select device via jobs node parameter so we can skip the numa affinity change, also properly delete jobs in destructor

offloading-cacher/cache.hpp

@@ -400,12 +400,6 @@ inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, con
     const size_t size = task->GetSize() / task_count;
     const size_t last_size = size + task->GetSize() % task_count;
 
-    // save the current numa node mask to restore later
-    // as executing the copy task will place this thread
-    // on a different node
-
-    bitmask* nodemask = numa_get_run_node_mask();
-
     auto handlers = new std::vector<dml_job_t*>();
 
     for (uint32_t i = 0; i < task_count; i++) {
@@ -418,18 +412,11 @@ inline void dsacache::Cache::SubmitTask(CacheData* task, const int dst_node, con
     }
 
     task->SetTaskHandlersAndCache(dst, handlers);
-
-    // restore the previous nodemask
-
-    numa_run_on_node_mask(nodemask);
-    numa_free_nodemask(nodemask);
 }
 
 inline dml_job_t* dsacache::Cache::ExecuteCopy(
         const uint8_t* src, uint8_t* dst, const size_t size, const int node
 ) const {
-    numa_run_on_node(node);
-
     uint32_t job_size = 0;
     dml_status_t status = dml_get_job_size(DML_PATH_HW, &job_size);
 
@@ -452,6 +439,7 @@ inline dml_job_t* dsacache::Cache::ExecuteCopy(
     job->destination_first_ptr = dst;
     job->source_length = size;
     job->flags |= DML_FLAG_BLOCK_ON_FAULT | DML_FLAG_COPY_ONLY;
+    job->numa_id = node;
 
     status = dml_submit_job(job);
 
@@ -642,9 +630,13 @@ inline dsacache::CacheData::~CacheData() {
 
         Deallocate();
 
+	std::vector<dml_job_t*>* handlers = handlers_->load();
+
+	if (handlers != nullptr && handlers != reinterpret_cast<std::vector<dml_job_t*>*>(maxptr)) {
 	    for (dml_job_t* job : *handlers_->load()) {
 	        if (job != nullptr) delete job;
 	    }
+        }
 
         delete active_;
         delete cache_;