Initial refactoring of AIE trace plugins

src/runtime_src/xdp/profile/database/static_info/aie_util.h

@@ -58,15 +58,38 @@ namespace xdp::aie {
   xdp::aie::driver_config
   getDriverConfig(const boost::property_tree::ptree& aie_meta,
                 const std::string& root);
+
   XDP_CORE_EXPORT
   uint16_t
   getAIETileRowOffset(const boost::property_tree::ptree& aie_meta,
                     const std::string& location);
+
   XDP_CORE_EXPORT
   std::vector<std::string>
   getValidGraphs(const boost::property_tree::ptree& aie_meta,
                 const std::string& root);
 
+  XDP_CORE_EXPORT
+  bool isInfoVerbosity();
+
+  XDP_CORE_EXPORT
+  bool isDebugVerbosity();
+
+  XDP_CORE_EXPORT
+  bool isInputSet(const module_type type, const std::string metricSet);
+
+  XDP_CORE_EXPORT
+  uint16_t getRelativeRow(uint16_t absRow, uint16_t rowOffset);
+
+  XDP_CORE_EXPORT
+  module_type getModuleType(uint16_t absRow, uint16_t rowOffset);
+
+  XDP_CORE_EXPORT
+  uint32_t bcIdToEvent(int bcId);
+
+  XDP_CORE_EXPORT
+  std::string getModuleName(module_type mod);
+
 } // namespace xdp::aie
 
 #endif

src/runtime_src/xdp/profile/device/device_trace_logger.cpp

@@ -422,7 +422,6 @@ namespace xdp {
       // start event, end event
       // start end must have created already
       // check if the memory ports on current cu has any event
-
       uint64_t cuLastTimestamp  = amLastTrans[amIndex];
 
       // get CU Id for the current slot
@@ -434,7 +433,6 @@ namespace xdp {
         // To reduce overhead, first check the timestamp.
         // If last activity timestamp on CU is earlier than current AIM, then only check
         // whether the current AIM is attached to the same CU.
-        //
         if(cuLastTimestamp >= aimLastTrans[aimIndex]) {
           continue;
         }
@@ -452,7 +450,6 @@ namespace xdp {
         // To reduce overhead, first check the timestamp.
         // If last activity timestamp on CU is earlier than current ASM, then only check
         // whether the current ASM is attached to the same CU.
-        ///
         if(cuLastTimestamp >= asmLastTrans[asmIndex]) {
           continue;
         }

src/runtime_src/xdp/profile/plugin/aie_profile/edge/aie_profile.cpp

@@ -92,8 +92,10 @@ namespace xdp {
                                    XAIE_EVENT_INSTR_CASCADE_PUT_CORE,    XAIE_EVENT_GROUP_CORE_STALL_CORE}},
       {"read_throughputs",        {XAIE_EVENT_ACTIVE_CORE,               XAIE_EVENT_INSTR_STREAM_GET_CORE,
                                    XAIE_EVENT_INSTR_CASCADE_GET_CORE,    XAIE_EVENT_GROUP_CORE_STALL_CORE}},
-      {"aie_trace",               {XAIE_EVENT_PORT_RUNNING_1_CORE,       XAIE_EVENT_PORT_STALLED_1_CORE,
-                                   XAIE_EVENT_PORT_RUNNING_0_CORE,       XAIE_EVENT_PORT_STALLED_0_CORE}},
+      {"s2mm_throughputs",        {XAIE_EVENT_PORT_RUNNING_0_CORE,       XAIE_EVENT_PORT_STALLED_0_CORE}},
+      {"mm2s_throughputs",        {XAIE_EVENT_PORT_RUNNING_0_CORE,       XAIE_EVENT_PORT_STALLED_0_CORE}},
+      {"aie_trace",               {XAIE_EVENT_PORT_RUNNING_0_CORE,       XAIE_EVENT_PORT_STALLED_0_CORE,
+                                   XAIE_EVENT_PORT_RUNNING_1_CORE,       XAIE_EVENT_PORT_STALLED_1_CORE}},
       {"events",                  {XAIE_EVENT_INSTR_EVENT_0_CORE,        XAIE_EVENT_INSTR_EVENT_1_CORE,
                                    XAIE_EVENT_USER_EVENT_0_CORE,         XAIE_EVENT_USER_EVENT_1_CORE}}
     };
@@ -104,8 +106,6 @@ namespace xdp {
     else {
       mCoreStartEvents["floating_point"]   = {XAIE_EVENT_FP_HUGE_CORE,        XAIE_EVENT_INT_FP_0_CORE, 
                                               XAIE_EVENT_FP_INVALID_CORE,     XAIE_EVENT_FP_INF_CORE};
-      mCoreStartEvents["s2mm_throughputs"] = {XAIE_EVENT_PORT_RUNNING_0_CORE, XAIE_EVENT_PORT_STALLED_0_CORE};
-      mCoreStartEvents["mm2s_throughputs"] = {XAIE_EVENT_PORT_RUNNING_0_CORE, XAIE_EVENT_PORT_STALLED_0_CORE};
     }
     mCoreEndEvents = mCoreStartEvents;
 
@@ -179,34 +179,32 @@ namespace xdp {
       {"mem_trace",               {XAIE_EVENT_PORT_RUNNING_0_MEM_TILE, 
                                    XAIE_EVENT_PORT_STALLED_0_MEM_TILE,
                                    XAIE_EVENT_PORT_IDLE_0_MEM_TILE,
-                                   XAIE_EVENT_PORT_TLAST_0_MEM_TILE}}
+                                   XAIE_EVENT_PORT_TLAST_0_MEM_TILE}},
+      {"input_throughputs",       {XAIE_EVENT_PORT_RUNNING_0_MEM_TILE,
+                                   XAIE_EVENT_DMA_S2MM_SEL0_STREAM_STARVATION_MEM_TILE,
+                                   XAIE_EVENT_DMA_S2MM_SEL0_MEMORY_BACKPRESSURE_MEM_TILE,
+                                   XAIE_EVENT_DMA_S2MM_SEL0_STALLED_LOCK_ACQUIRE_MEM_TILE}},
+      {"output_throughputs",      {XAIE_EVENT_PORT_RUNNING_0_MEM_TILE, 
+                                   XAIE_EVENT_DMA_MM2S_SEL0_STREAM_BACKPRESSURE_MEM_TILE,
+                                   XAIE_EVENT_DMA_MM2S_SEL0_MEMORY_STARVATION_MEM_TILE,
+                                   XAIE_EVENT_DMA_MM2S_SEL0_STALLED_LOCK_ACQUIRE_MEM_TILE}},
+      {"conflict_stats1",         {XAIE_EVENT_CONFLICT_DM_BANK_0_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_1_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_2_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_3_MEM_TILE}},
+      {"conflict_stats2",         {XAIE_EVENT_CONFLICT_DM_BANK_4_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_5_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_6_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_7_MEM_TILE}},
+      {"conflict_stats3",         {XAIE_EVENT_CONFLICT_DM_BANK_8_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_9_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_10_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_11_MEM_TILE}}, 
+      {"conflict_stats4",         {XAIE_EVENT_CONFLICT_DM_BANK_12_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_13_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_14_MEM_TILE,
+                                   XAIE_EVENT_CONFLICT_DM_BANK_15_MEM_TILE}}
     };
-    if (metadata->getHardwareGen() > 1) {
-      mMemTileStartEvents["input_throughputs"]  = {XAIE_EVENT_PORT_RUNNING_0_MEM_TILE,
-                                                   XAIE_EVENT_DMA_S2MM_SEL0_STREAM_STARVATION_MEM_TILE,
-                                                   XAIE_EVENT_DMA_S2MM_SEL0_MEMORY_BACKPRESSURE_MEM_TILE,
-                                                   XAIE_EVENT_DMA_S2MM_SEL0_STALLED_LOCK_ACQUIRE_MEM_TILE};
-      mMemTileStartEvents["output_throughputs"] = {XAIE_EVENT_PORT_RUNNING_0_MEM_TILE, 
-                                                   XAIE_EVENT_DMA_MM2S_SEL0_STREAM_BACKPRESSURE_MEM_TILE,
-                                                   XAIE_EVENT_DMA_MM2S_SEL0_MEMORY_STARVATION_MEM_TILE,
-                                                   XAIE_EVENT_DMA_MM2S_SEL0_STALLED_LOCK_ACQUIRE_MEM_TILE};
-      mMemTileStartEvents["conflict_stats1"]    = {XAIE_EVENT_CONFLICT_DM_BANK_0_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_1_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_2_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_3_MEM_TILE};
-      mMemTileStartEvents["conflict_stats2"]    = {XAIE_EVENT_CONFLICT_DM_BANK_4_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_5_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_6_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_7_MEM_TILE};
-      mMemTileStartEvents["conflict_stats3"]    = {XAIE_EVENT_CONFLICT_DM_BANK_8_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_9_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_10_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_11_MEM_TILE}; 
-      mMemTileStartEvents["conflict_stats4"]    = {XAIE_EVENT_CONFLICT_DM_BANK_12_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_13_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_14_MEM_TILE,
-                                                   XAIE_EVENT_CONFLICT_DM_BANK_15_MEM_TILE};
-    }
     mMemTileStartEvents["s2mm_channels"]         = mMemTileStartEvents["input_channels"];
     mMemTileStartEvents["s2mm_channels_details"] = mMemTileStartEvents["input_channels_details"];
     mMemTileStartEvents["s2mm_throughputs"]      = mMemTileStartEvents["input_throughputs"];
@@ -391,6 +389,7 @@ namespace xdp {
 
       bool newPort = false;
       auto portnum = getPortNumberFromEvent(startEvent);
+      uint8_t channel = (portnum == 0) ? channel0 : channel1;
 
       // New port needed: reserver, configure, and store
       if (switchPortMap.find(portnum) == switchPortMap.end()) {
@@ -401,10 +400,32 @@ namespace xdp {
         switchPortMap[portnum] = switchPortRsc;
 
         if (type == module_type::core) {
-          // AIE Tiles (e.g., trace streams)
-          // Define stream switch port to monitor core or memory trace
-          uint8_t traceSelect = (startEvent == XAIE_EVENT_PORT_RUNNING_0_CORE) ? 0 : 1;
-          switchPortRsc->setPortToSelect(XAIE_STRMSW_SLAVE, TRACE, traceSelect);
+          int channelNum = 0;
+          std::string portName;
+
+          // AIE Tiles
+          if (metricSet.find("trace") != std::string::npos) {
+            // Monitor memory or core trace (memory:1, core:0)
+            uint8_t traceSelect = (startEvent == XAIE_EVENT_PORT_RUNNING_0_CORE) ? 1 : 0;
+            switchPortRsc->setPortToSelect(XAIE_STRMSW_SLAVE, TRACE, traceSelect);
+
+            channelNum = traceSelect;
+            portName = (traceSelect == 0) ? "core trace" : "memory trace";
+          }
+          else {
+            auto slaveOrMaster = aie::isInputSet(type, metricSet) ? XAIE_STRMSW_SLAVE : XAIE_STRMSW_MASTER;
+            switchPortRsc->setPortToSelect(slaveOrMaster, DMA, channel);
+
+            channelNum = channel;
+            portName = aie::isInputSet(type, metricSet) ? "DMA MM2S" : "DMA S2MM";
+          }
+
+          if (aie::isDebugVerbosity()) {
+              std::stringstream msg;
+              msg << "Configured core module stream switch to monitor " << portName 
+                  << " for metric set " << metricSet << " and channel " << channelNum;
+              xrt_core::message::send(severity_level::debug, "XRT", msg.str());
+          }
         }
         else if (type == module_type::shim) {
           // Interface tiles (e.g., PLIO, GMIO)
@@ -419,7 +440,6 @@ namespace xdp {
             switchPortRsc->setPortToSelect(XAIE_STRMSW_SLAVE, TRACE, 0);
           }
           else {
-            uint8_t channel = (portnum == 0) ? channel0 : channel1;
             auto slaveOrMaster = isInputSet(type, metricSet) ? XAIE_STRMSW_MASTER : XAIE_STRMSW_SLAVE;
             switchPortRsc->setPortToSelect(slaveOrMaster, DMA, channel);
           }

src/runtime_src/xdp/profile/plugin/aie_trace/CMakeLists.txt

@@ -33,6 +33,10 @@ file(GLOB AIE_TRACE_PLUGIN_FILES
   "${IMPL_DIR}/*.h"
   "${IMPL_DIR}/*.cpp"
 )
+file(GLOB AIE_TRACE_UTIL_FILES
+  "${PROFILE_DIR}/plugin/aie_trace/util/*.h"
+  "${PROFILE_DIR}/plugin/aie_trace/util/*.cpp"
+)
 
 if (XDP_MINIMAL_BUILD STREQUAL "yes")
   add_library(xdp_aie_trace_plugin MODULE ${AIE_TRACE_PLUGIN_FILES})
@@ -63,7 +67,7 @@ elseif (${XRT_NATIVE_BUILD} STREQUAL "yes")
 
 elseif (DEFINED XRT_AIE_BUILD AND ${XRT_NATIVE_BUILD} STREQUAL "no")
 
-  add_library(xdp_aie_trace_plugin MODULE ${AIE_TRACE_PLUGIN_FILES})
+  add_library(xdp_aie_trace_plugin MODULE ${AIE_TRACE_PLUGIN_FILES} ${AIE_TRACE_UTIL_FILES})
 
   add_dependencies(xdp_aie_trace_plugin xdp_core xrt_coreutil)
   target_link_libraries(xdp_aie_trace_plugin PRIVATE xdp_core xrt_coreutil xaiengine)

src/runtime_src/xdp/profile/plugin/aie_trace/aie_trace_impl.h

@@ -22,38 +22,66 @@
 #include <memory>
 
 namespace xdp {
+
+  class VPDatabase;
 
-class VPDatabase;
-
-// AIE trace configurations can be done in different ways depending
-// on the platform.  For example, platforms like the VCK5000 or
-// discovery platform, where the host code runs on the x86 and the AIE
-// is not directly accessible, will require configuration be done via
-// PS kernel.
-class AieTraceImpl {
-
-protected:
-  VPDatabase *db = nullptr;
-  std::shared_ptr<AieTraceMetadata> metadata;
-
-public:
-  AieTraceImpl(VPDatabase *database, std::shared_ptr<AieTraceMetadata> metadata)
-      : db(database), metadata(metadata) {}
-
-  AieTraceImpl() = delete;
-  virtual ~AieTraceImpl(){};
-
-  virtual void updateDevice() = 0;
-  virtual void freeResources() = 0;
-  virtual void pollTimers(uint64_t index, void *handle) = 0;
-  virtual uint64_t checkTraceBufSize(uint64_t size) = 0;
-  /*
-   * If trace module is running, it might buffer partial trace.
-   * This leftover trace needs to be force flushed at the end using a custom end
-   * event. This applies to trace windowing on AIE1 and all scenarios on AIE2.
-   */
-  virtual void flushAieTileTraceModule() = 0;
-};
+  /** 
+   * @brief   Base class for AI Engine trace implementations
+   * @details Trace configurations can be done in different ways depending
+   *          on the platform.  For example, platforms like the VCK5000 or
+   *          discovery platform, where the host code runs on the x86 and the 
+   *          AIE is not directly accessible, will require configuration be 
+   *          done via PS kernel.
+   */ 
+  class AieTraceImpl {
+  public:
+    /**
+     * @brief AIE Trace implementation constructor
+     * @param database Profile database for storing results and configuation 
+     * @param metadata Design-specific AIE metadata typically taken from xclbin
+     */
+    AieTraceImpl(VPDatabase* database, std::shared_ptr<AieTraceMetadata> metadata)
+      :db(database), metadata(metadata) {}
+
+    AieTraceImpl() = delete;
+    /// @brief AIE Trace implementation destructor
+    virtual ~AieTraceImpl() {};
+
+  protected:
+    /// @brief Database for configuration and results
+    VPDatabase* db = nullptr;
+
+    /// @brief Trace metadata parsed from user settings
+    std::shared_ptr<AieTraceMetadata> metadata;
+
+  public:
+    /// @brief Update device (e.g., after loading xclbin)
+    virtual void updateDevice() = 0;
+
+    /// @brief Stop and release resources (e.g., counters, ports)
+    virtual void freeResources() = 0;
+
+    /**
+     * @brief Poll AIE timers (for system timeline only)
+     * @param index  Device index number
+     * @param handle Pointer to device handle
+     */
+    virtual void pollTimers(uint64_t index, void* handle) = 0;
+
+    /**
+     * @brief Verify correctness of trace buffer size
+     * @param size Requested size of trace buffer
+     */
+    virtual uint64_t checkTraceBufSize(uint64_t size) = 0;
+
+    /**
+     * @brief   Flush trace modules by forcing end events
+     * @details Trace modules buffer partial packets. At end of run, these need to be 
+     *          flushed using a custom end event. This applies to trace windowing and 
+     *          passive tiles like memory and interface.
+     */
+    virtual void flushTraceModules() = 0;
+  };
 
 } // namespace xdp
 

src/runtime_src/xdp/profile/plugin/aie_trace/aie_trace_plugin.cpp

@@ -404,7 +404,7 @@ void AieTracePluginUnified::flushAIEDevice(void *handle) {
     return;
 
   // Flush AIE then datamovers
-  AIEData.implementation->flushAieTileTraceModule();
+  AIEData.implementation->flushTraceModules();
   flushOffloader(AIEData.offloader, false);
 }
 
@@ -433,7 +433,7 @@ void AieTracePluginUnified::finishFlushAIEDevice(void *handle) {
   endPollforDevice(handle);
 
   // Flush AIE then datamovers
-  AIEData.implementation->flushAieTileTraceModule();
+  AIEData.implementation->flushTraceModules();
   flushOffloader(AIEData.offloader, true);
   XDPPlugin::endWrite();
   (db->getStaticInfo()).deleteCurrentlyUsedDeviceInterface(AIEData.deviceID);
@@ -453,7 +453,7 @@ void AieTracePluginUnified::writeAll(bool openNewFiles) {
     auto &AIEData = kv.second;
 
     if (AIEData.valid) {
-      AIEData.implementation->flushAieTileTraceModule();
+      AIEData.implementation->flushTraceModules();
       flushOffloader(AIEData.offloader, true);
     }
   }