Add HIP_DUMP_CODE_OBJECT

[ROCm/hip commit: f5d8842f6a]

projects/hip/src/hip_clang.cpp

@@ -22,6 +22,7 @@ THE SOFTWARE.
 
 #include <unordered_map>
 #include <string>
+#include <fstream>
 
 #include "hip/hip_runtime.h"
 #include "hip_hcc_internal.h"
@@ -112,6 +113,16 @@ __hipRegisterFatBinary(const void* data)
       if (module->executable.handle) {
         modules->at(deviceId) = module;
         tprintf(DB_FB, "Loaded code object for %s\n", name);
+        if (HIP_DUMP_CODE_OBJECT) {
+          char fname[30];
+          static std::atomic<int> index;
+          sprintf(fname, "__hip_dump_code_object%04d.o", index++);
+          tprintf(DB_FB, "Dump code object %s\n", fname);
+          std::ofstream ofs;
+          ofs.open(fname, std::ios::binary);
+          ofs << image;
+          ofs.close();
+        }
       } else {
         fprintf(stderr, "Failed to load code object for %s\n", name);
         abort();

projects/hip/src/hip_hcc.cpp

@@ -97,6 +97,8 @@ int HIP_INIT_ALLOC = -1;
 int HIP_SYNC_STREAM_WAIT = 0;
 int HIP_FORCE_NULL_STREAM = 0;
 
+int HIP_DUMP_CODE_OBJECT = 0;
+
 
 #if (__hcc_workweek__ >= 17300)
 // Make sure we have required bug fix in HCC
@@ -1294,6 +1296,10 @@ void HipReadEnv() {
                "overridden by specifying hipEventReleaseToSystem or hipEventReleaseToDevice flag "
                "when creating the event.");
 
+    READ_ENV_I(release, HIP_DUMP_CODE_OBJECT, 0,
+               "If set, dump code object as __hip_dump_code_object[nnnn].o in the current directory,"
+               "where nnnn is the index number.");
+
     // Some flags have both compile-time and runtime flags - generate a warning if user enables the
     // runtime flag but the compile-time flag is disabled.
     if (HIP_DB && !COMPILE_HIP_DB) {

projects/hip/src/hip_hcc_internal.h

@@ -83,11 +83,11 @@ extern int HIP_SYNC_NULL_STREAM;
 extern int HIP_INIT_ALLOC;
 extern int HIP_FORCE_NULL_STREAM;
 
+extern int HIP_DUMP_CODE_OBJECT;
 
 // TODO - remove when this is standard behavior.
 extern int HCC_OPT_FLUSH;
 
-
 // Class to assign a short TID to each new thread, for HIP debugging purposes.
 class TidInfo {
    public: