samples: tflite_ethosu: support numaker_m55m1

This supports nuvoton numaker_m55m1 board, including:
1. Choosing NPU buffer as cacheable or non-cacheable, dependent on
   CONFIG_NOCACHE_MEMORY. When cachealbe is chosen, data cache flush/
   invalidate weak functions (ethos-u core driver) are also overridden.
2. Add Vela-compiled model which matches this targets's NPU, Ethos-U55
   with 256 macs_per_cc

Besides, re-organize for new board support, including:
1. Refactor Vela-compiled models which must match target Ethos-U RTL config
2. Add non-Vela-compiled model for evaluation of no-Ethos-U platform

Signed-off-by: Chun-Chieh Li <[email protected]>

Author: ccli8

samples/modules/tflite-micro/tflm_ethosu/CMakeLists.txt

@@ -8,8 +8,17 @@ find_package(Zephyr REQUIRED HINTS $ENV{ZEPHYR_BASE})
 
 project(tflm_ethosu_app)
 
-target_include_directories(app PRIVATE src/models/keyword_spotting_cnn_small_int8)
+# Choose appropriate model
+if(CONFIG_TAINT_BLOBS_TFLM_ETHOSU)
+  # Vela-compiled model must match Ethos-U RTL config, e.g. MACs per cycle
+  target_include_directories_ifdef(CONFIG_BOARD_MPS3 app PRIVATE src/models/keyword_spotting_cnn_small_int8_ethos_u55_128)
+  target_include_directories_ifdef(CONFIG_BOARD_NUMAKER_M55M1 app PRIVATE src/models/keyword_spotting_cnn_small_int8_ethos_u55_256)
+elseif(CONFIG_TAINT_BLOBS_TFLM)
+  target_include_directories(app PRIVATE src/models/keyword_spotting_cnn_small_int8)
+endif()
 
 target_sources(app PRIVATE src/main.cpp src/inference_process.cpp)
+target_sources_ifdef(CONFIG_BOARD_NUMAKER_M55M1 app PRIVATE src/npu_cache.c)
 
-zephyr_linker_sources(SECTIONS linker.ld)
+# Add platform specific linker source snippet
+zephyr_linker_sources_ifdef(CONFIG_BOARD_MPS3 SECTIONS linker.ld)

samples/modules/tflite-micro/tflm_ethosu/Kconfig

@@ -1,9 +1,14 @@
 # Copyright 2022 Arm Limited and/or its affiliates <[email protected]>
 # SPDX-License-Identifier: Apache-2.0
 
-config TFLM_ETHOSU_TAINT_BLOBS
+config TAINT_BLOBS_TFLM
 	bool
 	default y
 	select TAINT_BLOBS
 
+config TAINT_BLOBS_TFLM_ETHOSU
+	bool "Choose Vela-compiled model targeting Ethos-U"
+	default y
+	depends on TAINT_BLOBS_TFLM
+
 source "Kconfig.zephyr"

samples/modules/tflite-micro/tflm_ethosu/boards/numaker_m55m1.overlay

@@ -0,0 +1,5 @@
+/* SPDX-License-Identifier: Apache-2.0 */
+
+&npu0 {
+	status = "okay";
+};

samples/modules/tflite-micro/tflm_ethosu/src/inference_process.cpp

@@ -118,9 +118,16 @@ bool InferenceProcess::runJob(InferenceJob &job)
 	}
 
 	/* Create the TFL micro interpreter */
+#ifdef CONFIG_TAINT_BLOBS_TFLM_ETHOSU
 	tflite::MicroMutableOpResolver <1> resolver;
 	resolver.AddEthosU();
-
+#else
+	tflite::MicroMutableOpResolver <4> resolver;
+	resolver.AddReshape();
+	resolver.AddConv2D();
+	resolver.AddFullyConnected();
+	resolver.AddSoftmax();
+#endif
 	tflite::MicroInterpreter interpreter(model, resolver, tensorArena, tensorArenaSize);
 
 	/* Allocate tensors */

samples/modules/tflite-micro/tflm_ethosu/src/main.cpp

@@ -98,7 +98,7 @@ volatile int totalCompletedJobs = 0;
 /* TensorArena static initialisation */
 const size_t arenaSize = TENSOR_ARENA_SIZE_PER_INFERENCE;
 
-__attribute__((section("tflm_arena"), aligned(16)))
+TENSOR_ARENA_ATTR
 uint8_t inferenceProcessTensorArena[NUM_INFERENCE_TASKS][arenaSize];
 
 /* Allocate and initialize heap */

samples/modules/tflite-micro/tflm_ethosu/src/models/keyword_spotting_cnn_small_int8/input.h

@@ -4,7 +4,7 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-__aligned(4) __attribute__((section("tflm_input"))) uint8_t inputData[] = {
+__aligned(4) __attribute__((section(".rodata.tflm_input"))) uint8_t inputData[] = {
 	0x2c, 0x8a, 0xff, 0x0c, 0xaf, 0x2a, 0x44, 0x17, 0xf5, 0x26, 0x96, 0x37, 0x40, 0x4c, 0xa1,
 	0x58, 0xc3, 0x33, 0xce, 0x1a, 0x7b, 0xd2, 0x22, 0x5b, 0x43, 0xf6, 0xfd, 0x0b, 0xe7, 0xfd,
 	0x65, 0x58, 0x89, 0x24, 0xf4, 0xec, 0x53, 0x5e, 0x21, 0x1f, 0x95, 0xd1, 0xd9, 0x25, 0x72,

samples/modules/tflite-micro/tflm_ethosu/src/models/keyword_spotting_cnn_small_int8/model.h

@@ -4,5 +4,5 @@
  * SPDX-License-Identifier: Apache-2.0
  */
 
-__aligned(4) __attribute__((section("tflm_output"))) uint8_t expectedOutputData[] = {
+__aligned(4) __attribute__((section(".rodata.tflm_output"))) uint8_t expectedOutputData[] = {
 	0x80, 0x23, 0x80, 0x80, 0x80, 0x80, 0x80, 0x83, 0x80, 0x80, 0x81, 0xd9 };

samples/modules/tflite-micro/tflm_ethosu/src/models/keyword_spotting_cnn_small_int8/output.h

@@ -0,0 +1,41 @@
+/*
+ * Copyright 2022 Arm Limited and/or its affiliates <[email protected]>
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+__aligned(4) __attribute__((section("tflm_input"))) uint8_t inputData[] = {
+	0x2c, 0x8a, 0xff, 0x0c, 0xaf, 0x2a, 0x44, 0x17, 0xf5, 0x26, 0x96, 0x37, 0x40, 0x4c, 0xa1,
+	0x58, 0xc3, 0x33, 0xce, 0x1a, 0x7b, 0xd2, 0x22, 0x5b, 0x43, 0xf6, 0xfd, 0x0b, 0xe7, 0xfd,
+	0x65, 0x58, 0x89, 0x24, 0xf4, 0xec, 0x53, 0x5e, 0x21, 0x1f, 0x95, 0xd1, 0xd9, 0x25, 0x72,
+	0x56, 0xe6, 0xe2, 0xa4, 0x37, 0x85, 0xf0, 0xd7, 0xba, 0xab, 0xcc, 0xc6, 0xbc, 0xcb, 0x64,
+	0x58, 0x3d, 0x04, 0x8e, 0xd8, 0x1a, 0x32, 0x76, 0x0c, 0x4d, 0x4c, 0xc5, 0xba, 0xb9, 0xa9,
+	0xe2, 0x41, 0xc2, 0xc8, 0xfa, 0x66, 0xfd, 0x2e, 0x4a, 0xa7, 0xca, 0x6a, 0x4f, 0xd7, 0x28,
+	0xe5, 0x07, 0x2e, 0x48, 0x5f, 0xfa, 0xd8, 0xde, 0xeb, 0x11, 0xd1, 0x0b, 0x0d, 0xe4, 0x25,
+	0x66, 0x73, 0x6c, 0x99, 0xc2, 0x89, 0x56, 0xcd, 0xeb, 0xaf, 0x92, 0xc8, 0x18, 0xdf, 0xd6,
+	0x89, 0x9b, 0xce, 0x96, 0x14, 0x17, 0x6e, 0x25, 0xf3, 0x88, 0xad, 0x85, 0x50, 0x93, 0xc5,
+	0xde, 0x73, 0x12, 0xa6, 0x55, 0x45, 0x9e, 0x88, 0x75, 0x7e, 0xc7, 0xb5, 0x47, 0xcf, 0x87,
+	0x6e, 0xa3, 0x2f, 0x9d, 0x39, 0x5e, 0x40, 0x89, 0x3e, 0x5e, 0xd2, 0x3b, 0x06, 0x7a, 0xe3,
+	0xe0, 0xbe, 0xf9, 0x58, 0x27, 0x15, 0x4c, 0x31, 0xfc, 0x88, 0x4d, 0x73, 0x02, 0xa3, 0xf6,
+	0x9d, 0x83, 0xbe, 0x05, 0x13, 0x42, 0x50, 0x47, 0x13, 0x8c, 0x4b, 0x2d, 0x0e, 0xbb, 0xc7,
+	0x9e, 0x27, 0xe1, 0xfb, 0x38, 0xa0, 0x5d, 0x51, 0x23, 0x41, 0xbf, 0x7f, 0x14, 0x89, 0x4d,
+	0x32, 0xa4, 0x39, 0xca, 0x11, 0x09, 0xff, 0x52, 0x55, 0x71, 0xa0, 0x22, 0xb4, 0x42, 0x9f,
+	0x5a, 0x17, 0x05, 0x4a, 0x42, 0x9c, 0x9b, 0x74, 0xc8, 0xa7, 0xea, 0x17, 0x60, 0xbd, 0xf9,
+	0x23, 0x6a, 0x39, 0xc3, 0x7e, 0xf9, 0x33, 0xaf, 0x4b, 0xad, 0xb4, 0x46, 0xf2, 0x2b, 0x26,
+	0x3c, 0x37, 0xc4, 0x46, 0xf4, 0x9c, 0x58, 0xef, 0xb7, 0xa2, 0x63, 0x04, 0x11, 0x00, 0x28,
+	0x1d, 0xa2, 0x00, 0xe8, 0x4f, 0x84, 0x24, 0x67, 0x7a, 0xd2, 0xb5, 0xdc, 0x9d, 0x1e, 0x05,
+	0x97, 0xd1, 0xa6, 0xa6, 0xce, 0x32, 0x1c, 0x68, 0x8d, 0x6a, 0xb8, 0x74, 0xe2, 0x6f, 0x1d,
+	0x91, 0x71, 0x70, 0xe2, 0xcf, 0xf9, 0x19, 0x71, 0x04, 0x31, 0x90, 0x67, 0xe9, 0x46, 0x0b,
+	0x2e, 0xaa, 0x6c, 0xaf, 0xf3, 0x3a, 0x35, 0x88, 0xdc, 0x9f, 0x3a, 0x71, 0x1c, 0xf8, 0xd3,
+	0x61, 0xef, 0x81, 0xd1, 0x80, 0x67, 0xc1, 0x75, 0x17, 0x32, 0x67, 0xef, 0xea, 0x99, 0x29,
+	0x6f, 0xea, 0x8f, 0xb9, 0xe8, 0xc5, 0x78, 0xa3, 0xb9, 0x31, 0x2a, 0xe6, 0xaf, 0xb4, 0x27,
+	0xf7, 0x6c, 0x2e, 0x2b, 0x8b, 0x3b, 0x40, 0xab, 0x2e, 0x21, 0x5b, 0xb5, 0xd2, 0x5c, 0x2f,
+	0xdb, 0xdb, 0xd0, 0x81, 0xa1, 0x00, 0x77, 0xac, 0x40, 0x0e, 0x69, 0x3f, 0xd0, 0xe3, 0x4a,
+	0x1a, 0x1b, 0xb5, 0xa6, 0x9c, 0xdd, 0x0c, 0xcb, 0xb6, 0xd3, 0xf9, 0xea, 0x78, 0x11, 0x2a,
+	0xbf, 0x56, 0x89, 0xd4, 0xb0, 0xc8, 0xf0, 0x4b, 0x48, 0xdf, 0x22, 0xc4, 0x91, 0x86, 0x7d,
+	0x86, 0x33, 0xfd, 0x75, 0x44, 0x1b, 0x9f, 0x9a, 0xaf, 0xc4, 0x9d, 0x27, 0xff, 0x72, 0xf8,
+	0xb5, 0x74, 0xb7, 0x13, 0x22, 0x03, 0x14, 0xcb, 0xa9, 0x4c, 0x73, 0x4b, 0x7e, 0xe4, 0x9b,
+	0xb2, 0x27, 0x34, 0xa4, 0x74, 0x14, 0x68, 0x37, 0x57, 0xc0, 0xce, 0x8f, 0x94, 0xea, 0x0f,
+	0xa4, 0xe0, 0xf7, 0x14, 0x46, 0x05, 0xf9, 0x63, 0x2c, 0x12, 0xbe, 0x3a, 0xdb, 0x1e, 0x1f,
+	0x97, 0xbb, 0x32, 0xa8, 0x4f, 0x2a, 0x07, 0xd6, 0x0d, 0x0b
+};