I am trying to track down a bug that happens inside a pytorch android binary (speed_benchmark_torch) for one of my models.
When I look at the backtrace on android logcat
, it looks like debug symbols are disabled, so I don’t know what line of code is failing. Also, I tried building with -DANDROID_DEBUG_SYMBOLS=1
, and I still couldn’t get a stacktrace with debug symbols.
I suspect that I will get a better stacktrace if I can compile with -O0
instead of the current -O2
. Towards this end, I changed one of the lines in pytorch/CMakeLists.txt
from this:
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O2 -fPIC")
to this:
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O0 -fPIC")
I also changed the O2
to O0
in the QNNPACK CMakeLists.
And, after making that change, I recompiled pytorch. When I look at pytorch/build/compile_commands.json
, I see we use -O0 in the pytorch build. Great.
Now, I run scripts/build_android.sh
, and to my dismay I see that each command in build_android/compile_commands.json
has both -O0 and -O2. That’s not good!
Does anyone know…
(a) Where might this stray -O2 be coming from, and
(b) How do I do an “only O0” build of the android binaries, particularly speed_benchmark_torch?
One idea I had is to just manually compile speed_benchmark_torch, since I have its compile command in build_android/compile_commands.json
. However, I don’t know the link command, and I imagine there’s a long list of flags on the link command.