Use CMSIS-NN optimized kernels for Arduino library

This change should provide significant performance improvements when running on the Nano BLE 33 Sense and other Cortex-M-based Arduino boards.

PiperOrigin-RevId: 316195847
Change-Id: Ie5197a2f2be417fca4efa5d1b7f0534ad0c47820

tensorflow/lite/micro/tools/ci_build/test_arduino.sh

@@ -25,14 +25,14 @@ cd "${ROOT_DIR}"
 
 source tensorflow/lite/micro/tools/ci_build/helper_functions.sh
 
-readable_run make -f tensorflow/lite/micro/tools/make/Makefile clean
+readable_run make -f tensorflow/lite/micro/tools/make/Makefile clean clean_downloads
 
 TARGET=arduino
 
 # TODO(b/143715361): parallel builds do not work with generated files right now.
 readable_run make -f tensorflow/lite/micro/tools/make/Makefile \
   TARGET=${TARGET} \
-  TAGS="portable_optimized" \
+  TAGS="cmsis-nn" \
   generate_arduino_zip
 
 readable_run tensorflow/lite/micro/tools/ci_build/install_arduino_cli.sh

tensorflow/lite/micro/tools/make/download_and_extract.sh

@@ -114,6 +114,15 @@ patch_cmsis() {
     -iname '*.*' -exec \
     sed -i -E $'s@#include "arm_nn_tables.h"@#include "cmsis/CMSIS/NN/Include/arm_nn_tables.h"@g' {} \;
 
+  # Until the fix for https://github.com/ARMmbed/mbed-os/issues/12568 is
+  # rolled into Mbed version used on the Arduino IDE, we have to replace
+  # one intrinsic with a patched equivalent.
+  sed -i -E 's@__SXTB16_RORn@__patched_SXTB16_RORn@g' \
+    tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c
+
+  sed -i -E $'33 a \\\n\\\n// Work around for https://github.com/ARMmbed/mbed-os/issues/12568\\\n__STATIC_FORCEINLINE uint32_t __patched_SXTB16_RORn(uint32_t op1, uint32_t rotate) {\\\n  uint32_t result;\\\n  __ASM ("sxtb16 %0, %1, ROR %2" : "=r" (result) : "r" (op1), "i" (rotate) );\\\n  return result;\\\n}' \
+    tensorflow/lite/micro/tools/make/downloads/cmsis/CMSIS/NN/Source/NNSupportFunctions/arm_nn_mat_mult_nt_t_s8.c
+
   echo "Finished patching CMSIS"
 }