Fix profiler to report time spent on GPU kernels again instead of on 'wait for parallel tasks'.

src/Profiling.cpp

@@ -133,6 +133,12 @@ class InjectProfiling : public IRMutator {
                             unconditionally_set_current_func(stack.back())});
     }
 
+    Stmt suspend_thread_but_keep_task_id(const Stmt &s) {
+        return Block::make({decr_active_threads(profiler_instance),
+                            s,
+                            incr_active_threads(profiler_instance)});
+    }
+
 private:
     using IRMutator::visit;
 
@@ -499,7 +505,11 @@ class InjectProfiling : public IRMutator {
         Stmt stmt = For::make(op->name, op->min, op->extent, op->for_type, op->partition_policy, op->device_api, body);
 
         if (update_active_threads) {
-            stmt = suspend_thread(stmt);
+            if (Internal::is_gpu(op->for_type)) {
+                stmt = suspend_thread_but_keep_task_id(stmt);
+            } else {
+                stmt = suspend_thread(stmt);
+            }
         }
 
         return stmt;

src/runtime/HalideRuntime.h

@@ -910,7 +910,8 @@ extern int halide_device_sync(void *user_context, struct halide_buffer_t *buf);
  * without specifying any buffer to synchronize on.
  * Calling this with a null device_interface is always illegal.
  */
-extern int halide_device_sync_global(void *user_context, const struct halide_device_interface_t *device_interface);
+extern int halide_device_sync_global(void *user_context,
+                                     const struct halide_device_interface_t *device_interface);
 
 /** Allocate device memory to back a halide_buffer_t. */
 extern int halide_device_malloc(void *user_context, struct halide_buffer_t *buf,

src/runtime/device_interface.cpp

@@ -237,11 +237,12 @@ WEAK int halide_device_sync(void *user_context, struct halide_buffer_t *buf) {
  * This variation of the synchronizing is useful when a synchronization is desirable
  * without specifying any buffer to synchronize on.
  */
-WEAK int halide_device_sync_global(void *user_context, const struct halide_device_interface_t *device_interface) {
+WEAK int halide_device_sync_global(void *user_context,
+                                   const struct halide_device_interface_t *device_interface) {
     if (device_interface == nullptr) {
         return halide_error_code_no_device_interface;
     }
-    // This function calls immediately the device_interface implementation to syncrhonize on
+    // This function calls immediately the device_interface implementation to synchronize on
     //  "no buffer" (i.e., nullptr buffer) to trigger a "global" device sync.
     return device_interface->impl->device_sync(user_context, nullptr);
 }

src/runtime/profiler_common.cpp

@@ -517,9 +517,10 @@ WEAK void halide_profiler_report_unlocked(void *user_context, halide_profiler_st
             for (int i = 0; i < p->num_funcs; i++) {
                 halide_profiler_func_stats *fs = p->funcs + i;
 
-                // The first id is always a catch-all overhead
-                // slot. Only report overhead time if it's non-zero
-                if (i == 0 && fs->time == 0) {
+                // The first id is always a catch-all overhead slot (notably containing the asserts).
+                // The second id is always the "wait for parallel tasks" slot.
+                // Only report these time if it's non-zero
+                if ((i == 0 || i == 1) && fs->time == 0) {
                     continue;
                 }