diff --git a/general/package/legacy/ambarella-osdrv-s3l/ambarella-osdrv-s3l.mk b/general/package/legacy/ambarella-osdrv-s3l/ambarella-osdrv-s3l.mk
index 79e63f6ca..0c896dc3a 100644
--- a/general/package/legacy/ambarella-osdrv-s3l/ambarella-osdrv-s3l.mk
+++ b/general/package/legacy/ambarella-osdrv-s3l/ambarella-osdrv-s3l.mk
@@ -4,11 +4,14 @@
 #
 ################################################################################
 
-AMBARELLA_OSDRV_S3L_VERSION =
-AMBARELLA_OSDRV_S3L_SITE =
+AMBARELLA_OSDRV_S3L_SITE_METHOD = local
+AMBARELLA_OSDRV_S3L_SITE = $(AMBARELLA_OSDRV_S3L_PKGDIR)/src
 AMBARELLA_OSDRV_S3L_LICENSE = MIT
-AMBARELLA_OSDRV_S3L_LICENSE_FILES = LICENSE
 
+# load_ucode puts the DSP microcode in place; the microcode itself is not shipped.
+define AMBARELLA_OSDRV_S3L_BUILD_CMDS
+	$(MAKE) $(TARGET_CONFIGURE_OPTS) -C $(@D) load_ucode
+endef
 
 define AMBARELLA_OSDRV_S3L_INSTALL_TARGET_CMDS
 	$(INSTALL) -m 755 -d $(TARGET_DIR)/etc/sensors
@@ -19,6 +22,7 @@ define AMBARELLA_OSDRV_S3L_INSTALL_TARGET_CMDS
 
 	$(INSTALL) -m 755 -d $(TARGET_DIR)/usr/bin
 	$(INSTALL) -m 755 -t $(TARGET_DIR)/usr/bin $(AMBARELLA_OSDRV_S3L_PKGDIR)/files/script/load*
+	$(INSTALL) -m 755 -t $(TARGET_DIR)/usr/bin $(@D)/load_ucode
 
 	$(INSTALL) -m 755 -d $(TARGET_DIR)/usr/lib/sensors
 	# $(INSTALL) -m 644 -t $(TARGET_DIR)/usr/lib/sensors $(AMBARELLA_OSDRV_S3L_PKGDIR)/files/sensor/*.so
diff --git a/general/package/legacy/ambarella-osdrv-s3l/files/script/load_ambarella b/general/package/legacy/ambarella-osdrv-s3l/files/script/load_ambarella
index 0a9fed01f..c51e468b7 100755
--- a/general/package/legacy/ambarella-osdrv-s3l/files/script/load_ambarella
+++ b/general/package/legacy/ambarella-osdrv-s3l/files/script/load_ambarella
@@ -7,11 +7,10 @@
 # them and the nodes have to be made here. The major comes from /proc/devices;
 # the minors are fixed.
 #
-# The microcode (orccode.bin, orcme.bin, default_binary.bin) and the load_ucode
-# tool are Ambarella's and are not part of this package, and the stock tool is
-# a glibc armhf binary that will not run on this musl image. Without them the
-# drivers load and the DSP does not start, so no video. Drop working ones in
-# place and this script uses them.
+# The microcode (orccode.bin, orcme.bin, default_binary.bin) is Ambarella's and
+# is not part of this package; load_ucode, which puts it in place, is built from
+# the package's src/. Without the microcode the drivers load and the DSP does
+# not start, so no video. Drop it in /lib/firmware and this script loads it.
 
 vendor=$(ipcinfo -v)
 ucode_dir=/lib/firmware
diff --git a/general/package/legacy/ambarella-osdrv-s3l/src/Makefile b/general/package/legacy/ambarella-osdrv-s3l/src/Makefile
new file mode 100644
index 000000000..8f9155cc9
--- /dev/null
+++ b/general/package/legacy/ambarella-osdrv-s3l/src/Makefile
@@ -0,0 +1,2 @@
+load_ucode: load_ucode.c
+	$(CC) $(CFLAGS) -Wall -Wextra -o $@ $< $(LDFLAGS)
diff --git a/general/package/legacy/ambarella-osdrv-s3l/src/load_ucode.c b/general/package/legacy/ambarella-osdrv-s3l/src/load_ucode.c
new file mode 100644
index 000000000..f69befe8f
--- /dev/null
+++ b/general/package/legacy/ambarella-osdrv-s3l/src/load_ucode.c
@@ -0,0 +1,188 @@
+// SPDX-License-Identifier: MIT
+/*
+ * Load the Ambarella S3L DSP microcode through /dev/ucode.
+ *
+ *   load_ucode DIR
+ *
+ * A drop-in for the vendor's load_ucode: dsp.ko describes where each file
+ * goes (UCODE_GET_INFO), the files are copied into the microcode window it
+ * maps, and the version the DSP will run is printed. The DSP itself is not
+ * started here; it boots when IAV first enters the idle state.
+ *
+ * The window is mapped uncached by dsp.ko, so no flush is needed.
+ */
+
+#include <errno.h>
+#include <fcntl.h>
+#include <stdint.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <sys/ioctl.h>
+#include <sys/mman.h>
+#include <sys/stat.h>
+#include <unistd.h>
+
+#define UCODE_DEV "/dev/ucode"
+#define UCODE_MAX_ITEMS 4
+
+struct ucode_item {
+	uint32_t addr_offset;
+	char filename[24];
+};
+
+struct ucode_info {
+	uint32_t map_size;
+	uint32_t nr_item;
+	struct ucode_item item[UCODE_MAX_ITEMS];
+};
+
+struct ucode_version {
+	uint16_t year;
+	uint8_t month;
+	uint8_t day;
+	uint32_t edition_num;
+	uint32_t edition_ver;
+	uint32_t arch;
+};
+
+#define UCODE_GET_INFO _IOR('u', 0, struct ucode_info)
+#define UCODE_GET_VERSION _IOR('u', 1, struct ucode_version)
+#define UCODE_LOADED _IOW('u', 2, uint32_t)
+
+struct ucode_file {
+	uint8_t *data;
+	size_t size;
+};
+
+static int read_file(const char *dir, const char *name, size_t room,
+		     struct ucode_file *f)
+{
+	char path[256];
+	struct stat st;
+	size_t done = 0;
+	int fd, ret = -1;
+
+	snprintf(path, sizeof(path), "%s/%s", dir, name);
+	if ((fd = open(path, O_RDONLY)) < 0 || fstat(fd, &st) < 0) {
+		fprintf(stderr, "load_ucode: %s: %s\n", path, strerror(errno));
+		goto out;
+	}
+	if (st.st_size <= 0 || (size_t)st.st_size > room) {
+		fprintf(stderr, "load_ucode: %s is %lld bytes, needs 1..%zu\n",
+			path, (long long)st.st_size, room);
+		goto out;
+	}
+	f->size = st.st_size;
+	if (!(f->data = malloc(f->size))) {
+		fprintf(stderr, "load_ucode: %s: %s\n", path, strerror(errno));
+		goto out;
+	}
+
+	while (done < f->size) {
+		ssize_t n = read(fd, f->data + done, f->size - done);
+		if (n <= 0) {
+			fprintf(stderr, "load_ucode: reading %s: %s\n", path,
+				n ? strerror(errno) : "file shrank");
+			goto out;
+		}
+		done += n;
+	}
+	ret = 0;
+out:
+	if (fd >= 0)
+		close(fd);
+	return ret;
+}
+
+int main(int argc, char **argv)
+{
+	struct ucode_info info;
+	struct ucode_version ver;
+	struct ucode_file files[UCODE_MAX_ITEMS] = { 0 };
+	uint8_t *map;
+	int fd, ret = 1;
+
+	if (argc != 2) {
+		fprintf(stderr, "Usage: %s DIR\n", argv[0]);
+		return 1;
+	}
+
+	if ((fd = open(UCODE_DEV, O_RDWR)) < 0) {
+		fprintf(stderr, "load_ucode: %s: %s\n", UCODE_DEV,
+			strerror(errno));
+		return 1;
+	}
+
+	memset(&info, 0, sizeof(info));
+	if (ioctl(fd, UCODE_GET_INFO, &info) < 0) {
+		fprintf(stderr, "load_ucode: UCODE_GET_INFO: %s\n",
+			strerror(errno));
+		goto out;
+	}
+	if (!info.map_size || !info.nr_item ||
+	    info.nr_item > UCODE_MAX_ITEMS) {
+		fprintf(stderr, "load_ucode: bad layout: %u items in 0x%x\n",
+			info.nr_item, info.map_size);
+		goto out;
+	}
+
+	/* Every file is read in full before the window is touched: a set that
+	 * failed half way would leave old and new microcode mixed in it. */
+	for (uint32_t i = 0; i < info.nr_item; i++) {
+		struct ucode_item *it = &info.item[i];
+		/* Each file may run up to the next one's offset, the last one
+		 * to the end of the window. */
+		uint32_t end = info.map_size;
+
+		for (uint32_t j = 0; j < info.nr_item; j++)
+			if (info.item[j].addr_offset > it->addr_offset &&
+			    info.item[j].addr_offset < end)
+				end = info.item[j].addr_offset;
+
+		it->filename[sizeof(it->filename) - 1] = '\0';
+		if (it->addr_offset >= info.map_size) {
+			fprintf(stderr, "load_ucode: %s at 0x%x is outside 0x%x\n",
+				it->filename, it->addr_offset, info.map_size);
+			goto out;
+		}
+		if (read_file(argv[1], it->filename, end - it->addr_offset,
+			      &files[i]))
+			goto out;
+	}
+
+	map = mmap(NULL, info.map_size, PROT_READ | PROT_WRITE, MAP_SHARED, fd,
+		   0);
+	if (map == MAP_FAILED) {
+		fprintf(stderr, "load_ucode: mmap 0x%x: %s\n", info.map_size,
+			strerror(errno));
+		goto out;
+	}
+
+	for (uint32_t i = 0; i < info.nr_item; i++) {
+		struct ucode_item *it = &info.item[i];
+
+		memcpy(map + it->addr_offset, files[i].data, files[i].size);
+		printf("%s: %zu bytes at +0x%06x\n", it->filename,
+		       files[i].size, it->addr_offset);
+	}
+
+	if (ioctl(fd, UCODE_LOADED, 0) < 0) {
+		fprintf(stderr, "load_ucode: UCODE_LOADED: %s\n",
+			strerror(errno));
+		goto unmap;
+	}
+
+	if (ioctl(fd, UCODE_GET_VERSION, &ver) == 0)
+		printf("ucode version %u/%u/%u %u.%u\n", ver.year, ver.month,
+		       ver.day, ver.edition_num, ver.edition_ver);
+	ret = 0;
+
+unmap:
+	munmap(map, info.map_size);
+out:
+	for (int i = 0; i < UCODE_MAX_ITEMS; i++)
+		free(files[i].data);
+	close(fd);
+	return ret;
+}