ZES_ENABLE_SYSMAN vs multiple layers using L0

Hello

I am porting hwloc to L0 for exposing the locality of GPU devices. Things worked fine with API 0.91 but porting to 1.0 doesn't look good.

I am listing devices with zeDriverGet() and zeDeviceGet() and then calling zesDevicePciGetProperties() and zesDeviceGetProperties() after casting into zes_device_handle_t. As documented in 1.0 doc, this only works if ZES_ENABLE_SYSMAN=1 in the environment. This looks like a terrible idea.

1. Requiring a third-party library to always modify the environment of a program isn't nice. It's ok when enabling some debugging in a library from time to time, but not when enabling a mandatory feature. I . Passing a flag to zeInit() would be better.

2. This won't work if multiple layers use L0 at the same time: an application using L0 (most likely a runtime system managing tasks) will call zeInit() without ZES_ENABLE_SYSMAN (because it doesn't need it). It will then call hwloc to get the locality of that some ze devices (to select the local one, etc). hwloc will fail because sysman wasn't enabled in L0 (it will actually segfault in the current implementation). hwloc could set the env var and call its own zeInit() too but it's likely too late anyway. Given that all drivers/devices are shared between all these layers (which is good), the behavior of L0 will basically change depending on which layer calls zeInit() first, that's unpredictable and bad.

0.95 had zetSysmanGet() for converting between ze and zet devices. 1.0 should just do the same: have an explicit function to enable sysman and "cast" a ze device handle to zes. The env var of init flags won't work just like multiple MPI_Init() don't work in MPI.

Start with zeInit(), zesDevicePciGetProperties(), zesDeviceGetProperties(), and the 0.95 zetSysmanGet() entry point; review the 1.0 specification and loader behavior around ZES_ENABLE_SYSMAN. Define an explicit initialization or conversion path that remains safe when another layer calls zeInit() first, and verify the current failure or segfault scenario is covered.