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

Implements SWDEV-213230.

[ROCm/clr commit: 888a7f2a90]

projects/clr/hipamd/CMakeLists.txt

@@ -336,9 +336,9 @@ if(HIP_PLATFORM STREQUAL "hcc")
         target_link_libraries(hip_hcc_static PRIVATE hc_am)
 
         add_library(hiprtc SHARED src/hiprtc.cpp src/code_object_bundle.cpp)
+        target_compile_options(hiprtc PRIVATE -DDISABLE_REDUCED_GPU_BLOB_COPY)
         set_property ( TARGET hiprtc PROPERTY VERSION "${HIP_LIB_VERSION_STRING}" )
         set_property ( TARGET hiprtc PROPERTY SOVERSION "${HIP_LIB_VERSION_MAJOR}" )
-
         target_include_directories(
             hiprtc SYSTEM
                 PRIVATE ${PROJECT_SOURCE_DIR}/include ${HSA_PATH}/include)

projects/clr/hipamd/include/hip/hcc_detail/code_object_bundle.hpp

@@ -31,9 +31,11 @@ THE SOFTWARE.
 #include <string>
 #include <utility>
 #include <vector>
-
+#include <unordered_set>
 namespace hip_impl {
-
+#if !defined(DISABLE_REDUCED_GPU_BLOB_COPY)
+std::unordered_set<std::string>& get_all_gpuarch();
+#endif
 inline
 std::string transmogrify_triple(const std::string& triple)
 {
@@ -43,7 +45,6 @@ std::string transmogrify_triple(const std::string& triple)
     if (triple.find(old_prefix) == 0) {
         return new_prefix + triple.substr(sizeof(old_prefix) - 1);
     }
-
     return (triple.find(new_prefix) == 0) ? triple : "";
 }
 
@@ -114,9 +115,7 @@ class Bundled_code_header {
     friend inline bool read(RandomAccessIterator f, RandomAccessIterator l,
                             Bundled_code_header& x) {
         if (f == l) return false;
-
         std::copy_n(f, sizeof(x.header_.cbuf_), x.header_.cbuf_);
-
         if (valid(x)) {
             x.bundles_.resize(x.header_.bundle_cnt_);
 
@@ -126,11 +125,16 @@ class Bundled_code_header {
                 it += sizeof(y.header.cbuf);
 
                 y.triple.assign(it, it + y.header.triple_sz);
-
+                #ifdef DISABLE_REDUCED_GPU_BLOB_COPY
                 std::copy_n(f + y.header.offset, y.header.bundle_sz, std::back_inserter(y.blob));
-
+                #else
+                auto& gpuArch = get_all_gpuarch();
+                auto itgpuArch = std::find(gpuArch.begin(),gpuArch.end(),y.triple);
+                if (itgpuArch != gpuArch.end()){
+                    std::copy_n(f + y.header.offset, y.header.bundle_sz, std::back_inserter(y.blob));
+                }
+                #endif
                 it += y.header.triple_sz;
-
                 x.bundled_code_size = std::max(x.bundled_code_size, 
                                                y.header.offset + y.header.bundle_sz);
             }

projects/clr/hipamd/lpl_ca/CMakeLists.txt

@@ -26,7 +26,7 @@ target_include_directories(ca PUBLIC ${PROJECT_SOURCE_DIR}/src)
 find_library(
     hsart NAMES libhsa-runtime64.so libhsa-runtime64.so.1 HINTS ${HSA_PATH}/lib)
 target_link_libraries(ca PUBLIC ${hsart})
-target_compile_options(ca PUBLIC -Wall)
+target_compile_options(ca PUBLIC -DDISABLE_REDUCED_GPU_BLOB_COPY -Wall)
 
 install(TARGETS ca RUNTIME DESTINATION bin)
-#-------------------------------------CA---------------------------------------#
+#-------------------------------------CA---------------------------------------#

projects/clr/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;

projects/clr/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;