levelzero: only use Sysman queries instead of similar Core API queries

[bgoglin]

This goes on top of #594 which might be merged once a compute-runtime release brings a working zesInit(). This PR requires zesInit() to be more widely available, especially the last commit which completely remove the old setting of ZES_ENABLE_SYSMAN=1 in the env.
Now that we require zesInit(), just use ZES queries instead of switching from/to the Core API depending on ZES being available or not.

[bgoglin]

This should be rebased/simplified on top of #695

[bgoglin]

@saik-intel Now that we use zesInit(), I guess there's no reason to query both zeDevicePciGetPropertiesExt() and zesDevicePciGetProperties() in case one fails but no the other. The later should always be supported now, right?
Same question for zeDeviceGetMemoryProperties() vs zesDeviceEnumMemoryModules()+zesMemoryGetProperties(), is there any reason to prefer one of the other? We just want to know how much memory each device or subdevice has, and if that memory is HBM/DRAM/etc ?

[bgoglin]

@TApplencourt Could you please run lstopo aurora.xml from this branch using the tarball from https://ci.inria.fr/hwloc/job/basic/view/change-requests/job/PR-595/ ? I'd like to double-check that we don't loose any info when removing non-Sysman queries.

[TApplencourt]

Of course! Here it's:
aurora.xml.txt

At first glance, Everything looks good to me.

$ applenco@x1921c3s2b0n0:~/hwloc-PR-595-20241113.1537.gitc2b2cfd7/build> ./ici/bin/lstopo  | grep "ze"
              CoProc(LevelZero) "ze0"
                CoProc(LevelZero) "ze0.0"
                CoProc(LevelZero) "ze0.1"
              CoProc(LevelZero) "ze1"
                CoProc(LevelZero) "ze1.0"
                CoProc(LevelZero) "ze1.1"
              CoProc(LevelZero) "ze2"
                CoProc(LevelZero) "ze2.0"
                CoProc(LevelZero) "ze2.1"
              CoProc(LevelZero) "ze3"
                CoProc(LevelZero) "ze3.0"
                CoProc(LevelZero) "ze3.1"
              CoProc(LevelZero) "ze4"
                CoProc(LevelZero) "ze4.0"
                CoProc(LevelZero) "ze4.1"
              CoProc(LevelZero) "ze5"
                CoProc(LevelZero) "ze5.0"
                CoProc(LevelZero) "ze5.1"
$applenco@x1921c3s2b0n0:~/hwloc-PR-595-20241113.1537.gitc2b2cfd7/build> ./ici/bin/lstopo  | grep "cl"
              CoProc(OpenCL) "opencl0d0"
              CoProc(OpenCL) "opencl0d1"
              CoProc(OpenCL) "opencl0d2"
              CoProc(OpenCL) "opencl0d3"
              CoProc(OpenCL) "opencl0d4"
              CoProc(OpenCL) "opencl0d5"

But Make check reported an error:

applenco@x1921c3s2b0n0:~/hwloc-PR-595-20241113.1537.gitc2b2cfd7/build> cat tests/hwloc/test-suite.log
========================================================================
   hwloc PR-595-20241113.1537.gitc2b2cfd7: tests/hwloc/test-suite.log
========================================================================

# TOTAL: 46
# PASS:  45
# SKIP:  0
# XFAIL: 0
# FAIL:  1
# XPASS: 0
# ERROR: 0

.. contents:: :depth: 2

FAIL: levelzero
===============

levelzero: ../../../tests/hwloc/levelzero.c:188: int main(void): Assertion `atoi(osdev->name+2) == (int) k' failed.
./wrapper.sh: line 32: 175099 Aborted                 "$@"
FAIL levelzero (exit status: 134)

Composite or FLat doesn't change anything:

applenco@x1921c3s2b0n0:~/hwloc-PR-595-20241113.1537.gitc2b2cfd7/build> ZE_FLAT_DEVICE_HIERARCHY=FLAT ./tests/hwloc/levelzero
testing ZE devices
found 1 L0 drivers
found 12 L0 devices in driver #0
found OSDev ze0
got cpuset 0x0fffffff,0xffffff00,,0x000fffff,0xffffffff for driver #0 device #0
found OSDev ze0
levelzero: ../../../tests/hwloc/levelzero.c:100: int main(void): Assertion `atoi(osdev->name+2) == (int) k' failed.
Aborted
applenco@x1921c3s2b0n0:~/hwloc-PR-595-20241113.1537.gitc2b2cfd7/build> ZE_FLAT_DEVICE_HIERARCHY=COMPOSITE ./tests/hwloc/levelzero
testing ZE devices
found 1 L0 drivers
found 6 L0 devices in driver #0
found OSDev ze0
got cpuset 0x0fffffff,0xffffff00,,0x000fffff,0xffffffff for driver #0 device #0
found OSDev ze1
got cpuset 0x0fffffff,0xffffff00,,0x000fffff,0xffffffff for driver #1 device #0
found OSDev ze2
got cpuset 0x0fffffff,0xffffff00,,0x000fffff,0xffffffff for driver #2 device #0
found OSDev ze3
got cpuset 0x0000ffff,0xffffffff,0xf0000000,0x000000ff,0xffffffff,0xfff00000,0x0 for driver #3 device #0
found OSDev ze4
got cpuset 0x0000ffff,0xffffffff,0xf0000000,0x000000ff,0xffffffff,0xfff00000,0x0 for driver #4 device #0
found OSDev ze5
got cpuset 0x0000ffff,0xffffffff,0xf0000000,0x000000ff,0xffffffff,0xfff00000,0x0 for driver #5 device #0
testing ZES devices
found 1 L0 ZES drivers
found 6 L0 ZES devices in driver #0
found OSDev ze1
levelzero: ../../../tests/hwloc/levelzero.c:188: int main(void): Assertion `atoi(osdev->name+2) == (int) k' failed.
Aborted

Using:

applenco@x1921c3s2b0n0:~/hwloc-PR-595-20241113.1537.gitc2b2cfd7/build> ze_info
Number of drivers                                 1
  Driver API Version                              1.5
  Driver Version                                  17004696
$intel_compute_runtime/release/996.26

[bgoglin]

Thanks. There's at least one bug in the test file, I'll try to debug more your report.

Aside of that it looks like ZES fails to report a valid PCI link speed (it shows 0.25GB/s instead of 63 in your case, and nothing on my machines). I'll revert that part and use ZE for this for now.

[bgoglin]

Could you comment out the assert on line 188 of tests/hwloc/levelzero.c, make -C tests/hwloc levelzero && tests/hwloc/levelzero? I'd like to confirm that devices are reported by ZES and ZE in different orders. That'd be funny, but easy to fix.