CMakeLists.txt

@@ -232,6 +232,13 @@ if(CHAMELEON_SCHED_STARPU)
         message("-- ${BoldGreen}CHAMELEON_SIMULATION is set to OFF, turn it ON to use"
             " SIMULATION mode (only with StarPU compiled with SimGrid)${ColourReset}")
     endif()
+    option(CHAMELEON_ENABLE_PRUNING_STATS "Enable pruning statistics" OFF)
+    if (NOT CHAMELEON_ENABLE_PRUNING_STATS)
+        message("-- ${BoldGreen}CHAMELEON_ENABLE_PRUNING_STATS is set to OFF, turn it ON to build pruning statistics${ColourReset}")
+    endif()
+    if(CHAMELEON_ENABLE_PRUNING_STATS)
+        add_definitions(-DCHAMELEON_ENABLE_PRUNING_STATS)
+    endif(CHAMELEON_ENABLE_PRUNING_STATS)
 endif()
 
 # Initially we need to generate files for different precisions

morse_cmake @ 0a974775

-Subproject commit 32c1ba6ae5d7e19d570ac3bd86a24acb82f2081e
+Subproject commit 0a974775b7192227b887dcba77515305083c1f13

compute/zgesvd.c

@@ -521,6 +521,8 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
     }
 
     morse_sequence_wait(morse, sequence);
+
+#if !defined(CHAMELEON_SIMULATION)
     info = LAPACKE_zgbbrd( LAPACK_COL_MAJOR,
                            gbbrd_vect,
                            M, N,
@@ -533,6 +535,7 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
     if (info != 0) {
         fprintf(stderr, "MORSE_zgesvd_Tile_Async: LAPACKE_zgbbrd = %d\n", info );
     }
+#endif /* !defined(CHAMELEON_SIMULATION) */
     morse_desc_mat_free(&descAB);
 
     /* Transform U and Vt into tile format */
@@ -594,6 +597,7 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
 
     /* Solve the bidiagonal SVD problem */
     /* On exit, U and VT are updated with bidiagonal matrix singular vectors */
+#if !defined(CHAMELEON_SIMULATION)
     info = LAPACKE_zbdsqr( LAPACK_COL_MAJOR, 'U',
                            MINMN, ncvt, nru, 0,
                            S, E,
@@ -601,6 +605,7 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
     if (info != 0) {
         fprintf(stderr, "MORSE_zgesvd_Tile_Async: LAPACKE_zbdsqr = %d\n", info );
     }
+#endif /* !defined(CHAMELEON_SIMULATION) */
 
     if (jobu != MorseNoVec)
         morse_desc_mat_free( &descU );

compute/zheevd.c

@@ -420,6 +420,7 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
 
 
     if (jobz == MorseNoVec){
+#if !defined(CHAMELEON_SIMULATION)
         /* Tridiagonal eigensolver */
         status = LAPACKE_zstedc( LAPACK_COL_MAJOR,
                                  morse_lapack_const(jobz),
@@ -428,6 +429,7 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
         if (status != 0) {
             morse_error("MORSE_zheevd_Tile_Async", "LAPACKE_zstedc failed");
         }
+#endif /* !defined(CHAMELEON_SIMULATION) */
         free(E);
         return MORSE_SUCCESS;
     }
@@ -441,8 +443,11 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
     /* For bug in lapacke */
     memset(V, 0, N * N * sizeof(MORSE_Complex64_t));
 
-    /* Tridiagonal eigensolver */
-    /* V contains the eigenvectors of the tridiagonal system on exit */
+    /*
+     * Tridiagonal eigensolver
+     * V contains the eigenvectors of the tridiagonal system on exit
+     */
+#if !defined(CHAMELEON_SIMULATION)
     status = LAPACKE_zstedc( LAPACK_COL_MAJOR,
                              morse_lapack_const(MorseIvec),
                              N, W, E,
@@ -450,6 +455,7 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
     if (status != 0) {
         morse_error("MORSE_zheevd_Tile_Async", "LAPACKE_zstedc failed");
     }
+#endif /* !defined(CHAMELEON_SIMULATION) */
 
     /* Convert matrices in tile format */
     /* A/T from MORSE_zhetrd   refer  to Q1 (tile   layout) */

compute/zhetrd.c

@@ -408,6 +408,7 @@ int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
     morse_sequence_wait(morse, sequence);
 
     /* Reduce band matrix to tridiagonal matrix */
+#if !defined(CHAMELEON_SIMULATION)
     status = LAPACKE_zhbtrd( LAPACK_COL_MAJOR,
                              morse_lapack_const(jobz),
                              morse_lapack_const(uplo),
@@ -417,6 +418,7 @@ int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
     if (status != 0) {
         morse_error("MORSE_zhetrd_Tile_Async", "LAPACKE_zhbtrd failed");
     }
+#endif /* !defined(CHAMELEON_SIMULATION) */
 
     morse_desc_mat_free(&descAB);
     return MORSE_SUCCESS;

include/runtime.h

@@ -126,6 +126,9 @@ double RUNTIME_get_time();
 void RUNTIME_start_profiling();
 void RUNTIME_stop_profiling();
 
+void RUNTIME_start_stats();
+void RUNTIME_stop_stats();
+
 #if defined(PRECISION_z)
 void RUNTIME_zdisplay_allprofile ();
 void RUNTIME_zdisplay_oneprofile (MORSE_kernel_t);

runtime/parsec/control/runtime_profiling.c

@@ -88,6 +88,16 @@ void RUNTIME_stop_profiling()
     morse_warning("RUNTIME_stop_profiling()", "FxT profiling is not available with PaRSEC\n");
 }
 
+void RUNTIME_start_stats()
+{
+    morse_warning("RUNTIME_start_stats()", "pruning stats are not available with PaRSEC\n");
+}
+
+void RUNTIME_stop_stats()
+{
+    morse_warning("RUNTIME_stop_stats()", "pruning stats are not available with PaRSEC\n");
+}
+
 void RUNTIME_schedprofile_display(void)
 {
     morse_warning("RUNTIME_schedprofile_display(parsec)", "Scheduler profiling is not available with PaRSEC\n");

runtime/quark/control/runtime_profiling.c

@@ -103,6 +103,16 @@ void RUNTIME_stop_profiling()
     morse_warning("RUNTIME_stop_profiling()", "FxT profiling is not available with Quark\n");
 }
 
+void RUNTIME_start_stats()
+{
+    morse_warning("RUNTIME_start_stats()", "pruning stats are not available with Quark\n");
+}
+
+void RUNTIME_stop_stats()
+{
+    morse_warning("RUNTIME_stop_stats()", "pruning stats are not available with Quark\n");
+}
+
 void RUNTIME_schedprofile_display(void)
 {
     morse_warning("RUNTIME_schedprofile_display(quark)", "Scheduler profiling is not available with Quark\n");

runtime/starpu/control/runtime_profiling.c

@@ -29,6 +29,12 @@
 #include <starpu_fxt.h>
 #endif
 
+#ifdef CHAMELEON_ENABLE_PRUNING_STATS
+unsigned long RUNTIME_total_tasks;
+unsigned long RUNTIME_exec_tasks;
+unsigned long RUNTIME_comm_tasks;
+unsigned long RUNTIME_changed_tasks;
+#endif
 
 double RUNTIME_get_time(){
     return starpu_timing_now()*1e-6;
@@ -50,6 +56,21 @@ void RUNTIME_stop_profiling(){
 #endif
 }
 
+void RUNTIME_start_stats(){
+#ifdef CHAMELEON_ENABLE_PRUNING_STATS
+    RUNTIME_total_tasks = 0;
+    RUNTIME_exec_tasks = 0;
+    RUNTIME_comm_tasks = 0;
+    RUNTIME_changed_tasks = 0;
+#endif
+}
+
+void RUNTIME_stop_stats(){
+#ifdef CHAMELEON_ENABLE_PRUNING_STATS
+    fprintf(stderr, "\ntasks: %u = exec: %u + comm: %u + changed: %u\n", RUNTIME_total_tasks, RUNTIME_exec_tasks, RUNTIME_comm_tasks, RUNTIME_changed_tasks);
+#endif
+}
+
 void RUNTIME_profiling_display_info(const char *kernel_name, measure_t perf[STARPU_NMAXWORKERS])
 {
     int header = 1;

runtime/starpu/include/morse_starpu.h

@@ -86,16 +86,53 @@ typedef struct starpu_conf starpu_conf_t;
 
 void RUNTIME_set_reduction_methods(starpu_data_handle_t handle, MORSE_enum dtyp);
 
-#define RUNTIME_BEGIN_ACCESS_DECLARATION
+#ifdef CHAMELEON_ENABLE_PRUNING_STATS
+
+#define RUNTIME_PRUNING_STATS_BEGIN_ACCESS_DECLARATION \
+    int __morse_exec = 0; \
+    int __morse_changed = 0;
+
+#define RUNTIME_PRUNING_STATS_ACCESS_W(A, Am, An) \
+    if (morse_desc_islocal(A, Am, An)) \
+        __morse_exec = 1;
+
+#define RUNTIME_PRUNING_STATS_END_ACCESS_DECLARATION \
+    RUNTIME_total_tasks++; \
+    if (__morse_exec) \
+        RUNTIME_exec_tasks++; \
+    else if (__morse_need_submit) \
+        RUNTIME_comm_tasks++; \
+    else if (__morse_changed) \
+        RUNTIME_changed_tasks++;
+
+#define RUNTIME_PRUNING_STATS_RANK_CHANGED(rank) \
+    int __morse_myrank; \
+    RUNTIME_comm_rank(&__morse_myrank); \
+    __morse_exec = (rank) == __morse_myrank; \
+    __morse_changed = 1; \
 
-#define RUNTIME_ACCESS_R(A, Am, An)
+#else
+#define RUNTIME_PRUNING_STATS_BEGIN_ACCESS_DECLARATION
+#define RUNTIME_PRUNING_STATS_ACCESS_W(A, Am, An)
+#define RUNTIME_PRUNING_STATS_END_ACCESS_DECLARATION
+#define RUNTIME_PRUNING_STATS_RANK_CHANGED(rank)
+#endif
+
+#define RUNTIME_BEGIN_ACCESS_DECLARATION \
+    RUNTIME_PRUNING_STATS_BEGIN_ACCESS_DECLARATION
+
+#define RUNTIME_ACCESS_R(A, Am, An) \
 
-#define RUNTIME_ACCESS_W(A, Am, An)
+#define RUNTIME_ACCESS_W(A, Am, An) \
+    RUNTIME_PRUNING_STATS_ACCESS_W(A, Am, An)
 
-#define RUNTIME_ACCESS_RW(A, Am, An)
+#define RUNTIME_ACCESS_RW(A, Am, An) \
+    RUNTIME_PRUNING_STATS_ACCESS_W(A, Am, An)
 
-#define RUNTIME_RANK_CHANGED(rank)
+#define RUNTIME_RANK_CHANGED(rank) \
+    RUNTIME_PRUNING_STATS_RANK_CHANGED(rank)
 
-#define RUNTIME_END_ACCESS_DECLARATION
+#define RUNTIME_END_ACCESS_DECLARATION \
+    RUNTIME_PRUNING_STATS_END_ACCESS_DECLARATION;
 
 #endif /* _MORSE_STARPU_H_ */

runtime/starpu/include/runtime_profiling.h

@@ -26,6 +26,13 @@
 #ifndef _PROFILING_H_
 #define _PROFILING_H_
 
+#ifdef CHAMELEON_ENABLE_PRUNING_STATS
+extern unsigned long RUNTIME_total_tasks;
+extern unsigned long RUNTIME_exec_tasks;
+extern unsigned long RUNTIME_comm_tasks;
+extern unsigned long RUNTIME_changed_tasks;
+#endif
+
 typedef struct measure_s {
     double sum;
     double sum2;

timing/timing.h

@@ -159,6 +159,7 @@ enum dparam_timing {
  *
  */
 #define START_TRACING()                        \
+    RUNTIME_start_stats();                     \
     if(iparam[IPARAM_TRACE] == 2) {            \
     	RUNTIME_start_profiling();             \
     }                                          \
@@ -167,6 +168,7 @@ enum dparam_timing {
     }
 
 #define STOP_TRACING()                         \
+    RUNTIME_stop_stats();                      \
     if(iparam[IPARAM_TRACE] == 2) {            \
     	RUNTIME_stop_profiling();              \
     }                                          \