diff --git a/runtime/src/kmp_gsupport.cpp b/runtime/src/kmp_gsupport.cpp
index dc52957018b282d8e9e3e0dcca3d2b5a9dd308eb..218957f8823b4521769135adf244503f4f298175 100644
--- a/runtime/src/kmp_gsupport.cpp
+++ b/runtime/src/kmp_gsupport.cpp
@@ -891,15 +891,12 @@ xexpand(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data, void (*copy_fu
ndeps = 0;
/* force allocation of all dependencies in task */
+ /* extra deps are allocated in taskdata struct if depsinalloc is set and VARDEP configured */
input_flags->depsinalloc = 1;
kmp_task_t *task = __kmp_task_alloc(&loc, gtid, input_flags,
sizeof(kmp_task_t), arg_size ? arg_size + arg_align - 1 : 0,
(kmp_routine_entry_t)func,
-#if LIBOMP_USE_VARDEP
- ndeps+thread->th.th_edps_size[0], thread->th.th_edps_size[1]
-#else
ndeps, 0
-#endif
);
kmp_taskdata_t *taskdata = KMP_TASK_TO_TASKDATA(task);
ndeps = taskdata->td_ndeps + taskdata->td_ndeps_noalias;
@@ -966,21 +963,17 @@ xexpand(KMP_API_NAME_GOMP_TASK)(void (*func)(void *), void *data, void (*copy_fu
dep_list[i].flags.alias = 0;
}
}
- /* consider that gcc does not fill ALL dependencies fields */
-/* BUG HERE: gcc move extra dep to taskdata->deps and omp_task_with_deps recompute the total
- number of deps using also extra deps stored in threaddata.
-*/
-#error "STOP"
- taskdata->td_flags.depsfill= 0;
+ /* consider that gcc has filled ALL dependencies fields (excluding extra deps) */
+ taskdata->td_flags.depsfill= 1;
if (if_cond)
__kmpc_omp_task_with_deps( &loc, gtid, task,
- taskdata->td_ndeps, taskdata->td_deps,
- taskdata->td_ndeps_noalias, taskdata->td_deps_noalias
+ depend == 0 ? 0 : (kmp_int32)(uintptr_t)depend[0], taskdata->td_deps,
+ 0, taskdata->td_deps_noalias
);
else {
__kmpc_omp_wait_deps( &loc, gtid,
- taskdata->td_ndeps, taskdata->td_deps,
- taskdata->td_ndeps_noalias, taskdata->td_deps_noalias
+ depend == 0 ? 0 : (kmp_int32)(uintptr_t)depend[0], taskdata->td_deps,
+ 0, taskdata->td_deps_noalias
);
#if OMPT_SUPPORT
ompt_thread_info_t oldInfo;
diff --git a/runtime/src/kmp_taskdeps.cpp b/runtime/src/kmp_taskdeps.cpp
index a6ad799cfa4c7ee66ae636ac940e802906f8e60b..95c0c38c309dd10ccb4766053653ee2d9476fa49 100644
--- a/runtime/src/kmp_taskdeps.cpp
+++ b/runtime/src/kmp_taskdeps.cpp
@@ -19,6 +19,7 @@
#include "kmp_io.h"
#include "kmp_wait_release.h"
+
#if LIBOMP_USE_PARALLEL_SPAWN
#include "kmp_atomic.h"
#endif
@@ -489,7 +490,7 @@ __kmp_process_deps(kmp_int32 gtid, kmp_depnode_t *node, kmp_dephash_t *hash,
KMP_DEBUG_ASSERT(dep->flags.in);
if (filter &&
-#if LIBOMP_USE_VARDEP
+#if LIBOMP_USE_VARDEP && OMP_40_ENABLED
dep->flags.alias == 1
#else
dep->base_addr == 0
@@ -668,7 +669,7 @@ static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
// Filter deps in dep_list
// TODO: Different algorithm for large dep_list ( > 10 ? )
for ( i = 0; i < ndeps; i ++ ) {
-#if LIBOMP_USE_VARDEP
+#if LIBOMP_USE_VARDEP && OMP_40_ENABLED
if (dep_list[i].flags.alias == 0)
{
#if LIBOMP_USE_CONCURRENT_WRITE
@@ -705,7 +706,7 @@ static bool __kmp_check_deps(kmp_int32 gtid, kmp_depnode_t *node,
dep_list[j].base_addr = 0; // Mark j element as void
}
}
-#endif // #if LIBOMP_USE_VARDEP
+#endif // #if LIBOMP_USE_VARDEP && OMP_40_ENABLED
}
#if LIBOMP_USE_CONCURRENT_WRITE
thread->th.th_commute_addr = 0;
@@ -1016,6 +1017,9 @@ kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
// );
#if LIBOMP_USE_VARDEP
+ /* - all deps including extra deps are allocated into task data struct if depsinalloc is set */
+ /* - ndeps and ndeps_noalias does not take into account extra deps */
+ /* - if depsfill == 1 (i.e. gcc) then deps are already copied into taskdata->td_deps array */
kmp_int32 ndeps_extra = thread->th.th_edps_size[0];
kmp_int32 ndeps_extra_noalias = thread->th.th_edps_size[1];
kmp_int32 total_ndeps = ndeps + ndeps_extra;
@@ -1024,24 +1028,27 @@ kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
kmp_int32 total_ndeps = ndeps;
kmp_int32 total_ndeps_noalias = ndeps_noalias;
#endif
- if (new_taskdata->td_flags.depsfill ==0) /* gcc already have stored dependencies */
+ /* allocate here if not done or not enough (+extra deps) in task_alloc */
+ if (new_taskdata->td_ndeps+new_taskdata->td_ndeps_noalias < total_ndeps+total_ndeps_noalias)
{
- /* allocate here if not done or not enough (+extra deps) in task_alloc */
- if (new_taskdata->td_ndeps+new_taskdata->td_ndeps_noalias < total_ndeps+total_ndeps_noalias)
- {
- kmp_depend_info_t* td_deps = (kmp_depend_info_t*)__kmp_fast_allocate( thread,
- (total_ndeps+total_ndeps_noalias)*sizeof(kmp_depend_info_t)
- );
- new_taskdata->td_flags.depsinalloc = 0;
- if (new_taskdata->td_deps)
- __kmp_fast_free(thread, new_taskdata->td_deps);
- // do not free td_deps_noalias
- new_taskdata->td_deps = td_deps;
- }
- new_taskdata->td_deps_noalias = new_taskdata->td_deps+total_ndeps;
+ kmp_depend_info_t* td_deps = (kmp_depend_info_t*)__kmp_fast_allocate( thread,
+ (total_ndeps + total_ndeps_noalias)*sizeof(kmp_depend_info_t)
+ );
+ new_taskdata->td_flags.depsinalloc = 0;
+ if (new_taskdata->td_flags.depsfill)
+ KMP_MEMCPY( td_deps, new_taskdata->td_deps, ndeps * sizeof(kmp_depend_info_t) );
+ if (new_taskdata->td_deps)
+ __kmp_fast_free(thread, new_taskdata->td_deps);
+ // do not free td_deps_noalias, always in same allocated bloc
+ new_taskdata->td_deps = td_deps;
+ new_taskdata->td_deps_noalias = td_deps+total_ndeps;
+ }
- kmp_int32 i;
- kmp_depend_info_t* deps = new_taskdata->td_deps;
+ kmp_int32 i;
+ kmp_depend_info_t* deps = new_taskdata->td_deps;
+ if (new_taskdata->td_flags.depsfill ==0) /* gcc already have stored dependencies */
+ {
+ /* if not allocated */
if (deps != dep_list)
for (i=0; i<ndeps; ++i) {
dep_list[i].flags.alias = 0;
@@ -1049,62 +1056,66 @@ kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
}
else
deps += ndeps;
+ }
#if LIBOMP_USE_VARDEP
- kmp_extra_depinfo_th_t* bloc = &thread->th.th_edeps[0];
- do {
- if (bloc->ed_size) {
- kmp_depend_info_flags_t flags;
- flags.in = 1; //(bloc->ed_mode & OMPEXT_MODE_READ ? 1: 0);
- flags.out= (bloc->ed_mode & OMPEXT_MODE_WRITE ? 1: 0);
- flags.cw = (bloc->ed_mode & OMPEXT_MODE_CONCURRENT ? 1: 0);
- flags.commute = 0;
- flags.alias = 0;
- for (i=0; i<bloc->ed_size; ++i, ++deps)
- {
- deps->base_addr = bloc->ed_deps[i];
- deps->len = 1;
- deps->flags = flags;
- }
+ /* deps with possibly alias */
+ kmp_extra_depinfo_th_t* bloc = &thread->th.th_edeps[0];
+ do {
+ if (bloc->ed_size) {
+ kmp_depend_info_flags_t flags;
+ flags.in = 1; //(bloc->ed_mode & OMPEXT_MODE_READ ? 1: 0);
+ flags.out= (bloc->ed_mode & OMPEXT_MODE_WRITE ? 1: 0);
+ flags.cw = (bloc->ed_mode & OMPEXT_MODE_CONCURRENT ? 1: 0);
+ flags.commute = 0;
+ flags.alias = 0;
+ for (i=0; i<bloc->ed_size; ++i, ++deps)
+ {
+ deps->base_addr = bloc->ed_deps[i];
+ deps->len = 1;
+ deps->flags = flags;
}
- kmp_extra_depinfo_th_t* blocn = bloc->ed_next;
- if (bloc != &thread->th.th_edeps[0])
- __kmp_fast_free(thread, bloc );
- bloc = blocn;
- } while (bloc != 0);
-#endif
- KMP_ASSERT( deps == new_taskdata->td_deps_noalias );
- for (i=0; i<ndeps_noalias; ++i) {
- noalias_dep_list[i].flags.alias = 0;
- *deps++ = noalias_dep_list[i];
}
+ kmp_extra_depinfo_th_t* blocn = bloc->ed_next;
+ if (bloc != &thread->th.th_edeps[0])
+ __kmp_fast_free(thread, bloc );
+ bloc = blocn;
+ } while (bloc != 0);
+#endif
+ KMP_ASSERT( deps == new_taskdata->td_deps_noalias || (total_ndeps_noalias==0) );
+ for (i=0; i<ndeps_noalias; ++i) {
+ noalias_dep_list[i].flags.alias = 0;
+ *deps++ = noalias_dep_list[i];
+ }
#if LIBOMP_USE_VARDEP
- bloc = &thread->th.th_edeps[1];
- do {
- if (bloc->ed_size) {
- kmp_depend_info_flags_t flags;
- flags.in = 1; //(bloc->ed_mode & OMPEXT_MODE_READ ? 1: 0);
- flags.out= (bloc->ed_mode & OMPEXT_MODE_WRITE ? 1: 0);
- flags.cw = (bloc->ed_mode & OMPEXT_MODE_CONCURRENT ? 1: 0);
- flags.commute = 0;
- flags.alias = 0;
- for (i=0; i<bloc->ed_size; ++i, ++deps)
- {
- deps->base_addr = bloc->ed_deps[i];
- deps->len = 1;
- deps->flags = flags;
- }
+ bloc = &thread->th.th_edeps[1];
+ do {
+ if (bloc->ed_size) {
+ kmp_depend_info_flags_t flags;
+ flags.in = 1; //(bloc->ed_mode & OMPEXT_MODE_READ ? 1: 0);
+ flags.out= (bloc->ed_mode & OMPEXT_MODE_WRITE ? 1: 0);
+ flags.cw = (bloc->ed_mode & OMPEXT_MODE_CONCURRENT ? 1: 0);
+ flags.commute = 0;
+ flags.alias = 0;
+ for (i=0; i<bloc->ed_size; ++i, ++deps)
+ {
+ deps->base_addr = bloc->ed_deps[i];
+ deps->len = 1;
+ deps->flags = flags;
}
- kmp_extra_depinfo_th_t* blocn = bloc->ed_next;
- if (bloc != &thread->th.th_edeps[1])
- __kmp_fast_free(thread, bloc );
- bloc = blocn;
- } while (bloc != 0);
- thread->th.th_edps_size[0] =0;
- thread->th.th_edps_size[1] =0;
- thread->th.th_edeps_tail[0] =0;
- thread->th.th_edeps_tail[1] =0;
-#endif
- }
+ }
+ kmp_extra_depinfo_th_t* blocn = bloc->ed_next;
+ if (bloc != &thread->th.th_edeps[1])
+ __kmp_fast_free(thread, bloc );
+ bloc = blocn;
+ } while (bloc != 0);
+ /* reset thread state about extra deps */
+ thread->th.th_edps_size[0] =0;
+ thread->th.th_edps_size[1] =0;
+ thread->th.th_edeps_tail[0] =0;
+ thread->th.th_edeps_tail[1] =0;
+#endif
+
+//kmp_depend_info_t *save_dep_list = dep_list;
dep_list = new_taskdata->td_deps;
ndeps = new_taskdata->td_ndeps = total_ndeps;
noalias_dep_list = new_taskdata->td_deps_noalias;
@@ -1170,8 +1181,22 @@ kmp_int32 __kmpc_omp_task_with_deps(ident_t *loc_ref, kmp_int32 gtid,
"dependencies: "
"loc=%p task=%p, return: TASK_CURRENT_NOT_QUEUED\n",
gtid, loc_ref, new_taskdata));
+#if 0
+ printf("%i:: Task %p #deps=%i @dep=%p/%p @=%p/%p has dependencies.\n", gtid, new_task, ndeps, dep_list, save_dep_list,
+ (ndeps >0 ? dep_list[0].base_addr : 0),
+ (ndeps >0 ? save_dep_list[0].base_addr : 0)
+ );
+#endif
return TASK_CURRENT_NOT_QUEUED;
}
+ else {
+#if 0
+ printf("%i:: Task %p #deps=%i @dep=%p/%p @=%p/%p is independent.\n", gtid, new_task, ndeps, dep_list, save_dep_list,
+ (ndeps >0 ? dep_list[0].base_addr : 0),
+ (ndeps >0 ? save_dep_list[0].base_addr : 0)
+ );
+#endif
+ }
#if LIBOMP_USE_CONCURRENT_WRITE
/* task ready: check to acquired commute ressources */
if ((node->dn./*list_*/cw) && __kmp_acquired_or_failed( gtid, node, node->dn./*list_*/cw) )
diff --git a/runtime/src/kmp_tasking.cpp b/runtime/src/kmp_tasking.cpp
index 804c19d6dc8ecaeb9a3bdf483d2fcdce758f0116..8d9a588d0cf9b8fc347549288170e8ccbdf27f8d 100644
--- a/runtime/src/kmp_tasking.cpp
+++ b/runtime/src/kmp_tasking.cpp
@@ -1115,10 +1115,13 @@ static size_t __kmp_round_up_to_val(size_t size, size_t val) {
// sizeof_shareds: Size in bytes of array of pointers to shared vars accessed
// in task.
// task_entry: Pointer to task code entry point generated by compiler.
+// ndeps,ndeps_noalias: size of deps + deps without alias to be stored in taskdata struct.
+// only to be considered if flag depsinalloc is set.
// returns: a pointer to the allocated kmp_task_t structure (task).
+// Not that variable length deps are allocated by task_alloc if flag depsinalloc is set
kmp_task_t* __kmp_task_alloc( ident_t *loc_ref, kmp_int32 gtid, kmp_tasking_flags_t *flags,
size_t sizeof_kmp_task_t, size_t sizeof_shareds,
- kmp_routine_entry_t task_entry, kmp_int32 deps, kmp_int32 deps_noalias )
+ kmp_routine_entry_t task_entry, kmp_int32 ndeps, kmp_int32 ndeps_noalias )
{
kmp_task_t *task;
kmp_taskdata_t *taskdata;
@@ -1128,9 +1131,9 @@ kmp_task_t* __kmp_task_alloc( ident_t *loc_ref, kmp_int32 gtid, kmp_tasking_flag
size_t shareds_offset;
KA_TRACE(10, ("__kmp_task_alloc(enter): T#%d loc=%p, flags=(0x%x) "
- "sizeof_task=%ld sizeof_shared=%ld entry=%p, deps=%d deps_noalias=%d\n",
+ "sizeof_task=%ld sizeof_shared=%ld entry=%p, ndeps=%d ndeps_noalias=%d\n",
gtid, loc_ref, *((kmp_int32 *)flags), sizeof_kmp_task_t,
- sizeof_shareds, task_entry, deps, deps_noalias) );
+ sizeof_shareds, task_entry, ndeps, ndeps_noalias) );
if ( parent_task->td_flags.final ) {
if (flags->merged_if0) {
@@ -1179,9 +1182,15 @@ kmp_task_t* __kmp_task_alloc( ident_t *loc_ref, kmp_int32 gtid, kmp_tasking_flag
#endif
#if OMP_40_ENABLED
- int ndeps;
if (flags->depsinalloc)
- ndeps = deps + deps_noalias;
+ {
+ ndeps = ndeps + ndeps_noalias;
+#if LIBOMP_USE_VARDEP
+ kmp_int32 ndeps_extra = thread->th.th_edps_size[0];
+ kmp_int32 ndeps_extra_noalias = thread->th.th_edps_size[1];
+ ndeps += ndeps_extra + ndeps_extra_noalias;
+#endif
+ }
else
#endif
ndeps = 0;
@@ -1230,11 +1239,11 @@ kmp_task_t* __kmp_task_alloc( ident_t *loc_ref, kmp_int32 gtid, kmp_tasking_flag
{
//taskdata->td_deps = (kmp_depend_info_t*)& ((char *) taskdata)[ shareds_offset - ndeps*sizeof(kmp_depend_info_t) ];
taskdata->td_deps = (kmp_depend_info_t*)(task+1);
- taskdata->td_ndeps = deps;
- taskdata->td_deps_noalias = taskdata->td_deps+deps;
- taskdata->td_ndeps_noalias = deps_noalias;
+ taskdata->td_ndeps = ndeps;
+ taskdata->td_deps_noalias = taskdata->td_deps+ndeps;
+ taskdata->td_ndeps_noalias = ndeps_noalias;
taskdata->td_flags.depsinalloc = 1;
- KMP_DEBUG_ASSERT( (void*)(taskdata->td_deps_noalias+deps_noalias) <= (void*)task->shareds || task->shareds == NULL);
+ KMP_DEBUG_ASSERT( (void*)(taskdata->td_deps_noalias+ndeps_noalias) <= (void*)task->shareds || task->shareds == NULL);
}
else {
taskdata->td_deps = 0;