levelzero: use zesInit() when available

When zesInit() is available (starting in specs 1.5), we don't need
all the stuff to set ZES_ENABLE_SYSMAN=1 in the environment anymore.

Unfortunately  zesInit() is available but returns an error in current
runtimes. Hence we'd need AC_TRY_RUN() to be sure we can enable it
and disable the ZES_ENABLE_SYSMAN=1 stuff.

Signed-off-by: Brice Goglin <Brice.Goglin@inria.fr>

config/hwloc.m4

@@ -1409,6 +1409,7 @@ return clGetDeviceIDs(0, 0, 0, NULL, NULL);
       HWLOC_PKG_CHECK_MODULES([LEVELZERO], [libze_loader], [zesDevicePciGetProperties], [level_zero/zes_api.h],
                               [hwloc_levelzero_happy=yes
                                HWLOC_LEVELZERO_REQUIRES=libze_loader
+			       AC_CHECK_LIB([ze_loader], [zesInit], [AC_DEFINE(HWLOC_HAVE_ZESINIT, 1, [Define to 1 if zesInit is available])])
 			       AC_CHECK_LIB([ze_loader], [zeDevicePciGetPropertiesExt], [AC_DEFINE(HWLOC_HAVE_ZEDEVICEPCIGETPROPERTIESEXT, 1, [Define to 1 if zeDevicePciGetPropertiesExt is available])])
                               ], [hwloc_levelzero_happy=no])
       if test x$hwloc_levelzero_happy = xno; then
@@ -1419,6 +1420,7 @@ return clGetDeviceIDs(0, 0, 0, NULL, NULL);
               AC_CHECK_LIB([ze_loader],
 	                   [zesDevicePciGetProperties],
 	                   [HWLOC_LEVELZERO_LIBS="-lze_loader"
+			    AC_CHECK_LIB([ze_loader], [zesInit], [AC_DEFINE(HWLOC_HAVE_ZESINIT, 1, [Define to 1 if zesInit is available])])
 			    AC_CHECK_LIB([ze_loader], [zeDevicePciGetPropertiesExt], [AC_DEFINE(HWLOC_HAVE_ZEDEVICEPCIGETPROPERTIESEXT, 1, [Define to 1 if zeDevicePciGetPropertiesExt is available])])
                            ], [hwloc_levelzero_happy=no])
             ], [hwloc_levelzero_happy=no])

hwloc/topology-levelzero.c

@@ -1,5 +1,5 @@
 /*
- * Copyright © 2020-2022 Inria.  All rights reserved.
+ * Copyright © 2020-2023 Inria.  All rights reserved.
  * See COPYING in top-level directory.
  */
 
@@ -604,8 +604,7 @@ hwloc_levelzero_discover(struct hwloc_backend *backend, struct hwloc_disc_status
   uint32_t nbdrivers, i, k, zeidx;
   struct hwloc_osdev_array oarray;
   struct hwloc_levelzero_ports hports;
-  int sysman_maybe_missing = 0; /* 1 if ZES_ENABLE_SYSMAN=1 was NOT set early, 2 if ZES_ENABLE_SYSMAN=0 */
-  char *env;
+  int sysman_maybe_missing = 0; /* 1 if ZES_ENABLE_SYSMAN=1 was NOT set early and zesInit() isn't available, 2 if ZES_ENABLE_SYSMAN=0 */
 
   hwloc__levelzero_osdev_array_init(&oarray);
 
@@ -617,13 +616,23 @@ hwloc_levelzero_discover(struct hwloc_backend *backend, struct hwloc_disc_status
   if (filter == HWLOC_TYPE_FILTER_KEEP_NONE)
     return 0;
 
+#ifdef HWLOC_HAVE_ZESINIT
+  res = zesInit(0);
+  if (res != ZE_RESULT_SUCCESS) {
+    if (HWLOC_SHOW_ALL_ERRORS()) {
+      fprintf(stderr, "Failed to initialize LevelZero in zesInit(): 0x%x\n", (unsigned)res);
+    }
+    return 0;
+  }
+#else /* !HWLOC_HAVE_ZESINIT */
   /* Tell L0 to create sysman devices.
    * If somebody already initialized L0 without Sysman,
    * zesDeviceGetProperties() will fail and warn in hwloc__levelzero_properties_get().
    * The lib constructor and Windows DllMain tried to set ZES_ENABLE_SYSMAN=1 early (see topology.c),
    * we try again in case they didn't.
    */
-  env = getenv("ZES_ENABLE_SYSMAN");
+ {
+  char *env = getenv("ZES_ENABLE_SYSMAN");
   if (!env) {
     /* setenv() is safer than putenv() but not available on Windows */
 #ifdef HWLOC_WIN_SYS
@@ -636,6 +645,8 @@ hwloc_levelzero_discover(struct hwloc_backend *backend, struct hwloc_disc_status
   } else if (!atoi(env)) {
     sysman_maybe_missing = 2;
   }
+ }
+#endif /* !HWLOC_HAVE_ZESINIT */
 
   res = zeInit(0);
   if (res != ZE_RESULT_SUCCESS) {

hwloc/topology.c

@@ -54,9 +54,10 @@
 #endif
 
 
-#ifdef HWLOC_HAVE_LEVELZERO
+#if (defined HWLOC_HAVE_LEVELZERO) && !(defined HWLOC_HAVE_ZESINIT)
 /*
  * Define ZES_ENABLE_SYSMAN=1 early so that the LevelZero backend gets Sysman enabled.
+ * This is only for old releases (<1.5) without zesInit().
  *
  * Only if the levelzero was enabled in this build so that we don't enable sysman
  * for external levelzero users when hwloc doesn't need it. If somebody ever loads
@@ -101,7 +102,7 @@ BOOL WINAPI DllMain(HINSTANCE hinstDLL, DWORD fdwReason, LPVOID lpReserved)
   return TRUE;
 }
 #endif
-#endif /* HWLOC_HAVE_LEVELZERO */
+#endif /* HWLOC_HAVE_LEVELZERO && !HWLOC_HAVE_ZESINIT */
 
 
 unsigned hwloc_get_api_version(void)

include/hwloc/levelzero.h

@@ -44,8 +44,9 @@ extern "C" {
  * the Level Zero device \p device.
  *
  * Topology \p topology and device \p device must match the local machine.
- * The Level Zero must have been initialized with Sysman enabled
- * (ZES_ENABLE_SYSMAN=1 in the environment).
+ * The Level Zero library must have been initialized with Sysman enabled
+ * (by calling zesInit(0) in versions >=1.5,
+ *  or by setting ZES_ENABLE_SYSMAN=1 in the environment).
  * I/O devices detection and the Level Zero component are not needed in the
  * topology.
  *

tests/hwloc/levelzero.c

@@ -1,5 +1,5 @@
 /*
- * Copyright © 2021 Inria.  All rights reserved.
+ * Copyright © 2021-2023 Inria.  All rights reserved.
  * See COPYING in top-level directory.
  */
 
@@ -22,7 +22,15 @@ int main(void)
   ze_result_t res;
   int err = 0;
 
+#ifdef HWLOC_HAVE_ZESINIT
+  zesInit(0);
+  if (res != ZE_RESULT_SUCCESS) {
+    fprintf(stderr, "Failed to initialize LevelZero in zesInit(): %d\n", (int)res);
+    return 0;
+  }
+#else
   putenv((char *) "ZES_ENABLE_SYSMAN=1");
+#endif
 
   res = zeInit(0);
   if (res != ZE_RESULT_SUCCESS) {