linux-dpdk: ml: implement on top of mldev

Add ML implementation, which utilizes DPDK's mldev.

Limitations:

- Only a single device can be used. The device is specified in the
  configuration file.

- Only a single inference at a time per model.

- The maximum number of concurrent inferences is limited by the number
  of queue pairs.

- Inference is always synchronous, regardless of completion mode.

Signed-off-by: Jere Leppänen <[email protected]>

config/odp-linux-dpdk.conf

@@ -16,7 +16,7 @@
 
 # Mandatory fields
 odp_implementation = "linux-dpdk"
-config_file_version = "0.1.26"
+config_file_version = "0.1.27"
 
 # System options
 system: {
@@ -337,32 +337,9 @@ dma: {
 }
 
 ml: {
-	# Enable onnxruntime profiling, when enabled, a json file will be
-	# generated after inference. chrome://tracing/ can be used to check
-	# the profiling. Use 0 to disable and 1 to enable profiling.
-	enable_profiling = 0
-
-	# Choose onnxruntime execution mode, which can be "SEQUENTIAL" or
-	# "PARALLEL"
-	execution_mode = "SEQUENTIAL"
-
-	# Set the number of threads used to parallelize the execution of the
-	# graph across nodes. A value of 0 means onnxruntime will pick a default.
-	inter_op_num_threads = 0
-
-	# Set the number of threads used to parallelize the execution within
-	# a node. A value of 0 means onnxruntime will pick a default.
-	intra_op_num_threads = 0
-
-	# Set graph optimization level. Valid values are:
-	#	DISABLE_ALL: disables all optimizations
-	#	ENABLE_BASIC: enables basic optimizations
-	#	ENABLE_EXTENDED: enables basic and extended optimizations
-	#	ENABLE_ALL:  enables all available optimizations including layout optimization
-	graph_optimization_level = "ENABLE_ALL"
-
-	# Serialize the optimized model to disk. When initializing a session
-	# with the same model, no need to apply optimization anymore, thus
-	# reducing model startup time.
-	optimized_model_filepath = ""
+	# Device ID
+	dev_id = 0
+
+	# Number of queue pairs
+	num_queue_pairs = 16
 }

platform/linux-dpdk/Makefile.am

@@ -13,7 +13,6 @@ AM_CPPFLAGS +=  -I$(top_srcdir)/platform/$(with_platform)/arch/@ARCH_DIR@
 AM_CPPFLAGS +=  -I$(top_srcdir)/platform/$(with_platform)/arch/default
 
 AM_CPPFLAGS +=  $(OPENSSL_CPPFLAGS)
-AM_CPPFLAGS +=  $(ORT_CPPFLAGS)
 
 AM_CFLAGS +=  $(DPDK_CFLAGS)
 AM_CFLAGS +=  $(LIBCONFIG_CFLAGS)
@@ -262,7 +261,7 @@ __LIB__libodp_dpdk_la_SOURCES = \
 
 if WITH_ML
 __LIB__libodp_dpdk_la_SOURCES += \
-			   ../linux-generic/odp_ml.c
+			   odp_ml.c
 else
 __LIB__libodp_dpdk_la_SOURCES += \
 			   ../linux-generic/odp_ml_null.c
@@ -423,7 +422,7 @@ __LIB__libodp_dpdk_la_LIBADD += $(LIBCONFIG_LIBS)
 __LIB__libodp_dpdk_la_LIBADD += $(DPDK_LIBS_LIBODP)
 __LIB__libodp_dpdk_la_LIBADD += $(PTHREAD_LIBS)
 __LIB__libodp_dpdk_la_LIBADD += $(TIMER_LIBS)
-__LIB__libodp_dpdk_la_LIBADD += $(ORT_LIBS)
+__LIB__libodp_dpdk_la_LIBADD += $(ML_LIBS)
 
 CHECK_GLOBALS_REGEX = " (odp_|_odp_|_deprecated_odp_|miniz_|mz_|tdefl_|tinfl_|mp_hdlr_init_odp_pool_ops|__odr_asan)"
 

platform/linux-dpdk/include/odp_config_internal.h

@@ -73,8 +73,10 @@ extern "C" {
 
 /*
  * Stashes reserved for internal usage
+ *
+ * One stash reserved per DMA session, and one reserved for ML.
  */
-#define CONFIG_INTERNAL_STASHES CONFIG_MAX_DMA_SESSIONS
+#define CONFIG_INTERNAL_STASHES (CONFIG_MAX_DMA_SESSIONS + 1)
 
 /*
  * Maximum number of stashes

platform/linux-dpdk/libodp-dpdk.pc.in

@@ -8,5 +8,5 @@ Description: The ODP packet processing engine
 Version: @PKGCONFIG_VERSION@
 Requires.private: libconfig
 Libs: -L${libdir} -l@ODP_LIB_NAME@ @DPDK_LIBS_NON_ABI_COMPAT@ @ATOMIC_LIBS_NON_ABI_COMPAT@
-Libs.private: @DPDK_LIBS_ABI_COMPAT@ @OPENSSL_STATIC_LIBS@ @PTHREAD_LIBS@ @TIMER_LIBS@ -lpthread @ATOMIC_LIBS_ABI_COMPAT@ @ORT_LIBS@
+Libs.private: @DPDK_LIBS_ABI_COMPAT@ @OPENSSL_STATIC_LIBS@ @PTHREAD_LIBS@ @TIMER_LIBS@ -lpthread @ATOMIC_LIBS_ABI_COMPAT@ @ML_LIBS@
 Cflags: -I${includedir} @DPDK_CFLAGS@

platform/linux-dpdk/m4/configure.m4

@@ -71,7 +71,11 @@ esac
 # Required for experimental rte_event_port_unlinks_in_progress() API
 DPDK_CFLAGS="${DPDK_CFLAGS} -DALLOW_EXPERIMENTAL_API"
 
-AS_VAR_APPEND([PLAT_DEP_LIBS], ["${ATOMIC_LIBS} ${LIBCONFIG_LIBS} ${OPENSSL_LIBS} ${DPDK_LIBS_LT} ${LIBCLI_LIBS} ${ORT_LIBS}"])
+# ML availability check requires DPDK include path, and must therefore be done
+# after DPDK check.
+ODP_ML
+
+AS_VAR_APPEND([PLAT_DEP_LIBS], ["${ATOMIC_LIBS} ${LIBCONFIG_LIBS} ${OPENSSL_LIBS} ${DPDK_LIBS_LT} ${LIBCLI_LIBS}"])
 
 # Add text to the end of configure with platform specific settings.
 # Make sure it's aligned same as other lines in configure.ac.

platform/linux-dpdk/m4/odp_libconfig.m4

@@ -7,7 +7,7 @@
 ##########################################################################
 m4_define([_odp_config_version_generation], [0])
 m4_define([_odp_config_version_major], [1])
-m4_define([_odp_config_version_minor], [26])
+m4_define([_odp_config_version_minor], [27])
 
 m4_define([_odp_config_version],
           [_odp_config_version_generation._odp_config_version_major._odp_config_version_minor])

platform/linux-dpdk/m4/odp_ml.m4

@@ -0,0 +1,36 @@
+##########################################################################
+# DPDK ML API
+##########################################################################
+AC_DEFUN([ODP_ML], [dnl
+
+ml_support=no
+ml_option=yes
+AC_ARG_ENABLE([dpdk-ml],
+	[AS_HELP_STRING([--disable-dpdk-ml],
+			[disable ML support]
+			[[default=enabled] (linux-dpdk)])],
+	[if test "x$enableval" = "xno"; then
+		ml_option=no
+	fi])
+
+AS_IF([test "x$ml_option" != "xno"], [dnl
+
+#########################################################################
+# If dpdk mldev header is available, enable ML API
+#########################################################################
+OLD_CPPFLAGS=$CPPFLAGS
+CPPFLAGS="$DPDK_CFLAGS $CPPFLAGS"
+AC_CHECK_HEADERS([rte_mldev.h], [ml_support=yes], [], [])
+CPPFLAGS=$OLD_CPPFLAGS
+
+])
+
+AC_CONFIG_COMMANDS_PRE([dnl
+AM_CONDITIONAL([WITH_ML], [test x$ml_support = xyes])
+])
+
+# Even if ML is not enabled, -lm is needed by quantization.
+ML_LIBS="-lm"
+AC_SUBST([ML_LIBS])
+
+]) # ODP_ML