PSM3 OFI Provider from IEFS 11_5_1_1_1 (#9)

Ensure 11.5.1.1 release matches SRPM found in IEFS release

Signed-off-by: Scott Breyer <[email protected]>

VERSION

@@ -1 +1 @@
-11_5_1_0
+11_5_1_1

configure

@@ -1,6 +1,6 @@
 #! /bin/sh
 # Guess values for system-dependent variables and create Makefiles.
-# Generated by GNU Autoconf 2.69 for libpsm3-fi 11.5.1.0.
+# Generated by GNU Autoconf 2.69 for libpsm3-fi 11.5.1.1.
 #
 #
 # Copyright (C) 1992-1996, 1998-2012 Free Software Foundation, Inc.
@@ -587,8 +587,8 @@ MAKEFLAGS=
 # Identity of this package.
 PACKAGE_NAME='libpsm3-fi'
 PACKAGE_TARNAME='libpsm3-fi'
-PACKAGE_VERSION='11.5.1.0'
-PACKAGE_STRING='libpsm3-fi 11.5.1.0'
+PACKAGE_VERSION='11.5.1.1'
+PACKAGE_STRING='libpsm3-fi 11.5.1.1'
 PACKAGE_BUGREPORT=''
 PACKAGE_URL=''
 
@@ -1366,7 +1366,7 @@ if test "$ac_init_help" = "long"; then
   # Omit some internal or obsolete options to make the list less imposing.
   # This message is too long to be a string in the A/UX 3.1 sh.
   cat <<_ACEOF
-\`configure' configures libpsm3-fi 11.5.1.0 to adapt to many kinds of systems.
+\`configure' configures libpsm3-fi 11.5.1.1 to adapt to many kinds of systems.
 
 Usage: $0 [OPTION]... [VAR=VALUE]...
 
@@ -1436,7 +1436,7 @@ fi
 
 if test -n "$ac_init_help"; then
   case $ac_init_help in
-     short | recursive ) echo "Configuration of libpsm3-fi 11.5.1.0:";;
+     short | recursive ) echo "Configuration of libpsm3-fi 11.5.1.1:";;
    esac
   cat <<\_ACEOF
 
@@ -1589,7 +1589,7 @@ fi
 test -n "$ac_init_help" && exit $ac_status
 if $ac_init_version; then
   cat <<\_ACEOF
-libpsm3-fi configure 11.5.1.0
+libpsm3-fi configure 11.5.1.1
 generated by GNU Autoconf 2.69
 
 Copyright (C) 2012 Free Software Foundation, Inc.
@@ -2000,7 +2000,7 @@ cat >config.log <<_ACEOF
 This file contains any messages produced by compilers while
 running configure, to aid debugging if configure makes a mistake.
 
-It was created by libpsm3-fi $as_me 11.5.1.0, which was
+It was created by libpsm3-fi $as_me 11.5.1.1, which was
 generated by GNU Autoconf 2.69.  Invocation command line was
 
   $ $0 $@
@@ -2869,7 +2869,7 @@ fi
 
 # Define the identity of the package.
  PACKAGE='libpsm3-fi'
- VERSION='11.5.1.0'
+ VERSION='11.5.1.1'
 
 
 cat >>confdefs.h <<_ACEOF
@@ -14045,7 +14045,7 @@ Usage: $0 [OPTIONS]
 Report bugs to <[email protected]>."
 
 lt_cl_version="\
-libpsm3-fi config.lt 11.5.1.0
+libpsm3-fi config.lt 11.5.1.1
 configured by $0, generated by GNU Autoconf 2.69.
 
 Copyright (C) 2011 Free Software Foundation, Inc.
@@ -17319,7 +17319,7 @@ cat >>$CONFIG_STATUS <<\_ACEOF || ac_write_fail=1
 # report actual input values of CONFIG_FILES etc. instead of their
 # values after options handling.
 ac_log="
-This file was extended by libpsm3-fi $as_me 11.5.1.0, which was
+This file was extended by libpsm3-fi $as_me 11.5.1.1, which was
 generated by GNU Autoconf 2.69.  Invocation command line was
 
   CONFIG_FILES    = $CONFIG_FILES
@@ -17385,7 +17385,7 @@ _ACEOF
 cat >>$CONFIG_STATUS <<_ACEOF || ac_write_fail=1
 ac_cs_config="`$as_echo "$ac_configure_args" | sed 's/^ //; s/[\\""\`\$]/\\\\&/g'`"
 ac_cs_version="\\
-libpsm3-fi config.status 11.5.1.0
+libpsm3-fi config.status 11.5.1.1
 configured by $0, generated by GNU Autoconf 2.69,
   with options \\"\$ac_cs_config\\"
 

debian/changelog

@@ -1,4 +1,4 @@
-libpsm3-fi (11.5.1.0-3) unstable; urgency=medium
+libpsm3-fi (11.5.1.1-1) unstable; urgency=medium
 
   * Initial release 
 

libpsm3-fi.spec

@@ -3,8 +3,8 @@
 %{!?provider_formal: %define provider_formal PSM3}
 
 Name: lib%{provider}-fi
-Version: 11.5.1.0
-Release: 3
+Version: 11.5.1.1
+Release: 1
 Summary: Dynamic %{provider_formal} provider for Libfabric
 
 Group: System Environment/Libraries

man/man7/fi_psm3.7

@@ -1,6 +1,6 @@
 .\" Automatically generated by Pandoc 2.9.2.1
 .\"
-.TH "fi_psm3" "7" "2022\-12\-09" "Libfabric Programmer\[cq]s Manual" "Libfabric v11.5.1.0"
+.TH "fi_psm3" "7" "2022\-12\-09" "Libfabric Programmer\[cq]s Manual" "Libfabric v11.5.1.1"
 .hy
 .SH NAME
 .PP

psm3/hal_sockets/sockets_ep.c

@@ -816,6 +816,9 @@ static psm2_error_t open_rv(psm2_ep_t ep, psm2_uuid_t const job_key)
 
 	// GPU Direct is enabled and we need a GPU Cache
 	loc_info.rdma_mode = RV_RDMA_MODE_GPU_ONLY;
+#ifdef PSM_ONEAPI
+	psm3_oneapi_ze_can_use_zemem();
+#endif
 
 	// need portnum for rdma_mode KERNEL or (USER|GPU)
 	loc_info.port_num = ep->portnum;

psm3/hal_verbs/verbs_ep.c

@@ -796,6 +796,15 @@ psm2_error_t psm_verbs_alloc_send_pool(psm2_ep_t ep, struct ibv_pd *pd,
 			_HFI_ERROR( "can't alloc send buffers");
 			goto fail;
 		}
+#if defined(PSM_ONEAPI) && !defined(PSM3_NO_ONEAPI_IMPORT)
+		// By registering memory with Level Zero, we make
+		// zeCommandListAppendMemoryCopy run faster for copies from
+		// GPU to the send buffer.
+		if (PSMI_IS_GPU_ENABLED)
+			PSMI_ONEAPI_ZE_CALL(zexDriverImportExternalPointer,
+				ze_driver, pool->send_buffers,
+				pool->send_total*pool->send_buffer_size);
+#endif
 
 		_HFI_PRDBG("send pool: buffers: %p size %u\n",  pool->send_buffers, pool->send_buffer_size);
 		pool->send_bufs = (struct verbs_sbuf *)psmi_calloc(ep, NETWORK_BUFFERS,
@@ -883,6 +892,15 @@ psm2_error_t psm_verbs_alloc_recv_pool(psm2_ep_t ep, struct ibv_qp *qp,
 				_HFI_ERROR( "can't alloc recv buffers");
 				goto fail;
 			}
+#if defined(PSM_ONEAPI) && !defined(PSM3_NO_ONEAPI_IMPORT)
+			// By registering memory with Level Zero, we make
+			// zeCommandListAppendMemoryCopy run faster for copies from
+			// recv buffer to GPU
+			if (PSMI_IS_GPU_ENABLED)
+				PSMI_ONEAPI_ZE_CALL(zexDriverImportExternalPointer,
+					ze_driver, pool->recv_buffers,
+					pool->recv_total*pool->recv_buffer_size);
+#endif
 			//printf("recv pool: buffers: %p size %u\n",  pool->recv_buffers, pool->recv_buffer_size);
 #ifdef USE_RC
 			pool->recv_bufs = (struct verbs_rbuf *)psmi_calloc(ep, NETWORK_BUFFERS,
@@ -971,6 +989,11 @@ void psm_verbs_free_send_pool(psm3_verbs_send_pool_t pool)
 		pool->send_bufs = NULL;
 	}
 	if (pool->send_buffers) {
+#if defined(PSM_ONEAPI) && !defined(PSM3_NO_ONEAPI_IMPORT)
+		if (PSMI_IS_GPU_ENABLED)
+			PSMI_ONEAPI_ZE_CALL(zexDriverReleaseImportedPointer,
+					 ze_driver, pool->send_buffers);
+#endif
 		psmi_free(pool->send_buffers);
 		pool->send_buffers = NULL;
 	}
@@ -991,6 +1014,11 @@ void psm_verbs_free_recv_pool(psm3_verbs_recv_pool_t pool)
 	}
 #endif
 	if (pool->recv_buffers) {
+#if defined(PSM_ONEAPI) && !defined(PSM3_NO_ONEAPI_IMPORT)
+		if (PSMI_IS_GPU_ENABLED)
+			PSMI_ONEAPI_ZE_CALL(zexDriverReleaseImportedPointer,
+					 ze_driver, pool->recv_buffers);
+#endif
 		psmi_free(pool->recv_buffers);
 		pool->recv_buffers = NULL;
 	}
@@ -1968,6 +1996,9 @@ static psm2_error_t open_rv(psm2_ep_t ep, psm2_uuid_t const job_key)
  		if (psmi_parse_gpudirect()) {
 			// When GPU Direct is enabled we need a GPU Cache
 			loc_info.rdma_mode |= RV_RDMA_MODE_GPU;
+#ifdef PSM_ONEAPI
+			psm3_oneapi_ze_can_use_zemem();
+#endif
 			if ((ep->rdmamode & IPS_PROTOEXP_FLAG_ENABLED)
 				&& (psmi_parse_gpudirect_rdma_send_limit(1)
 				|| psmi_parse_gpudirect_rdma_recv_limit(1))) {

psm3/psm.c

@@ -366,6 +366,11 @@ static void psmi_cuda_stats_register()
 ze_result_t (*psmi_zeInit)(ze_init_flags_t flags);
 ze_result_t (*psmi_zeDriverGet)(uint32_t *pCount, ze_driver_handle_t *phDrivers);
 ze_result_t (*psmi_zeDeviceGet)(ze_driver_handle_t hDriver, uint32_t *pCount, ze_device_handle_t *phDevices);
+#ifndef PSM3_NO_ONEAPI_IMPORT
+ze_result_t (*psmi_zeDriverGetExtensionFunctionAddress)(ze_driver_handle_t hDriver, const char *name, void **ppFunctionAddress);
+ze_result_t (*psmi_zexDriverImportExternalPointer)(ze_driver_handle_t hDriver, void *ptr, size_t size);
+ze_result_t (*psmi_zexDriverReleaseImportedPointer)(ze_driver_handle_t hDriver, void *ptr);
+#endif
 ze_result_t (*psmi_zeContextCreate)(ze_driver_handle_t hDriver, const ze_context_desc_t *desc, ze_context_handle_t *phContext);
 ze_result_t (*psmi_zeContextDestroy)(ze_context_handle_t hContext);
 ze_result_t (*psmi_zeCommandQueueCreate)(ze_context_handle_t hContext, ze_device_handle_t hDevice,const ze_command_queue_desc_t *desc, ze_command_queue_handle_t *phCommandQueue);
@@ -406,6 +411,11 @@ ze_result_t (*psmi_zelLoaderGetVersions)(size_t *num_elems, zel_component_versio
 uint64_t psmi_count_zeInit;
 uint64_t psmi_count_zeDriverGet;
 uint64_t psmi_count_zeDeviceGet;
+#ifndef PSM3_NO_ONEAPI_IMPORT
+uint64_t psmi_count_zeDriverGetExtensionFunctionAddress;
+uint64_t psmi_count_zexDriverImportExternalPointer;
+uint64_t psmi_count_zexDriverReleaseImportedPointer;
+#endif
 uint64_t psmi_count_zeContextCreate;
 uint64_t psmi_count_zeContextDestroy;
 uint64_t psmi_count_zeCommandQueueCreate;
@@ -463,6 +473,9 @@ int psmi_oneapi_ze_load()
 	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeInit);
 	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeDriverGet);
 	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeDeviceGet);
+#ifndef PSM3_NO_ONEAPI_IMPORT
+	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeDriverGetExtensionFunctionAddress);
+#endif
 	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeContextCreate);
 	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeContextDestroy);
 	PSMI_ONEAPI_ZE_DLSYM(psmi_oneapi_ze_lib, zeCommandQueueCreate);
@@ -522,6 +535,11 @@ static void psmi_oneapi_ze_stats_register()
 		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeInit),
 		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeDriverGet),
 		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeDeviceGet),
+#ifndef PSM3_NO_ONEAPI_IMPORT
+		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeDriverGetExtensionFunctionAddress),
+		PSMI_ONEAPI_ZE_COUNT_DECLU64(zexDriverImportExternalPointer),
+		PSMI_ONEAPI_ZE_COUNT_DECLU64(zexDriverReleaseImportedPointer),
+#endif
 		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeContextCreate),
 		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeContextDestroy),
 		PSMI_ONEAPI_ZE_COUNT_DECLU64(zeCommandQueueCreate),
@@ -738,39 +756,38 @@ static void psmi_oneapi_find_copy_only_engine(ze_device_handle_t dev,
 	psmi_free(props);
 }
 
+// create command queue for use in psmi_oneapi_ze_memcpy for sync memcpy
 static void psmi_oneapi_cmd_create(ze_device_handle_t dev, struct ze_dev_ctxt *ctxt)
 {
 	ze_command_queue_desc_t ze_cq_desc = {
 		.stype = ZE_STRUCTURE_TYPE_COMMAND_QUEUE_DESC,
 		.flags = 0,
-#ifdef PSM3_USE_ONEAPI_IMMEDIATE
-		.mode = ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS,
-#else
-		.mode = ZE_COMMAND_QUEUE_MODE_DEFAULT,
-#endif
+		//.mode set below
 		.priority = ZE_COMMAND_QUEUE_PRIORITY_NORMAL,
 	};
-#ifndef PSM3_USE_ONEAPI_IMMEDIATE
-	ze_command_list_desc_t ze_cl_desc = {
-		.stype = ZE_STRUCTURE_TYPE_COMMAND_LIST_DESC,
-		.flags = 0
-	};
-#endif
 
 	psmi_oneapi_find_copy_only_engine(dev, ctxt);
 	ze_cq_desc.ordinal = ctxt->ordinal;
 	ze_cq_desc.index = ctxt->index;
-#ifdef PSM3_USE_ONEAPI_IMMEDIATE
-	PSMI_ONEAPI_ZE_CALL(zeCommandListCreateImmediate, ze_context, dev,
-			    &ze_cq_desc, &ctxt->cl);
-#else
-	PSMI_ONEAPI_ZE_CALL(zeCommandQueueCreate, ze_context, dev,
-			    &ze_cq_desc, &ctxt->cq);
 
-	ze_cl_desc.commandQueueGroupOrdinal = ctxt->ordinal;
-	PSMI_ONEAPI_ZE_CALL(zeCommandListCreate, ze_context, dev, &ze_cl_desc,
-			    &ctxt->cl);
-#endif
+	if (psm3_oneapi_immed_sync_copy) {
+		ze_cq_desc.mode = ZE_COMMAND_QUEUE_MODE_SYNCHRONOUS;
+		PSMI_ONEAPI_ZE_CALL(zeCommandListCreateImmediate, ze_context,
+			 dev, &ze_cq_desc, &ctxt->cl);
+	} else {
+		ze_command_list_desc_t ze_cl_desc = {
+			.stype = ZE_STRUCTURE_TYPE_COMMAND_LIST_DESC,
+			.flags = 0
+		};
+		ze_cq_desc.mode = ZE_COMMAND_QUEUE_MODE_DEFAULT;
+
+		PSMI_ONEAPI_ZE_CALL(zeCommandQueueCreate, ze_context,
+			dev, &ze_cq_desc, &ctxt->cq);
+
+		ze_cl_desc.commandQueueGroupOrdinal = ctxt->ordinal;
+		PSMI_ONEAPI_ZE_CALL(zeCommandListCreate, ze_context,
+			dev, &ze_cl_desc, &ctxt->cl);
+	}
 	ctxt->dev = dev;
 }
 
@@ -787,11 +804,7 @@ void psmi_oneapi_cmd_create_all(void)
 	for (i = 0; i < num_ze_devices; i++) {
 		ctxt = &ze_devices[i];
 
-#ifdef PSM3_USE_ONEAPI_IMMEDIATE
 		if (!ctxt->cl)
-#else
-		if (!ctxt->cq || !ctxt->cl)
-#endif
 			psmi_oneapi_cmd_create(ctxt->dev, ctxt);
 	}
 	if (num_ze_devices > 0)
@@ -810,12 +823,10 @@ void psmi_oneapi_cmd_destroy_all(void)
 			PSMI_ONEAPI_ZE_CALL(zeCommandListDestroy, ctxt->cl);
 			ctxt->cl = NULL;
 		}
-#ifndef PSM3_USE_ONEAPI_IMMEDIATE
 		if (ctxt->cq) {
 			PSMI_ONEAPI_ZE_CALL(zeCommandQueueDestroy, ctxt->cq);
 			ctxt->cq = NULL;
 		}
-#endif
 	}
 	cur_ze_dev = NULL;
 
@@ -835,6 +846,7 @@ int psmi_oneapi_ze_initialize()
 	zel_component_version_t *zel_comps = NULL;
 	size_t num_zel_comps;
 	int i;
+	union psmi_envvar_val env;
 
 	PSM2_LOG_MSG("entering");
 	_HFI_VDBG("Init Level Zero library.\n");
@@ -844,6 +856,19 @@ int psmi_oneapi_ze_initialize()
 	if (err != PSM2_OK)
 		goto fail;
 
+	psm3_getenv("PSM3_ONEAPI_IMMED_SYNC_COPY",
+				"Use Immediate CommandList for synchronous copy to/from GPU]",
+				PSMI_ENVVAR_LEVEL_HIDDEN, PSMI_ENVVAR_TYPE_INT,
+				(union psmi_envvar_val)1, &env);
+	psm3_oneapi_immed_sync_copy = env.e_int;
+
+	psm3_getenv("PSM3_ONEAPI_IMMED_ASYNC_COPY",
+				"Use Immediate CommandList for asynchronous pipeline copy to/from GPU]",
+				PSMI_ENVVAR_LEVEL_HIDDEN, PSMI_ENVVAR_TYPE_INT,
+				(union psmi_envvar_val)1, &env);
+	psm3_oneapi_immed_async_copy = env.e_int;
+
+
 	PSMI_ONEAPI_ZE_CALL(zeInit, ZE_INIT_FLAG_GPU_ONLY);
 
 	/* Need to query count before alloc array */
@@ -874,6 +899,10 @@ int psmi_oneapi_ze_initialize()
 	}
 
 	PSMI_ONEAPI_ZE_CALL(zeDriverGet, &ze_driver_count, &ze_driver);
+#ifndef PSM3_NO_ONEAPI_IMPORT
+	PSMI_ONEAPI_ZE_CALL(zeDriverGetExtensionFunctionAddress, ze_driver, "zexDriverImportExternalPointer", (void **)&psmi_zexDriverImportExternalPointer);
+	PSMI_ONEAPI_ZE_CALL(zeDriverGetExtensionFunctionAddress, ze_driver, "zexDriverReleaseImportedPointer", (void **)&psmi_zexDriverReleaseImportedPointer);
+#endif
 
 	PSMI_ONEAPI_ZE_CALL(zeDeviceGet, ze_driver, &ze_device_count, NULL);
 	if (ze_device_count > MAX_ZE_DEVICES)