The opencl-rocm runtime hasn’t been tested on the public gem5 model, so there are likely going to be issues like this.
Running clinfo on the host system may be working fine, but the environment for the simulator may not reflect what is on the host system. In order to support the off-the-shelf ROCm runtime we utilize a file-system redirection scheme so that we can provide the correct FS interface to the runtime in a deterministic and host-agnostic way. You can see the basic FS config in configs/common/FileSystemConfig.py. There we provide definitions for certain files that the runtime uses, for example some files in /proc/. You may need to add more support for some special files for opencl. A SyscallAll trace should give you an idea if there are any open calls to files that are not being supported, as well as which libraries etc. are being opened during runtime initialization. If that doesn’t shed any light you can run a remote gdb connection into gem5 to try to figure out what the runtime is trying to do at the time it fails. Once we have merged our staging branch into master, we hope to streamline the process of getting ROCm working on gem5. Tony From: gem5-users <[email protected]> On Behalf Of rahul shrivastava Sent: Thursday, May 23, 2019 12:26 AM To: [email protected] Subject: [gem5-users] AMD ROCM GCN3 simulation using opencl-runtime has assertion errors [CAUTION: External Email] Hi All, I am trying to perform GCN3 simulation by following the steps mentioned in http://gem5.org/GPU_Models -- Step 1) Compiled and installed version 1.6.x of ROCR runtime and thunk interface, Step 2) Instead of HCC, I have compiled and installed version 1.6.x of opecl-rocm-runtime. Step 3) Using the gem5 version mentioned in http://gem5.org/GPU_Models, I am trying to simulate clinfo. ISSUE When the icd_dispatcher does a dlopen on amdocl (which is basically a opecl-rocm-runtime), the assertion at line no 505 of the file amd_blit_kernel.cpp (part of rocr-runtime) fails and the program ends. Upon insepection I found that the contents of the static variable kBlitKernelSource is getting corrupted. Running clinfo on the system itself does not have any issue. It looks like when the call to dlopen is made, while setting up the storage for all the global and static variables of opecl-rocm-runtime and rocr-runtime, some corruption is happening. Could you please shed some light on this. Regards Rahul
_______________________________________________ gem5-users mailing list [email protected] http://m5sim.org/cgi-bin/mailman/listinfo/gem5-users
