Fix: unify tensor dump control under profiling flags

  - add `enable_profiling_flag` to the AICPU/AICore handshake and
    initialize the dump bit in onboard and sim device runners
  - replace `PTO2_DUMP_TENSOR` guards with `PTO2_PROFILING` and remove
    the old per-runtime dump macro definitions
  - add an AICore pipe barrier before completion when dumping tensors to
    preserve write visibility for dumps

Author: ChaoZheng109

src/a2a3/platform/include/common/kernel_args.h

@@ -72,7 +72,7 @@ struct KernelArgs {
     __may_used_by_aicore__ Runtime *runtime_args{nullptr};  // Task runtime in device memory
     uint64_t regs{0};                                       // Per-core register base address array (platform-specific)
     uint64_t ffts_base_addr{0};                             // FFTS base address for AICore
-    uint64_t dump_data_base{0};                             // Dump shared memory base address, zero when unused
+    uint64_t dump_data_base{0};  // Dump shared memory base address; use explicit flags to detect enablement
 };
 
 #ifdef __cplusplus

src/a2a3/platform/include/common/platform_config.h

@@ -142,6 +142,13 @@ inline double cycles_to_us(uint64_t cycles) {
     return (static_cast<double>(cycles) / PLATFORM_PROF_SYS_CNT_FREQ) * 1000000.0;
 }
 
+// Profiling-related runtime flags shared through AICPU-AICore handshake.
+#define PROFILING_FLAG_NONE 0u
+#define PROFILING_FLAG_DUMP_TENSOR (1u << 0)
+#define GET_PROFILING_FLAG(flags, bit) ((((uint32_t)(flags)) & ((uint32_t)(bit))) != 0u)
+#define SET_PROFILING_FLAG(flags, bit) ((flags) |= (uint32_t)(bit))
+#define CLEAR_PROFILING_FLAG(flags, bit) ((flags) &= ~((uint32_t)(bit)))
+
 // =============================================================================
 // Tensor Dump Configuration
 // =============================================================================

src/a2a3/platform/include/common/tensor_dump.h

@@ -17,7 +17,7 @@
  * Fully decoupled from profiling — uses its own ready queues, buffer states,
  * and memory manager thread.
  *
- * Memory layout (Dump SHM, allocated only when PTO2_DUMP_TENSOR=1):
+ * Memory layout (Dump SHM, allocated only when PTO2_PROFILING=1):
  * ┌─────────────────────────────────────────────────────────────┐
  * │ DumpDataHeader (fixed header)                               │
  * │  - Per-thread ready queues (circular FIFOs)                 │

src/a2a3/platform/onboard/aicpu/kernel.cpp

@@ -80,9 +80,11 @@ extern "C" __attribute__((visibility("default"))) int DynTileFwkBackendKernelSer
     }
 
     // Store platform regs before calling aicpu_execute
+    // Dump enable is an execution control flag propagated via handshake.
+    // The dump base address is only the backing storage location.
     set_platform_regs(k_args->regs);
     set_platform_dump_base(k_args->dump_data_base);
-    set_enable_dump_tensor(k_args->dump_data_base != 0);
+    set_enable_dump_tensor(GET_PROFILING_FLAG(runtime->workers[0].enable_profiling_flag, PROFILING_FLAG_DUMP_TENSOR));
 
     // Affinity gate: drop excess threads before entering runtime
     if (!platform_aicpu_affinity_gate(runtime->sche_cpu_num, PLATFORM_MAX_AICPU_THREADS_JUST_FOR_LAUNCH)) {

src/a2a3/platform/onboard/host/device_runner.cpp

@@ -421,6 +421,7 @@ int DeviceRunner::run(
 
     // Calculate number of AIC cores (1/3 of total)
     int num_aic = block_dim;  // Round up for 1/3
+    const uint32_t enable_profiling_flag = enable_dump_tensor ? PROFILING_FLAG_DUMP_TENSOR : PROFILING_FLAG_NONE;
 
     for (int i = 0; i < num_aicore; i++) {
         runtime.workers[i].aicpu_ready = 0;
@@ -430,6 +431,7 @@ int DeviceRunner::run(
         runtime.workers[i].task_status = 0;
         // Set core type: first 1/3 are AIC, remaining 2/3 are AIV
         runtime.workers[i].core_type = (i < num_aic) ? CoreType::AIC : CoreType::AIV;
+        runtime.workers[i].enable_profiling_flag = enable_profiling_flag;
         runtime.workers[i].perf_records_addr = static_cast<uint64_t>(0);
         runtime.workers[i].perf_buffer_status = 0;
     }

src/a2a3/platform/sim/aicore/inner_kernel.h

@@ -57,6 +57,10 @@ typedef int mem_dsb_t;
 #define SINGLE_CACHE_LINE 0
 #define CACHELINE_OUT 0
 
+// pipe_barrier - memory barrier in simulation (hardware pipeline synchronization)
+#define PIPE_ALL 0
+#define pipe_barrier(pipe) __sync_synchronize()
+
 // SPIN_WAIT_HINT - CPU pause hint + OS yield for idle polling loops in simulation.
 // In simulation, all AICore/AICPU threads share a small number of host CPU cores.
 // The CPU hint (pause/yield) reduces pipeline waste, and sched_yield() lets the OS

src/a2a3/platform/sim/host/device_runner.cpp

@@ -286,6 +286,7 @@ int DeviceRunner::run(
 
     // Calculate number of AIC cores
     int num_aic = block_dim;
+    const uint32_t enable_profiling_flag = enable_dump_tensor ? PROFILING_FLAG_DUMP_TENSOR : PROFILING_FLAG_NONE;
 
     for (int i = 0; i < num_aicore; i++) {
         runtime.workers[i].aicpu_ready = 0;
@@ -295,6 +296,7 @@ int DeviceRunner::run(
         runtime.workers[i].task_status = 0;
         // First 1/3 are AIC, remaining 2/3 are AIV
         runtime.workers[i].core_type = (i < num_aic) ? CoreType::AIC : CoreType::AIV;
+        runtime.workers[i].enable_profiling_flag = enable_profiling_flag;
     }
 
     // Set function_bin_addr for each task: func_id_to_addr_[] stores CoreCallable

src/a2a3/runtime/aicpu_build_graph/aicore/aicore_executor.cpp

@@ -86,6 +86,7 @@ __aicore__ __attribute__((weak)) void aicore_execute(__gm__ Runtime *runtime, in
     __gm__ PTO2DispatchPayload *payload = reinterpret_cast<__gm__ PTO2DispatchPayload *>(my_hank->task);
 
     bool profiling_enabled = runtime->enable_profiling;
+    bool dump_tensor_enabled = GET_PROFILING_FLAG(my_hank->enable_profiling_flag, PROFILING_FLAG_DUMP_TENSOR);
 
     // Phase 4: Main execution loop - poll register for tasks until exit signal
     // Register encoding: AICPU_IDLE_TASK_ID=idle, task_id=task, AICORE_EXIT_SIGNAL=exit
@@ -120,6 +121,10 @@ __aicore__ __attribute__((weak)) void aicore_execute(__gm__ Runtime *runtime, in
             // Execute the task
             execute_task(payload);
 
+            if (dump_tensor_enabled) {
+                pipe_barrier(PIPE_ALL);
+            }
+
             // Performance profiling: record task execution
             // (func_id and core_type are filled by AICPU at completion time)
             if (profiling_enabled) {

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

@@ -406,7 +406,7 @@ struct AicpuExecutor {
                 );
                 cur_thread_completed++;
                 if (mixed_complete) {
-#if PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                     if (get_enable_dump_tensor()) {
                         dump_tensors_for_task<PTO2_SUBTASK_SLOT_COUNT>(
                             thread_idx, slot_state, TensorDumpStage::AFTER_COMPLETION,
@@ -515,7 +515,7 @@ struct AicpuExecutor {
         ,
         bool profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
         ,
         int32_t thread_idx
 #endif
@@ -944,7 +944,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
             perf_aicpu_set_orch_thread_idx(sched_thread_num_);
         }
 #endif
-#if PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
         if (get_enable_dump_tensor()) {
             dump_tensor_init(orch_to_sched_ ? thread_num_ : sched_thread_num_);
         }
@@ -1163,7 +1163,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                     uint64_t t_setup_start = get_sys_cnt_aicpu();
 #endif
                     ResourceCount rc = shape_resource_count(shape);
-#if PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                     if (get_enable_dump_tensor()) {
                         dump_tensors_for_task<PTO2_SUBTASK_SLOT_COUNT>(
                             thread_idx, *slot_state, TensorDumpStage::BEFORE_DISPATCH,
@@ -1183,7 +1183,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                             ,
                             profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                             ,
                             thread_idx
 #endif
@@ -1197,7 +1197,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                             ,
                             profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                             ,
                             thread_idx
 #endif
@@ -1210,7 +1210,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                             ,
                             profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                             ,
                             thread_idx
 #endif
@@ -1272,7 +1272,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
 #endif
                 Cluster &c = tracker.clusters[ci];
                 ResourceCount rc = shape_resource_count(shape);
-#if PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                 if (get_enable_dump_tensor()) {
                     dump_tensors_for_task<PTO2_SUBTASK_SLOT_COUNT>(
                         thread_idx, *slot_state, TensorDumpStage::BEFORE_DISPATCH,
@@ -1292,7 +1292,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                         ,
                         profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                         ,
                         thread_idx
 #endif
@@ -1306,7 +1306,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                         ,
                         profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                         ,
                         thread_idx
 #endif
@@ -1319,7 +1319,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
                         ,
                         profiling_enabled
 #endif
-#if PTO2_PROFILING || PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
                         ,
                         thread_idx
 #endif
@@ -1650,7 +1650,7 @@ int32_t AicpuExecutor::resolve_and_dispatch_pto2(Runtime *runtime, int32_t threa
         perf_aicpu_flush_phase_buffers(thread_idx);
     }
 #endif
-#if PTO2_DUMP_TENSOR
+#if PTO2_PROFILING
     if (get_enable_dump_tensor()) {
         dump_tensor_flush(thread_idx);
     }

src/a2a3/runtime/aicpu_build_graph/runtime/pto_runtime2_types.h

@@ -58,14 +58,6 @@
 #error "PTO2_SCHED_PROFILING requires PTO2_PROFILING=1"
 #endif
 
-// =============================================================================
-// Dump Tensor Configuration
-// =============================================================================
-
-#ifndef PTO2_DUMP_TENSOR
-#define PTO2_DUMP_TENSOR 1
-#endif
-
 // =============================================================================
 // AICPU Error Codes (written to shared memory for Host-side diagnosis)
 // =============================================================================