Bug fix in Ethos runtime backend (#9517)

Summary:

See the comment in source for explanation.

Also change the default MACs for U85 to 256, which is default on FVP SSE-320.

Differential Revision: D71658384

Author: limintang

backends/arm/runtime/EthosUBackend.cpp

@@ -83,6 +83,13 @@ void __attribute__((weak)) EthosUBackend_execute_begin() {}
 void __attribute__((weak)) EthosUBackend_execute_end() {}
 }
 
+// If the address is >= 0x80000000, cast to uint64_t signed extends it to 0xFFFFFFFFXXXXXXXX,
+// which causes address validity check in Ethos driver to fail. Given Cortex-M is 32-bit, it
+// is safe to clear the upper 32 bits after the cast and keep the address valid.
+static inline uint64_t cast_address_to_uint64(void* p) {
+  return (0xFFFFFFFF & reinterpret_cast<uint64_t>(p));
+}
+
 class EthosUBackendExecuteCallbacks {
  public:
   EthosUBackendExecuteCallbacks() {
@@ -282,7 +289,8 @@ class EthosUBackend final : public ::executorch::runtime::BackendInterface {
     // constant weight data, then scratch (which contains input and output)
     // scratch is written above in this function.
     uint64_t bases[2] = {
-        (uint64_t)handles.weight_data, (uint64_t)handles.scratch_data};
+        cast_address_to_uint64(handles.weight_data),
+        cast_address_to_uint64(handles.scratch_data)};
     size_t bases_size[2] = {
         handles.weight_data_size, handles.scratch_data_size};
     int result = 0;

backends/arm/test/common.py

@@ -110,7 +110,7 @@ def get_u55_compile_spec(
 
 
 def get_u85_compile_spec(
-    macs: int = 128,
+    macs: int = 256,
     system_config="Ethos_U85_SYS_DRAM_Mid",
     memory_mode="Shared_Sram",
     extra_flags="--output-format=raw",