[SYCL][HIP] Memory access fault by GPU on address (nil)

[zjin-lcf]

https://github.com/zjin-lcf/oneAPI-DirectProgramming/tree/master/ced-sycl

./ced -a 0

Running the program shows the following error on an AMD GPU. Could you reproduce the error ? Thanks.

Memory access fault by GPU node-2 (Agent handle: 0x51b550) on address (nil). Reason: Page not present or supervisor privilege.
bt

gdb message:
Thread 2 "ced" received signal SIGABRT, Aborted.
[Switching to Thread 0x7ffff2ec5700 (LWP 2852315)]
0x00007ffff766c18b in raise () from /lib/x86_64-linux-gnu/libc.so.6
(gdb) bt
#0 0x00007ffff766c18b in raise () from /lib/x86_64-linux-gnu/libc.so.6
#1 0x00007ffff764b859 in abort () from /lib/x86_64-linux-gnu/libc.so.6
#2 0x00007ffff3344a7f in rocr::core::Runtime::VMFaultHandler(long, void*) () from /opt/rocm/hip/lib/../../lib/libhsa-runtime64.so.1
#3 0x00007ffff334753b in rocr::core::Runtime::AsyncEventsLoop(void*) () from /opt/rocm/hip/lib/../../lib/libhsa-runtime64.so.1
#4 0x00007ffff32ef497 in rocr::os::ThreadTrampoline(void*) () from /opt/rocm/hip/lib/../../lib/libhsa-runtime64.so.1
#5 0x00007ffff7fa1609 in start_thread () from /lib/x86_64-linux-gnu/libpthread.so.0
#6 0x00007ffff7748293 in clone () from /lib/x86_64-linux-gnu/libc.so.6

[AidanBeltonS]

Hello,

I encountered a similar issue not long ago. A PR to fix this was merged a few days ago after this issue was posted.
Could you check if this is still an issue for a tip build of llvm, so I know if it is the same issue or not?

Many Thanks,
Aidan

[bader]

@AidanBeltonS, which PR fixed this?

[AidanBeltonS]

PR #4604 resolved a similar error.
The memory access errors are quite generic though, so it could be a separate issue.

[zjin-lcf]

HI Aidan,

Thank you for your work. I typed 'git pull' in the llvm directory, and then built the compiler with HIP support from scratch. The message is:

Memory access fault by GPU node-2 (Agent handle: 0x519c50) on address (nil). Reason: Page not present or supervisor privilege.

Thread 7 "ced" received signal SIGSEGV, Segmentation fault.
[Switching to AMDGPU Thread 2:4:1:1 (0,0,0)/0]
main::{lambda()#1}::operator()() const::{lambda(cl::sycl::handler&)#5}::operator()(cl::sycl::handler&) const::{lambda(cl::sycl::nd_item<2>)#1}::operator()(cl::sycl::nd_item<2>) const (
/home/release/git/aomp13/ROCgdb/gdb/dwarf2/frame.c:1029: internal-error: Unknown CFA rule.
A problem internal to GDB has been detected,
further debugging may prove unreliable.

[AidanBeltonS]

Okay thank you for testing that out.
Then this is a separate issue. That helps narrow this down.

[zjin-lcf]

You are welcome.

[zjin-lcf]

The gdb error was reported to AMD. ROCm/ROCgdb#9

[npmiller]

Hello @zjin-lcf are you still seeing this issue? I'm not able to reproduce the crash with the latest dpc++ and benchmark on MI100.

[zjin-lcf]

Could you please explain which changes you made fix the related issues ? Thanks.

[npmiller]

Oooh, nevermind, I can actually reproduce it, I didn't realize the binary had changed name and is main now, it seems that it was working with the old one I had though.

[npmiller]

Okay, so I can now confirm that the following patch fixes ced:

• [SYCL][HIP] Disable partial global offset support #4905

Basically this is an issue if you have multiple kernels and one has less arguments than the previous one.

At the moment the global offset value is loaded from an address right after the kernel arguments, this is initialized to 0 so for a single kernel it works, but when there is multiple kernel there might be something different in the memory so it leads to random global offset values. Which causes crashes like you were seeing in this sample.

But with the patch it should run fine.