GDB appears to hang at synchronization points

Synchronization points (e.g., a semaphore) may be implicitly inserted by the compiler at kernel boundaries. For example, consider the following code:

1        #include <omp.h>
2
3        int main(void)
4        {
5
6          const int N = 8;
7          float sum = 0.f;
8          const float alpha = 2.f;
9
10         #pragma omp parallel for simd reduction(+: sum)
11         for(size_t j = 0; j < N; j++)
12           sum += alpha;
13
14         return 0;
15       }

You can run this code on your supported GPU device:

OMP_TARGET_OFFLOAD=MANDATORY LIBOMPTARGET_DEVICETYPE=GPU LIBOMPTARGET_PLUGIN=LEVEL0 gdb-oneapi -q omp_test

A barrier is implicitly inserted by the compiler between lines 10 and 11. You can see this by inspecting the disassembly:

(gdb) info line 9
Line 9 of "main.cpp" starts at address 0xfffdf000 <main.extracted(void)> and ends at 0xfffdf460 <main.extracted(void)+1120>.
(gdb) info line 11
Line 11 of "main.cpp" starts at address 0xfffe4a70 <main.extracted(void)+23152> and ends at 0xfffe4bb0 <main.extracted(void)+23472>.
(gdb) disassemble /m 0xfffdf000,0xfffe4bb0
0x00000000fffe0980 <main.extracted(void)+6528>:      (W)     send.gtwy (1|M0)    null  r4   null 0x0  0x02000004  {@1,$11} // wr:1+0, rd:0; signal barrier

Inserting a breakpoint at line 11 and stepping into the for-loop via set scheduler-locking step and step may cause the debugger to appear ‘hanging’ as only the current thread is resumed and other threads, which are expected to signal the barrier, are not resumed. To avoid this, make sure to set the scheduler-locking mode to replay or off at synchronization points.

Another way to work around this issue and be able to step a single thread over a synchronization point is using the non-stop mode and inserting a temporary breakpoint at the synchronization point. Temporary breakpoints are only ever hit by one thread (and are deleted afterwards) which, together with ‘non-stop mode’, ensures that only the first thread that hits the breakpoint is stopped while other threads may continue to the synchronization point. This enables you to step the thread that hit the temporary breakpoint over the synchronization point.

Kernel Stops Responding

If the kernel that is offloaded to a GPU stops responding:

1. Check whether there are any stray `gdbserver-ze` processes running in the background:

   ps -u $USER | grep gdbserver-ze

2. Stop background `gdbserver-ze` processes, if there are any:

   killall -9 gdbserver-ze

If the breakpoints defined inside the kernel are not hit when running on a GPU, and Virtualization technology for directed I/O (VT-d) is enabled, disable VT-d through the BIOS menu.

Conditional Breakpoints

All-stop mode

In all-stop mode, GDB may appear to be hanging if a conditional breakpoint is defined inside a kernel. This is because thousands of GPU threads may hit the breakpoint but may have to be resumed one by one if they do not satisfy the breakpoint condition. If feasible, modify the kernel code to have a statement that is guarded by the breakpoint condition, so that only the desired threads would hit the breakpoint in the code, e.g., move the breakpoint condition into an if-clause and rebuild the application. Alternatively, consider using the non-stop mode if applicable for your debug session.

Non-stop mode

GDB may report a timeout if a conditional breakpoint is defined inside a kernel. In non-stop mode, GDB resumes all threads that do not satisfy the breakpoint condition and resuming a large number of threads may delay the response from gdbserver. You may want to increase the default timeout if you encounter this problem:

(gdb) set remotetimeout 10

Seeing a SIGSEGV in a Unified Shared Memory (USM) Program

Applications that use unified shared memory (USM) may appear as raising a SIGSEGV when a USM-allocated memory is being accessed. This is a mechanism used by the runtime to trigger memory migration. In such cases, sending the signal back to the application resumes the program. For this, use the GDB continue or signal SIGSEGV command.