Fix the issue where ipc_shmem_->read_index goes out of bounds when waiting for an IPC event.

[limou102]

In the hipEvent_t stream-wait operation using shared-memory (IPC) events, the calculation of

offset = ipc_shmem_->read_index

does not apply a modulo with IPC_SIGNALS_PER_EVENT (which is fixed at 32), causing subsequent access to

ipc_shmem_->signal[offset]

to go out of bounds.

rocm-systems/projects/clr/hipamd/src/hip_event_ipc.cpp

Line 107 in f733871

&(ipc_evt_.ipc_shmem_->signal[offset]), 0, 1, 1, sizeof(uint32_t));

Looking at the initialization code, ipc_shmem_->signal[] is only initialized with IPC_SIGNALS_PER_EVENT = 32 elements:

rocm-systems/projects/clr/hipamd/src/hip_event_ipc.cpp

Line 67 in f733871

ihipHostRegister(&ipc_evt_.ipc_shmem_->signal, sizeof(uint32_t) * IPC_SIGNALS_PER_EVENT, 0);

Therefore, when an IPC event’s stream-wait is invoked more than 32 times, it triggers an out-of-bounds access and results in an abnormal termination — which indeed matches what we encountered in practice.