Update full tflite to 2.16.1, add native Mac ARM64 libs (#22)

* updated source * updated compiled libs * bash script and readme fixes * Makefile and readme changes for mac arm64 and TI

Author: AIWintermuteAI

Makefile

@@ -83,29 +83,32 @@ CFLAGS += -Itensorflow-lite/
 CCSOURCES += $(wildcard edge-impulse-sdk/tensorflow/lite/kernels/custom/*.cc)
 
 ifeq (${TARGET_LINUX_ARMV7},1)
-LDFLAGS += -L./tflite/linux-armv7 -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lflatbuffers -lruy -lXNNPACK -lpthreadpool -lpthread -lcpuinfo -lclog
+LDFLAGS += -L./tflite/linux-armv7 -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lflatbuffers -lruy -lXNNPACK -lpthreadpool -lpthread -lcpuinfo
 endif # TARGET_LINUX_ARMV7
 ifeq (${TARGET_LINUX_AARCH64},1)
 CFLAGS += -DDISABLEFLOAT16
-LDFLAGS += -L./tflite/linux-aarch64 -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lclog -lpthread
+LDFLAGS += -L./tflite/linux-aarch64 -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lpthread
 endif # TARGET_LINUX_AARCH64
 ifeq (${TARGET_LINUX_X86},1)
-LDFLAGS += -L./tflite/linux-x86 -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lclog -lpthread
+LDFLAGS += -L./tflite/linux-x86 -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lpthread
 endif # TARGET_LINUX_X86
 ifeq (${TARGET_MAC_X86_64},1)
-LDFLAGS += -L./tflite/mac-x86_64 -ltensorflow-lite -lcpuinfo -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lpthreadpool -lclog
+LDFLAGS += -L./tflite/mac-x86_64 -ltensorflow-lite -lcpuinfo -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lpthreadpool
 endif # TARGET_MAC_X86_64
+ifeq (${TARGET_MAC_ARM64},1)
+LDFLAGS += -L./tflite/mac-arm64 -ltensorflow-lite -lcpuinfo -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lpthreadpool
+endif # TARGET_MAC_ARM64
 
 ifeq (${LINK_TFLITE_FLEX_LIBRARY},1)
-LDFLAGS += -ltensorflowlite_flex_2.6.5
+LDFLAGS += -ltensorflowlite_flex_2.16.1
 endif
 
 else ifeq (${USE_AKIDA},1) # USE_FULL_TFLITE
 CFLAGS += -DEI_CLASSIFIER_USE_FULL_TFLITE=1
 CFLAGS += -DPYBIND11_DETAILED_ERROR_MESSAGES # add more detailed pybind error descriptions
 CFLAGS += -Itensorflow-lite
 CFLAGS += -Iedge-impulse-sdk/third_party/gemmlowp
-LDFLAGS += -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lclog -lpthread
+LDFLAGS += -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lpthread
 ifeq (${TARGET_LINUX_AARCH64},1)
 CFLAGS += $(shell $(PYTHON_CROSS_PATH)python3-config --cflags)
 LDFLAGS += -L./tflite/linux-aarch64
@@ -125,6 +128,8 @@ CCSOURCES += $(wildcard edge-impulse-sdk/tensorflow/lite/kernels/*.cc) $(wildcar
 endif # not USE_FULL_TFLITE
 
 ifeq (${USE_MEMRYX},1)
+CFLAGS += -Iedge-impulse-sdk/third_party/gemmlowp
+LDFLAGS += -Wl,--no-as-needed -ldl -ltensorflow-lite -lfarmhash -lfft2d_fftsg -lfft2d_fftsg2d -lruy -lXNNPACK -lcpuinfo -lpthreadpool -lpthread
 ifeq (${TARGET_LINUX_AARCH64},1)
 $(error MemryX drivers and runtime do not support AARCH64)
 else ifeq (${TARGET_LINUX_X86},1)
@@ -133,6 +138,7 @@ CFLAGS += $(shell python3-config --cflags)
 CFLAGS += -DPYBIND11_DETAILED_ERROR_MESSAGES
 LDFLAGS += -rdynamic $(shell python3-config --ldflags --embed)
 else
+LDFLAGS += -L./tflite/linux-x86
 LDFLAGS += -lmemx
 endif # USE_MEMRYX_SOFTWARE
 endif # USE_MEMRYX && TARGET_LINUX_X86

README.md

@@ -126,11 +126,9 @@ $ APP_CUSTOM=1 TARGET_MAC_X86_64=1 USE_FULL_TFLITE=1 make -j
 Build with the following flags:
 
 ```
-$ APP_CUSTOM=1 TARGET_MAC_X86_64=1 USE_FULL_TFLITE=1 arch -x86_64 /usr/bin/make -j
+$ APP_CUSTOM=1 TARGET_MAC_ARM64=1 USE_FULL_TFLITE=1 /usr/bin/make -j
 ```
 
-Note that this does build an x86 binary, but it runs very fast through Rosetta.
-
 ### AARCH64 with AI Acceleration
 
 #### NVIDIA Jetson Orin - TensorRT (JetPack 5.1.2)
@@ -224,13 +222,16 @@ In case of any issues during runtime, check [Troubleshooting](https://docs.edgei
 
 You can also build with support for TIDL, this fully leverages the Deep Learning Accelerator on the Texas Instruments TDA4VM (AM68PA), AM62A, AM68A.
 
+> [!IMPORTANT]
+> Texas Instruments boards are legacy-supported. Current version of Tensorflow Lite (2.16.1) is not supported, you need to checkout the earlier commit before proceeding.
+> `git checkout ab4fa0758093a0ebb713e49462ae3c615d19bfa1`
+
 ##### TDA4VM (AM68PA)
 
 1. Go to the **Deployment** page in the Edge Impulse Studio.
-1. Select the 'TIDL-RT Library', and the 'float32' optimizations.
-1. Build the library and copy the folders into this repository.
-1. Build your (.eim) application:
-
+2. Select the 'TIDL-RT Library', and the 'float32' optimizations.
+3. Build the library and copy the folders into this repository.
+4. Build your (.eim) application:
  ```
  $ APP_EIM=1 TARGET_TDA4VM=1 make -j
  ```
@@ -244,9 +245,9 @@ $ APP_EIM=1 TARGET_TDA4VM=1 USE_ONNX=1 make -j
 ##### TI AM62A
 
 1. Go to the **Deployment** page in the Edge Impulse Studio.
-1. Select the 'TIDL-RT Library (AM62A)', and the 'float32' optimizations.
-1. Build the library and copy the folders into this repository.
-1. Build your (.eim) application:
+2. Select the 'TIDL-RT Library (AM62A)', and the 'float32' optimizations.
+3. Build the library and copy the folders into this repository.
+4. Build your (.eim) application:
 
  ```
  $ APP_EIM=1 TARGET_AM62A=1 make -j
@@ -261,9 +262,9 @@ $ APP_EIM=1 TARGET_AM62A=1 USE_ONNX=1 make -j
 ##### TI AM68A
 
 1. Go to the **Deployment** page in the Edge Impulse Studio.
-1. Select the 'TIDL-RT Library (AM68A)', and the 'float32' optimizations.
-1. Build the library and copy the folders into this repository.
-1. Build your (.eim) application:
+2. Select the 'TIDL-RT Library (AM68A)', and the 'float32' optimizations.
+3. Build the library and copy the folders into this repository.
+4. Build your (.eim) application:
 
  ```
  $ APP_EIM=1 TARGET_AM68A=1 make -j

tensorflow-lite/flatbuffers/allocator.h

@@ -0,0 +1,68 @@
+/*
+ * Copyright 2021 Google Inc. 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 FLATBUFFERS_ALLOCATOR_H_
+#define FLATBUFFERS_ALLOCATOR_H_
+
+#include "tensorflow-lite/flatbuffers/base.h"
+
+namespace flatbuffers {
+
+// Allocator interface. This is flatbuffers-specific and meant only for
+// `vector_downward` usage.
+class Allocator {
+ public:
+ virtual ~Allocator() {}
+
+ // Allocate `size` bytes of memory.
+ virtual uint8_t *allocate(size_t size) = 0;
+
+ // Deallocate `size` bytes of memory at `p` allocated by this allocator.
+ virtual void deallocate(uint8_t *p, size_t size) = 0;
+
+ // Reallocate `new_size` bytes of memory, replacing the old region of size
+ // `old_size` at `p`. In contrast to a normal realloc, this grows downwards,
+ // and is intended specifcally for `vector_downward` use.
+ // `in_use_back` and `in_use_front` indicate how much of `old_size` is
+ // actually in use at each end, and needs to be copied.
+ virtual uint8_t *reallocate_downward(uint8_t *old_p, size_t old_size,
+ size_t new_size, size_t in_use_back,
+ size_t in_use_front) {
+ FLATBUFFERS_ASSERT(new_size > old_size); // vector_downward only grows
+ uint8_t *new_p = allocate(new_size);
+ memcpy_downward(old_p, old_size, new_p, new_size, in_use_back,
+ in_use_front);
+ deallocate(old_p, old_size);
+ return new_p;
+ }
+
+ protected:
+ // Called by `reallocate_downward` to copy memory from `old_p` of `old_size`
+ // to `new_p` of `new_size`. Only memory of size `in_use_front` and
+ // `in_use_back` will be copied from the front and back of the old memory
+ // allocation.
+ void memcpy_downward(uint8_t *old_p, size_t old_size, uint8_t *new_p,
+ size_t new_size, size_t in_use_back,
+ size_t in_use_front) {
+ memcpy(new_p + new_size - in_use_back, old_p + old_size - in_use_back,
+ in_use_back);
+ memcpy(new_p, old_p, in_use_front);
+ }
+};
+
+} // namespace flatbuffers
+
+#endif // FLATBUFFERS_ALLOCATOR_H_