!17009 [PROF] Profiler support AicMetrics MemoryAccess feature

Merge pull request !17009 from wangjie/memory_access_feature_master

third_party/acl/inc/acl/acl_prof.h

@@ -49,6 +49,7 @@ typedef enum {
     ACL_AICORE_RESOURCE_CONFLICT_RATIO = 4,
     ACL_AICORE_MEMORY_UB = 5,
     ACL_AICORE_L2_CACHE = 6,
+    ACL_AICORE_MEMORY_ACCESS = 8,
     ACL_AICORE_NONE = 0xFF
 } aclprofAicoreMetrics;
 

torch_npu/csrc/profiler/feature_mgr.cpp

@@ -12,11 +12,13 @@ const static char* VERSION = "master\0";
 
 static std::unordered_map<std::string, FeatureType> NAME_TABLE = {
     {"ATTR", FeatureType::FEATURE_ATTR},
+    {"MemoryAccess", FeatureType::FEATURE_MEMORY_ACCESS}
 };
 
 // featureName, featureVersion
 static std::unordered_map<FeatureType, std::string> FMK_FEATURES = {
     {FeatureType::FEATURE_ATTR, "1"},
+    {FeatureType::FEATURE_MEMORY_ACCESS, "1"}
 };
 }
 

torch_npu/csrc/profiler/feature_mgr.h

@@ -17,6 +17,7 @@ namespace profiler {
 enum class FeatureType {
     FEATURE_MIN = 0,
     FEATURE_ATTR,
+    FEATURE_MEMORY_ACCESS,
     FEATURE_MAX,
 };
 using torch_npu::toolkit::profiler::Utils;

torch_npu/csrc/profiler/profiler_mgr.cpp

@@ -23,6 +23,7 @@ std::map<std::string, aclprofAicoreMetrics> ProfilerMgr::npu_metrics_map_ = {
     {"ACL_AICORE_RESOURCE_CONFLICT_RATIO", ACL_AICORE_RESOURCE_CONFLICT_RATIO},
     {"ACL_AICORE_MEMORY_UB", ACL_AICORE_MEMORY_UB},
     {"ACL_AICORE_L2_CACHE", ACL_AICORE_L2_CACHE},
+    {"ACL_AICORE_MEMORY_ACCESS", ACL_AICORE_MEMORY_ACCESS},
     {"ACL_AICORE_NONE", ACL_AICORE_NONE},
 };
 
@@ -34,7 +35,19 @@ std::map<std::string, uint64_t> ProfilerMgr::trace_level_map_ = {
 };
 
 constexpr uint32_t capacity_ = 1048576;          // 2^20, Experience value for default ringbuffer size for single data
-constexpr uint32_t trace_capacity_ = 128;       // 2^7, Experience value for python trace data ringbuffer size for batch data
+constexpr uint32_t trace_capacity_ = 128;        // 2^7, Experience value for python trace data ringbuffer size for batch data
+
+aclprofAicoreMetrics CheckAicMetricsFeature(aclprofAicoreMetrics aic_metrics, int8_t level)
+{
+    if (aic_metrics == ACL_AICORE_MEMORY_ACCESS &&
+        !FeatureMgr::GetInstance()->IsSupportFeature(FeatureType::FEATURE_MEMORY_ACCESS)) {
+        ASCEND_LOGW("AiCMetrics is not supported to set to MemoryAccess.");
+        printf("[WARN]%s,%s:%u:AiCMetrics is not supported to set to MemoryAccess, reset to default.\n",
+               __FUNCTION__, __FILENAME__, __LINE__);
+        return (level >= 1 ? ACL_AICORE_PIPE_UTILIZATION : ACL_AICORE_NONE);
+    }
+    return aic_metrics;
+}
 
 ProfilerMgr::ProfilerMgr()
     : report_enable_(false),
@@ -77,12 +90,14 @@ void ProfilerMgr::Start(const NpuTraceConfig &npu_config, bool cpu_trace)
     c10_npu::npuSynchronizeDevice();
     if (npu_trace_.load() == true) {
         aclprofAicoreMetrics aic_metrics = ACL_AICORE_NONE;
+        int8_t level_int = trace_level_to_int_.find(npu_config.trace_level) != trace_level_to_int_.end() ?
+            trace_level_to_int_[npu_config.trace_level] : -1;
         auto level_iter = trace_level_map_.find(npu_config.trace_level);
         uint64_t datatype_config = (level_iter == trace_level_map_.end()) ? Level0 : trace_level_map_[npu_config.trace_level];
         auto metrics_iter = npu_metrics_map_.find(npu_config.metrics);
         if (metrics_iter != npu_metrics_map_.end() && npu_config.metrics.compare("ACL_AICORE_NONE") != 0) {
             datatype_config |= ACL_PROF_AICORE_METRICS;
-            aic_metrics = npu_metrics_map_[npu_config.metrics];
+            aic_metrics = CheckAicMetricsFeature(npu_metrics_map_[npu_config.metrics], level_int);
         }
         if (npu_config.l2_cache) {
             datatype_config |= ACL_PROF_L2CACHE;
@@ -111,8 +126,6 @@ void ProfilerMgr::Start(const NpuTraceConfig &npu_config, bool cpu_trace)
         const uint32_t deviceNum = 1;
         uint32_t deviceIdList[deviceNum] = {deviceId};
         EnableMsProfiler(deviceIdList, deviceNum, aic_metrics, datatype_config);
-        int8_t level_int = trace_level_to_int_.find(npu_config.trace_level) != trace_level_to_int_.end() ?
-            trace_level_to_int_[npu_config.trace_level] : -1;
         trace_level_.store(level_int);
     }
 

torch_npu/profiler/analysis/prof_common_func/_constant.py

@@ -118,6 +118,7 @@ class Constant(object):
     AicMemoryUB = "ACL_AICORE_MEMORY_UB"
     AicResourceConflictRatio = "ACL_AICORE_RESOURCE_CONFLICT_RATIO"
     AicL2Cache = "ACL_AICORE_L2_CACHE"
+    AicMemoryAccess = "ACL_AICORE_MEMORY_ACCESS"
     AicMetricsNone = "ACL_AICORE_NONE"
     Db = "db"
     Text = "text"

torch_npu/profiler/experimental_config.py

@@ -21,7 +21,7 @@ def supported_profiler_level():
 def supported_ai_core_metrics():
     return set((AiCMetrics.AiCoreNone, AiCMetrics.PipeUtilization, AiCMetrics.ArithmeticUtilization,
                 AiCMetrics.Memory, AiCMetrics.MemoryL0, AiCMetrics.MemoryUB,
-                AiCMetrics.ResourceConflictRatio, AiCMetrics.L2Cache))
+                AiCMetrics.ResourceConflictRatio, AiCMetrics.L2Cache, AiCMetrics.MemoryAccess))
 
 
 def supported_export_type():
@@ -43,6 +43,7 @@ class AiCMetrics:
     MemoryUB = Constant.AicMemoryUB
     ResourceConflictRatio = Constant.AicResourceConflictRatio
     L2Cache = Constant.AicL2Cache
+    MemoryAccess = Constant.AicMemoryAccess
     AiCoreNone = Constant.AicMetricsNone
 
 
@@ -120,10 +121,10 @@ def _check_params(self):
         if self._aic_metrics not in (
                 AiCMetrics.L2Cache, AiCMetrics.MemoryL0, AiCMetrics.Memory, AiCMetrics.MemoryUB,
                 AiCMetrics.PipeUtilization, AiCMetrics.ArithmeticUtilization, AiCMetrics.ResourceConflictRatio,
-                Constant.AicMetricsNone):
+                AiCMetrics.MemoryAccess, AiCMetrics.AiCoreNone):
             print_warn_msg("Invalid parameter aic_metrics, reset it to default.")
             if self._profiler_level == ProfilerLevel.Level0:
-                self._aic_metrics = Constant.AicMetricsNone
+                self._aic_metrics = AiCMetrics.AiCoreNone
             else:
                 self._aic_metrics = AiCMetrics.PipeUtilization
         if not isinstance(self._op_attr, bool):