Arm backend: Allocate buffers with alignment (pytorch#13412)

Ethos-u driver needs 16 bit alignment for command and scratch buffers.

Signed-off-by: Erik Lundell <[email protected]>

Author: Erik-Lundell

backends/arm/runtime/EthosUBackend.cpp

@@ -192,8 +192,9 @@ class EthosUBackend final : public ::executorch::runtime::BackendInterface {
     // Use a temporary allocator for the intermediate tensors of the
     // computation. The allocator is released in runtime/executor/method.cpp at
     // the end of the execution of the Ethos-U custom delegate
-    char* ethosu_scratch =
-        static_cast<char*>(temp_allocator->allocate(handles.scratch_data_size));
+    // Ethos-U driver requires 16 bit alignment.
+    char* ethosu_scratch = static_cast<char*>(
+        temp_allocator->allocate(handles.scratch_data_size, 16UL));
     if (ethosu_scratch == nullptr) {
       ET_LOG(
           Error,

examples/arm/executor_runner/arm_executor_runner.cpp

@@ -521,8 +521,9 @@ void runner_init(
     ET_LOG(Info, "Setting up planned buffer %zu, size %zu.", id, buffer_size);
 
     /* Move to it's own allocator when MemoryPlanner is in place. */
-    uint8_t* buffer =
-        reinterpret_cast<uint8_t*>(ctx.method_allocator->allocate(buffer_size));
+    /* Ethos-U driver requires 16 bit alignment. */
+    uint8_t* buffer = reinterpret_cast<uint8_t*>(
+        ctx.method_allocator->allocate(buffer_size, 16UL));
     ET_CHECK_MSG(
         buffer != nullptr,
         "Could not allocate memory for memory planned buffer size %zu",