route EthosU input/output memcpy through overridable hook (#19264)

backends/arm/runtime/EthosUBackend.cpp

@@ -26,6 +26,12 @@
 #include <executorch/runtime/core/exec_aten/util/dim_order_util.h>
 #include <executorch/runtime/core/exec_aten/util/scalar_type_util.h>
 
+// Overridable memcpy used by the EthosU backend for input/output scratch
+// shuffling. Default (weak) implementation in EthosUBackend_IoMemcpy.cpp does
+// std::memcpy. Firmware targets can supply a strong override (e.g. routing
+// through a DMA engine) to reduce CPU memcpy load on the host MCU.
+extern "C" void io_memcpy(void* dst, const void* src, size_t size);
+
 using namespace std;
 
 using executorch::aten::ScalarType;
@@ -237,8 +243,9 @@ class EthosUBackend final : public ::executorch::runtime::BackendInterface {
       if (both_char || both_int || both_short || both_bool) {
         EXECUTORCH_PROF_SCOPE(
             event_tracer, "+EthosUBackend::execute()handles.input.memcpy()");
-        // Sizes match and elt size matches so memcpy
-        memcpy(
+        // Sizes match and elt size matches so memcpy.
+        // Routed through io_memcpy so firmware can DMA-accelerate.
+        io_memcpy(
             scratch_addr,
             tensor_in.mutable_data_ptr<char>(),
             tensor_in.nbytes());
@@ -389,7 +396,8 @@ Error copy_with_layout_adjustment(
   }
   const char* src_bytes = src;
   for (size_t chunk_idx = 0; chunk_idx < chunk_count; ++chunk_idx) {
-    memcpy(dest, src_bytes, chunk_size);
+    // Routed through io_memcpy so firmware can DMA-accelerate.
+    io_memcpy(dest, src_bytes, chunk_size);
     src_bytes += vela_chunk_size;
     dest += chunk_size;
   }

backends/arm/runtime/EthosUBackend_Cortex_M.cpp

@@ -42,6 +42,12 @@ extern "C" __attribute__((weak)) struct ethosu_driver* ethosu_reserve_driver_ex(
   return ethosu_reserve_driver();
 }
 
+// Overridable memcpy used by the EthosU backend for output scratch
+// shuffling. Default (weak) implementation in EthosUBackend_IoMemcpy.cpp does
+// std::memcpy. Firmware targets can supply a strong override (e.g. routing
+// through a DMA engine) to reduce CPU memcpy load on the host MCU.
+extern "C" void io_memcpy(void* dst, const void* src, size_t size);
+
 namespace executorch {
 namespace backends {
 namespace arm {
@@ -136,7 +142,8 @@ Error platform_execute(
       }
       io_bytes_total += tensor_bytes;
     } else {
-      memcpy(
+      // Routed through io_memcpy so firmware can DMA-accelerate.
+      io_memcpy(
           tensor_out.mutable_data_ptr<char>(),
           static_cast<const char*>(output_addr),
           tensor_bytes);

backends/arm/runtime/EthosUBackend_IoMemcpy.cpp

@@ -0,0 +1,19 @@
+/*
+ * Copyright 2026 Arm Limited and/or its affiliates.
+ *
+ * This source code is licensed under the BSD-style license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <cstddef>
+#include <cstring>
+
+// Weak default for io_memcpy. Firmware targets can provide a
+// strong-symbol override (e.g. routing through DMA on Cortex-M55) without
+// touching the upstream EthosUBackend code. Lives in its own translation
+// unit so the compiler in the call-site TUs cannot inline this body and
+// bypass the link-time override (same trick as bolt_arm_memcpy_external).
+extern "C" __attribute__((weak)) void
+io_memcpy(void* dst, const void* src, size_t size) {
+  std::memcpy(dst, src, size);
+}

backends/arm/runtime/targets.bzl

@@ -15,6 +15,7 @@ def define_common_targets():
         srcs = [
             "EthosUBackend.cpp",
             "EthosUBackend_Cortex_M.cpp",
+            "EthosUBackend_IoMemcpy.cpp",
         ],
         headers = ["EthosUBackend_Internal.h"],
         compatible_with = ["ovr_config//cpu:arm32-embedded", "ovr_config//cpu:arm32-embedded-fpu"],