No perfcounters in the generated csv about radeon_compute_profiler HOT 4 OPEN

Using old driver 17.50 (18.10 won't do), the perfcounters appear. I hope this info helps.

from radeon_compute_profiler.

If you get a chance, can you please try this with RCP v5.6 which was released last week, and let me know if this issues still reproduces. There were a few issues fixed with regards to using legacy OpenCL in amdgpu-pro

Thanks

from radeon_compute_profiler.

I have issues building it. I had the same issues with previous builds, but I don't remember how to fix them.
First I have to remove all -Werror (new gccs add new warnings all the time).
Second sprofile fails to build with the error
ParseCmdLine.cpp:(.text.startup+0x2): undefined reference to boost::system::generic_category()' /usr/bin/ld: ParseCmdLine.cpp:(.text.startup+0x7): undefined reference to boost::system::generic_category()'
/usr/bin/ld: ParseCmdLine.cpp:(.text.startup+0xc): undefined reference to `boost::system::system_category()'

from radeon_compute_profiler.

I have issues building it. I had the same issues with previous builds, but I don't remember how to fix them.
First I have to remove all -Werror (new gccs add new warnings all the time).
Second sprofile fails to build with the error:

ParseCmdLine.cpp:(.text.startup+0x2): undefined reference to boost::system::generic_category()'
/usr/bin/ld: ParseCmdLine.cpp:(.text.startup+0x7): undefined reference to boost::system::generic_category()'
/usr/bin/ld: ParseCmdLine.cpp:(.text.startup+0xc): undefined reference to `boost::system::system_category()'

I encountered the same problem. It seems that this software package is no longer maintained, there are many outdated identifiers, missing library includes, and other problems. These problems are relatively easy to fix, but I was completely puzzled by the final Boost linking undefined reference problem, it's the most tricky one that took me several hours to find the answer: it turned out to be a C++ ABI compatibility problem because rcprof was building using the bundled Boost headers from Boost 1.59, but at link time I was using the system Boost library binaries, creating ABI conflicts. This is clearly not guaranteed to work. To fix that, there are two ways.

1. The proper way. Change BOOST_DIR = $(COMMON_LIB_EXT)/Boost/boost_1_59_0 to the Boost source tree, the version should be the same as your system's version, for example BOOST_DIR = /home/user/boost_1_83_0/. If your system doesn't provide static .a files for Boost, just compile Boost from source, then set BOOST_LIB_DIR to be the build output of the same Boost version, such as BOOST_LIB_DIR = /home/user/boost_1_83_0/stage/lib. Otherwise, there will be undefined references to boost::filesystem::path_traits.

2. The coincidence. My experiment showed that if Boost 1.76 is built using --std=c++11, linking against Boost 1.76 would work even if the source is compiled with Boost 1.59. But this is entirely a coincidence. If Boost is built using a higher C++ standard, there will be undefined reference to boost::system::generic_category().

Unfortunately, upon further investigation, even the "fixed" build still has problems. Only HSA profiling works, but not OpenCL or occupancy profiling. It turned out that rocprof is just a wrapper for various underlying modules, such as HSAFdnTrace, HSAFdnPMC, CLOccupancyAgent, CLProfileAgent, CLTraceAgent - all rocprof does is registers these modules and writes a configuration file before the user application runs, it's up to the modules themselves to do the actual work. This is why rocprof only reports Failed to generate profile result at the very end - it only reads an output file and has no idea about how the modules work by themselves - in this case, they're never executed.

For HSA and OpenCL, different loading mechanisms are used. For OpenCL and occupancy, it uses a method similar to LD_PRELOAD by replacing the OpenCL functions with instrumented versions. However, for reason unknown, the entry point is no longer invoked (possibility due to incompatible SDK change)? So now CLOccupancyAgent, CLProfileAgent, CLTraceAgent can no longer be invoked anywhere, their entry point cl_int CL_CALLBACK clAgent_OnLoad(cl_agent* agent) never executes.

Without someone who has a deeper understand about how the library preloading actually works, I'm afraid that this problem cannot be fixed. If anyone's interested and wishes to continue the investigation, reply to this thread and I can send a list of patches to replicate my build.

from radeon_compute_profiler.