argonne-lcf · Kerilk · Sep 8, 2023 · Sep 1, 2023 · Sep 6, 2023 · Sep 6, 2023
diff --git a/Makefile.am b/Makefile.am
@@ -1,4 +1,4 @@
 ACLOCAL_AMFLAGS = -I m4
-SUBDIRS = utils opencl ze xprof cuda omp hip
+SUBDIRS = utils sampling opencl ze xprof cuda omp hip
 
 EXTRA_DIST = autogen.sh README.md
diff --git a/configure.ac b/configure.ac
@@ -98,6 +98,7 @@ AC_CHECK_FUNCS([atexit clock_gettime ftruncate memmove memset strdup strstr strt
 
 AC_CONFIG_FILES([Makefile
                  utils/Makefile
+                 sampling/Makefile
                  xprof/xprof.sh.erb
                  opencl/Makefile
                  ze/Makefile
@@ -113,6 +114,7 @@ AC_CONFIG_FILES([ze/test_wrapper.sh], [chmod +x ze/test_wrapper.sh])
 AC_CONFIG_FILES([cuda/tracer_cuda.sh], [chmod +x cuda/tracer_cuda.sh])
 AC_CONFIG_FILES([cuda/test_wrapper.sh], [chmod +x cuda/test_wrapper.sh])
 AC_CONFIG_FILES([xprof/test_wrapper.sh], [chmod +x xprof/test_wrapper.sh])
+AC_CONFIG_FILES([utils/babeltrace_energy], [chmod +x utils/babeltrace_energy])
 AC_CONFIG_FILES([utils/babeltrace_thapi], [chmod +x utils/babeltrace_thapi])
 AC_CONFIG_FILES([omp/tracer_omp.sh], [chmod +x omp/tracer_omp.sh])
 AC_CONFIG_FILES([hip/tracer_hip.sh], [chmod +x hip/tracer_hip.sh])

diff --git a/cuda/Makefile.am b/cuda/Makefile.am
@@ -269,19 +269,19 @@ nodist_libTracerCUDA_la_SOURCES = \
 	$(CUDA_PROBES_INCL) \
 	tracer_cuda.c
 
-libTracerCUDA_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(srcdir)/include -I../utils -I./
+libTracerCUDA_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(top_srcdir)/sampling -I$(srcdir)/include -I../utils -I./
 libTracerCUDA_la_CFLAGS = -Wall -Wextra $(WERROR) $(LIBFFI_CFLAGS) $(LTTNG_UST_CFLAGS)
-libTracerCUDA_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS)
+libTracerCUDA_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS) ../sampling/libThapiSampling.la
 libTracerCUDA_la_LDFLAGS += -version-info 1:0:0
 libTracerCUDA_la_LIBADD = libcudatracepoints.la
 
 nodist_libTracerCUDART_la_SOURCES = \
 	$(CUDART_PROBES_INCL) \
 	tracer_cudart.c
 
-libTracerCUDART_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(srcdir)/include -I../utils -I./
+libTracerCUDART_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(top_srcdir)/sampling -I$(srcdir)/include -I../utils -I./
 libTracerCUDART_la_CFLAGS = -Wall -Wextra $(WERROR) $(LIBFFI_CFLAGS) $(LTTNG_UST_CFLAGS)
-libTracerCUDART_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS)
+libTracerCUDART_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS) ../sampling/libThapiSampling.la
 libTracerCUDART_la_LDFLAGS += -version-number 12:1:55 -Wl,--version-script,tracer_cudart.map
 libTracerCUDART_la_LIBADD = libcudarttracepoints.la
 

diff --git a/cuda/tracer_cuda.sh.in b/cuda/tracer_cuda.sh.in
@@ -43,6 +43,7 @@ display_help() {
     echo "  -e, --exports                 Trace export functions"
     echo "  -v, --visualize               Visualize trace on thefly"
     echo "  --properties                  Dump devices infos"
+    echo "  --sample                      Sample performance counters"
     exit 1
 }
 
@@ -61,6 +62,7 @@ while true; do
     -e | --exports   ) shift; exports=1;;
     -v | --visualize ) shift; lttng_view=1;;
     --properties ) shift; properties=1;;
+    --sample ) shift; sample=1;;
     -- ) shift; break ;;
     * ) break ;;
   esac
@@ -101,6 +103,12 @@ if [ ! -z "$properties" ]
 then
   lttng enable-event --channel=blocking-channel --userspace lttng_ust_cuda_properties:*
 fi
+if [ ! -z "$sample" ]
+then
+  export LTTNG_UST_SAMPLING=1
+  lttng enable-channel --userspace nonblocking-channel
+  lttng enable-event --channel=nonblocking-channel --userspace lttng_ust_sampling:*
+fi
 if [ -z "$LTTNG_UST_CUDA_LIBCUDA" ]
 then
   LTTNG_UST_CUDA_LIBCUDA=$(whichlib64_head libcuda.so)

diff --git a/cuda/tracer_cuda_helpers.include.c b/cuda/tracer_cuda_helpers.include.c
@@ -1,3 +1,5 @@
+#include "thapi_sampling.h"
+
 //pthread_mutex_t cuda_closures_mutex = PTHREAD_MUTEX_INITIALIZER;
 //
 //struct cuda_closure {
@@ -485,6 +487,8 @@ static void _load_tracer(void) {
   void *handle = NULL;
   int verbose = 0;
 
+  thapi_sampling_init();
+
   s = getenv("LTTNG_UST_CUDA_LIBCUDA");
   if (s)
     handle = dlopen(s, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);

diff --git a/cuda/tracer_cudart_helpers.include.c b/cuda/tracer_cudart_helpers.include.c
@@ -1,3 +1,5 @@
+#include "thapi_sampling.h"
+
 static pthread_once_t _init = PTHREAD_ONCE_INIT;
 static __thread volatile int in_init = 0;
 static volatile int _initialized = 0;
@@ -7,6 +9,8 @@ static void _load_tracer(void) {
   void *handle = NULL;
   int verbose = 0;
 
+  thapi_sampling_init();
+
   s = getenv("LTTNG_UST_CUDART_LIBCUDART");
   if (s)
       handle = dlopen(s, RTLD_LAZY | RTLD_LOCAL);

diff --git a/hip/Makefile.am b/hip/Makefile.am
@@ -277,9 +277,9 @@ nodist_libTracerHIP_la_SOURCES = \
 	$(HIP_PROBES_INCL) \
 	tracer_hip.c
 
-libTracerHIP_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(srcdir)/include -I./utils -I./
+libTracerHIP_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(top_srcdir)/sampling -I$(srcdir)/include -I./utils -I./
 libTracerHIP_la_CFLAGS = -Wall -Wextra $(WERROR) $(LIBFFI_CFLAGS) $(LTTNG_UST_CFLAGS)
-libTracerHIP_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS)
+libTracerHIP_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS) ../sampling/libThapiSampling.la
 libTracerHIP_la_LDFLAGS += -Wl,--version-script,$(srcdir)/hip.map -version-number 5:4:50400
 libTracerHIP_la_DEPENDS = $(srcdir)/hip.map
 libTracerHIP_la_LIBADD = libhiptracepoints.la

diff --git a/omp/Makefile.am b/omp/Makefile.am
@@ -129,9 +129,9 @@ nodist_libTracerOMPT_la_SOURCES = \
 	$(OMP_PROBES_INCL) \
 	tracer_ompt.c
 
-libTracerOMPT_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I./modified_include -I../utils -I./
+libTracerOMPT_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(top_srcdir)/sampling -I./modified_include -I../utils -I./
 libTracerOMPT_la_CFLAGS = -Wall -Wextra $(WERROR) $(LTTNG_UST_CFLAGS)
-libTracerOMPT_la_LDFLAGS = $(LTTNG_UST_LIBS) -avoid-version -module
+libTracerOMPT_la_LDFLAGS = $(LTTNG_UST_LIBS) -avoid-version -module ../sampling/libThapiSampling.la
 libTracerOMPT_la_LIBADD = libompttracepoints.la
 
 install-exec-hook:

diff --git a/omp/tracer_omp.sh.in b/omp/tracer_omp.sh.in
@@ -12,6 +12,7 @@ display_help() {
     echo "  --help                        Show this screen"
     echo "  --version                     Print the version string"
     echo "  --disable-intel-extensions    Disable Intel extensions"
+    echo "  --sample                      Sample performance counters"
     exit 1
 }
 
@@ -26,6 +27,7 @@ while true; do
     --help                   ) display_help; exit;;
     --version                ) display_version; exit;;
     --disable-intel-extensions) intel_extensions=false shift;;
+    --sample ) shift; sample=1;;
     -- ) shift; break ;;
     * ) break ;;
   esac
@@ -53,6 +55,12 @@ export LTTNG_UST_ALLOW_BLOCKING=1
 if [ "$intel_extensions" = true ] ; then
     export LTTNG_UST_OMP_INTEL=1
 fi
+if [ ! -z "$sample" ]
+then
+  export LTTNG_UST_SAMPLING=1
+  lttng enable-channel --userspace nonblocking-channel
+  lttng enable-event --channel=nonblocking-channel --userspace lttng_ust_sampling:*
+fi
 
 lttng start
 

diff --git a/omp/tracer_ompt_helpers.include.c.erb b/omp/tracer_ompt_helpers.include.c.erb
@@ -1,3 +1,5 @@
+#include "thapi_sampling.h"
+
 <% require "yaml" %>
 
 #define _OMPT_SET_CALLBACK(value, name) \
@@ -33,6 +35,8 @@ static int _ompt_initialize(ompt_function_lookup_t lookup,
     int do_callbacks_intel = 0;
     int verbose = 0;
 
+    thapi_sampling_init();
+
     if (getenv("LTTNG_UST_OMP_INTEL"))
         do_callbacks_intel = 1;
     if (getenv("LTTNG_UST_OMP_VERBOSE"))

diff --git a/opencl/Makefile.am b/opencl/Makefile.am
@@ -157,9 +157,9 @@ nodist_libTracerOpenCL_la_SOURCES = \
 	tracer_opencl.h \
 	tracer_opencl.c
 
-libTracerOpenCL_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(srcdir)/include -I../utils -I./
+libTracerOpenCL_la_CPPFLAGS = -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(top_srcdir)/sampling -I$(srcdir)/include -I../utils -I./
 libTracerOpenCL_la_CFLAGS = -Wall -Wextra -Wno-unused-parameter $(WERROR) $(LIBFFI_CFLAGS) $(LTTNG_UST_CFLAGS)
-libTracerOpenCL_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS)
+libTracerOpenCL_la_LDFLAGS = $(LTTNG_UST_LIBS) -ldl -lpthread $(LIBFFI_LIBS) ../sampling/libThapiSampling.la
 libTracerOpenCL_la_LDFLAGS += -Wl,--version-script,$(srcdir)/tracer_opencl.map -version-info 1:0:0
 libTracerOpenCL_la_DEPENDS = $(srcdir)/tracer_opencl.map
 libTracerOpenCL_la_LIBADD = libtracepoints.la

diff --git a/opencl/tracer_opencl.sh.in b/opencl/tracer_opencl.sh.in
@@ -49,6 +49,7 @@ display_help() {
     echo "  -e, --iteration-end VALUE     Dump inputs and outputs for kernels until enqueue counter VALUE"
     echo "  -v, --visualize               Visualize trace on thefly"
     echo "  --devices                     Dump devices information"
+    echo "  --sample                      Sample performance counters"
     exit 1
 }
 
@@ -73,6 +74,7 @@ while true; do
     -e | --iteration-end ) shift; export LTTNG_UST_OPENCL_DUMP=1; export LTTNG_UST_OPENCL_DUMP_END=$1; shift ;;
     -v | --visualize) shift; lttng_view=1;;
     --devices ) shift; devices=1;;
+    --sample ) shift; sample=1;;
     -- ) shift; break ;;
     * ) break ;;
   esac
@@ -117,6 +119,12 @@ if [ ! -z "$build" ]
 then
   lttng enable-event --channel=blocking-channel --userspace lttng_ust_opencl_build:*
 fi
+if [ ! -z "$sample" ]
+then
+  export LTTNG_UST_SAMPLING=1
+  lttng enable-channel --userspace nonblocking-channel
+  lttng enable-event --channel=nonblocking-channel --userspace lttng_ust_sampling:*
+fi
 if [ ! -z "$LTTNG_UST_OPENCL_DUMP" ]
 then
   lttng enable-event --channel=blocking-channel --userspace lttng_ust_opencl_dump:*

diff --git a/opencl/tracer_opencl_helpers.include.c b/opencl/tracer_opencl_helpers.include.c
@@ -1,3 +1,5 @@
+#include "thapi_sampling.h"
+
 void CL_CALLBACK  event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) {
   (void)user_data;
   if (tracepoint_enabled(lttng_ust_opencl_profiling, event_profiling_results)) {
@@ -1180,6 +1182,8 @@ static void _load_tracer(void) {
   void * handle = NULL;
   int verbose = 0;
 
+  thapi_sampling_init();
+
   s = getenv("LTTNG_UST_OPENCL_LIBOPENCL");
   if (s)
     handle = dlopen(s, RTLD_LAZY | RTLD_LOCAL | RTLD_DEEPBIND);

diff --git a/sampling/Makefile.am b/sampling/Makefile.am
@@ -0,0 +1,59 @@
+if STRICT
+  WERROR = -Werror
+else
+  WERROR =
+endif
+LTTNG_FLAGS = -fPIC -Wall -Wextra -Wno-unused-parameter -Wno-type-limits -Wno-sign-compare $(WERROR) -I$(top_srcdir)/utils -I$(top_srcdir)/utils/include -I$(srcdir)/include -I../utils -I./
+
+
+TRACEPOINT_GEN = \
+	$(srcdir)/sampling_events.yaml
+
+SAMPLING_STATIC_PROBES = \
+	sampling
+
+SAMPLING_STATIC_PROBES_TP = $(SAMPLING_STATIC_PROBES:=.tp)
+
+SAMPLING_STATIC_PROBES_INCL = $(SAMPLING_STATIC_PROBES:=.h)
+
+SAMPLING_STATIC_PROBES_SRC = $(SAMPLING_STATIC_PROBES:=.c)
+
+$(SAMPLING_STATIC_PROBES_TP): %.tp: $(srcdir)/gen_sampling_custom_probes.rb $(TRACEPOINT_GEN)
+	SRC_DIR=$(srcdir) $(RUBY) $< lttng_ust_$* > $@
+
+%.h %.c: %.tp
+	$(LTTNG_GEN_TP) $< -o $*.c -o $*.h
+
+noinst_LTLIBRARIES = libtracepoints.la
+
+nodist_libtracepoints_la_SOURCES = \
+	$(SAMPLING_STATIC_PROBES_INCL) \
+	$(SAMPLING_STATIC_PROBES_SRC)
+
+libtracepoints_la_CFLAGS = $(LTTNG_FLAGS) $(LTTNG_UST_CFLAGS)
+libtracepoints_la_LDFLAGS = $(LTTNG_UST_LIBS)
+
+EXTRA_DIST = \
+	sampling_events.yaml \
+	gen_sampling_custom_probes.rb
+
+CLEANFILES = \
+	$(SAMPLING_STATIC_PROBES_INCL) \
+	$(SAMPLING_STATIC_PROBES_TP) \
+	$(SAMPLING_STATIC_PROBES_SRC)
+
+BUILT_SOURCES = \
+	$(SAMPLING_STATIC_PROBES_INCL)
+
+nodist_libThapiSampling_la_SOURCES = \
+	$(SAMPLING_STATIC_PROBES_INCL)
+
+libThapiSampling_la_SOURCES = \
+	thapi_sampling.h \
+	thapi_sampling.c
+
+libThapiSampling_la_CFLAGS = -Wall -Wextra -Wno-unused-parameter $(WERROR) -I$(top_srcdir)/utils/include
+libThapiSampling_la_LDFLAGS = -lpthread -version-info 1:0:0
+libThapiSampling_la_LIBADD = libtracepoints.la
+
+lib_LTLIBRARIES = libThapiSampling.la
diff --git a/sampling/gen_sampling_custom_probes.rb b/sampling/gen_sampling_custom_probes.rb
@@ -0,0 +1,32 @@
+require 'yaml'
+require_relative '../utils/LTTng'
+
+if ENV["SRC_DIR"]
+  SRC_DIR = ENV["SRC_DIR"]
+else
+  SRC_DIR = "."
+end
+
+namespace = ARGV[0]
+
+raise "No namespace provided!" unless namespace
+
+h = YAML::load_file(File.join(SRC_DIR,"sampling_events.yaml"))[namespace]
+
+raise "Invalid namespace!" unless h
+
+puts <<EOF
+#include "lttng/tracepoint_gen.h"
+
+EOF
+enums = h["enums"]
+
+if enums
+  enums.each { |e|
+    LTTng.print_enum(namespace, e)
+  }
+end
+
+h["events"].each { |e|
+  LTTng.print_tracepoint(namespace, e)
+}
diff --git a/sampling/sampling_events.yaml b/sampling/sampling_events.yaml
@@ -0,0 +1,11 @@
+---
+lttng_ust_sampling:
+  events:
+  - name: heartbeat
+    args:
+    - [uint64_t, foo]
+    fields:
+    - [ctf_integer, uint64_t, foo, foo]
+  - name: heartbeat2
+    args: []
+    fields: []