added support for patching cmsis lib to use builtin memcpy and memset

tensorflow/lite/micro/tools/rpi0_pip_package/build_pip_package.sh

@@ -21,7 +21,7 @@ TF_DIR=$ROOT_DIR/tensorflow
 TFLITE_DIR=$TF_DIR/lite 
 TFLM_DIR=$TFLITE_DIR/micro
 
-export PACKAGE_VERSION=1.0.0
+export PACKAGE_VERSION=1.1.0
 PYTHON=python3.7
 BUILD_DIR=/tmp/tflite_micro_runtime_build
 GIT_HASH=`git rev-parse HEAD`

tensorflow/lite/micro/tools/rpi0_pip_package/install_tflite_micro.sh

@@ -16,5 +16,6 @@ $TFLM_DIR/tools/make/flatbuffers_download.sh $TFLM_DIR/tools/make/downloads
 $TFLM_DIR/tools/make/ruy_download.sh $TFLM_DIR/tools/make/downloads
 $TFLM_DIR/tools/make/ext_libs/cmsis_download.sh $TFLM_DIR/tools/make/downloads
 
+python3 $SCRIPT_DIR/patch_cmsis.py
 
 

tensorflow/lite/micro/tools/rpi0_pip_package/patch_cmsis.py

@@ -0,0 +1,24 @@
+import os 
+import sys 
+
+curdir = os.path.dirname(os.path.abspath(__file__))
+cmsis_arm_math_types_h_path = f'{curdir}/../make/downloads/cmsis/CMSIS/DSP/Include/arm_math_types.h'
+
+
+with open(cmsis_arm_math_types_h_path, 'r') as fp:
+    data = fp.read()
+
+if '#define memset __builtin_memset' in data:
+    sys.exit()
+
+
+# Ensure the builtin memset and memcpy
+# are used. This greatly reduce the overhead of the kernels.
+patch_data = '#define _ARM_MATH_TYPES_H_\n\n'
+patch_data += '#define memset __builtin_memset\n'
+patch_data += '#define memcpy __builtin_memcpy\n\n'
+
+data = data.replace('#define _ARM_MATH_TYPES_H_', patch_data)
+
+with open(cmsis_arm_math_types_h_path, 'w') as fp:
+    fp.write(data)