From 6cad7d210b3f4ba76de1c78533c5321915def92b Mon Sep 17 00:00:00 2001
From: Aurelien Bouteiller <bouteill@icl.utk.edu>
Date: Thu, 22 Jun 2023 14:05:09 -0400
Subject: [PATCH 1/2] Use new feature parsec_devices_print_stats to print
 per-tp stats if verbose >=4

Signed-off-by: Aurelien Bouteiller <bouteill@icl.utk.edu>
---
 tests/common.c | 4 ++++
 tests/common.h | 6 ++++++
 2 files changed, 10 insertions(+)

diff --git a/tests/common.c b/tests/common.c
index 78a3d865..3d78ec0b 100644
--- a/tests/common.c
+++ b/tests/common.c
@@ -67,6 +67,8 @@ double time_elapsed = 0.0;
 double sync_time_elapsed = 0.0;
 double alpha = 1.;
 
+uint64_t *dev_stats = NULL;
+
 /**********************************
  * Command line arguments
  **********************************/
@@ -734,6 +736,8 @@ void cleanup_parsec(parsec_context_t* parsec, int *iparam)
     parsec_info_unregister(&parsec_per_stream_infos, CuHI, NULL);
 #endif
 
+    if(NULL != dev_stats) parsec_devices_free_statistics(&dev_stats);
+
     parsec_fini(&parsec);
 
 #ifdef PARSEC_HAVE_MPI
diff --git a/tests/common.h b/tests/common.h
index 6c3d23e9..c51a3711 100644
--- a/tests/common.h
+++ b/tests/common.h
@@ -35,6 +35,7 @@
 extern char *PARSEC_SCHED_NAME[];
 extern int unix_timestamp;
 extern char cwd[];
+extern uint64_t *dev_stats;
 
 /* Update PASTE_CODE_PROGRESS_KERNEL below if you change this list */
 enum iparam_t {
@@ -231,6 +232,7 @@ static inline int min(int a, int b) { return a < b ? a : b; }
     PROFILING_SAVE_iINFO("PARAM_SCHEDULER", iparam[IPARAM_SCHEDULER]);  
 
 #define PASTE_CODE_PROGRESS_KERNEL(PARSEC, KERNEL)                      \
+    parsec_devices_save_statistics(&dev_stats);                         \
     SYNC_TIME_START();                                                  \
     PARSEC_CHECK_ERROR(parsec_context_start(PARSEC), "parsec_context_start"); \
     TIME_START();                                                       \
@@ -239,6 +241,7 @@ static inline int min(int a, int b) { return a < b ? a : b; }
                            P, Q, gpus, NB, N,                           \
                            gflops=(flops/1e9)/sync_time_elapsed));      \
     PASTE_PROF_INFO;                                                    \
+    if(loud >= 4) parsec_devices_print_statistics(PARSEC, dev_stats);   \
     if(loud >= 5 && rank == 0) {                                        \
         printf("<DartMeasurement name=\"performance\" type=\"numeric/double\"\n" \
                "                 encoding=\"none\" compression=\"none\">\n" \
@@ -246,6 +249,7 @@ static inline int min(int a, int b) { return a < b ? a : b; }
                "</DartMeasurement>\n",                                  \
                gflops);                                                 \
     }                                                                   \
+    if(rank==0) fflush(stdout);                                         \
     (void)gflops;
 
 
@@ -254,6 +258,7 @@ static inline int min(int a, int b) { return a < b ? a : b; }
     parsec_taskpool_t* PARSEC_##KERNEL = dplasma_##KERNEL##_New PARAMS; \
     PARSEC_CHECK_ERROR(parsec_context_add_taskpool(PARSEC, PARSEC_##KERNEL), "parsec_context_add_taskpool");\
     SYNC_TIME_STOP();                                                   \
+    parsec_devices_save_statistics(&dev_stats);                         \
     double stime_A = sync_time_elapsed;                                 \
     SYNC_TIME_START();                                                  \
     PARSEC_CHECK_ERROR(parsec_context_start(PARSEC), "parsec_context_start");\
@@ -276,6 +281,7 @@ static inline int min(int a, int b) { return a < b ? a : b; }
                           stime_A,stime_C);                             \
     }                                                                   \
     PASTE_PROF_INFO;                                                    \
+    if(loud >= 4) parsec_devices_print_statistics(PARSEC, dev_stats);   \
     if(loud >= 5 && rank == 0) {                                        \
         printf("<DartMeasurement name=\"performance\" type=\"numeric/double\"\n" \
                "                 encoding=\"none\" compression=\"none\">\n" \

From 0d1e9e7524cc618de3db3dfb032d2716c06df541 Mon Sep 17 00:00:00 2001
From: Aurelien Bouteiller <bouteill@icl.utk.edu>
Date: Wed, 31 Jan 2024 13:22:58 -0500
Subject: [PATCH 2/2] Update parsec submodule tracker for a version with
 per-tp-stats

---
 parsec | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/parsec b/parsec
index e718c6cf..adabbd4d 160000
--- a/parsec
+++ b/parsec
@@ -1 +1 @@
-Subproject commit e718c6cfd0a6d7f5c8431cba305a3fc138200585
+Subproject commit adabbd4d1fb580358a32d489df19fa9c05a316e1