Fix: initialise dep_gen on the orchestrator, not behind the handshake

src/a2a3/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp

@@ -695,9 +695,20 @@ int32_t AicpuExecutor::run(Runtime *runtime) {
 
 #if SIMPLER_DFX
             // dep_gen plugs into the orchestrator thread (single-instance subsystem):
-            // record the per-thread ready_queue index before any submit_task fires
-            // inside orch_func_.
+            // resolve its buffer state and record the per-thread ready_queue index
+            // before any submit_task fires inside orch_func_. The init belongs to
+            // this thread, not to the scheduler cold path: dep_gen needs nothing
+            // from the AICore handshake, while the orchestrator skips that
+            // handshake entirely on the decoupled path and starts submitting
+            // immediately. Initialising it behind the handshake makes the first
+            // submits race a barrier they never joined — and the wider the device,
+            // the longer that handshake, so the orchestrator wins more of the race
+            // the more clusters there are.
+            //
+            // The free_queue is SPSC: this must remain the only site that pops
+            // from it on the device side.
             if (is_dep_gen_enabled()) {
+                dep_gen_aicpu_init();
                 dep_gen_aicpu_set_orch_thread_idx(thread_idx);
             }
 

src/a2a3/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp

@@ -981,9 +981,7 @@ void SchedulerContext::post_handshake_profiling_init() {
         pmu_aicpu_init(physical_core_ids_, cores_total_num_);
         LOG_INFO_V0("PMU profiling started on %d cores", cores_total_num_);
     }
-    if (is_dep_gen_enabled()) {
-        dep_gen_aicpu_init();
-    }
+    if (is_dep_gen_enabled()) {}
 #endif
 }
 
@@ -1224,9 +1222,7 @@ int32_t SchedulerContext::post_handshake_init(Runtime *runtime) {
     // init() only pops the initial buffer from instance 0's free_queue; the
     // orchestrator thread still records its idx via
     // dep_gen_aicpu_set_orch_thread_idx() before the first record_submit.
-    if (is_dep_gen_enabled()) {
-        dep_gen_aicpu_init();
-    }
+    if (is_dep_gen_enabled()) {}
 #endif
 
     // total_tasks_ is read in pre_handshake_init (before the orchestrator's early

src/a5/runtime/tensormap_and_ringbuffer/aicpu/aicpu_executor.cpp

@@ -696,9 +696,20 @@ int32_t AicpuExecutor::run(Runtime *runtime) {
             scope_stats_aicpu_set_orch_thread_idx(thread_idx);
 
             // dep_gen plugs into the orchestrator thread (single-instance subsystem):
-            // record the per-thread ready_queue index before any submit_task fires
-            // inside orch_func_.
+            // resolve its buffer state and record the per-thread ready_queue index
+            // before any submit_task fires inside orch_func_. The init belongs to
+            // this thread, not to the scheduler cold path: dep_gen needs nothing
+            // from the AICore handshake, while the orchestrator skips that
+            // handshake entirely on the decoupled path and starts submitting
+            // immediately. Initialising it behind the handshake makes the first
+            // submits race a barrier they never joined — and the wider the device,
+            // the longer that handshake, so the orchestrator wins more of the race
+            // the more clusters there are.
+            //
+            // The free_queue is SPSC: this must remain the only site that pops
+            // from it on the device side.
             if (is_dep_gen_enabled()) {
+                dep_gen_aicpu_init();
                 dep_gen_aicpu_set_orch_thread_idx(thread_idx);
             }
 #endif

src/a5/runtime/tensormap_and_ringbuffer/runtime/scheduler/scheduler_cold_path.cpp

@@ -973,9 +973,7 @@ void SchedulerContext::post_handshake_profiling_init() {
         pmu_aicpu_init(physical_core_ids_, cores_total_num_);
         LOG_INFO_V0("PMU profiling started on %d cores", cores_total_num_);
     }
-    if (is_dep_gen_enabled()) {
-        dep_gen_aicpu_init();
-    }
+    if (is_dep_gen_enabled()) {}
 #endif
 }
 
@@ -1220,9 +1218,7 @@ int32_t SchedulerContext::post_handshake_init(Runtime *runtime) {
     // init() only pops the initial buffer from instance 0's free_queue; the
     // orchestrator thread still records its idx via
     // dep_gen_aicpu_set_orch_thread_idx() before the first record_submit.
-    if (is_dep_gen_enabled()) {
-        dep_gen_aicpu_init();
-    }
+    if (is_dep_gen_enabled()) {}
 #endif
 
     // total_tasks_ is read in pre_handshake_init (before the orchestrator's early

src/common/platform/shared/aicpu/dep_gen_collector_aicpu.cpp

@@ -42,6 +42,12 @@ static bool g_enable_dep_gen = false;
 static DepGenDataHeader *s_dep_gen_header = nullptr;
 static DepGenBufferState *s_dep_gen_state = nullptr;
 static int s_orch_thread_idx = -1;  // set via dep_gen_aicpu_set_orch_thread_idx
+// A record arriving before dep_gen_aicpu_init() cannot be accounted: the
+// counter reconcile_counters() reads lives in the very BufferState that is
+// still null, so the host would see total == collected == dropped == 0 and
+// call an empty deps.json clean. One-shot so a per-task path never floods the
+// AICPU log (see .claude/rules/codestyle.md).
+static bool s_pre_init_drop_reported = false;
 
 static constexpr uint64_t kDepGenQueueBackpressureWaitCycles = PLATFORM_DFX_BACKPRESSURE_TIMEOUT_CYCLES;
 
@@ -133,6 +139,7 @@ void dep_gen_aicpu_init() {
         LOG_ERROR("dep_gen_aicpu_init: dep_gen_data_base is NULL");
         return;
     }
+    s_pre_init_drop_reported = false;
     s_dep_gen_header = get_dep_gen_header(base);
     s_dep_gen_state = get_dep_gen_buffer_state(base, /*instance_index=*/0);
 
@@ -156,7 +163,17 @@ void dep_gen_aicpu_record_submit(
     const uint8_t *arg_types, int explicit_dep_count, const uint64_t *explicit_deps_raw, int block_num,
     const int32_t kernel_ids[3]
 ) {
-    if (!g_enable_dep_gen || s_dep_gen_state == nullptr) {
+    if (!g_enable_dep_gen) {
+        return;
+    }
+    if (s_dep_gen_state == nullptr) {
+        if (!s_pre_init_drop_reported) {
+            s_pre_init_drop_reported = true;
+            LOG_ERROR(
+                "dep_gen: submit recorded before dep_gen_aicpu_init(); this and any further "
+                "pre-init records are lost and cannot be reconciled on the host"
+            );
+        }
         return;
     }