.github/workflows/xtensa_postmerge.yml

@@ -31,7 +31,7 @@ jobs:
       - run: |
           rm -rf .git
           echo ${{ secrets.tflm-bot-token }} | docker login ghcr.io -u tflm-bot --password-stdin
-          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa:0.6 \
+          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa_xplorer_13:0.1 \
           /bin/bash -c \
           "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_fusion_f1.sh EXTERNAL tflite-micro/"
 
@@ -46,7 +46,7 @@ jobs:
       - run: |
           rm -rf .git
           echo ${{ secrets.tflm-bot-token }} | docker login ghcr.io -u tflm-bot --password-stdin
-          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa:0.6 \
+          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa_xplorer_13:0.1 \
           /bin/bash -c \
           "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_vision_p6.sh RUN_TESTS tflite-micro/"
 
@@ -61,6 +61,6 @@ jobs:
       - run: |
           rm -rf .git
           echo ${{ secrets.tflm-bot-token }} | docker login ghcr.io -u tflm-bot --password-stdin
-          docker run --env XTENSA_TOOLS_VERSION=RI-2019.2-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa:0.6 \
+          docker run --env XTENSA_TOOLS_VERSION=RI-2019.2-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa_xplorer_11:0.1 \
           /bin/bash -c \
-          "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_hifimini.sh tflite-micro/"
\ No newline at end of file
+          "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_hifimini.sh tflite-micro/"

.github/workflows/xtensa_presubmit.yml

@@ -23,7 +23,7 @@ jobs:
 
   vision_p6_presubmit:
     runs-on: ubuntu-latest
-      
+
     name: Vision P6 Build (presubmit)
     steps:
       - uses: actions/checkout@v2
@@ -32,7 +32,7 @@ jobs:
       - run: |
           rm -rf .git
           echo ${{ secrets.tflm-bot-token }} | docker login ghcr.io -u tflm-bot --password-stdin
-          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa:0.6 \
+          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa_xplorer_13:0.1 \
           /bin/bash -c \
           "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_vision_p6.sh RUN_NO_TESTS tflite-micro/"
 
@@ -47,7 +47,7 @@ jobs:
       - run: |
           rm -rf .git
           echo ${{ secrets.tflm-bot-token }} | docker login ghcr.io -u tflm-bot --password-stdin
-          docker run --env XTENSA_TOOLS_VERSION=RI-2022.9-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa:0.6 \
+          docker run --env XTENSA_TOOLS_VERSION=RI-2022.9-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa_hifi5:0.1 \
           /bin/bash -c \
           "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_hifi5.sh tflite-micro/"
 
@@ -62,6 +62,6 @@ jobs:
       - run: |
           rm -rf .git
           echo ${{ secrets.tflm-bot-token }} | docker login ghcr.io -u tflm-bot --password-stdin
-          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa:0.6 \
+          docker run --env XTENSA_TOOLS_VERSION=RI-2020.4-linux --rm -v `pwd`:/opt/tflite-micro ghcr.io/tflm-bot/xtensa_xplorer_13:0.1 \
           /bin/bash -c \
-          "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_hifi3z.sh EXTERNAL tflite-micro/"
\ No newline at end of file
+          "cd /opt && tflite-micro/tensorflow/lite/micro/tools/ci_build/test_xtensa_hifi3z.sh EXTERNAL tflite-micro/"

python/tflite_micro/signal/BUILD

@@ -27,6 +27,7 @@ py_library(
         "ops/__init__.py",
     ],
     srcs_version = "PY3",
+    visibility = ["//python/tflite_micro/signal/utils:__subpackages__"],
     deps = [
         ":fft_ops",
         ":window_op",

python/tflite_micro/signal/utils/BUILD

@@ -13,6 +13,7 @@ py_library(
     srcs = ["util.py"],
     deps = [
         "//python/tflite_micro:runtime",
+        "//python/tflite_micro/signal:ops",
         requirement("tensorflow-cpu"),
     ],
 )

signal/micro/kernels/rfft.cc

@@ -142,16 +142,16 @@ void* InitAll(TfLiteContext* context, const char* buffer, size_t length) {
 
   switch (tensor_type) {
     case TensorType_INT16: {
-      return Init<int16_t, RfftInt16GetNeededMemory, RfftInt16Init>(
-          context, buffer, length);
+      return Init<int16_t, ::tflm_signal::RfftInt16GetNeededMemory,
+                  ::tflm_signal::RfftInt16Init>(context, buffer, length);
     }
     case TensorType_INT32: {
-      return Init<int32_t, RfftInt32GetNeededMemory, RfftInt32Init>(
-          context, buffer, length);
+      return Init<int32_t, ::tflm_signal::RfftInt32GetNeededMemory,
+                  ::tflm_signal::RfftInt32Init>(context, buffer, length);
     }
     case TensorType_FLOAT32: {
-      return Init<float, RfftFloatGetNeededMemory, RfftFloatInit>(
-          context, buffer, length);
+      return Init<float, ::tflm_signal::RfftFloatGetNeededMemory,
+                  ::tflm_signal::RfftFloatInit>(context, buffer, length);
     }
     default:
       return nullptr;
@@ -183,13 +183,13 @@ TfLiteStatus EvalAll(TfLiteContext* context, TfLiteNode* node) {
 
   switch (params->fft_type) {
     case kTfLiteInt16: {
-      return Eval<int16_t, RfftInt16Apply>(context, node);
+      return Eval<int16_t, ::tflm_signal::RfftInt16Apply>(context, node);
     }
     case kTfLiteInt32: {
-      return Eval<int32_t, RfftInt32Apply>(context, node);
+      return Eval<int32_t, ::tflm_signal::RfftInt32Apply>(context, node);
     }
     case kTfLiteFloat32: {
-      return Eval<float, RfftFloatApply>(context, node);
+      return Eval<float, ::tflm_signal::RfftFloatApply>(context, node);
     }
     default:
       return kTfLiteError;
@@ -208,22 +208,28 @@ TFLMRegistration* Register_RFFT() {
 
 TFLMRegistration* Register_RFFT_FLOAT() {
   static TFLMRegistration r = tflite::micro::RegisterOp(
-      Init<float, RfftFloatGetNeededMemory, RfftFloatInit>,
-      Prepare<float, kTfLiteFloat32>, Eval<float, RfftFloatApply>);
+      Init<float, ::tflm_signal::RfftFloatGetNeededMemory,
+           ::tflm_signal::RfftFloatInit>,
+      Prepare<float, kTfLiteFloat32>,
+      Eval<float, ::tflm_signal::RfftFloatApply>);
   return &r;
 }
 
 TFLMRegistration* Register_RFFT_INT16() {
   static TFLMRegistration r = tflite::micro::RegisterOp(
-      Init<int16_t, RfftInt16GetNeededMemory, RfftInt16Init>,
-      Prepare<int16_t, kTfLiteInt16>, Eval<int16_t, RfftInt16Apply>);
+      Init<int16_t, ::tflm_signal::RfftInt16GetNeededMemory,
+           ::tflm_signal::RfftInt16Init>,
+      Prepare<int16_t, kTfLiteInt16>,
+      Eval<int16_t, ::tflm_signal::RfftInt16Apply>);
   return &r;
 }
 
 TFLMRegistration* Register_RFFT_INT32() {
   static TFLMRegistration r = tflite::micro::RegisterOp(
-      Init<int32_t, RfftInt32GetNeededMemory, RfftInt32Init>,
-      Prepare<int32_t, kTfLiteInt32>, Eval<int32_t, RfftInt32Apply>);
+      Init<int32_t, ::tflm_signal::RfftInt32GetNeededMemory,
+           ::tflm_signal::RfftInt32Init>,
+      Prepare<int32_t, kTfLiteInt32>,
+      Eval<int32_t, ::tflm_signal::RfftInt32Apply>);
   return &r;
 }
 

signal/src/rfft.h

@@ -21,6 +21,9 @@ limitations under the License.
 
 #include "signal/src/complex.h"
 
+// TODO(b/286250473): remove namespace once de-duped libraries
+namespace tflm_signal {
+
 // RFFT (Real Fast Fourier Transform)
 // FFT for real valued time domain inputs.
 
@@ -77,4 +80,6 @@ void* RfftFloatInit(int32_t fft_length, void* state, size_t state_size);
 // * `output` must be of size (`fft_length` * 2) + 1 elements
 void RfftFloatApply(void* state, const float* input, Complex<float>* output);
 
+}  // namespace tflm_signal
+
 #endif  // SIGNAL_SRC_RFFT_H_

signal/src/rfft_float.cc

@@ -20,6 +20,9 @@ limitations under the License.
 #include "signal/src/kiss_fft_wrappers/kiss_fft_float.h"
 #include "signal/src/rfft.h"
 
+// TODO(b/286250473): remove namespace once de-duped libraries
+namespace tflm_signal {
+
 size_t RfftFloatGetNeededMemory(int32_t fft_length) {
   size_t state_size = 0;
   kiss_fft_float::kiss_fftr_alloc(fft_length, 0, nullptr, &state_size);
@@ -36,3 +39,5 @@ void RfftFloatApply(void* state, const float* input, Complex<float>* output) {
       reinterpret_cast<const kiss_fft_scalar*>(input),
       reinterpret_cast<kiss_fft_float::kiss_fft_cpx*>(output));
 }
+
+}  // namespace tflm_signal

signal/src/rfft_int16.cc

@@ -20,6 +20,9 @@ limitations under the License.
 #include "signal/src/kiss_fft_wrappers/kiss_fft_int16.h"
 #include "signal/src/rfft.h"
 
+// TODO(b/286250473): remove namespace once de-duped libraries
+namespace tflm_signal {
+
 size_t RfftInt16GetNeededMemory(int32_t fft_length) {
   size_t state_size = 0;
   kiss_fft_fixed16::kiss_fftr_alloc(fft_length, 0, nullptr, &state_size);
@@ -37,3 +40,5 @@ void RfftInt16Apply(void* state, const int16_t* input,
       reinterpret_cast<const kiss_fft_scalar*>(input),
       reinterpret_cast<kiss_fft_fixed16::kiss_fft_cpx*>(output));
 }
+
+}  // namespace tflm_signal

signal/src/rfft_int32.cc

@@ -20,6 +20,9 @@ limitations under the License.
 #include "signal/src/kiss_fft_wrappers/kiss_fft_int32.h"
 #include "signal/src/rfft.h"
 
+// TODO(b/286250473): remove namespace once de-duped libraries
+namespace tflm_signal {
+
 size_t RfftInt32GetNeededMemory(int32_t fft_length) {
   size_t state_size = 0;
   kiss_fft_fixed32::kiss_fftr_alloc(fft_length, 0, nullptr, &state_size);
@@ -36,4 +39,6 @@ void RfftInt32Apply(void* state, const int32_t* input,
       static_cast<kiss_fft_fixed32::kiss_fftr_cfg>(state),
       reinterpret_cast<const kiss_fft_scalar*>(input),
       reinterpret_cast<kiss_fft_fixed32::kiss_fft_cpx*>(output));
-}
\ No newline at end of file
+}
+
+}  // namespace tflm_signal

signal/tensorflow_core/kernels/fft_kernels.cc

@@ -82,21 +82,24 @@ class RfftOp : public tensorflow::OpKernel {
 };
 
 // TODO(b/286250473): change back name after name clash resolved
-REGISTER_KERNEL_BUILDER(Name("SignalRfft")
-                            .Device(tensorflow::DEVICE_CPU)
-                            .TypeConstraint<float>("T"),
-                        RfftOp<float, DT_FLOAT, RfftFloatGetNeededMemory,
-                               RfftFloatInit, RfftFloatApply>);
-REGISTER_KERNEL_BUILDER(Name("SignalRfft")
-                            .Device(tensorflow::DEVICE_CPU)
-                            .TypeConstraint<int16>("T"),
-                        RfftOp<int16_t, DT_INT16, RfftInt16GetNeededMemory,
-                               RfftInt16Init, RfftInt16Apply>);
-REGISTER_KERNEL_BUILDER(Name("SignalRfft")
-                            .Device(tensorflow::DEVICE_CPU)
-                            .TypeConstraint<int32>("T"),
-                        RfftOp<int32_t, DT_INT32, RfftInt32GetNeededMemory,
-                               RfftInt32Init, RfftInt32Apply>);
+REGISTER_KERNEL_BUILDER(
+    Name("SignalRfft")
+        .Device(tensorflow::DEVICE_CPU)
+        .TypeConstraint<float>("T"),
+    RfftOp<float, DT_FLOAT, ::tflm_signal::RfftFloatGetNeededMemory,
+           ::tflm_signal::RfftFloatInit, ::tflm_signal::RfftFloatApply>);
+REGISTER_KERNEL_BUILDER(
+    Name("SignalRfft")
+        .Device(tensorflow::DEVICE_CPU)
+        .TypeConstraint<int16>("T"),
+    RfftOp<int16_t, DT_INT16, ::tflm_signal::RfftInt16GetNeededMemory,
+           ::tflm_signal::RfftInt16Init, ::tflm_signal::RfftInt16Apply>);
+REGISTER_KERNEL_BUILDER(
+    Name("SignalRfft")
+        .Device(tensorflow::DEVICE_CPU)
+        .TypeConstraint<int32>("T"),
+    RfftOp<int32_t, DT_INT32, ::tflm_signal::RfftInt32GetNeededMemory,
+           ::tflm_signal::RfftInt32Init, ::tflm_signal::RfftInt32Apply>);
 
 }  // namespace signal
 }  // namespace tensorflow
\ No newline at end of file

tensorflow/lite/array.h

@@ -82,23 +82,56 @@ inline FloatArrayUniquePtr BuildTfLiteArray<float>(const int size) {
   return FloatArrayUniquePtr(TfLiteFloatArrayCreate(size));
 }
 
-// Allocates a TFLiteArray of given size and initializes it.
+// Allocates a TFLiteArray of given size and initializes it with the given
+// values.
 //
 // `values` is expected to holds `size` elements.
-template <class T>
-TfLiteArrayUniquePtr<T> BuildTfLiteArray(const int size,
-                                         const T* const values) {
-  auto array = BuildTfLiteArray<T>(size);
-  if (array) {
-    memcpy(array->data, values, size * sizeof(T));
+//
+// If T is explicitely specified and the type of values is not the same as T,
+// then a static_cast is performed.
+template <class T = void, class U,
+          class Type = std::conditional_t<std::is_same<T, void>::value, U, T>>
+TfLiteArrayUniquePtr<Type> BuildTfLiteArray(const int size,
+                                            const U* const values) {
+  TfLiteArrayUniquePtr<Type> array = BuildTfLiteArray<Type>(size);
+  // If size is 0, the array pointer may be null.
+  if (array && values) {
+    if (std::is_same<Type, U>::value) {
+      memcpy(array->data, values, size * sizeof(Type));
+    } else {
+      for (int i = 0; i < size; ++i) {
+        array->data[i] = static_cast<Type>(values[i]);
+      }
+    }
   }
   return array;
 }
 
+// Allocates a TFLiteArray and initializes it with the given array.
+//
+// `values` is expected to holds `size` elements.
+template <class T, size_t N>
+TfLiteArrayUniquePtr<T> BuildTfLiteArray(const T (&values)[N]) {
+  return BuildTfLiteArray<T>(static_cast<int>(N), values);
+}
+
 // Allocates a TFLiteArray and initializes it with the given values.
-template <class T>
-TfLiteArrayUniquePtr<T> BuildTfLiteArray(const std::vector<T>& values) {
-  return BuildTfLiteArray(static_cast<int>(values.size()), values.data());
+//
+// This uses SFINAE to only be picked up by for types that implement `data()`
+// and `size()` member functions. We cannot reuse detection facilities provided
+// by Abseil in this code.
+//
+// To conform with the other overloads, we allow specifying the type of the
+// array as well as deducing it from the container.
+template <
+    class T = void, class Container,
+    class ElementType =
+        std::decay_t<decltype(*std::declval<Container>().data())>,
+    class SizeType = std::decay_t<decltype(std::declval<Container>().size())>,
+    class Type =
+        std::conditional_t<std::is_same<T, void>::value, ElementType, T>>
+TfLiteArrayUniquePtr<Type> BuildTfLiteArray(const Container& values) {
+  return BuildTfLiteArray<Type>(static_cast<int>(values.size()), values.data());
 }
 
 // Allocates a TFLiteArray and initializes it with the given values.

tensorflow/lite/core/c/c_api_types.h

@@ -13,14 +13,16 @@ See the License for the specific language governing permissions and
 limitations under the License.
 ==============================================================================*/
 
-// This file declares types used by the pure C inference API defined in c_api.h,
-// some of which are also used in the C++ and C kernel and interpreter APIs.
-
-/// WARNING: Users of TensorFlow Lite should not include this file directly,
-/// but should instead include
-/// "third_party/tensorflow/lite/c/c_api_types.h".
-/// Only the TensorFlow Lite implementation itself should include this
-/// file directly.
+/// This file declares types used by the pure C inference API defined in
+/// c_api.h, some of which are also used in the C++ and C kernel and interpreter
+/// APIs.
+
+// WARNING: Users of TensorFlow Lite should not include this file directly,
+// but should instead include
+// "third_party/tensorflow/lite/c/c_api_types.h".
+// Only the TensorFlow Lite implementation itself should include this
+// file directly.
+
 // IWYU pragma: private, include "third_party/tensorflow/lite/c/c_api_types.h"
 
 #ifndef TENSORFLOW_LITE_CORE_C_C_API_TYPES_H_
@@ -32,6 +34,10 @@ limitations under the License.
 extern "C" {
 #endif
 
+/** \addtogroup c_api_types tensorflow/lite/c/c_api_types.h
+ *  @{
+ */
+
 // Define TFL_CAPI_EXPORT macro to export a function properly with a shared
 // library.
 #ifdef SWIG
@@ -50,50 +56,51 @@ 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.
+/// 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 {
+  /// Success
   kTfLiteOk = 0,
 
-  // Generally referring to an error in the runtime (i.e. interpreter)
+  /// Generally referring to an error in the runtime (i.e. interpreter)
   kTfLiteError = 1,
 
-  // Generally referring to an error from a TfLiteDelegate itself.
+  /// Generally referring to an error from a TfLiteDelegate itself.
   kTfLiteDelegateError = 2,
 
-  // 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 TF Lite delegate onto a model graph that's already
-  // immutable.
+  /// 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 TF Lite delegate onto a model graph that's already
+  /// immutable.
   kTfLiteApplicationError = 3,
 
-  // Generally referring to serialized delegate data not being found.
-  // See tflite::delegates::Serialization.
+  /// Generally referring to serialized delegate data not being found.
+  /// See tflite::delegates::Serialization.
   kTfLiteDelegateDataNotFound = 4,
 
-  // Generally referring to data-writing issues in delegate serialization.
-  // See tflite::delegates::Serialization.
+  /// Generally referring to data-writing issues in delegate serialization.
+  /// See tflite::delegates::Serialization.
   kTfLiteDelegateDataWriteError = 5,
 
-  // Generally referring to data-reading issues in delegate serialization.
-  // See tflite::delegates::Serialization.
+  /// Generally referring to data-reading issues in delegate serialization.
+  /// See tflite::delegates::Serialization.
   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.
+  /// 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,
 
-  // Generally referring to invocation cancelled by the user.
-  // See `interpreter::Cancel`.
+  /// Generally referring to invocation cancelled by the user.
+  /// See `interpreter::Cancel`.
   // TODO(b/194915839): Implement `interpreter::Cancel`.
   // TODO(b/250636993): Cancellation triggered by `SetCancellationFunction`
   // should also return this status code.
   kTfLiteCancelled = 8,
 } TfLiteStatus;
 
-// Types supported by tensor
+/// Types supported by tensor
 typedef enum {
   kTfLiteNoType = 0,
   kTfLiteFloat32 = 1,
@@ -116,12 +123,12 @@ typedef enum {
   kTfLiteInt4 = 18,
 } TfLiteType;
 
-// Legacy. Will be deprecated in favor of TfLiteAffineQuantization.
-// If per-layer quantization is specified this field will still be populated in
-// addition to TfLiteAffineQuantization.
-// Parameters for asymmetric quantization. Quantized values can be converted
-// back to float using:
-//     real_value = scale * (quantized_value - zero_point)
+/// Legacy. Will be deprecated in favor of TfLiteAffineQuantization.
+/// If per-layer quantization is specified this field will still be populated in
+/// addition to TfLiteAffineQuantization.
+/// Parameters for asymmetric quantization. Quantized values can be converted
+/// back to float using:
+///     real_value = scale * (quantized_value - zero_point)
 typedef struct TfLiteQuantizationParams {
   float scale;
   int32_t zero_point;
@@ -130,39 +137,41 @@ typedef struct TfLiteQuantizationParams {
 // --------------------------------------------------------------------------
 // Opaque types used by c_api.h, c_api_opaque.h and common.h.
 
-// TfLiteOpaqueContext is an opaque version of TfLiteContext;
+/// TfLiteOpaqueContext is an opaque version of TfLiteContext;
 typedef struct TfLiteOpaqueContext TfLiteOpaqueContext;
 
-// TfLiteOpaqueNode is an opaque version of TfLiteNode;
+/// TfLiteOpaqueNode is an opaque version of TfLiteNode;
 typedef struct TfLiteOpaqueNode TfLiteOpaqueNode;
 
-// TfLiteOpaqueTensor is an opaque version of TfLiteTensor;
+/// TfLiteOpaqueTensor is an opaque version of TfLiteTensor;
 typedef struct TfLiteOpaqueTensor TfLiteOpaqueTensor;
 
-// TfLiteDelegate: allows delegation of nodes to alternative backends.
-// Forward declaration of concrete type declared in common.h.
+/// TfLiteDelegate: allows delegation of nodes to alternative backends.
+/// Forward declaration of concrete type declared in common.h.
 typedef struct TfLiteDelegate TfLiteDelegate;
 
-// TfLiteOpaqueDelegateStruct: unconditionally opaque version of
-// TfLiteDelegate; allows delegation of nodes to alternative backends.
-//
-// This is an abstract type that is intended to have the same
-// role as TfLiteDelegate, but without exposing the implementation
-// details of how delegates are implemented.
-// WARNING: This is an experimental type and subject to change.
+/// TfLiteOpaqueDelegateStruct: unconditionally opaque version of
+/// TfLiteDelegate; allows delegation of nodes to alternative backends.
+///
+/// This is an abstract type that is intended to have the same
+/// role as TfLiteDelegate, but without exposing the implementation
+/// details of how delegates are implemented.
+/// WARNING: This is an experimental type and subject to change.
 typedef struct TfLiteOpaqueDelegateStruct TfLiteOpaqueDelegateStruct;
 
-// TfLiteOpaqueDelegate: conditionally opaque version of
-// TfLiteDelegate; allows delegation of nodes to alternative backends.
-// For TF Lite in Play Services, this is an opaque type,
-// but for regular TF Lite, this is just a typedef for TfLiteDelegate.
-// WARNING: This is an experimental type and subject to change.
+/// TfLiteOpaqueDelegate: conditionally opaque version of
+/// TfLiteDelegate; allows delegation of nodes to alternative backends.
+/// For TF Lite in Play Services, this is an opaque type,
+/// but for regular TF Lite, this is just a typedef for TfLiteDelegate.
+/// WARNING: This is an experimental type and subject to change.
 #if TFLITE_WITH_STABLE_ABI || TFLITE_USE_OPAQUE_DELEGATE
 typedef TfLiteOpaqueDelegateStruct TfLiteOpaqueDelegate;
 #else
 typedef TfLiteDelegate TfLiteOpaqueDelegate;
 #endif
 
+/** @} */
+
 #ifdef __cplusplus
 }  // extern C
 #endif

tensorflow/lite/core/c/common.cc

@@ -262,10 +262,19 @@ TfLiteStatus TfLiteTensorCopy(const TfLiteTensor* src, TfLiteTensor* dst) {
   if (dst->dims) TfLiteIntArrayFree(dst->dims);
   dst->dims = TfLiteIntArrayCopy(src->dims);
   if (src->allocation_type == kTfLiteVariantObject) {
-    if (dst->allocation_type != kTfLiteVariantObject) return kTfLiteError;
+    // An edge case exists in control flow ops when they copy inputs to outputs
+    // before invoking any body, in this case the `dst` will not have its
+    // `allocation_type` set properly, so we handle here for now.
+    if (dst->allocation_type != kTfLiteVariantObject) {
+      TfLiteTensorDataFree(dst);
+      dst->allocation_type = kTfLiteVariantObject;
+    }
     auto* dst_vd = static_cast<VariantData*>(dst->data.data);
     auto* src_vd = static_cast<VariantData*>(src->data.data);
-    // Implicitly casted via return from `CloneTo`. Don't need static cast here.
+
+    // `CloneTo` will handle the case when `dst_vd` is nullptr, so it is safe
+    // to `CloneTo` something which was "freed". Also, returning from `CloneTo`
+    // will implicitly cast to `VariantData`; don't need static cast here.
     dst->data.data = src_vd->CloneTo(dst_vd);
   } else {
     memcpy(dst->data.raw, src->data.raw, src->bytes);

tensorflow/lite/core/c/common.h

@@ -860,7 +860,7 @@ typedef struct TfLiteContext {
   // }
   //
   // NOTE: The context owns the memory referenced by partition_params_array. It
-  // will be cleared with another call to PreviewDelegateParitioning, or after
+  // will be cleared with another call to PreviewDelegatePartitioning, or after
   // TfLiteDelegateParams::Prepare returns.
   //
   // WARNING: This is an experimental interface that is subject to change.
@@ -920,6 +920,64 @@ typedef struct TfLiteContext {
 // field is the exactly the same as with `TfLiteRegistration`.
 typedef struct TfLiteRegistrationExternal TfLiteRegistrationExternal;
 
+// The valid values of the `inplace_operator` field in `TfLiteRegistration`.
+// This allow an op to signal to the runtime that the same data pointer
+// may be passed as an input and output without impacting the result.
+// This does not mean that the memory can safely be reused, it is up to the
+// runtime to determine this, e.g. if another op consumes the same input or not
+// or if an input tensor has sufficient memory allocated to store the output
+// data.
+//
+// Setting these flags authorizes the runtime to set the data pointers of an
+// input and output tensor to the same value. In such cases, the memory required
+// by the output must be less than or equal to that required by the shared
+// input, never greater. If kTfLiteInplaceOpDataUnmodified is set, then the
+// runtime can share the same input tensor with multiple operator's outputs,
+// provided that kTfLiteInplaceOpDataUnmodified is set for all of them.
+// Otherwise, if an input tensor is consumed by multiple operators, it may only
+// be shared with the operator which is the last to consume it.
+//
+// Note that this is a bitmask, so the values should be 1, 2, 4, 8, ...etc.
+typedef enum {
+  // The default value. This indicates that the same data pointer cannot safely
+  // be passed as an op's input and output.
+  kTfLiteInplaceOpNone = 0,
+  // This indicates that an op's first output's data is identical to its first
+  // input's data, for example Reshape.
+  kTfLiteInplaceOpDataUnmodified = 1,
+  // Setting kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput means
+  // that InputN may be shared with OutputN instead of with the first output.
+  // This flag requires one or more of kTfLiteInplaceOpInputNShared to be set.
+  kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput = 2,
+  // kTfLiteInplaceOpInputNShared indicates that it is safe for an op to share
+  // InputN's data pointer with an output tensor. If
+  // kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput is set then
+  // kTfLiteInplaceOpInputNShared indicates that InputN may be shared
+  // with OutputN, otherwise kTfLiteInplaceOpInputNShared indicates that InputN
+  // may be shared with the first output.
+  //
+  // Indicates that an op's first input may be shared with the first output
+  // tensor. kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput has
+  // no impact on the behavior allowed by this flag.
+  kTfLiteInplaceOpInput0Shared = 4,
+  // Indicates that an op's second input may be shared with the first output
+  // if kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput is not set
+  // or second output if kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput
+  // is set.
+  kTfLiteInplaceOpInput1Shared = 8,
+  // Indicates that an op's third input may be shared with the first output
+  // if kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput is not set
+  // or third output if kTfLiteInplaceInputCanBeSharedWithCorrespondingOutput is
+  // set.
+  kTfLiteInplaceOpInput2Shared = 16,
+  // Placeholder to ensure that enum can hold 64 bit values to accommodate
+  // future fields.
+  kTfLiteInplaceOpMaxValue = UINT64_MAX,
+} TfLiteInPlaceOp;
+
+// The number of shareable inputs supported.
+static const int kTfLiteMaxSharableOpInputs = 3;
+
 typedef struct TfLiteRegistration {
   // Initializes the op from serialized data.
   // Called only *once* for the lifetime of the op, so any one-time allocations
@@ -1000,8 +1058,37 @@ typedef struct TfLiteRegistration {
   // does not support asynchronous execution for this `node`.
   struct TfLiteAsyncKernel* (*async_kernel)(TfLiteContext* context,
                                             TfLiteNode* node);
+
+  // Indicates if an operator's output may safely overwrite its inputs.
+  // See the comments in `TfLiteInPlaceOp`.
+  uint64_t inplace_operator;
 } TfLiteRegistration;
 
+/// \private
+// Old version of `TfLiteRegistration` to maintain binary backward
+// compatibility.
+// The legacy registration type must be a POD struct type whose field types must
+// be a prefix of the field types in TfLiteRegistration, and offset of the first
+// field in TfLiteRegistration that is not present in the legacy registration
+// type must be greater than or equal to the size of the legacy registration
+// type.
+// WARNING: This structure is deprecated / not an official part of the
+// API. It should be only used for binary backward compatibility.
+typedef struct TfLiteRegistration_V3 {
+  void* (*init)(TfLiteContext* context, const char* buffer, size_t length);
+  void (*free)(TfLiteContext* context, void* buffer);
+  TfLiteStatus (*prepare)(TfLiteContext* context, TfLiteNode* node);
+  TfLiteStatus (*invoke)(TfLiteContext* context, TfLiteNode* node);
+  const char* (*profiling_string)(const TfLiteContext* context,
+                                  const TfLiteNode* node);
+  int32_t builtin_code;
+  const char* custom_name;
+  int version;
+  TfLiteRegistrationExternal* registration_external;
+  struct TfLiteAsyncKernel* (*async_kernel)(TfLiteContext* context,
+                                            TfLiteNode* node);
+} TfLiteRegistration_V3;
+
 /// \private
 // Old version of `TfLiteRegistration` to maintain binary backward
 // compatibility.
@@ -1314,9 +1401,9 @@ class AbstractVariantData : public VariantData {
       // If the output is still allocated, then its object may still be
       // in its life time and the destructor must be called before re-using the
       // buffer.
-      //     This may actual have a non-negligle effect on perfomance if the
+      //     This may actual have a non-negligible effect on performance if the
       // destructor is complex. A future iteration may
-      // introduce copy or move asignment semantics, allowing for the
+      // introduce copy or move assignment semantics, allowing for the
       // underlying implementation to optimize for this case.
       auto* derived = static_cast<ErasedDerived*>(maybe_alloc);
       derived->~ErasedDerived();

tensorflow/lite/micro/examples/micro_speech/train/train_micro_speech_model.ipynb

@@ -162,7 +162,6 @@
       },
       "outputs": [],
       "source": [
-        "%tensorflow_version 1.x\n",
         "import tensorflow as tf"
       ]
     },
@@ -420,7 +419,7 @@
       },
       "outputs": [],
       "source": [
-        "with tf.Session() as sess:\n",
+        "with tf.compat.v1.Session() as sess:\n",
         "  float_converter = tf.lite.TFLiteConverter.from_saved_model(SAVED_MODEL)\n",
         "  float_tflite_model = float_converter.convert()\n",
         "  float_tflite_model_size = open(FLOAT_MODEL_TFLITE, \"wb\").write(float_tflite_model)\n",
@@ -428,8 +427,8 @@
         "\n",
         "  converter = tf.lite.TFLiteConverter.from_saved_model(SAVED_MODEL)\n",
         "  converter.optimizations = [tf.lite.Optimize.DEFAULT]\n",
-        "  converter.inference_input_type = tf.lite.constants.INT8\n",
-        "  converter.inference_output_type = tf.lite.constants.INT8\n",
+        "  converter.inference_input_type = tf.int8\n",
+        "  converter.inference_output_type = tf.int8\n",
         "  def representative_dataset_gen():\n",
         "    for i in range(100):\n",
         "      data, _ = audio_processor.get_data(1, i*1, model_settings,\n",
@@ -472,7 +471,7 @@
         "def run_tflite_inference(tflite_model_path, model_type=\"Float\"):\n",
         "  # Load test data\n",
         "  np.random.seed(0) # set random seed for reproducible test results.\n",
-        "  with tf.Session() as sess:\n",
+        "  with tf.compat.v1.Session() as sess:\n",
         "    test_data, test_labels = audio_processor.get_data(\n",
         "        -1, 0, model_settings, BACKGROUND_FREQUENCY, BACKGROUND_VOLUME_RANGE,\n",
         "        TIME_SHIFT_MS, 'testing', sess)\n",

tensorflow/lite/micro/kernels/Makefile.inc

@@ -48,13 +48,13 @@ $(eval $(call microlite_test,unidirectional_sequence_lstm_test,\
   $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/testdata/lstm_test_data.cc,\
   $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/testdata/lstm_test_data.h))
 
-$(eval $(call microlite_test,kernel_fft_test,\
+$(eval $(call microlite_test,kernel_signal_fft_test,\
   $(TENSORFLOW_ROOT)signal/micro/kernels/fft_test.cc \
   $(TENSORFLOW_ROOT)signal/micro/kernels/fft_flexbuffers_generated_data.cc \
   $(TENSORFLOW_ROOT)signal/testdata/fft_test_data.cc, \
   $(TENSORFLOW_ROOT)signal/micro/kernels/fft_flexbuffers_generated_data.h))
 
-$(eval $(call microlite_test,kernel_window_test,\
+$(eval $(call microlite_test,kernel_signal_window_test,\
   $(TENSORFLOW_ROOT)signal/micro/kernels/window_test.cc \
   $(TENSORFLOW_ROOT)signal/micro/kernels/window_flexbuffers_generated_data.cc, \
   $(TENSORFLOW_ROOT)signal/micro/kernels/window_flexbuffers_generated_data.h))

tensorflow/lite/micro/kernels/if.cc

-/* Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+/* Copyright 2023 The TensorFlow Authors. All Rights Reserved.
 
 Licensed under the Apache License, Version 2.0 (the "License");
 you may not use this file except in compliance with the License.
@@ -31,6 +31,8 @@ namespace tflite {
 
 namespace {
 
+constexpr int kCondTensor = 0;
+
 struct OpData {
   int then_subgraph_index;
   int else_subgraph_index;
@@ -52,7 +54,8 @@ TfLiteStatus Prepare(TfLiteContext* context, TfLiteNode* node) {
 
   // The first input is the condition.
   tflite::MicroContext* micro_context = tflite::GetMicroContext(context);
-  TfLiteTensor* cond = micro_context->AllocateTempInputTensor(node, 0);
+  TfLiteTensor* cond =
+      micro_context->AllocateTempInputTensor(node, kCondTensor);
 
   TF_LITE_ENSURE(context, cond != nullptr);
   TF_LITE_ENSURE_EQ(context, cond->type, kTfLiteBool);
@@ -86,11 +89,10 @@ TfLiteStatus Eval(TfLiteContext* context, TfLiteNode* node) {
   const OpData* op_data = reinterpret_cast<OpData*>(node->user_data);
 
   tflite::MicroContext* micro_context = tflite::GetMicroContext(context);
-  TfLiteTensor* cond = micro_context->AllocateTempInputTensor(node, 0);
-
+  const TfLiteEvalTensor* cond =
+      tflite::micro::GetEvalInput(context, node, kCondTensor);
   TF_LITE_ENSURE(context, cond != nullptr);
-  bool cond_value = cond->data.b[0];
-  micro_context->DeallocateTempTfLiteTensor(cond);
+  const bool cond_value = cond->data.b[0];
 
   MicroGraph* graph_info = &micro_context->graph();
   // Currently we copy the input / output between the subgraphs.

tensorflow/lite/micro/riscv32_mcu/README.md

-# RISC-V MCU
-
-This folder contains TFLite kernel operations optimized for RISC-V micro
-controllers.
-
-It is designed to be portable even to 'bare metal', so it follows the same
-design goals as the micro experimental port.

tensorflow/lite/micro/riscv32_mcu/debug_log.cc

-/* 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.
-==============================================================================*/
-// TODO(b/121324430): Add test for DebugLog functions
-// TODO(b/121275099): Remove dependency on debug_log once the platform supports
-// printf
-
-#include <stdio.h>
-
-extern "C" void DebugLog(const char* s) { puts(s); }

tensorflow/lite/micro/testing/riscv32_mcu.resc

-# 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.
-# ==============================================================================
-
-using sysbus
-
-mach create
-
-# This is a little hack to increase our ram size from 16k to 256k, since some
-# tests require a larger ram size. The platform's linker script is also modified
-# to account for the larger ram size, see patch_sifive_sdk() in download_and_extract.sh
-set platform
-"""
-using "platforms/cpus/sifive-fe310.repl"
-
-dtim:
-    size: 0x40000
-"""
-
-machine LoadPlatformDescriptionFromString $platform
-
-sysbus Tag <0x10008000 4> "PRCI_HFROSCCFG" 0xFFFFFFFF
-sysbus Tag <0x10008008 4> "PRCI_PLLCFG" 0xFFFFFFFF
-
-showAnalyzer uart0 Antmicro.Renode.Analyzers.LoggingUartAnalyzer
-uart0 CreateFileBackend $logfile true
-
-cpu PerformanceInMips 320

tensorflow/lite/micro/testing/riscv32_mcu.resource.txt

-*** Variables ***
-${UART}                       sysbus.uart0

tensorflow/lite/micro/tools/make/Makefile

@@ -314,9 +314,6 @@ $(TENSORFLOW_ROOT)tensorflow/lite/micro/memory_planner/non_persistent_buffer_pla
 MICROLITE_CC_KERNEL_SRCS := \
 $(TENSORFLOW_ROOT)signal/micro/kernels/rfft.cc \
 $(TENSORFLOW_ROOT)signal/micro/kernels/window.cc \
-$(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/kiss_fft_float.cc \
-$(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/kiss_fft_int16.cc \
-$(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/kiss_fft_int32.cc \
 $(TENSORFLOW_ROOT)signal/src/rfft_float.cc \
 $(TENSORFLOW_ROOT)signal/src/rfft_int16.cc \
 $(TENSORFLOW_ROOT)signal/src/rfft_int32.cc \
@@ -456,7 +453,6 @@ $(TFL_CC_SRCS)
 MICROLITE_CC_HDRS := \
 $(wildcard $(TENSORFLOW_ROOT)signal/micro/kernels/*.h) \
 $(wildcard $(TENSORFLOW_ROOT)signal/src/*.h) \
-$(wildcard $(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/*.h) \
 $(wildcard $(TENSORFLOW_ROOT)tensorflow/lite/micro/*.h) \
 $(wildcard $(TENSORFLOW_ROOT)tensorflow/lite/micro/benchmarks/*model_data.h) \
 $(wildcard $(TENSORFLOW_ROOT)tensorflow/lite/micro/kernels/*.h) \

tensorflow/lite/micro/tools/make/additional_kernels.inc

+# TODO(b/288938993): de-dupe internal wrapper directory and remove this
+MICROLITE_CC_KERNEL_SRCS += \
+$(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/kiss_fft_float.cc \
+$(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/kiss_fft_int16.cc \
+$(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/kiss_fft_int32.cc
+
+MICROLITE_CC_HDRS += \
+$(wildcard $(TENSORFLOW_ROOT)signal/src/kiss_fft_wrappers/*.h) \

tensorflow/lite/micro/tools/make/targets/hexagon_makefile.inc

@@ -53,9 +53,7 @@ PLATFORM_ARGS = \
   -DPTHREAD_STUBS \
   -DUSE_PREALLOCATED_BUFFER \
   -D_HAS_C9X \
-  -DTF_LITE_USE_CTIME \
   -MMD \
-  -DHEXAGON \
   -Wall \
   -Wextra \
   -Wno-missing-field-initializers \