Merge branch 'develop' of github.com:khuck/xpress-apex into develop

UO-OACISS · May 13, 2022 · 022f162 · 022f162
2 parents 3e37861 + 5cc7fac
commit 022f162
Show file tree

Hide file tree

Showing 4 changed files with 10 additions and 4 deletions.
diff --git a/src/apex/apex_kokkos.cpp b/src/apex/apex_kokkos.cpp
@@ -96,11 +96,17 @@ inline ExecutionSpaceIdentifier identifier_from_devid(const uint32_t in) {
           ((in & 0x00FFFFFF) >> 17),  // next 7 bits
            (in & 0x0001FFFF)}; // last 17 bits
            */
+#if 0
     constexpr const uint32_t shift = num_avail_bits - num_type_bits;
     return {devicetype_from_uint32t(in >> shift), /*First 8 bits*/
            (~((uint32_t(-1)) << num_device_bits)) &
             (in >> num_instance_bits), /*Next 7 bits */
            (~((uint32_t(-1)) << num_instance_bits)) & in}; /*Last 17 bits*/
+#else
+    constexpr const uint32_t shift = num_instance_bits;
+    return {devicetype_from_uint32t(in >> shift), /*First 8 bits*/
+           0, (~((uint32_t(-1)) << num_instance_bits)) & in}; /*Last 17 bits*/
+#endif
 }
 
 /* "Top 8 bits represent the device type. Next 7 are the device id (think

diff --git a/src/apex/dependency_tree.cpp b/src/apex/dependency_tree.cpp
@@ -73,7 +73,7 @@ void Node::writeNode(std::ofstream& outfile, double total) {
         outfile << std::endl;
     }
 
-    double acc = (data == task_identifier::get_main_task_id()) ?
+    double acc = (data == task_identifier::get_main_task_id() || accumulated == 0.0) ?
         total : accumulated;
     node_color * c = get_node_color_visible(acc, 0.0, total);
     double ncalls = (calls == 0) ? 1 : calls;

diff --git a/src/apex/hip_trace.cpp b/src/apex/hip_trace.cpp
@@ -261,7 +261,7 @@ void handle_roctx(uint32_t domain, uint32_t cid, const void* callback_data, void
         case ROCTX_API_ID_roctxRangePushA:
             {
                 std::stringstream ss;
-                ss << "roctx: " << data->args.message;
+                ss /* << "roctx: " */ << data->args.message;
                 timer_stack.push(apex::start(ss.str()));
                 break;
             }
@@ -274,7 +274,7 @@ void handle_roctx(uint32_t domain, uint32_t cid, const void* callback_data, void
         case ROCTX_API_ID_roctxRangeStartA:
             {
                 std::stringstream ss;
-                ss << "roctx: " << data->args.message;
+                ss /* << "roctx: " */ << data->args.message;
                 apex::profiler* p = apex::start(ss.str());
                 const std::lock_guard<std::mutex> guard(map_lock);
                 timer_map.insert(

diff --git a/src/openmp/ompt_target_matmult.c b/src/openmp/ompt_target_matmult.c
@@ -13,7 +13,7 @@
 #include <stdlib.h>
 
 #ifndef MATRIX_SIZE
-#define MATRIX_SIZE 4096
+#define MATRIX_SIZE 1024
 #endif
 
 #define MAX_ITERATIONS 3