For some reason, starting tests before gms makes DM run much faster.

dm/DM.cpp

 // Main binary for DM.
 // For a high-level overview, please see dm/README.
 
 #include "CrashHandler.h"
 #include "SkCommonFlags.h"
 #include "SkForceLinking.h"
 #include "SkGraphics.h"
 #include "SkPicture.h"
 #include "SkString.h"
 #include "Test.h"
@@ skipping 216 matching lines @@
         append_matching_factories<Test>(TestRegistry::Head(), &tests);
     }
 
     SkTArray<SkString> skps;
     find_skps(&skps);
 
     SkDebugf("%d GMs x %d configs, %d tests, %d pictures\n",
              gms.count(), configs.count(), tests.count(), skps.count());
     DM::Reporter reporter;
     DM::TaskRunner tasks(FLAGS_threads, FLAGS_gpuThreads);
+    kick_off_tests(tests, &reporter, &tasks);
     kick_off_gms(gms, configs, gpuAPI, *expectations, &reporter, &tasks);
-    kick_off_tests(tests, &reporter, &tasks);
     kick_off_skps(skps, &reporter, &tasks);
     tasks.wait();
 
     SkDebugf("\n");
 #ifdef SK_DEBUG
     if (FLAGS_portableFonts && FLAGS_reportUsedChars) {
         sk_tool_utils::report_used_chars();
     }
 #endif
 
     SkTArray<SkString> failures;
     reporter.getFailures(&failures);
     report_failures(failures);
     return failures.count() > 0;
 }
 
 #if !defined(SK_BUILD_FOR_IOS) && !defined(SK_BUILD_FOR_NACL)
 int main(int argc, char** argv) {
     SkCommandLineFlags::Parse(argc, argv);
     return dm_main();
 }
 #endif