Reduce GPU copying based on arch it runs on (#1751)

Implements SWDEV-213230.

hipamd/src/hip_hcc.cpp

@@ -39,6 +39,7 @@ THE SOFTWARE.
 #include <algorithm>
 #include <atomic>
 #include <mutex>
+#include <unordered_set>
 
 #include <hc.hpp>
 #include <hc_am.hpp>
@@ -2542,6 +2543,16 @@ hipError_t hipHccGetAcceleratorView(hipStream_t stream, hc::accelerator_view** a
 // TODO - add a contect sequence number for debug. Print operator<< ctx:0.1 (device.ctx)
 
 namespace hip_impl {
+    std::unordered_set<std::string>& get_all_gpuarch() {
+        static std::unordered_set<std::string> r{};
+        static std::once_flag init;
+        std::call_once(init, []() {
+            for (int i=0; i < g_deviceCnt; i++){
+                r.insert("hcc-amdgcn-amd-amdhsa--gfx"+std::to_string(g_deviceArray[i]->_props.gcnArch));
+        }});
+        return r;
+    }
+
     std::vector<hsa_agent_t> all_hsa_agents() {
         std::vector<hsa_agent_t> r{};
         std::vector<hc::accelerator> visible_accelerators;

hipamd/src/program_state.inl

@@ -244,7 +244,8 @@ public:
                     if (!valid(tmp)) break;
 
                     for (auto&& bundle : bundles(tmp)) {
-                        impl.code_object_blobs.second[elf][triple_to_hsa_isa(bundle.triple)].push_back(bundle.blob);
+                        if(bundle.blob.size())
+                            impl.code_object_blobs.second[elf][triple_to_hsa_isa(bundle.triple)].push_back(bundle.blob);
                     }
 
                     blob_it += tmp.bundled_code_size;