {
@@ -1856,6 +1970,14 @@ TfLiteStatus ParseSin(const Operator*, ErrorReporter*, BuiltinDataAllocator*,
   return kTfLiteOk;
 }
 
+// We have this parse function instead of directly returning kTfLiteOk from the
+// switch-case in ParseOpData because this function is used as part of the
+// selective registration for the OpResolver implementation in micro.
+TfLiteStatus ParseSlice(const Operator*, ErrorReporter*, BuiltinDataAllocator*,
+                        void**) {
+  return kTfLiteOk;
+}
+
 TfLiteStatus ParseSoftmax(const Operator* op, ErrorReporter* error_reporter,
                           BuiltinDataAllocator* allocator,
                           void** builtin_data) {
@@ -1962,6 +2084,29 @@ TfLiteStatus ParseSplitV(const Operator* op, ErrorReporter* error_reporter,
   return kTfLiteOk;
 }
 
+TfLiteStatus ParseUnidirectionalSequenceLSTM(const Operator* op,
+                                             ErrorReporter* error_reporter,
+                                             BuiltinDataAllocator* allocator,
+                                             void** builtin_data) {
+  CheckParsePointerParams(op, error_reporter, allocator, builtin_data);
+  SafeBuiltinDataAllocator safe_allocator(allocator);
+  auto params =
+      safe_allocator.Allocate<TfLiteUnidirectionalSequenceLSTMParams>();
+  TF_LITE_ENSURE(error_reporter, params != nullptr);
+  if (const auto* seq_lstm_params =
+          op->builtin_options_as_UnidirectionalSequenceLSTMOptions()) {
+    params->activation =
+        ConvertActivation(seq_lstm_params->fused_activation_function());
+    params->cell_clip = seq_lstm_params->cell_clip();
+    params->proj_clip = seq_lstm_params->proj_clip();
+    params->time_major = seq_lstm_params->time_major();
+    params->asymmetric_quantize_inputs =
+        seq_lstm_params->asymmetric_quantize_inputs();
+  }
+  *builtin_data = params.release();
+  return kTfLiteOk;
+}
+
 TfLiteStatus ParseSqueeze(const Operator* op, ErrorReporter* error_reporter,
                           BuiltinDataAllocator* allocator,
                           void** builtin_data) {
@@ -2161,6 +2306,37 @@ TfLiteStatus ParseUnpack(const Operator* op, ErrorReporter* error_reporter,
   return kTfLiteOk;
 }
 
+TfLiteStatus ParseVarHandle(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data) {
+  CheckParsePointerParams(op, error_reporter, allocator, builtin_data);
+
+  SafeBuiltinDataAllocator safe_allocator(allocator);
+  std::unique_ptr<TfLiteVarHandleParams,
+                  SafeBuiltinDataAllocator::BuiltinDataDeleter>
+      params = safe_allocator.Allocate<TfLiteVarHandleParams>();
+  TF_LITE_ENSURE(error_reporter, params != nullptr);
+
+  const VarHandleOptions* schema_params =
+      op->builtin_options_as_VarHandleOptions();
+
+  if (schema_params != nullptr) {
+    if (schema_params->container()) {
+      params->container = schema_params->container()->c_str();
+    }
+    if (schema_params->shared_name()) {
+      params->shared_name = schema_params->shared_name()->c_str();
+    }
+  } else {
+    // TODO(b/157480169): We should either return kTfLiteError or fill in some
+    // reasonable defaults in the params struct. We are not doing so until we
+    // better undertand the ramifications of changing the legacy behavior.
+  }
+
+  *builtin_data = params.release();
+  return kTfLiteOk;
+}
+
 // We have this parse function instead of directly returning kTfLiteOk from the
 // switch-case in ParseOpData because this function is used as part of the
 // selective registration for the OpResolver implementation in micro.

code/components/tflite-lib/tensorflow/lite/core/api/flatbuffer_conversions.h

@@ -1,4 +1,4 @@
-/* Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2021 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.
@@ -84,6 +84,11 @@ TfLiteStatus ParseArgMax(const Operator* op, ErrorReporter* error_reporter,
 TfLiteStatus ParseArgMin(const Operator* op, ErrorReporter* error_reporter,
                          BuiltinDataAllocator* allocator, void** builtin_data);
 
+TfLiteStatus ParseAssignVariable(const Operator* op,
+                                 ErrorReporter* error_reporter,
+                                 BuiltinDataAllocator* allocator,
+                                 void** builtin_data);
+
 TfLiteStatus ParseBatchMatMul(const Operator* op, ErrorReporter* error_reporter,
                               BuiltinDataAllocator* allocator,
                               void** builtin_data);
@@ -93,6 +98,10 @@ TfLiteStatus ParseBatchToSpaceNd(const Operator* op,
                                  BuiltinDataAllocator* allocator,
                                  void** builtin_data);
 
+TfLiteStatus ParseCallOnce(const Operator* op, ErrorReporter* error_reporter,
+                           BuiltinDataAllocator* allocator,
+                           void** builtin_data);
+
 TfLiteStatus ParseCeil(const Operator* op, ErrorReporter* error_reporter,
                        BuiltinDataAllocator* allocator, void** builtin_data);
 
@@ -229,6 +238,10 @@ TfLiteStatus ParseMaximum(const Operator* op, ErrorReporter* error_reporter,
 TfLiteStatus ParseMinimum(const Operator* op, ErrorReporter* error_reporter,
                           BuiltinDataAllocator* allocator, void** builtin_data);
 
+TfLiteStatus ParseMirrorPad(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data);
+
 TfLiteStatus ParseMul(const Operator* op, ErrorReporter* error_reporter,
                       BuiltinDataAllocator* allocator, void** builtin_data);
 
@@ -261,6 +274,11 @@ TfLiteStatus ParseQuantize(const Operator* op, ErrorReporter* error_reporter,
                            BuiltinDataAllocator* allocator,
                            void** builtin_data);
 
+TfLiteStatus ParseReadVariable(const Operator* op,
+                               ErrorReporter* error_reporter,
+                               BuiltinDataAllocator* allocator,
+                               void** builtin_data);
+
 TfLiteStatus ParseReducer(const Operator* op, ErrorReporter* error_reporter,
                           BuiltinDataAllocator* allocator, void** builtin_data);
 
@@ -295,6 +313,9 @@ TfLiteStatus ParseShape(const Operator* op, ErrorReporter* error_reporter,
 TfLiteStatus ParseSin(const Operator* op, ErrorReporter* error_reporter,
                       BuiltinDataAllocator* allocator, void** builtin_data);
 
+TfLiteStatus ParseSlice(const Operator* op, ErrorReporter* error_reporter,
+                        BuiltinDataAllocator* allocator, void** builtin_data);
+
 TfLiteStatus ParseSoftmax(const Operator* op, ErrorReporter* error_reporter,
                           BuiltinDataAllocator* allocator, void** builtin_data);
 
@@ -349,6 +370,15 @@ TfLiteStatus ParseTransposeConv(const Operator* op,
 TfLiteStatus ParseUnpack(const Operator* op, ErrorReporter* error_reporter,
                          BuiltinDataAllocator* allocator, void** builtin_data);
 
+TfLiteStatus ParseUnidirectionalSequenceLSTM(const Operator* op,
+                                             ErrorReporter* error_reporter,
+                                             BuiltinDataAllocator* allocator,
+                                             void** builtin_data);
+
+TfLiteStatus ParseVarHandle(const Operator* op, ErrorReporter* error_reporter,
+                            BuiltinDataAllocator* allocator,
+                            void** builtin_data);
+
 TfLiteStatus ParseZerosLike(const Operator* op, ErrorReporter* error_reporter,
                             BuiltinDataAllocator* allocator,
                             void** builtin_data);

code/components/tflite-lib/tensorflow/lite/core/api/op_resolver.h

@@ -15,6 +15,7 @@ limitations under the License.
 #ifndef TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_
 #define TENSORFLOW_LITE_CORE_API_OP_RESOLVER_H_
 
+#include <functional>
 #include <memory>
 #include <vector>
 
@@ -36,15 +37,26 @@ class OpResolver {
   virtual const TfLiteRegistration* FindOp(const char* op,
                                            int version) const = 0;
 
+  using TfLiteDelegatePtrVector =
+      std::vector<std::unique_ptr<TfLiteDelegate, void (*)(TfLiteDelegate*)>>;
   // Returns optional delegates for resolving and handling ops in the flatbuffer
   // model. This may be used in addition to the standard TfLiteRegistration
   // lookup for graph resolution.
-  using TfLiteDelegatePtrVector =
+  // WARNING: This API is deprecated, GetDelegateCreators is preferred.
-      std::vector<std::unique_ptr<TfLiteDelegate, void (*)(TfLiteDelegate*)>>;
   virtual TfLiteDelegatePtrVector GetDelegates(int num_threads) const {
-    return TfLiteDelegatePtrVector();
+    return {};
   }
 
+  // Represent a function that creates a TfLite delegate instance.
+  using TfLiteDelegateCreator =
+      std::function<std::unique_ptr<TfLiteDelegate, void (*)(TfLiteDelegate*)>(
+          int /*num_threads*/)>;
+  using TfLiteDelegateCreators = std::vector<TfLiteDelegateCreator>;
+  // Returns a vector of delegate creators to create optional delegates for
+  // resolving and handling ops in the flatbuffer model. This may be used in
+  // addition to the standard TfLiteRegistration lookup for graph resolution.
+  virtual TfLiteDelegateCreators GetDelegateCreators() const { return {}; }
+
   virtual ~OpResolver() {}
 
  private:

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/bits.h

@@ -0,0 +1,102 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_
+
+#ifdef __cplusplus
+#include <cstdint>
+
+extern "C" {
+#endif
+
+static inline int CountLeadingZeros32Slow(uint64_t n) {
+  int zeroes = 28;
+  if (n >> 16) zeroes -= 16, n >>= 16;
+  if (n >> 8) zeroes -= 8, n >>= 8;
+  if (n >> 4) zeroes -= 4, n >>= 4;
+  return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[n] + zeroes;
+}
+
+static inline int CountLeadingZeros32(uint32_t n) {
+#if defined(_MSC_VER)
+  unsigned long result = 0;  // NOLINT(runtime/int)
+  if (_BitScanReverse(&result, n)) {
+    return 31 - result;
+  }
+  return 32;
+#elif defined(__GNUC__)
+
+  // Handle 0 as a special case because __builtin_clz(0) is undefined.
+  if (n == 0) {
+    return 32;
+  }
+  return __builtin_clz(n);
+#else
+  return CountLeadingZeros32Slow(n);
+#endif
+}
+
+static inline int MostSignificantBit32(uint32_t n) {
+  return 32 - CountLeadingZeros32(n);
+}
+
+static inline int CountLeadingZeros64Slow(uint64_t n) {
+  int zeroes = 60;
+  if (n >> 32) zeroes -= 32, n >>= 32;
+  if (n >> 16) zeroes -= 16, n >>= 16;
+  if (n >> 8) zeroes -= 8, n >>= 8;
+  if (n >> 4) zeroes -= 4, n >>= 4;
+  return "\4\3\2\2\1\1\1\1\0\0\0\0\0\0\0"[n] + zeroes;
+}
+
+static inline int CountLeadingZeros64(uint64_t n) {
+#if defined(_MSC_VER) && defined(_M_X64)
+  // MSVC does not have __builtin_clzll. Use _BitScanReverse64.
+  unsigned long result = 0;  // NOLINT(runtime/int)
+  if (_BitScanReverse64(&result, n)) {
+    return 63 - result;
+  }
+  return 64;
+#elif defined(_MSC_VER)
+  // MSVC does not have __builtin_clzll. Compose two calls to _BitScanReverse
+  unsigned long result = 0;  // NOLINT(runtime/int)
+  if ((n >> 32) && _BitScanReverse(&result, n >> 32)) {
+    return 31 - result;
+  }
+  if (_BitScanReverse(&result, n)) {
+    return 63 - result;
+  }
+  return 64;
+#elif defined(__GNUC__)
+
+  // Handle 0 as a special case because __builtin_clzll(0) is undefined.
+  if (n == 0) {
+    return 64;
+  }
+  return __builtin_clzll(n);
+#else
+  return CountLeadingZeros64Slow(n);
+#endif
+}
+
+static inline int MostSignificantBit64(uint64_t n) {
+  return 64 - CountLeadingZeros64(n);
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_BITS_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/fft.cc

@@ -0,0 +1,52 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/fft.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/kiss_fft_int16.h"
+
+#include <string.h>
+
+
+void FftCompute(struct FftState* state, const int16_t* input,
+                int input_scale_shift) {
+  const size_t input_size = state->input_size;
+  const size_t fft_size = state->fft_size;
+
+  int16_t* fft_input = state->input;
+  // First, scale the input by the given shift.
+  size_t i;
+  for (i = 0; i < input_size; ++i) {
+    fft_input[i] = static_cast<int16_t>(static_cast<uint16_t>(input[i])
+                                        << input_scale_shift);
+  }
+  // Zero out whatever else remains in the top part of the input.
+  for (; i < fft_size; ++i) {
+    fft_input[i] = 0;
+  }
+
+  // Apply the FFT.
+  kissfft_fixed16::kiss_fftr(
+    reinterpret_cast<kissfft_fixed16::kiss_fftr_cfg>(state->scratch),
+    state->input,
+    reinterpret_cast<kissfft_fixed16::kiss_fft_cpx*>(state->output));
+}
+
+void FftInit(struct FftState* state) {
+  // All the initialization is done in FftPopulateState()
+}
+
+void FftReset(struct FftState* state) {
+  memset(state->input, 0, state->fft_size * sizeof(*state->input));
+  memset(state->output, 0, (state->fft_size / 2 + 1) * sizeof(*state->output));
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/fft.h

@@ -0,0 +1,50 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct complex_int16_t {
+  int16_t real;
+  int16_t imag;
+};
+
+struct FftState {
+  int16_t* input;
+  struct complex_int16_t* output;
+  size_t fft_size;
+  size_t input_size;
+  void* scratch;
+  size_t scratch_size;
+};
+
+void FftCompute(struct FftState* state, const int16_t* input,
+                int input_scale_shift);
+
+void FftInit(struct FftState* state);
+
+void FftReset(struct FftState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/fft_util.cc

@@ -0,0 +1,69 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/fft_util.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/kiss_fft_int16.h"
+
+#include <stdio.h>
+
+int FftPopulateState(struct FftState* state, size_t input_size) {
+  state->input_size = input_size;
+  state->fft_size = 1;
+  while (state->fft_size < state->input_size) {
+    state->fft_size <<= 1;
+  }
+
+  state->input = reinterpret_cast<int16_t*>(
+      malloc(state->fft_size * sizeof(*state->input)));
+  if (state->input == nullptr) {
+    fprintf(stderr, "Failed to alloc fft input buffer\n");
+    return 0;
+  }
+
+  state->output = reinterpret_cast<complex_int16_t*>(
+      malloc((state->fft_size / 2 + 1) * sizeof(*state->output) * 2));
+  if (state->output == nullptr) {
+    fprintf(stderr, "Failed to alloc fft output buffer\n");
+    return 0;
+  }
+
+  // Ask kissfft how much memory it wants.
+  size_t scratch_size = 0;
+  kissfft_fixed16::kiss_fftr_cfg kfft_cfg = kissfft_fixed16::kiss_fftr_alloc(
+      state->fft_size, 0, nullptr, &scratch_size);
+  if (kfft_cfg != nullptr) {
+    fprintf(stderr, "Kiss memory sizing failed.\n");
+    return 0;
+  }
+  state->scratch = malloc(scratch_size);
+  if (state->scratch == nullptr) {
+    fprintf(stderr, "Failed to alloc fft scratch buffer\n");
+    return 0;
+  }
+  state->scratch_size = scratch_size;
+  // Let kissfft configure the scratch space we just allocated
+  kfft_cfg = kissfft_fixed16::kiss_fftr_alloc(state->fft_size, 0,
+                                              state->scratch, &scratch_size);
+  if (kfft_cfg != state->scratch) {
+    fprintf(stderr, "Kiss memory preallocation strategy failed.\n");
+    return 0;
+  }
+  return 1;
+}
+
+void FftFreeStateContents(struct FftState* state) {
+  free(state->input);
+  free(state->output);
+  free(state->scratch);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/fft_util.h

@@ -0,0 +1,34 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/fft.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Prepares and FFT for the given input size.
+int FftPopulateState(struct FftState* state, size_t input_size);
+
+// Frees any allocated buffers.
+void FftFreeStateContents(struct FftState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FFT_UTIL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/filterbank.c

@@ -0,0 +1,134 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/filterbank.h"
+
+#include <string.h>
+
+#include "tensorflow/lite/experimental/microfrontend/lib/bits.h"
+
+void FilterbankConvertFftComplexToEnergy(struct FilterbankState* state,
+                                         struct complex_int16_t* fft_output,
+                                         int32_t* energy) {
+  const int end_index = state->end_index;
+  int i;
+  energy += state->start_index;
+  fft_output += state->start_index;
+  for (i = state->start_index; i < end_index; ++i) {
+    const int32_t real = fft_output->real;
+    const int32_t imag = fft_output->imag;
+    fft_output++;
+    const uint32_t mag_squared = (real * real) + (imag * imag);
+    *energy++ = mag_squared;
+  }
+}
+
+void FilterbankAccumulateChannels(struct FilterbankState* state,
+                                  const int32_t* energy) {
+  uint64_t* work = state->work;
+  uint64_t weight_accumulator = 0;
+  uint64_t unweight_accumulator = 0;
+
+  const int16_t* channel_frequency_starts = state->channel_frequency_starts;
+  const int16_t* channel_weight_starts = state->channel_weight_starts;
+  const int16_t* channel_widths = state->channel_widths;
+
+  int num_channels_plus_1 = state->num_channels + 1;
+  int i;
+  for (i = 0; i < num_channels_plus_1; ++i) {
+    const int32_t* magnitudes = energy + *channel_frequency_starts++;
+    const int16_t* weights = state->weights + *channel_weight_starts;
+    const int16_t* unweights = state->unweights + *channel_weight_starts++;
+    const int width = *channel_widths++;
+    int j;
+    for (j = 0; j < width; ++j) {
+      weight_accumulator += *weights++ * ((uint64_t)*magnitudes);
+      unweight_accumulator += *unweights++ * ((uint64_t)*magnitudes);
+      ++magnitudes;
+    }
+    *work++ = weight_accumulator;
+    weight_accumulator = unweight_accumulator;
+    unweight_accumulator = 0;
+  }
+}
+
+static uint16_t Sqrt32(uint32_t num) {
+  if (num == 0) {
+    return 0;
+  }
+  uint32_t res = 0;
+  int max_bit_number = 32 - MostSignificantBit32(num);
+  max_bit_number |= 1;
+  uint32_t bit = 1U << (31 - max_bit_number);
+  int iterations = (31 - max_bit_number) / 2 + 1;
+  while (iterations--) {
+    if (num >= res + bit) {
+      num -= res + bit;
+      res = (res >> 1U) + bit;
+    } else {
+      res >>= 1U;
+    }
+    bit >>= 2U;
+  }
+  // Do rounding - if we have the bits.
+  if (num > res && res != 0xFFFF) {
+    ++res;
+  }
+  return res;
+}
+
+static uint32_t Sqrt64(uint64_t num) {
+  // Take a shortcut and just use 32 bit operations if the upper word is all
+  // clear. This will cause a slight off by one issue for numbers close to 2^32,
+  // but it probably isn't going to matter (and gives us a big performance win).
+  if ((num >> 32) == 0) {
+    return Sqrt32((uint32_t)num);
+  }
+  uint64_t res = 0;
+  int max_bit_number = 64 - MostSignificantBit64(num);
+  max_bit_number |= 1;
+  uint64_t bit = 1ULL << (63 - max_bit_number);
+  int iterations = (63 - max_bit_number) / 2 + 1;
+  while (iterations--) {
+    if (num >= res + bit) {
+      num -= res + bit;
+      res = (res >> 1U) + bit;
+    } else {
+      res >>= 1U;
+    }
+    bit >>= 2U;
+  }
+  // Do rounding - if we have the bits.
+  if (num > res && res != 0xFFFFFFFFLL) {
+    ++res;
+  }
+  return res;
+}
+
+uint32_t* FilterbankSqrt(struct FilterbankState* state, int scale_down_shift) {
+  const int num_channels = state->num_channels;
+  const uint64_t* work = state->work + 1;
+  // Reuse the work buffer since we're fine clobbering it at this point to hold
+  // the output.
+  uint32_t* output = (uint32_t*)state->work;
+  int i;
+  for (i = 0; i < num_channels; ++i) {
+    *output++ = Sqrt64(*work++) >> scale_down_shift;
+  }
+  return (uint32_t*)state->work;
+}
+
+void FilterbankReset(struct FilterbankState* state) {
+  memset(state->work, 0, (state->num_channels + 1) * sizeof(*state->work));
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/filterbank.h

@@ -0,0 +1,63 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#include "tensorflow/lite/experimental/microfrontend/lib/fft.h"
+
+#define kFilterbankBits 12
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct FilterbankState {
+  int num_channels;
+  int start_index;
+  int end_index;
+  int16_t* channel_frequency_starts;
+  int16_t* channel_weight_starts;
+  int16_t* channel_widths;
+  int16_t* weights;
+  int16_t* unweights;
+  uint64_t* work;
+};
+
+// Converts the relevant complex values of an FFT output into energy (the
+// square magnitude).
+void FilterbankConvertFftComplexToEnergy(struct FilterbankState* state,
+                                         struct complex_int16_t* fft_output,
+                                         int32_t* energy);
+
+// Computes the mel-scale filterbank on the given energy array. Output is cached
+// internally - to fetch it, you need to call FilterbankSqrt.
+void FilterbankAccumulateChannels(struct FilterbankState* state,
+                                  const int32_t* energy);
+
+// Applies an integer square root to the 64 bit intermediate values of the
+// filterbank, and returns a pointer to them. Memory will be invalidated the
+// next time FilterbankAccumulateChannels is called.
+uint32_t* FilterbankSqrt(struct FilterbankState* state, int scale_down_shift);
+
+void FilterbankReset(struct FilterbankState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.c

@@ -0,0 +1,220 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h"
+
+#include <assert.h>
+#include <math.h>
+#include <stdio.h>
+
+#define kFilterbankIndexAlignment 4
+#define kFilterbankChannelBlockSize 4
+
+void FilterbankFillConfigWithDefaults(struct FilterbankConfig* config) {
+  config->num_channels = 32;
+  config->lower_band_limit = 125.0f;
+  config->upper_band_limit = 7500.0f;
+  config->output_scale_shift = 7;
+}
+
+static float FreqToMel(float freq) { return 1127.0 * log1p(freq / 700.0); }
+
+static void CalculateCenterFrequencies(const int num_channels,
+                                       const float lower_frequency_limit,
+                                       const float upper_frequency_limit,
+                                       float* center_frequencies) {
+  assert(lower_frequency_limit >= 0.0f);
+  assert(upper_frequency_limit > lower_frequency_limit);
+
+  const float mel_low = FreqToMel(lower_frequency_limit);
+  const float mel_hi = FreqToMel(upper_frequency_limit);
+  const float mel_span = mel_hi - mel_low;
+  const float mel_spacing = mel_span / ((float)num_channels);
+  int i;
+  for (i = 0; i < num_channels; ++i) {
+    center_frequencies[i] = mel_low + (mel_spacing * (i + 1));
+  }
+}
+
+static void QuantizeFilterbankWeights(const float float_weight, int16_t* weight,
+                                      int16_t* unweight) {
+  *weight = floor(float_weight * (1 << kFilterbankBits) + 0.5);
+  *unweight = floor((1.0 - float_weight) * (1 << kFilterbankBits) + 0.5);
+}
+
+int FilterbankPopulateState(const struct FilterbankConfig* config,
+                            struct FilterbankState* state, int sample_rate,
+                            int spectrum_size) {
+  state->num_channels = config->num_channels;
+  const int num_channels_plus_1 = config->num_channels + 1;
+
+  // How should we align things to index counts given the byte alignment?
+  const int index_alignment =
+      (kFilterbankIndexAlignment < sizeof(int16_t)
+           ? 1
+           : kFilterbankIndexAlignment / sizeof(int16_t));
+
+  state->channel_frequency_starts =
+      malloc(num_channels_plus_1 * sizeof(*state->channel_frequency_starts));
+  state->channel_weight_starts =
+      malloc(num_channels_plus_1 * sizeof(*state->channel_weight_starts));
+  state->channel_widths =
+      malloc(num_channels_plus_1 * sizeof(*state->channel_widths));
+  state->work = malloc(num_channels_plus_1 * sizeof(*state->work));
+
+  float* center_mel_freqs =
+      malloc(num_channels_plus_1 * sizeof(*center_mel_freqs));
+  int16_t* actual_channel_starts =
+      malloc(num_channels_plus_1 * sizeof(*actual_channel_starts));
+  int16_t* actual_channel_widths =
+      malloc(num_channels_plus_1 * sizeof(*actual_channel_widths));
+
+  if (state->channel_frequency_starts == NULL ||
+      state->channel_weight_starts == NULL || state->channel_widths == NULL ||
+      center_mel_freqs == NULL || actual_channel_starts == NULL ||
+      actual_channel_widths == NULL) {
+    free(center_mel_freqs);
+    free(actual_channel_starts);
+    free(actual_channel_widths);
+    fprintf(stderr, "Failed to allocate channel buffers\n");
+    return 0;
+  }
+
+  CalculateCenterFrequencies(num_channels_plus_1, config->lower_band_limit,
+                             config->upper_band_limit, center_mel_freqs);
+
+  // Always exclude DC.
+  const float hz_per_sbin = 0.5 * sample_rate / ((float)spectrum_size - 1);
+  state->start_index = 1.5 + config->lower_band_limit / hz_per_sbin;
+  state->end_index = 0;  // Initialized to zero here, but actually set below.
+
+  // For each channel, we need to figure out what frequencies belong to it, and
+  // how much padding we need to add so that we can efficiently multiply the
+  // weights and unweights for accumulation. To simplify the multiplication
+  // logic, all channels will have some multiplication to do (even if there are
+  // no frequencies that accumulate to that channel) - they will be directed to
+  // a set of zero weights.
+  int chan_freq_index_start = state->start_index;
+  int weight_index_start = 0;
+  int needs_zeros = 0;
+
+  int chan;
+  for (chan = 0; chan < num_channels_plus_1; ++chan) {
+    // Keep jumping frequencies until we overshoot the bound on this channel.
+    int freq_index = chan_freq_index_start;
+    while (FreqToMel((freq_index)*hz_per_sbin) <= center_mel_freqs[chan]) {
+      ++freq_index;
+    }
+
+    const int width = freq_index - chan_freq_index_start;
+    actual_channel_starts[chan] = chan_freq_index_start;
+    actual_channel_widths[chan] = width;
+
+    if (width == 0) {
+      // This channel doesn't actually get anything from the frequencies, it's
+      // always zero. We need then to insert some 'zero' weights into the
+      // output, and just redirect this channel to do a single multiplication at
+      // this point. For simplicity, the zeros are placed at the beginning of
+      // the weights arrays, so we have to go and update all the other
+      // weight_starts to reflect this shift (but only once).
+      state->channel_frequency_starts[chan] = 0;
+      state->channel_weight_starts[chan] = 0;
+      state->channel_widths[chan] = kFilterbankChannelBlockSize;
+      if (!needs_zeros) {
+        needs_zeros = 1;
+        int j;
+        for (j = 0; j < chan; ++j) {
+          state->channel_weight_starts[j] += kFilterbankChannelBlockSize;
+        }
+        weight_index_start += kFilterbankChannelBlockSize;
+      }
+    } else {
+      // How far back do we need to go to ensure that we have the proper
+      // alignment?
+      const int aligned_start =
+          (chan_freq_index_start / index_alignment) * index_alignment;
+      const int aligned_width = (chan_freq_index_start - aligned_start + width);
+      const int padded_width =
+          (((aligned_width - 1) / kFilterbankChannelBlockSize) + 1) *
+          kFilterbankChannelBlockSize;
+
+      state->channel_frequency_starts[chan] = aligned_start;
+      state->channel_weight_starts[chan] = weight_index_start;
+      state->channel_widths[chan] = padded_width;
+      weight_index_start += padded_width;
+    }
+    chan_freq_index_start = freq_index;
+  }
+
+  // Allocate the two arrays to store the weights - weight_index_start contains
+  // the index of what would be the next set of weights that we would need to
+  // add, so that's how many weights we need to allocate.
+  state->weights = calloc(weight_index_start, sizeof(*state->weights));
+  state->unweights = calloc(weight_index_start, sizeof(*state->unweights));
+
+  // If the alloc failed, we also need to nuke the arrays.
+  if (state->weights == NULL || state->unweights == NULL) {
+    free(center_mel_freqs);
+    free(actual_channel_starts);
+    free(actual_channel_widths);
+    fprintf(stderr, "Failed to allocate weights or unweights\n");
+    return 0;
+  }
+
+  // Next pass, compute all the weights. Since everything has been memset to
+  // zero, we only need to fill in the weights that correspond to some frequency
+  // for a channel.
+  const float mel_low = FreqToMel(config->lower_band_limit);
+  for (chan = 0; chan < num_channels_plus_1; ++chan) {
+    int frequency = actual_channel_starts[chan];
+    const int num_frequencies = actual_channel_widths[chan];
+    const int frequency_offset =
+        frequency - state->channel_frequency_starts[chan];
+    const int weight_start = state->channel_weight_starts[chan];
+    const float denom_val = (chan == 0) ? mel_low : center_mel_freqs[chan - 1];
+
+    int j;
+    for (j = 0; j < num_frequencies; ++j, ++frequency) {
+      const float weight =
+          (center_mel_freqs[chan] - FreqToMel(frequency * hz_per_sbin)) /
+          (center_mel_freqs[chan] - denom_val);
+
+      // Make the float into an integer for the weights (and unweights).
+      const int weight_index = weight_start + frequency_offset + j;
+      QuantizeFilterbankWeights(weight, state->weights + weight_index,
+                                state->unweights + weight_index);
+    }
+    if (frequency > state->end_index) {
+      state->end_index = frequency;
+    }
+  }
+
+  free(center_mel_freqs);
+  free(actual_channel_starts);
+  free(actual_channel_widths);
+  if (state->end_index >= spectrum_size) {
+    fprintf(stderr, "Filterbank end_index is above spectrum size.\n");
+    return 0;
+  }
+  return 1;
+}
+
+void FilterbankFreeStateContents(struct FilterbankState* state) {
+  free(state->channel_frequency_starts);
+  free(state->channel_weight_starts);
+  free(state->channel_widths);
+  free(state->weights);
+  free(state->unweights);
+  free(state->work);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h

@@ -0,0 +1,50 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/filterbank.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct FilterbankConfig {
+  // number of frequency channel buckets for filterbank
+  int num_channels;
+  // maximum frequency to include
+  float upper_band_limit;
+  // minimum frequency to include
+  float lower_band_limit;
+  // unused
+  int output_scale_shift;
+};
+
+// Fills the frontendConfig with "sane" defaults.
+void FilterbankFillConfigWithDefaults(struct FilterbankConfig* config);
+
+// Allocates any buffers.
+int FilterbankPopulateState(const struct FilterbankConfig* config,
+                            struct FilterbankState* state, int sample_rate,
+                            int spectrum_size);
+
+// Frees any allocated buffers.
+void FilterbankFreeStateContents(struct FilterbankState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FILTERBANK_UTIL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/frontend.c

@@ -0,0 +1,72 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/frontend.h"
+
+#include "tensorflow/lite/experimental/microfrontend/lib/bits.h"
+
+struct FrontendOutput FrontendProcessSamples(struct FrontendState* state,
+                                             const int16_t* samples,
+                                             size_t num_samples,
+                                             size_t* num_samples_read) {
+  struct FrontendOutput output;
+  output.values = NULL;
+  output.size = 0;
+
+  // Try to apply the window - if it fails, return and wait for more data.
+  if (!WindowProcessSamples(&state->window, samples, num_samples,
+                            num_samples_read)) {
+    return output;
+  }
+
+  // Apply the FFT to the window's output (and scale it so that the fixed point
+  // FFT can have as much resolution as possible).
+  int input_shift =
+      15 - MostSignificantBit32(state->window.max_abs_output_value);
+  FftCompute(&state->fft, state->window.output, input_shift);
+
+  // We can re-ruse the fft's output buffer to hold the energy.
+  int32_t* energy = (int32_t*)state->fft.output;
+
+  FilterbankConvertFftComplexToEnergy(&state->filterbank, state->fft.output,
+                                      energy);
+
+  FilterbankAccumulateChannels(&state->filterbank, energy);
+  uint32_t* scaled_filterbank = FilterbankSqrt(&state->filterbank, input_shift);
+
+  // Apply noise reduction.
+  NoiseReductionApply(&state->noise_reduction, scaled_filterbank);
+
+  if (state->pcan_gain_control.enable_pcan) {
+    PcanGainControlApply(&state->pcan_gain_control, scaled_filterbank);
+  }
+
+  // Apply the log and scale.
+  int correction_bits =
+      MostSignificantBit32(state->fft.fft_size) - 1 - (kFilterbankBits / 2);
+  uint16_t* logged_filterbank =
+      LogScaleApply(&state->log_scale, scaled_filterbank,
+                    state->filterbank.num_channels, correction_bits);
+
+  output.size = state->filterbank.num_channels;
+  output.values = logged_filterbank;
+  return output;
+}
+
+void FrontendReset(struct FrontendState* state) {
+  WindowReset(&state->window);
+  FftReset(&state->fft);
+  FilterbankReset(&state->filterbank);
+  NoiseReductionReset(&state->noise_reduction);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/frontend.h

@@ -0,0 +1,64 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#include "tensorflow/lite/experimental/microfrontend/lib/fft.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/filterbank.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/log_scale.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/window.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct FrontendState {
+  struct WindowState window;
+  struct FftState fft;
+  struct FilterbankState filterbank;
+  struct NoiseReductionState noise_reduction;
+  struct PcanGainControlState pcan_gain_control;
+  struct LogScaleState log_scale;
+};
+
+struct FrontendOutput {
+  const uint16_t* values;
+  size_t size;
+};
+
+// Main entry point to processing frontend samples. Updates num_samples_read to
+// contain the number of samples that have been consumed from the input array.
+// Returns a struct containing the generated output. If not enough samples were
+// added to generate a feature vector, the returned size will be 0 and the
+// values pointer will be NULL. Note that the output pointer will be invalidated
+// as soon as FrontendProcessSamples is called again, so copy the contents
+// elsewhere if you need to use them later.
+struct FrontendOutput FrontendProcessSamples(struct FrontendState* state,
+                                             const int16_t* samples,
+                                             size_t num_samples,
+                                             size_t* num_samples_read);
+
+void FrontendReset(struct FrontendState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/frontend_util.c

@@ -0,0 +1,85 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/frontend_util.h"
+
+#include <stdio.h>
+#include <string.h>
+
+#include "tensorflow/lite/experimental/microfrontend/lib/bits.h"
+
+void FrontendFillConfigWithDefaults(struct FrontendConfig* config) {
+  WindowFillConfigWithDefaults(&config->window);
+  FilterbankFillConfigWithDefaults(&config->filterbank);
+  NoiseReductionFillConfigWithDefaults(&config->noise_reduction);
+  PcanGainControlFillConfigWithDefaults(&config->pcan_gain_control);
+  LogScaleFillConfigWithDefaults(&config->log_scale);
+}
+
+int FrontendPopulateState(const struct FrontendConfig* config,
+                          struct FrontendState* state, int sample_rate) {
+  memset(state, 0, sizeof(*state));
+
+  if (!WindowPopulateState(&config->window, &state->window, sample_rate)) {
+    fprintf(stderr, "Failed to populate window state\n");
+    return 0;
+  }
+
+  if (!FftPopulateState(&state->fft, state->window.size)) {
+    fprintf(stderr, "Failed to populate fft state\n");
+    return 0;
+  }
+  FftInit(&state->fft);
+
+  if (!FilterbankPopulateState(&config->filterbank, &state->filterbank,
+                               sample_rate, state->fft.fft_size / 2 + 1)) {
+    fprintf(stderr, "Failed to populate filterbank state\n");
+    return 0;
+  }
+
+  if (!NoiseReductionPopulateState(&config->noise_reduction,
+                                   &state->noise_reduction,
+                                   state->filterbank.num_channels)) {
+    fprintf(stderr, "Failed to populate noise reduction state\n");
+    return 0;
+  }
+
+  int input_correction_bits =
+      MostSignificantBit32(state->fft.fft_size) - 1 - (kFilterbankBits / 2);
+  if (!PcanGainControlPopulateState(
+          &config->pcan_gain_control, &state->pcan_gain_control,
+          state->noise_reduction.estimate, state->filterbank.num_channels,
+          state->noise_reduction.smoothing_bits, input_correction_bits)) {
+    fprintf(stderr, "Failed to populate pcan gain control state\n");
+    return 0;
+  }
+
+  if (!LogScalePopulateState(&config->log_scale, &state->log_scale)) {
+    fprintf(stderr, "Failed to populate log scale state\n");
+    return 0;
+  }
+
+  FrontendReset(state);
+
+  // All good, return a true value.
+  return 1;
+}
+
+void FrontendFreeStateContents(struct FrontendState* state) {
+  WindowFreeStateContents(&state->window);
+  FftFreeStateContents(&state->fft);
+  FilterbankFreeStateContents(&state->filterbank);
+  NoiseReductionFreeStateContents(&state->noise_reduction);
+  PcanGainControlFreeStateContents(&state->pcan_gain_control);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/frontend_util.h

@@ -0,0 +1,52 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/fft_util.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/filterbank_util.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/frontend.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/window_util.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct FrontendConfig {
+  struct WindowConfig window;
+  struct FilterbankConfig filterbank;
+  struct NoiseReductionConfig noise_reduction;
+  struct PcanGainControlConfig pcan_gain_control;
+  struct LogScaleConfig log_scale;
+};
+
+// Fills the frontendConfig with "sane" defaults.
+void FrontendFillConfigWithDefaults(struct FrontendConfig* config);
+
+// Allocates any buffers.
+int FrontendPopulateState(const struct FrontendConfig* config,
+                          struct FrontendState* state, int sample_rate);
+
+// Frees any allocated buffers.
+void FrontendFreeStateContents(struct FrontendState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_FRONTEND_UTIL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/kiss_fft_common.h

@@ -0,0 +1,48 @@
+/* Copyright 2019 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_KISS_FFT_COMMON_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_KISS_FFT_COMMON_H_
+
+// This header file should be included in all variants of kiss_fft_$type.{h,cc}
+// so that their sub-included source files do not mistakenly wrap libc header
+// files within their kissfft_$type namespaces.
+// E.g, This header avoids kissfft_int16.h containing:
+//   namespace kiss_fft_int16 {
+//     #include "kiss_fft.h"
+//   }
+// where kiss_fft_.h contains:
+//   #include <math.h>
+//
+// TRICK: By including the following header files here, their preprocessor
+// header guards prevent them being re-defined inside of the kiss_fft_$type
+// namespaces declared within the kiss_fft_$type.{h,cc} sources.
+// Note that the original kiss_fft*.h files are untouched since they
+// may be used in libraries that include them directly.
+
+#include <limits.h>
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#ifdef FIXED_POINT
+#include <sys/types.h>
+#endif
+
+#ifdef USE_SIMD
+#include <xmmintrin.h>
+#endif
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_KISS_FFT_COMMON_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/kiss_fft_int16.cc

@@ -0,0 +1,8 @@
+#include "tensorflow/lite/experimental/microfrontend/lib/kiss_fft_common.h"
+
+#define FIXED_POINT 16
+namespace kissfft_fixed16 {
+#include "kiss_fft.c"
+#include "tools/kiss_fftr.c"
+}  // namespace kissfft_fixed16
+#undef FIXED_POINT

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/kiss_fft_int16.h

@@ -0,0 +1,34 @@
+/* Copyright 2019 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_KISS_FFT_INT16_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_KISS_FFT_INT16_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/kiss_fft_common.h"
+
+// Wrap 16-bit kiss fft in its own namespace. Enables us to link an application
+// with different kiss fft resultions (16/32 bit interger, float, double)
+// without getting a linker error.
+#define FIXED_POINT 16
+namespace kissfft_fixed16 {
+#include "kiss_fft.h"
+#include "tools/kiss_fftr.h"
+}  // namespace kissfft_fixed16
+#undef FIXED_POINT
+#undef kiss_fft_scalar
+#undef KISS_FFT_H
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_KISS_FFT_INT16_H_
+

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/log_lut.c

@@ -0,0 +1,30 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/log_lut.h"
+const uint16_t kLogLut[]
+#ifndef _MSC_VER
+    __attribute__((aligned(4)))
+#endif  // _MSV_VER
+    = {0,    224,  442,  654,  861,  1063, 1259, 1450, 1636, 1817, 1992, 2163,
+       2329, 2490, 2646, 2797, 2944, 3087, 3224, 3358, 3487, 3611, 3732, 3848,
+       3960, 4068, 4172, 4272, 4368, 4460, 4549, 4633, 4714, 4791, 4864, 4934,
+       5001, 5063, 5123, 5178, 5231, 5280, 5326, 5368, 5408, 5444, 5477, 5507,
+       5533, 5557, 5578, 5595, 5610, 5622, 5631, 5637, 5640, 5641, 5638, 5633,
+       5626, 5615, 5602, 5586, 5568, 5547, 5524, 5498, 5470, 5439, 5406, 5370,
+       5332, 5291, 5249, 5203, 5156, 5106, 5054, 5000, 4944, 4885, 4825, 4762,
+       4697, 4630, 4561, 4490, 4416, 4341, 4264, 4184, 4103, 4020, 3935, 3848,
+       3759, 3668, 3575, 3481, 3384, 3286, 3186, 3084, 2981, 2875, 2768, 2659,
+       2549, 2437, 2323, 2207, 2090, 1971, 1851, 1729, 1605, 1480, 1353, 1224,
+       1094, 963,  830,  695,  559,  421,  282,  142,  0,    0};

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/log_lut.h

@@ -0,0 +1,40 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Number of segments in the log lookup table. The table will be kLogSegments+1
+// in length (with some padding).
+#define kLogSegments 128
+#define kLogSegmentsLog2 7
+
+// Scale used by lookup table.
+#define kLogScale 65536
+#define kLogScaleLog2 16
+#define kLogCoeff 45426
+
+extern const uint16_t kLogLut[];
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_LUT_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/log_scale.c

@@ -0,0 +1,83 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/log_scale.h"
+
+#include "tensorflow/lite/experimental/microfrontend/lib/bits.h"
+#include "tensorflow/lite/experimental/microfrontend/lib/log_lut.h"
+
+#define kuint16max 0x0000FFFF
+
+// The following functions implement integer logarithms of various sizes. The
+// approximation is calculated according to method described in
+//       www.inti.gob.ar/electronicaeinformatica/instrumentacion/utic/
+//       publicaciones/SPL2007/Log10-spl07.pdf
+// It first calculates log2 of the input and then converts it to natural
+// logarithm.
+
+static uint32_t Log2FractionPart(const uint32_t x, const uint32_t log2x) {
+  // Part 1
+  int32_t frac = x - (1LL << log2x);
+  if (log2x < kLogScaleLog2) {
+    frac <<= kLogScaleLog2 - log2x;
+  } else {
+    frac >>= log2x - kLogScaleLog2;
+  }
+  // Part 2
+  const uint32_t base_seg = frac >> (kLogScaleLog2 - kLogSegmentsLog2);
+  const uint32_t seg_unit =
+      (((uint32_t)1) << kLogScaleLog2) >> kLogSegmentsLog2;
+
+  const int32_t c0 = kLogLut[base_seg];
+  const int32_t c1 = kLogLut[base_seg + 1];
+  const int32_t seg_base = seg_unit * base_seg;
+  const int32_t rel_pos = ((c1 - c0) * (frac - seg_base)) >> kLogScaleLog2;
+  return frac + c0 + rel_pos;
+}
+
+static uint32_t Log(const uint32_t x, const uint32_t scale_shift) {
+  const uint32_t integer = MostSignificantBit32(x) - 1;
+  const uint32_t fraction = Log2FractionPart(x, integer);
+  const uint32_t log2 = (integer << kLogScaleLog2) + fraction;
+  const uint32_t round = kLogScale / 2;
+  const uint32_t loge = (((uint64_t)kLogCoeff) * log2 + round) >> kLogScaleLog2;
+  // Finally scale to our output scale
+  const uint32_t loge_scaled = ((loge << scale_shift) + round) >> kLogScaleLog2;
+  return loge_scaled;
+}
+
+uint16_t* LogScaleApply(struct LogScaleState* state, uint32_t* signal,
+                        int signal_size, int correction_bits) {
+  const int scale_shift = state->scale_shift;
+  uint16_t* output = (uint16_t*)signal;
+  uint16_t* ret = output;
+  int i;
+  for (i = 0; i < signal_size; ++i) {
+    uint32_t value = *signal++;
+    if (state->enable_log) {
+      if (correction_bits < 0) {
+        value >>= -correction_bits;
+      } else {
+        value <<= correction_bits;
+      }
+      if (value > 1) {
+        value = Log(value, scale_shift);
+      } else {
+        value = 0;
+      }
+    }
+    *output++ = (value < kuint16max) ? value : kuint16max;
+  }
+  return ret;
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/log_scale.h

@@ -0,0 +1,39 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct LogScaleState {
+  int enable_log;
+  int scale_shift;
+};
+
+// Applies a fixed point logarithm to the signal and converts it to 16 bit. Note
+// that the signal array will be modified.
+uint16_t* LogScaleApply(struct LogScaleState* state, uint32_t* signal,
+                        int signal_size, int correction_bits);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.c

@@ -1,4 +1,4 @@
-/* Copyright 2020 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2018 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.
@@ -12,11 +12,16 @@ 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 "tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h"
 
-#ifndef TENSORFLOW_LITE_MICRO_BENCHMARKS_KEYWORD_SCRAMBLED_MODEL_DATA_H_
+void LogScaleFillConfigWithDefaults(struct LogScaleConfig* config) {
-#define TENSORFLOW_LITE_MICRO_BENCHMARKS_KEYWORD_SCRAMBLED_MODEL_DATA_H_
+  config->enable_log = 1;
+  config->scale_shift = 6;
+}
 
-extern const unsigned char g_keyword_scrambled_model_data[];
+int LogScalePopulateState(const struct LogScaleConfig* config,
-extern const unsigned int g_keyword_scrambled_model_data_length;
+                          struct LogScaleState* state) {
-
+  state->enable_log = config->enable_log;
-#endif  // TENSORFLOW_LITE_MICRO_BENCHMARKS_KEYWORD_SCRAMBLED_MODEL_DATA_H_
+  state->scale_shift = config->scale_shift;
+  return 1;
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/log_scale_util.h

@@ -0,0 +1,45 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#include "tensorflow/lite/experimental/microfrontend/lib/log_scale.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct LogScaleConfig {
+  // set to false (0) to disable this module
+  int enable_log;
+  // scale results by 2^(scale_shift)
+  int scale_shift;
+};
+
+// Populates the LogScaleConfig with "sane" default values.
+void LogScaleFillConfigWithDefaults(struct LogScaleConfig* config);
+
+// Allocates any buffers.
+int LogScalePopulateState(const struct LogScaleConfig* config,
+                          struct LogScaleState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_LOG_SCALE_UTIL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.c

@@ -0,0 +1,51 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h"
+
+#include <string.h>
+
+void NoiseReductionApply(struct NoiseReductionState* state, uint32_t* signal) {
+  int i;
+  for (i = 0; i < state->num_channels; ++i) {
+    const uint32_t smoothing =
+        ((i & 1) == 0) ? state->even_smoothing : state->odd_smoothing;
+    const uint32_t one_minus_smoothing = (1 << kNoiseReductionBits) - smoothing;
+
+    // Update the estimate of the noise.
+    const uint32_t signal_scaled_up = signal[i] << state->smoothing_bits;
+    uint32_t estimate =
+        (((uint64_t)signal_scaled_up * smoothing) +
+         ((uint64_t)state->estimate[i] * one_minus_smoothing)) >>
+        kNoiseReductionBits;
+    state->estimate[i] = estimate;
+
+    // Make sure that we can't get a negative value for the signal - estimate.
+    if (estimate > signal_scaled_up) {
+      estimate = signal_scaled_up;
+    }
+
+    const uint32_t floor =
+        ((uint64_t)signal[i] * state->min_signal_remaining) >>
+        kNoiseReductionBits;
+    const uint32_t subtracted =
+        (signal_scaled_up - estimate) >> state->smoothing_bits;
+    const uint32_t output = subtracted > floor ? subtracted : floor;
+    signal[i] = output;
+  }
+}
+
+void NoiseReductionReset(struct NoiseReductionState* state) {
+  memset(state->estimate, 0, sizeof(*state->estimate) * state->num_channels);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h

@@ -0,0 +1,46 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_
+
+#define kNoiseReductionBits 14
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct NoiseReductionState {
+  int smoothing_bits;
+  uint16_t even_smoothing;
+  uint16_t odd_smoothing;
+  uint16_t min_signal_remaining;
+  int num_channels;
+  uint32_t* estimate;
+};
+
+// Removes stationary noise from each channel of the signal using a low pass
+// filter.
+void NoiseReductionApply(struct NoiseReductionState* state, uint32_t* signal);
+
+void NoiseReductionReset(struct NoiseReductionState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.c

@@ -0,0 +1,45 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h"
+
+#include <stdio.h>
+
+void NoiseReductionFillConfigWithDefaults(struct NoiseReductionConfig* config) {
+  config->smoothing_bits = 10;
+  config->even_smoothing = 0.025;
+  config->odd_smoothing = 0.06;
+  config->min_signal_remaining = 0.05;
+}
+
+int NoiseReductionPopulateState(const struct NoiseReductionConfig* config,
+                                struct NoiseReductionState* state,
+                                int num_channels) {
+  state->smoothing_bits = config->smoothing_bits;
+  state->odd_smoothing = config->odd_smoothing * (1 << kNoiseReductionBits);
+  state->even_smoothing = config->even_smoothing * (1 << kNoiseReductionBits);
+  state->min_signal_remaining =
+      config->min_signal_remaining * (1 << kNoiseReductionBits);
+  state->num_channels = num_channels;
+  state->estimate = calloc(state->num_channels, sizeof(*state->estimate));
+  if (state->estimate == NULL) {
+    fprintf(stderr, "Failed to alloc estimate buffer\n");
+    return 0;
+  }
+  return 1;
+}
+
+void NoiseReductionFreeStateContents(struct NoiseReductionState* state) {
+  free(state->estimate);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/noise_reduction_util.h

@@ -0,0 +1,50 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/noise_reduction.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct NoiseReductionConfig {
+  // scale the signal up by 2^(smoothing_bits) before reduction
+  int smoothing_bits;
+  // smoothing coefficient for even-numbered channels
+  float even_smoothing;
+  // smoothing coefficient for odd-numbered channels
+  float odd_smoothing;
+  // fraction of signal to preserve (1.0 disables this module)
+  float min_signal_remaining;
+};
+
+// Populates the NoiseReductionConfig with "sane" default values.
+void NoiseReductionFillConfigWithDefaults(struct NoiseReductionConfig* config);
+
+// Allocates any buffers.
+int NoiseReductionPopulateState(const struct NoiseReductionConfig* config,
+                                struct NoiseReductionState* state,
+                                int num_channels);
+
+// Frees any allocated buffers.
+void NoiseReductionFreeStateContents(struct NoiseReductionState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_NOISE_REDUCTION_UTIL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.c

@@ -0,0 +1,56 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h"
+
+#include "tensorflow/lite/experimental/microfrontend/lib/bits.h"
+
+int16_t WideDynamicFunction(const uint32_t x, const int16_t* lut) {
+  if (x <= 2) {
+    return lut[x];
+  }
+
+  const int16_t interval = MostSignificantBit32(x);
+  lut += 4 * interval - 6;
+
+  const int16_t frac =
+      ((interval < 11) ? (x << (11 - interval)) : (x >> (interval - 11))) &
+      0x3FF;
+
+  int32_t result = ((int32_t)lut[2] * frac) >> 5;
+  result += (int32_t)((uint32_t)lut[1] << 5);
+  result *= frac;
+  result = (result + (1 << 14)) >> 15;
+  result += lut[0];
+  return (int16_t)result;
+}
+
+uint32_t PcanShrink(const uint32_t x) {
+  if (x < (2 << kPcanSnrBits)) {
+    return (x * x) >> (2 + 2 * kPcanSnrBits - kPcanOutputBits);
+  } else {
+    return (x >> (kPcanSnrBits - kPcanOutputBits)) - (1 << kPcanOutputBits);
+  }
+}
+
+void PcanGainControlApply(struct PcanGainControlState* state,
+                          uint32_t* signal) {
+  int i;
+  for (i = 0; i < state->num_channels; ++i) {
+    const uint32_t gain =
+        WideDynamicFunction(state->noise_estimate[i], state->gain_lut);
+    const uint32_t snr = ((uint64_t)signal[i] * gain) >> state->snr_shift;
+    signal[i] = PcanShrink(snr);
+  }
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h

@@ -0,0 +1,47 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#define kPcanSnrBits 12
+#define kPcanOutputBits 6
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Details at https://research.google/pubs/pub45911.pdf
+struct PcanGainControlState {
+  int enable_pcan;
+  uint32_t* noise_estimate;
+  int num_channels;
+  int16_t* gain_lut;
+  int32_t snr_shift;
+};
+
+int16_t WideDynamicFunction(const uint32_t x, const int16_t* lut);
+
+uint32_t PcanShrink(const uint32_t x);
+
+void PcanGainControlApply(struct PcanGainControlState* state, uint32_t* signal);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.c

@@ -0,0 +1,92 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h"
+
+#include <math.h>
+#include <stdio.h>
+
+#define kint16max 0x00007FFF
+
+void PcanGainControlFillConfigWithDefaults(
+    struct PcanGainControlConfig* config) {
+  config->enable_pcan = 0;
+  config->strength = 0.95;
+  config->offset = 80.0;
+  config->gain_bits = 21;
+}
+
+int16_t PcanGainLookupFunction(const struct PcanGainControlConfig* config,
+                               int32_t input_bits, uint32_t x) {
+  const float x_as_float = ((float)x) / ((uint32_t)1 << input_bits);
+  const float gain_as_float =
+      ((uint32_t)1 << config->gain_bits) *
+      powf(x_as_float + config->offset, -config->strength);
+
+  if (gain_as_float > kint16max) {
+    return kint16max;
+  }
+  return (int16_t)(gain_as_float + 0.5f);
+}
+
+int PcanGainControlPopulateState(const struct PcanGainControlConfig* config,
+                                 struct PcanGainControlState* state,
+                                 uint32_t* noise_estimate,
+                                 const int num_channels,
+                                 const uint16_t smoothing_bits,
+                                 const int32_t input_correction_bits) {
+  state->enable_pcan = config->enable_pcan;
+  if (!state->enable_pcan) {
+    return 1;
+  }
+  state->noise_estimate = noise_estimate;
+  state->num_channels = num_channels;
+  state->gain_lut = malloc(kWideDynamicFunctionLUTSize * sizeof(int16_t));
+  if (state->gain_lut == NULL) {
+    fprintf(stderr, "Failed to allocate gain LUT\n");
+    return 0;
+  }
+  state->snr_shift = config->gain_bits - input_correction_bits - kPcanSnrBits;
+
+  const int32_t input_bits = smoothing_bits - input_correction_bits;
+  state->gain_lut[0] = PcanGainLookupFunction(config, input_bits, 0);
+  state->gain_lut[1] = PcanGainLookupFunction(config, input_bits, 1);
+  state->gain_lut -= 6;
+  int interval;
+  for (interval = 2; interval <= kWideDynamicFunctionBits; ++interval) {
+    const uint32_t x0 = (uint32_t)1 << (interval - 1);
+    const uint32_t x1 = x0 + (x0 >> 1);
+    const uint32_t x2 =
+        (interval == kWideDynamicFunctionBits) ? x0 + (x0 - 1) : 2 * x0;
+
+    const int16_t y0 = PcanGainLookupFunction(config, input_bits, x0);
+    const int16_t y1 = PcanGainLookupFunction(config, input_bits, x1);
+    const int16_t y2 = PcanGainLookupFunction(config, input_bits, x2);
+
+    const int32_t diff1 = (int32_t)y1 - y0;
+    const int32_t diff2 = (int32_t)y2 - y0;
+    const int32_t a1 = 4 * diff1 - diff2;
+    const int32_t a2 = diff2 - a1;
+
+    state->gain_lut[4 * interval] = y0;
+    state->gain_lut[4 * interval + 1] = (int16_t)a1;
+    state->gain_lut[4 * interval + 2] = (int16_t)a2;
+  }
+  state->gain_lut += 6;
+  return 1;
+}
+
+void PcanGainControlFreeStateContents(struct PcanGainControlState* state) {
+  free(state->gain_lut);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control_util.h

@@ -0,0 +1,57 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/pcan_gain_control.h"
+
+#define kWideDynamicFunctionBits 32
+#define kWideDynamicFunctionLUTSize (4 * kWideDynamicFunctionBits - 3)
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct PcanGainControlConfig {
+  // set to false (0) to disable this module
+  int enable_pcan;
+  // gain normalization exponent (0.0 disables, 1.0 full strength)
+  float strength;
+  // positive value added in the normalization denominator
+  float offset;
+  // number of fractional bits in the gain
+  int gain_bits;
+};
+
+void PcanGainControlFillConfigWithDefaults(
+    struct PcanGainControlConfig* config);
+
+int16_t PcanGainLookupFunction(const struct PcanGainControlConfig* config,
+                               int32_t input_bits, uint32_t x);
+
+int PcanGainControlPopulateState(const struct PcanGainControlConfig* config,
+                                 struct PcanGainControlState* state,
+                                 uint32_t* noise_estimate,
+                                 const int num_channels,
+                                 const uint16_t smoothing_bits,
+                                 const int32_t input_correction_bits);
+
+void PcanGainControlFreeStateContents(struct PcanGainControlState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_PCAN_GAIN_CONTROL_UTIL_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/window.c

@@ -0,0 +1,70 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/window.h"
+
+#include <string.h>
+
+int WindowProcessSamples(struct WindowState* state, const int16_t* samples,
+                         size_t num_samples, size_t* num_samples_read) {
+  const int size = state->size;
+
+  // Copy samples from the samples buffer over to our local input.
+  size_t max_samples_to_copy = state->size - state->input_used;
+  if (max_samples_to_copy > num_samples) {
+    max_samples_to_copy = num_samples;
+  }
+  memcpy(state->input + state->input_used, samples,
+         max_samples_to_copy * sizeof(*samples));
+  *num_samples_read = max_samples_to_copy;
+  state->input_used += max_samples_to_copy;
+
+  if (state->input_used < state->size) {
+    // We don't have enough samples to compute a window.
+    return 0;
+  }
+
+  // Apply the window to the input.
+  const int16_t* coefficients = state->coefficients;
+  const int16_t* input = state->input;
+  int16_t* output = state->output;
+  int i;
+  int16_t max_abs_output_value = 0;
+  for (i = 0; i < size; ++i) {
+    int16_t new_value =
+        (((int32_t)*input++) * *coefficients++) >> kFrontendWindowBits;
+    *output++ = new_value;
+    if (new_value < 0) {
+      new_value = -new_value;
+    }
+    if (new_value > max_abs_output_value) {
+      max_abs_output_value = new_value;
+    }
+  }
+  // Shuffle the input down by the step size, and update how much we have used.
+  memmove(state->input, state->input + state->step,
+          sizeof(*state->input) * (state->size - state->step));
+  state->input_used -= state->step;
+  state->max_abs_output_value = max_abs_output_value;
+
+  // Indicate that the output buffer is valid for the next stage.
+  return 1;
+}
+
+void WindowReset(struct WindowState* state) {
+  memset(state->input, 0, state->size * sizeof(*state->input));
+  memset(state->output, 0, state->size * sizeof(*state->output));
+  state->input_used = 0;
+  state->max_abs_output_value = 0;
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/window.h

@@ -0,0 +1,49 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_
+
+#include <stdint.h>
+#include <stdlib.h>
+
+#define kFrontendWindowBits 12
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct WindowState {
+  size_t size;
+  int16_t* coefficients;
+  size_t step;
+
+  int16_t* input;
+  size_t input_used;
+  int16_t* output;
+  int16_t max_abs_output_value;
+};
+
+// Applies a window to the samples coming in, stepping forward at the given
+// rate.
+int WindowProcessSamples(struct WindowState* state, const int16_t* samples,
+                         size_t num_samples, size_t* num_samples_read);
+
+void WindowReset(struct WindowState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_H_

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/window_util.c

@@ -0,0 +1,73 @@
+/* Copyright 2018 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.
+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 "tensorflow/lite/experimental/microfrontend/lib/window_util.h"
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+// Some platforms don't have M_PI
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+
+void WindowFillConfigWithDefaults(struct WindowConfig* config) {
+  config->size_ms = 25;
+  config->step_size_ms = 10;
+}
+
+int WindowPopulateState(const struct WindowConfig* config,
+                        struct WindowState* state, int sample_rate) {
+  state->size = config->size_ms * sample_rate / 1000;
+  state->step = config->step_size_ms * sample_rate / 1000;
+
+  state->coefficients = malloc(state->size * sizeof(*state->coefficients));
+  if (state->coefficients == NULL) {
+    fprintf(stderr, "Failed to allocate window coefficients\n");
+    return 0;
+  }
+
+  // Populate the window values.
+  const float arg = M_PI * 2.0 / ((float)state->size);
+  int i;
+  for (i = 0; i < state->size; ++i) {
+    float float_value = 0.5 - (0.5 * cos(arg * (i + 0.5)));
+    // Scale it to fixed point and round it.
+    state->coefficients[i] =
+        floor(float_value * (1 << kFrontendWindowBits) + 0.5);
+  }
+
+  state->input_used = 0;
+  state->input = malloc(state->size * sizeof(*state->input));
+  if (state->input == NULL) {
+    fprintf(stderr, "Failed to allocate window input\n");
+    return 0;
+  }
+
+  state->output = malloc(state->size * sizeof(*state->output));
+  if (state->output == NULL) {
+    fprintf(stderr, "Failed to allocate window output\n");
+    return 0;
+  }
+
+  return 1;
+}
+
+void WindowFreeStateContents(struct WindowState* state) {
+  free(state->coefficients);
+  free(state->input);
+  free(state->output);
+}

code/components/tflite-lib/tensorflow/lite/experimental/microfrontend/lib/window_util.h

@@ -0,0 +1,45 @@
+/* Copyright 2018 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.
+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 TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_
+#define TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_
+
+#include "tensorflow/lite/experimental/microfrontend/lib/window.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct WindowConfig {
+  // length of window frame in milliseconds
+  size_t size_ms;
+  // length of step for next frame in milliseconds
+  size_t step_size_ms;
+};
+
+// Populates the WindowConfig with "sane" default values.
+void WindowFillConfigWithDefaults(struct WindowConfig* config);
+
+// Allocates any buffers.
+int WindowPopulateState(const struct WindowConfig* config,
+                        struct WindowState* state, int sample_rate);
+
+// Frees any allocated buffers.
+void WindowFreeStateContents(struct WindowState* state);
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // TENSORFLOW_LITE_EXPERIMENTAL_MICROFRONTEND_LIB_WINDOW_UTIL_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/common.h

@@ -75,6 +75,7 @@ float ActivationFunction(float x) {
 inline void BiasAndClamp(float clamp_min, float clamp_max, int bias_size,
                          const float* bias_data, int array_size,
                          float* array_data) {
+  if (bias_size == 0) return;
   // Note: see b/132215220: in May 2019 we thought it would be OK to replace
   // this with the Eigen one-liner:
   //   return (array.colwise() + bias).cwiseMin(clamp_max).cwiseMin(clamp_max).
@@ -138,6 +139,100 @@ inline void BiasAndClamp(float clamp_min, float clamp_max, int bias_size,
 #endif
 }
 
+// Single-rounding MultiplyByQuantizedMultiplier
+#if TFLITE_SINGLE_ROUNDING
+inline int32_t MultiplyByQuantizedMultiplier(int32_t x,
+                                             int32_t quantized_multiplier,
+                                             int shift) {
+  TFLITE_DCHECK(quantized_multiplier >= 0);
+  TFLITE_DCHECK(shift >= -31 && shift <= 30);
+
+  const int64_t total_shift = 31 - shift;
+  const int64_t round = static_cast<int64_t>(1) << (total_shift - 1);
+  int64_t result = x * static_cast<int64_t>(quantized_multiplier) + round;
+  result = result >> total_shift;
+
+  TFLITE_DCHECK(result >= std::numeric_limits<int32_t>::min() &&
+                result <= std::numeric_limits<int32_t>::max());
+  return static_cast<int32_t>(result);
+}
+
+inline int32_t MultiplyByQuantizedMultiplierSmallerThanOneExp(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  TFLITE_DCHECK_LE(shift, 0);
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+
+inline int32_t MultiplyByQuantizedMultiplierGreaterThanOne(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  TFLITE_DCHECK_GE(shift, 0);
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+
+inline int32_t MultiplyByQuantizedMultiplier(int64_t x,
+                                             int32_t quantized_multiplier,
+                                             int shift) {
+  // Inputs:
+  // - quantized_multiplier has fixed point at bit 31
+  // - shift is -31 to +7 (negative for right shift)
+  //
+  // Assumptions: The following input ranges are assumed
+  // - quantize_scale>=0  (the usual range is (1<<30) to (1>>31)-1)
+  // - scaling is chosen so final scaled result fits in int32_t
+  // - input x is in the range -(1<<47) <= x < (1<<47)
+  TFLITE_DCHECK(quantized_multiplier >= 0);
+  TFLITE_DCHECK(shift >= -31 && shift < 8);
+  TFLITE_DCHECK(x >= -(static_cast<int64_t>(1) << 47) &&
+                x < (static_cast<int64_t>(1) << 47));
+
+  const int32_t reduced_multiplier =
+      (quantized_multiplier < 0x7FFF0000)
+          ? ((quantized_multiplier + (1 << 15)) >> 16)
+          : 0x7FFF;
+  const int64_t total_shift = 15 - shift;
+  const int64_t round = static_cast<int64_t>(1) << (total_shift - 1);
+  int64_t result = x * static_cast<int64_t>(reduced_multiplier) + round;
+  result = result >> total_shift;
+
+  TFLITE_DCHECK(result >= std::numeric_limits<int32_t>::min() &&
+                result <= std::numeric_limits<int32_t>::max());
+  return static_cast<int32_t>(result);
+}
+
+#ifdef USE_NEON
+inline int32x4x4_t MultiplyByQuantizedMultiplier4Rows(
+    int32x4x4_t input_val, int32_t quantized_multiplier, int shift) {
+  TFLITE_DCHECK(quantized_multiplier >= 0);
+
+  const int right_shift = std::min(-1, shift);
+  const int left_shift = shift - right_shift;
+
+  const int32x4_t multiplier_dup = vdupq_n_s32(quantized_multiplier);
+  const int32x4_t left_shift_dup = vdupq_n_s32(left_shift);
+  const int32x4_t right_shift_dup = vdupq_n_s32(right_shift);
+
+  int32x4x4_t result;
+  result.val[0] = vrshlq_s32(
+      vqdmulhq_s32(vshlq_s32(input_val.val[0], left_shift_dup), multiplier_dup),
+      right_shift_dup);
+
+  result.val[1] = vrshlq_s32(
+      vqdmulhq_s32(vshlq_s32(input_val.val[1], left_shift_dup), multiplier_dup),
+      right_shift_dup);
+
+  result.val[2] = vrshlq_s32(
+      vqdmulhq_s32(vshlq_s32(input_val.val[2], left_shift_dup), multiplier_dup),
+      right_shift_dup);
+
+  result.val[3] = vrshlq_s32(
+      vqdmulhq_s32(vshlq_s32(input_val.val[3], left_shift_dup), multiplier_dup),
+      right_shift_dup);
+
+  return result;
+}
+#endif  // USE_NEON
+// Double-rounding MultiplyByQuantizedMultiplier
+#else
 inline int32_t MultiplyByQuantizedMultiplierSmallerThanOneExp(
     int32_t x, int32_t quantized_multiplier, int left_shift) {
   using gemmlowp::RoundingDivideByPOT;
@@ -224,7 +319,8 @@ inline int32x4x4_t MultiplyByQuantizedMultiplier4Rows(
 
   return result;
 }
-#endif
+#endif  // USE_NEON
+#endif  // TFLITE_SINGLE_ROUNDING
 
 template <typename T>
 int CountLeadingZeros(T integer_input) {

code/components/tflite-lib/tensorflow/lite/kernels/internal/portable_tensor_utils.h

@@ -0,0 +1,124 @@
+/* Copyright 2021 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.
+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 TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_UTILS_H_
+#define TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_UTILS_H_
+
+#include <algorithm>
+#include <cmath>
+#include <cstdint>
+
+#include "tensorflow/lite/c/builtin_op_data.h"
+#include "tensorflow/lite/c/common.h"
+
+#if defined(_MSC_VER)
+#define __restrict__ __restrict
+#endif
+
+namespace tflite {
+
+namespace tensor_utils {
+
+// Multiplies a matrix with a scalar and reduce the result on each row to a
+// scalar.
+// Parameters:
+//     - matrix: matrix of size n_row * n_col
+//     - scalar: the scalar that is multiplied to each element in the matrix
+//     - n_row:  the row count of the matrix
+//     - n_col:  the column count of the matrix
+//     - output: the 32bit output
+// Note: We do not need saturation because the int8 * int8 is safe from overflow
+// in (2^31-1) / (2^14) = 131072, which is bigger than the n_row. Non-zero
+// initial output value is not exceptionally large.
+void MatrixScalarMultiplyAccumulate(const int8_t* matrix, int32_t scalar,
+                                    int32_t n_row, int32_t n_col,
+                                    int32_t* output);
+
+// Add another vector for each batch in the batch vector.
+template <typename T>
+void VectorBatchVectorAdd(const T* vector, int v_size, int n_batch,
+                          T* batch_vector) {
+  for (int b = 0; b < n_batch; b++) {
+    for (int i = 0; i < v_size; ++i) {
+      batch_vector[i] += vector[i];
+    }
+    batch_vector += v_size;
+  }
+}
+
+// Cwise product of two vectors.
+template <typename T>
+inline void VectorVectorCwiseProduct(const T* __restrict__ vector1,
+                                     const T* __restrict__ vector2, int v_size,
+                                     T* __restrict__ result) {
+  for (int v = 0; v < v_size; v++) {
+    *result++ = *vector1++ * *vector2++;
+  }
+}
+
+// Cwise product of a vector and a batch-vector.
+template <typename T>
+inline void VectorBatchVectorCwiseProduct(const T* vector, int v_size,
+                                          const T* batch_vector, int n_batch,
+                                          T* result) {
+  for (int b = 0; b < n_batch; b++) {
+    VectorVectorCwiseProduct(vector, batch_vector, v_size, result);
+    // Update the pointers.
+    result += v_size;
+    batch_vector += v_size;
+  }
+}
+
+// Cwise product and accumulate of two vectors. Since it's a MAC operation, the
+// assumption here is that result array is initialized to valid values.
+template <typename T>
+inline void VectorVectorCwiseProductAccumulate(const T* __restrict__ vector1,
+                                               const T* __restrict__ vector2,
+                                               int v_size,
+                                               T* __restrict__ result) {
+  for (int v = 0; v < v_size; v++) {
+    *result++ += *vector1++ * *vector2++;
+  }
+}
+
+// Cwise product and accumulate of a vector and a batch-vector. Since it's a MAC
+// operation, the assumption here is that result array is initialized to valid
+// values.
+template <typename T>
+inline void VectorBatchVectorCwiseProductAccumulate(const T* vector, int v_size,
+                                                    const T* batch_vector,
+                                                    int n_batch, T* result) {
+  for (int b = 0; b < n_batch; b++) {
+    VectorVectorCwiseProductAccumulate(vector, batch_vector, v_size, result);
+    // Update the pointers.
+    result += v_size;
+    batch_vector += v_size;
+  }
+}
+
+// Batch vector initialization with another vector.
+template <typename T>
+void VectorBatchVectorAssign(const T* vector, int v_size, int n_batch,
+                             T* batch_vector) {
+  for (int b = 0; b < n_batch; b++) {
+    std::copy_n(vector, v_size, batch_vector + b * v_size);
+  }
+}
+
+}  // namespace tensor_utils
+
+}  // namespace tflite
+
+#endif  // TENSORFLOW_LITE_KERNELS_INTERNAL_PORTABLE_TENSOR_UTILS_H_

code/components/tflite-lib/tensorflow/lite/kernels/internal/quantization_util.cc

@@ -52,6 +52,11 @@ constexpr uint32_t kFractionRoundingThreshold = 0x00200000;
 
 void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
                         int* shift) {
+#if TFLITE_SINGLE_ROUNDING
+  // Single-rounding MultiplyByQuantizedMultiplier only supports positive
+  // multipliers.
+  // TFLITE_DCHECK(double_multiplier >= 0);
+#endif
   if (double_multiplier == 0.) {
     *quantized_multiplier = 0;
     *shift = 0;
@@ -65,10 +70,10 @@ void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
   int64_t q_fixed = IntegerFrExp(double_multiplier, shift);
 #else   // TFLITE_EMULATE_FLOAT
   const double q = std::frexp(double_multiplier, shift);
-  auto q_fixed = static_cast<int64_t>(TfLiteRound(q * (1ll << 31)));
+  auto q_fixed = static_cast<int64_t>(TfLiteRound(q * (1LL << 31)));
 #endif  // TFLITE_EMULATE_FLOAT
-  TFLITE_CHECK(q_fixed <= (1ll << 31));
+  TFLITE_CHECK(q_fixed <= (1LL << 31));
-  if (q_fixed == (1ll << 31)) {
+  if (q_fixed == (1LL << 31)) {
     q_fixed /= 2;
     ++*shift;
   }
@@ -87,6 +92,14 @@ void QuantizeMultiplier(double double_multiplier, int32_t* quantized_multiplier,
     *shift = 0;
     q_fixed = 0;
   }
+#if TFLITE_SINGLE_ROUNDING
+  // Single-rounding MultiplyByQuantizedMultiplier doesn't support a shift > 30,
+  // saturate it.
+  if (*shift > 30) {
+    *shift = 30;
+    q_fixed = (1LL << 31) - 1;
+  }
+#endif
   *quantized_multiplier = static_cast<int32_t>(q_fixed);
 }
 
@@ -278,6 +291,12 @@ void PreprocessSoftmaxScaling(double beta, double input_scale,
   // result is double equivalent of Q0.31 (actually with more precision). Thus
   // this generates a Q(input_integer_bits).(31-input_integer_bits)
   // representation.
+#if TFLITE_SINGLE_ROUNDING
+  const double max_real_multiplier = (1LL << 30) - 1.0;
+#else
+  const double max_real_multiplier = (1LL << 31) - 1.0;
+#endif
+
 #ifdef TFLITE_EMULATE_FLOAT
   const double input_beta = IntegerDoubleMultiply(beta, input_scale);
   int shift;
@@ -285,12 +304,14 @@ void PreprocessSoftmaxScaling(double beta, double input_scale,
   shift += (31 - input_integer_bits);
   double input_beta_real_multiplier =
       DoubleFromFractionAndShift(fraction, shift);
-  if (IntegerDoubleCompare(input_beta_real_multiplier, (1ll << 31) - 1.0) > 0) {
+  if (IntegerDoubleCompare(input_beta_real_multiplier, max_real_multiplier) >
-    input_beta_real_multiplier = (1ll << 31) - 1.0;
+      0) {
+    input_beta_real_multiplier = max_real_multiplier;
   }
 #else   // TFLITE_EMULATE_FLOAT
-  const double input_beta_real_multiplier = std::min<double>(
+  const double input_beta_real_multiplier =
-      beta * input_scale * (1 << (31 - input_integer_bits)), (1ll << 31) - 1.0);
+      std::min<double>(beta * input_scale * (1 << (31 - input_integer_bits)),
+                       max_real_multiplier);
 #endif  // TFLITE_EMULATE_FLOAT
 
   QuantizeMultiplierGreaterThanOne(input_beta_real_multiplier,
@@ -324,8 +345,8 @@ int CalculateInputRadius(int input_integer_bits, int input_left_shift,
 #else   // TFLITE_EMULATE_FLOAT
   const double max_input_rescaled =
       1.0 * ((1 << input_integer_bits) - 1) *
-      (1ll << (total_signed_bits - input_integer_bits)) /
+      (1LL << (total_signed_bits - input_integer_bits)) /
-      (1ll << input_left_shift);
+      (1LL << input_left_shift);
   // Tighten bound using floor.  Suppose that we could use the exact value.
   // After scaling the difference, the result would be at the maximum.  Thus we
   // must ensure that our value has lower magnitude.

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/binary_function.h

@@ -43,16 +43,27 @@ inline void BroadcastBinaryFunction4DSlow(
   NdArrayDescsForElementwiseBroadcast(unextended_input1_shape,
                                       unextended_input2_shape, &desc1, &desc2);
 
+  const int* dims_data =
+      reinterpret_cast<const int*>(output_shape.DimsDataUpTo5D());
   for (int b = 0; b < output_shape.Dims(0); ++b) {
+    int out_idx_b = b * dims_data[1];
+    int in_idx1_b = desc1.strides[0] * b;
+    int in_idx2_b = desc2.strides[0] * b;
     for (int y = 0; y < output_shape.Dims(1); ++y) {
+      int out_idx_y = (out_idx_b + y) * dims_data[2];
+      int in_idx1_y = in_idx1_b + desc1.strides[1] * y;
+      int in_idx2_y = in_idx2_b + desc2.strides[1] * y;
       for (int x = 0; x < output_shape.Dims(2); ++x) {
+        int out_idx_x = (out_idx_y + x) * dims_data[3];
+        int in1_idx = in_idx1_y + desc1.strides[2] * x;
+        int in2_idx = in_idx2_y + desc2.strides[2] * x;
         for (int c = 0; c < output_shape.Dims(3); ++c) {
-          auto out_idx = Offset(output_shape, b, y, x, c);
+          auto out_idx = out_idx_x + c;
-          auto in1_idx = SubscriptToIndex(desc1, b, y, x, c);
-          auto in2_idx = SubscriptToIndex(desc2, b, y, x, c);
           auto in1_val = input1_data[in1_idx];
           auto in2_val = input2_data[in2_idx];
           output_data[out_idx] = func(in1_val, in2_val);
+          in1_idx += desc1.strides[3];
+          in2_idx += desc2.strides[3];
         }
       }
     }

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/depthwiseconv_uint8.h

@@ -68,6 +68,27 @@ inline int32_t DepthwiseConvRound(int32_t x, int32_t quantized_multiplier,
   return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
 }
 
+// Single-rounding MultiplyByQuantizedMultiplier
+#if TFLITE_SINGLE_ROUNDING
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kAwayFromZero>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  using gemmlowp::RoundingDivideByPOT;
+  using gemmlowp::SaturatingRoundingDoublingHighMul;
+  int left_shift = shift > 0 ? shift : 0;
+  int right_shift = shift > 0 ? 0 : -shift;
+  return RoundingDivideByPOT(SaturatingRoundingDoublingHighMul(
+                                 x * (1 << left_shift), quantized_multiplier),
+                             right_shift);
+}
+
+template <>
+inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kUpward>(
+    int32_t x, int32_t quantized_multiplier, int shift) {
+  return MultiplyByQuantizedMultiplier(x, quantized_multiplier, shift);
+}
+// Double-rounding MultiplyByQuantizedMultiplier
+#else
 template <>
 inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kAwayFromZero>(
     int32_t x, int32_t quantized_multiplier, int shift) {
@@ -86,6 +107,7 @@ inline int32_t DepthwiseConvRound<DepthwiseConvOutputRounding::kUpward>(
           rounding_offset) >>
          right_shift;
 }
+#endif  // TFLITE_SINGLE_ROUNDING
 
 template <DepthwiseConvOutputRounding output_rounding>
 struct DepthwiseConvBasicKernel {

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/fully_connected.h

@@ -15,6 +15,7 @@ limitations under the License.
 #ifndef TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_
 #define TENSORFLOW_LITE_KERNELS_INTERNAL_REFERENCE_FULLY_CONNECTED_H_
 
+#include "ruy/profiler/instrumentation.h"  // from @ruy
 #include "tensorflow/lite/kernels/internal/common.h"
 #include "tensorflow/lite/kernels/internal/cppmath.h"
 #include "tensorflow/lite/kernels/internal/quantization_util.h"

code/components/tflite-lib/tensorflow/lite/kernels/internal/reference/integer_ops/fully_connected.h

@@ -74,12 +74,13 @@ inline void FullyConnected(
   const int32_t output_activation_min = params.quantized_activation_min;
   const int32_t output_activation_max = params.quantized_activation_max;
   TFLITE_DCHECK_GE(filter_shape.DimensionsCount(), 2);
-  TFLITE_DCHECK_EQ(output_shape.DimensionsCount(), 2);
+  TFLITE_DCHECK_GE(output_shape.DimensionsCount(), 1);
 
   TFLITE_DCHECK_LE(output_activation_min, output_activation_max);
   const int filter_dim_count = filter_shape.DimensionsCount();
-  const int batches = output_shape.Dims(0);
+  const int output_dim_count = output_shape.DimensionsCount();
-  const int output_depth = output_shape.Dims(1);
+  const int batches = FlatSizeSkipDim(output_shape, output_dim_count - 1);
+  const int output_depth = output_shape.Dims(output_dim_count - 1);
   TFLITE_DCHECK_LE(output_depth, filter_shape.Dims(filter_dim_count - 2));
   const int accum_depth = filter_shape.Dims(filter_dim_count - 1);
   for (int b = 0; b < batches; ++b) {