drivers: misc: ethos_u: support nuvoton numaker m55m1x

MAINTAINERS.yml

@@ -5081,6 +5081,7 @@
     - wearyzen
   collaborators:
     - ithinuel
+    - ccli8
   files:
     - drivers/misc/ethos_u/
     - modules/hal_ethos_u/
@@ -5115,7 +5116,7 @@
 
 "West project: hal_microchip":
   status: maintained
   maintainers:
     - jvasanth1
   collaborators:
     - VenkatKotakonda

drivers/misc/ethos_u/CMakeLists.txt

@@ -5,3 +5,4 @@
 zephyr_library()
 zephyr_library_sources(ethos_u_common.c)
 zephyr_library_sources_ifdef(CONFIG_ETHOS_U_ARM ethos_u_arm.c)
+zephyr_library_sources_ifdef(CONFIG_ETHOS_U_NUMAKER ethos_u_numaker.c)

drivers/misc/ethos_u/Kconfig

@@ -5,10 +5,16 @@ choice
 	prompt "Select vendor Ethos-U NPU driver"
 	depends on ETHOS_U
 	default ETHOS_U_ARM if DT_HAS_ARM_ETHOS_U_ENABLED
+	default ETHOS_U_NUMAKER if DT_HAS_NUVOTON_NUMAKER_NPU_ENABLED
 
 config ETHOS_U_ARM
 	bool "Arm Ethos-U NPU driver"
 	help
 	  Enables Arm Ethos-U NPU driver.
 
+config ETHOS_U_NUMAKER
+	bool "Nuvoton NuMaker Ethos-U NPU driver"
+	help
+	  Enables Nuvoton NuMaker frontend of Arm Ethos-U NPU driver
+
 endchoice

drivers/misc/ethos_u/ethos_u_numaker.c

@@ -0,0 +1,130 @@
+/*
+ * Copyright (c) 2025 Nuvoton Technology Corporation.
+ *
+ * SPDX-License-Identifier: Apache-2.0
+ */
+
+#define DT_DRV_COMPAT nuvoton_numaker_npu
+
+#include <zephyr/kernel.h>
+#include <zephyr/drivers/clock_control.h>
+#include <zephyr/drivers/clock_control/clock_control_numaker.h>
+#include <zephyr/drivers/reset.h>
+
+#include <ethosu_driver.h>
+
+#include <zephyr/logging/log.h>
+LOG_MODULE_REGISTER(ethos_u_numaker, CONFIG_ETHOS_U_LOG_LEVEL);
+
+#include <soc.h>
+
+struct ethos_u_numaker_config {
+	void *base_addr;
+	const struct device *clkctrl_dev;
+	struct numaker_scc_subsys pcc;
+	struct reset_dt_spec reset;
+	void (*irq_config)(const struct device *dev);
+	bool secure_enable;
+	bool privilege_enable;
+	uint8_t flush_mask;
+	uint8_t invalidate_mask;
+};
+
+struct ethos_u_numaker_data {
+	struct ethosu_driver drv;
+};
+
+static void ethos_u_numaker_irq_handler(const struct device *dev)
+{
+	struct ethos_u_numaker_data *data = dev->data;
+	struct ethosu_driver *drv = &data->drv;
+
+	ethosu_irq_handler(drv);
+}
+
+static int ethos_u_numaker_init(const struct device *dev)
+{
+	const struct ethos_u_numaker_config *config = dev->config;
+	struct ethos_u_numaker_data *data = dev->data;
+	struct ethosu_driver *drv = &data->drv;
+	int rc;
+
+	/* Invoke Clock controller to enable module clock */
+
+	/* Equivalent to CLK_EnableModuleClock() */
+	rc = clock_control_on(config->clkctrl_dev, (clock_control_subsys_t)&config->pcc);
+	if (rc < 0) {
+		return rc;
+	}
+	rc = clock_control_on(config->clkctrl_dev, (clock_control_subsys_t)&config->pcc);
+	if (rc < 0) {
+		return rc;
+	}
+
+	/* Equivalent to CLK_SetModuleClock() */
+	rc = clock_control_configure(config->clkctrl_dev, (clock_control_subsys_t)&config->pcc,
+				     NULL);
+	if (rc < 0) {
+		return rc;
+	}
+
+	/* Invoke Reset controller to reset module to default state */
+	/* Equivalent to SYS_ResetModule() */
+	rc = reset_line_toggle_dt(&config->reset);
+	if (rc < 0) {
+		return rc;
+	}
+
+	LOG_DBG("Ethos-U DTS info: base_addr=0x%p, secure_enable=%u, privilege_enable=%u",
+		config->base_addr, config->secure_enable, config->privilege_enable);
+
+	if (ethosu_init(drv, config->base_addr, NULL, 0, config->secure_enable,
+			config->privilege_enable)) {
+		LOG_ERR("Failed to initialize NPU with ethosu_init().");
+		return -EINVAL;
+	}
+
+	ethosu_set_basep_cache_mask(drv, config->flush_mask, config->invalidate_mask);
+
+	config->irq_config(dev);
+
+	return 0;
+}
+
+/* Peripheral Clock Control */
+#define NUMAKER_PCC_INST_GET(inst)                                                                 \
+	{                                                                                          \
+		.subsys_id = NUMAKER_SCC_SUBSYS_ID_PCC,                                            \
+		.pcc.clk_modidx = DT_INST_CLOCKS_CELL(inst, clock_module_index),                   \
+		.pcc.clk_src = DT_INST_CLOCKS_CELL(inst, clock_source),                            \
+		.pcc.clk_div = DT_INST_CLOCKS_CELL(inst, clock_divider),                           \
+	}
+
+#define NUMAKER_ETHOS_U_INIT(inst)                                                                 \
+	static void ethos_u_numaker_irq_config_##inst(const struct device *dev)                    \
+	{                                                                                          \
+		IRQ_CONNECT(DT_INST_IRQ(inst, irq), DT_INST_IRQ(inst, priority),                   \
+			    ethos_u_numaker_irq_handler, DEVICE_DT_INST_GET(inst), 0);             \
+                                                                                                   \
+		irq_enable(DT_INST_IRQ(inst, irq));                                                \
+	}                                                                                          \
+                                                                                                   \
+	static const struct ethos_u_numaker_config ethos_u_numaker_config_##inst = {               \
+		.base_addr = (void *)DT_INST_REG_ADDR(inst),                                       \
+		.clkctrl_dev = DEVICE_DT_GET(DT_PARENT(DT_INST_CLOCKS_CTLR(inst))),                \
+		.pcc = NUMAKER_PCC_INST_GET(inst),                                                 \
+		.reset = RESET_DT_SPEC_INST_GET(inst),                                             \
+		.irq_config = ethos_u_numaker_irq_config_##inst,                                   \
+		.secure_enable = DT_INST_PROP(inst, secure_enable),                                \
+		.privilege_enable = DT_INST_PROP(inst, privilege_enable),                          \
+		.flush_mask = DT_INST_PROP(inst, flush_mask),                                      \
+		.invalidate_mask = DT_INST_PROP(inst, invalidate_mask),                            \
+	};                                                                                         \
+                                                                                                   \
+	static struct ethos_u_numaker_data ethos_u_numaker_data_##inst;                            \
+                                                                                                   \
+	DEVICE_DT_INST_DEFINE(inst, ethos_u_numaker_init, NULL, &ethos_u_numaker_data_##inst,      \
+			      &ethos_u_numaker_config_##inst, POST_KERNEL,                         \
+			      CONFIG_KERNEL_INIT_PRIORITY_DEFAULT, NULL);
+
+DT_INST_FOREACH_STATUS_OKAY(NUMAKER_ETHOS_U_INIT);

dts/arm/nuvoton/m55m1h2l.dtsi

@@ -31,5 +31,16 @@
 				reg = <0x100000 DT_SIZE_K(2048)>;
 			};
 		};
+
+		npu0: npu@40003000 {
+			compatible = "nuvoton,numaker-npu";
+			reg = <0x40003000 0x1000>;
+			interrupts = <13 0>;
+			resets = <&rst NUMAKER_SYS_NPURST>;
+			clocks = <&pcc NUMAKER_NPU0_MODULE 0 0>;
+			secure-enable;
+			privilege-enable;
+			status = "disabled";
+		};
 	};
 };

dts/bindings/arm/nuvoton,numaker-npu.yaml

@@ -0,0 +1,37 @@
+# Copyright (c) 2025 Nuvoton Technology Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+description: Nuvoton NuMaker frontend of Arm Ethos-U NPU driver
+
+compatible: "nuvoton,numaker-npu"
+
+include: ["arm,ethos-u.yaml", reset-device.yaml]
+
+properties:
+  reg:
+    required: true
+
+  interrupts:
+    required: true
+
+  resets:
+    required: true
+
+  clocks:
+    required: true
+
+  flush-mask:
+    type: int
+    default: 2
+    description: |
+      Base pointer cache flush mask passed to ethos-u core driver
+      ethosu_set_basep_cache_mask(). Default is to follow ethos-u
+      core driver for scratch region.
+
+  invalidate-mask:
+    type: int
+    default: 2
+    description: |
+      Base pointer cache invalidation mask passed to ethos-u core driver
+      ethosu_set_basep_cache_mask(). Default is to follow ethos-u
+      core driver for scratch region.

lib/libc/newlib/libc-hooks.c

@@ -307,6 +307,19 @@ __weak void _exit(int status)
 	}
 }
 
+/* Address undefined _fini for the exit call
+ *
+ * With gcc command-line options -nostartfiles or -nostdlib, the gcc
+ * built-in object files crti.o/crtn.o which implement _init/_fini
+ * won't get linked in automatically and will cause undefined reference
+ * to _fini error when exit is invoked.
+ *
+ * This is fixed by providing one dummey _fini to let linker pass.
+ */
+__weak void _fini(void)
+{
+}
+
 #ifndef CONFIG_NEWLIB_LIBC_CUSTOM_SBRK
 void *_sbrk(intptr_t count)
 {

modules/hal_ethos_u/Kconfig

@@ -16,6 +16,7 @@ choice  ETHOS_U_NPU_CONFIG
 	default ETHOS_U55_128 if SOC_SERIES_MPS3
 	default ETHOS_U65_256 if SOC_MPS4_CORSTONE315
 	default ETHOS_U85_256 if SOC_MPS4_CORSTONE320
+	default ETHOS_U55_256 if SOC_SERIES_M55M1X
 config ETHOS_U55_64
 	bool "using Ethos-U55 with 64 macs"
 config ETHOS_U55_128

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

@@ -8,8 +8,22 @@ 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. U55/U65/U85 and 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)
+
+  # No Vela-compiled model for these boards, fall back to not optimized
+  target_include_directories_ifdef(CONFIG_BOARD_MPS4_CORSTONE315_FVP app PRIVATE src/models/keyword_spotting_cnn_small_int8)
+  target_include_directories_ifdef(CONFIG_BOARD_MPS4_CORSTONE320_FVP app PRIVATE src/models/keyword_spotting_cnn_small_int8)
+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)
+zephyr_linker_sources_ifdef(CONFIG_BOARD_MPS4 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 */
@@ -225,18 +225,15 @@ int main()
 	k_queue_init(&inferenceQueue);
 
 	/* inferenceSender tasks to create and queue the jobs */
+	const size_t inferenceSenderStackSize = 2048;
+	static K_THREAD_STACK_ARRAY_DEFINE(inferenceSenderStacks, NUM_JOB_TASKS,
+					   inferenceSenderStackSize);
 	for (int n = 0; n < NUM_JOB_TASKS; n++) {
-		const size_t stackSize = 2048;
-		k_thread_stack_t *stack = static_cast<k_thread_stack_t *>(k_malloc(stackSize));
-		if (stack == nullptr) {
-			printf("Failed to allocate stack to 'inferenceSenderTask%i'\n", n);
-			exit(1);
-		}
-
 		auto &thread = threads[nthreads];
 		string *name = new string("sender " + to_string(n));
 
-		thread.id = k_thread_create(&thread.thread, stack, stackSize, inferenceSenderTask,
+		thread.id = k_thread_create(&thread.thread, inferenceSenderStacks[n],
+					    inferenceSenderStackSize, inferenceSenderTask,
 					    name, heapPtr, &inferenceQueue, 3, 0, K_FOREVER);
 		if (thread.id == 0) {
 			printf("Failed to create 'inferenceSenderTask%i'\n", n);
@@ -248,21 +245,18 @@ int main()
 
 	/* Create inferenceProcess tasks to process the queued jobs */
 	InferenceProcessParams taskParams[NUM_INFERENCE_TASKS];
+	const size_t inferenceProcessStackSize = 8192;
+	static K_THREAD_STACK_ARRAY_DEFINE(inferenceProcessStacks, NUM_INFERENCE_TASKS,
+					   inferenceProcessStackSize);
 	for (int n = 0; n < NUM_INFERENCE_TASKS; n++) {
-		const size_t stackSize = 8192;
-		k_thread_stack_t *stack = static_cast<k_thread_stack_t *>(k_malloc(stackSize));
-		if (stack == nullptr) {
-			printf("Failed to allocate stack to 'inferenceSenderTask%i'\n", n);
-			exit(1);
-		}
-
 		auto &thread = threads[nthreads];
 		auto &taskParam = taskParams[n];
 		taskParam = InferenceProcessParams(&inferenceQueue, inferenceProcessTensorArena[n],
 						   arenaSize);
 		string *name = new string("runner " + to_string(n));
 
-		thread.id = k_thread_create(&thread.thread, stack, stackSize, inferenceProcessTask,
+		thread.id = k_thread_create(&thread.thread, inferenceProcessStacks[n],
+					    inferenceProcessStackSize, inferenceProcessTask,
 					    name, heapPtr, &taskParam, 2, 0, K_FOREVER);
 		if (thread.id == 0) {
 			printf("Failed to create 'inferenceProcessTask%i'\n", n);

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,