[ETHOSN] Fix for the mock inference after NPU driver update (apache#1…

…3650)

NPU driver 22.11 supports buffer and network creation
with only process memory allocator. This change however
should be limited to the cases when the HW is available.

NPU driver update: apache#13637

src/runtime/contrib/ethosn/ethosn_device.cc

@@ -272,7 +272,7 @@ TVM_REGISTER_GLOBAL("relay.ethos-n.test.infra.inference_result")
 // Allow the ethos-n support code to be tested without a device
 bool Inference(tvm::runtime::TVMArgs args,
 #ifdef _ETHOSN_API_VERSION_3_2_0
-               dl::ProcMemAllocator* proc_mem_alloc,
+               dl::ProcMemAllocator* /*proc_mem_alloc*/,
 #endif
                dl::Network* /* npu */, const std::vector<uint32_t>& input_order,
                const std::vector<uint32_t>& output_order, const std::vector<uint32_t>& input_sizes,

src/runtime/contrib/ethosn/ethosn_runtime.cc

@@ -114,9 +114,6 @@ Module EthosnModule::LoadFromBinary(void* strm) {
   cmms.resize(func_count);
   for (unsigned int i = 0; i < func_count; i++) {
     OrderedCompiledNetwork& compiled = cmms[i];
-#ifdef _ETHOSN_API_VERSION_3_2_0
-    compiled.proc_mem_alloc = std::make_unique<dl::ProcMemAllocator>();
-#endif
     std::string ext_symbol;
     std::string cmm;
     uint64_t input_size;
@@ -130,6 +127,7 @@ Module EthosnModule::LoadFromBinary(void* strm) {
     // If hardware unavaiable use the mock inference functionality. If hardware is
     // avaiable, deserialize the compiled graph.
 #ifdef _ETHOSN_API_VERSION_3_2_0
+    compiled.proc_mem_alloc = std::make_unique<dl::ProcMemAllocator>();
     compiled.runtime_cmm = std::make_unique<dl::Network>(
         compiled.proc_mem_alloc->CreateNetwork(cmm.c_str(), cmm.size()));
 #else