[Performance] Wireless (Polling-based) task readiness detection brings 15~40% speedup

.github/workflows/sanitizers.yml

@@ -11,8 +11,8 @@ name: Sanitizers
 # parallelism-limited subset to dodge the sim-oversubscription livelock; see the
 # run step. detect_leaks=0 until LSan suppressions exist for the device arenas.
 on:
-  schedule:
-    - cron: "0 18 * * *"  # 02:00 Beijing
+  pull_request:
+    branches: [main]
 
 concurrency:
   group: sanitizers-${{ github.ref }}

examples/a2a3/tensormap_and_ringbuffer/paged_attention/kernels/orchestration/paged_attention_orch.cpp

@@ -106,8 +106,6 @@ __attribute__((visibility("default"))) void aicpu_orchestration_entry(const Chip
     uint64_t q_loop = (q_head_num + q_tile - 1) / q_tile;
     CYCLE_COUNT_LAP(prof_param_extract);
 
-    LOG_INFO_V9(">>>>>> batch = %" PRIu64, batch);
-
     // Reshape tensors for kernel consumption (2D flattened)
     void *query_ptr = orch_args.tensor(0).data_as<void>();
     void *kc_ptr = orch_args.tensor(1).data_as<void>();
@@ -251,43 +249,6 @@ __attribute__((visibility("default"))) void aicpu_orchestration_entry(const Chip
             CYCLE_COUNT_LAP(prof_scope);
         }
     }
-
-#ifdef ENABLE_PROFILING
-    uint64_t total = prof_param_extract + prof_ext_tensor + prof_make_tensor + prof_tensor_view + prof_param_setup +
-                     prof_submit_task + prof_scope;
-    LOG_INFO_V9(
-        "=== PagedAttn Orch Profiling: %d submits, %d makes, %d views, total=%.3fus ===", prof_submit_count,
-        prof_make_count, prof_view_count, cycles_to_us(total)
-    );
-    if (total > 0) {
-        LOG_INFO_V9(
-            "  param_extract    : %7.3fus (%5.1f%%)", cycles_to_us(prof_param_extract),
-            prof_param_extract * 100.0 / total
-        );
-        LOG_INFO_V9(
-            "  ext_tensor(x4)   : %7.3fus (%5.1f%%)", cycles_to_us(prof_ext_tensor), prof_ext_tensor * 100.0 / total
-        );
-        LOG_INFO_V9(
-            "  create_info(x%d) : %7.3fus (%5.1f%%)  avg=%.3fus", prof_make_count, cycles_to_us(prof_make_tensor),
-            prof_make_tensor * 100.0 / total,
-            prof_make_count > 0 ? cycles_to_us(prof_make_tensor) / prof_make_count : 0.0
-        );
-        LOG_INFO_V9(
-            "  tensor_view(x%d) : %7.3fus (%5.1f%%)  avg=%.3fus", prof_view_count, cycles_to_us(prof_tensor_view),
-            prof_tensor_view * 100.0 / total,
-            prof_view_count > 0 ? cycles_to_us(prof_tensor_view) / prof_view_count : 0.0
-        );
-        LOG_INFO_V9(
-            "  param_setup      : %7.3fus (%5.1f%%)", cycles_to_us(prof_param_setup), prof_param_setup * 100.0 / total
-        );
-        LOG_INFO_V9("  scope            : %7.3fus (%5.1f%%)", cycles_to_us(prof_scope), prof_scope * 100.0 / total);
-        LOG_INFO_V9(
-            "  submit_task(x%d) : %7.3fus (%5.1f%%)  avg=%.3fus", prof_submit_count, cycles_to_us(prof_submit_task),
-            prof_submit_task * 100.0 / total,
-            prof_submit_count > 0 ? cycles_to_us(prof_submit_task) / prof_submit_count : 0.0
-        );
-    }
-#endif
 }
 
 }  // extern "C"

examples/a2a3/tensormap_and_ringbuffer/paged_attention/test_paged_attention.py

@@ -108,6 +108,22 @@ class TestPagedAttention(SceneTestCase):
                 "dtype": "bfloat16",
             },
         },
+        {
+            "name": "Case4",
+            "platforms": ["a2a3"],
+            "config": {"aicpu_thread_num": 4, "block_dim": 24},
+            "manual": True,
+            "params": {
+                "batch": 16,
+                "num_heads": 16,
+                "kv_head_num": 1,
+                "head_dim": 16,
+                "block_size": 16,
+                "context_len": 256,
+                "max_model_len": 2048,
+                "dtype": "bfloat16",
+            },
+        },
         {
             "name": "CaseSmall1",
             "platforms": ["a2a3sim", "a2a3"],

src/a2a3/runtime/host_build_graph/runtime/pto_runtime2_types.h

@@ -20,7 +20,11 @@
 
 // Tensor dump uses these defaults to size its selective mask table so task-id
 // ring/slot lookup stays aligned with PTO2 task id layout.
+#ifndef PTO2_TASK_WINDOW_SIZE
 #define PTO2_TASK_WINDOW_SIZE 16384  // Default per-ring task window size (power of 2)
+#endif
+#ifndef PTO2_MAX_RING_DEPTH
 #define PTO2_MAX_RING_DEPTH 4        // Number of task-id ring layers
+#endif
 
 #endif  // SRC_A2A3_RUNTIME_HOST_BUILD_GRAPH_RUNTIME_PTO_RUNTIME2_TYPES_H_