Initial support for tracing DMA events in AIE tile memory modules

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

@@ -211,9 +211,16 @@ enum class module_type {
       uint32_t broadcast_mask_west = BROADCAST_MASK_DEFAULT;
       uint32_t broadcast_mask_east = BROADCAST_MASK_DEFAULT;
       uint32_t internal_events_broadcast[NUM_BROADCAST_EVENTS] = {};
+      bool port_trace_is_master[NUM_SWITCH_MONITOR_PORTS];
+      int8_t port_trace_ids[NUM_SWITCH_MONITOR_PORTS];
       std::vector<aie_cfg_counter> pc;
 
-      aie_cfg_base(uint32_t count) : pc(count) {};
+      aie_cfg_base(uint32_t count) : pc(count) {
+        for (uint32_t i=0; i < NUM_SWITCH_MONITOR_PORTS; ++i) {
+          port_trace_is_master[i] = false;
+          port_trace_ids[i] = -1;
+        }
+      };
   };
 
   /*
@@ -226,7 +233,6 @@ enum class module_type {
   {
   public:
     uint32_t trace_mode = 1;
-    std::string port_trace = "null";
     aie_cfg_core() : aie_cfg_base(4)
     {
       group_event_config = {
@@ -261,16 +267,9 @@ enum class module_type {
   class aie_cfg_peripheral_tile : public aie_cfg_base
   {
   public:
-    bool port_trace_is_master[NUM_SWITCH_MONITOR_PORTS];
-    int8_t port_trace_ids[NUM_SWITCH_MONITOR_PORTS];
     int8_t s2mm_channels[NUM_CHANNEL_SELECTS] = {-1, -1};
     int8_t mm2s_channels[NUM_CHANNEL_SELECTS] = {-1, -1};
-    aie_cfg_peripheral_tile() : aie_cfg_base(4) {
-      for (uint32_t i=0; i < NUM_SWITCH_MONITOR_PORTS; ++i) {
-        port_trace_is_master[i] = false;
-        port_trace_ids[i] = -1;
-      }
-    }
+    aie_cfg_peripheral_tile() : aie_cfg_base(4) {}
   };
 
   /*

src/runtime_src/xdp/profile/device/aieTraceS2MM.cpp

@@ -59,7 +59,7 @@ void AIETraceS2MM::init(uint64_t bo_size, int64_t bufaddr, bool circular)
       uint32_t regValue = 0;
       read(TS2MM_AP_CTRL, BYTES_PER_WORD, &regValue);
       std::stringstream msg;
-      msg << "AIE TraceS2MM AP control register after first start: 0x" << std::hex << regValue;
+      msg << "AIE TraceS2MM AP control register after second start: 0x" << std::hex << regValue;
       xrt_core::message::send(severity_level::debug, "XRT", msg.str());
     }
     // End of temporary code

src/runtime_src/xdp/profile/device/traceS2MM.cpp

@@ -91,7 +91,7 @@ void TraceS2MM::init(uint64_t bo_size, int64_t bufaddr, bool circular)
       uint32_t regValue = 0;
       read(TS2MM_AP_CTRL, BYTES_PER_WORD, &regValue);
       std::stringstream msg;
-      msg << "AIE TraceS2MM AP control register after first start: 0x" << std::hex << regValue;
+      msg << "AIE TraceS2MM AP control register after second start: 0x" << std::hex << regValue;
       xrt_core::message::send(severity_level::debug, "XRT", msg.str());
     }
     // End of temporary code

src/runtime_src/xdp/profile/device/tracedefs.h

@@ -117,22 +117,34 @@ constexpr unsigned int BYTES_128BIT = 16;
 constexpr uint32_t NUM_TRACE_EVENTS = 8;
 constexpr uint32_t NUM_OUTPUT_TRACE_EVENTS = 9;
 constexpr uint32_t NUM_BROADCAST_EVENTS = 16;
-constexpr uint32_t EVENT_CORE_ACTIVE = 28;
-constexpr uint32_t EVENT_CORE_DISABLED = 29;
-constexpr uint16_t EVENT_MEM_TILE_DMA_MM2S_SEL0_STALLED_LOCK = 35;
-constexpr uint16_t EVENT_MEM_TILE_DMA_S2MM_SEL0_STREAM_STARVATION = 37;
-constexpr uint32_t BROADCAST_MASK_DEFAULT = 65535;
 constexpr uint32_t NUM_TRACE_PCS = 4;
 constexpr uint32_t NUM_MEM_TRACE_PCS = 2;
 constexpr uint32_t NUM_COMBO_EVENT_CONTROL = 3;
 constexpr uint32_t NUM_COMBO_EVENT_INPUT = 4;
 constexpr uint32_t NUM_SWITCH_MONITOR_PORTS = 8;
 constexpr uint32_t NUM_CHANNEL_SELECTS = 2;
 
+constexpr uint32_t BROADCAST_MASK_DEFAULT = 65535;
 constexpr uint32_t CORE_BROADCAST_EVENT_BASE = 107;
 constexpr uint32_t ES1_TRACE_COUNTER = 1020;
 constexpr uint32_t ES2_TRACE_COUNTER = 0x3FF00;
 
+constexpr uint32_t EVENT_CORE_ACTIVE = 28;
+constexpr uint32_t EVENT_CORE_DISABLED = 29;
+constexpr uint16_t EVENT_MEM_DMA_MM2S_0_STALLED_LOCK = 33;
+constexpr uint16_t EVENT_MEM_DMA_MM2S_1_STALLED_LOCK = 34;
+constexpr uint16_t EVENT_MEM_DMA_S2MM_0_STREAM_STARVATION = 35;
+constexpr uint16_t EVENT_MEM_DMA_S2MM_1_STREAM_STARVATION = 36;
+constexpr uint16_t EVENT_MEM_TILE_DMA_MM2S_SEL0_STALLED_LOCK = 35;
+constexpr uint16_t EVENT_MEM_TILE_DMA_S2MM_SEL0_STREAM_STARVATION = 37;
+
+constexpr uint32_t GROUP_CORE_STALL_MASK            = 0x0000000F;
+constexpr uint32_t GROUP_CORE_FUNCTIONS_MASK        = 0x0000000C;
+constexpr uint32_t GROUP_STREAM_SWITCH_RUNNING_MASK = 0x00002222;
+
+constexpr uint64_t AIE_OFFSET_EDGE_CONTROL_MEM_TILE = 0x94408;
+constexpr uint64_t AIE_OFFSET_EDGE_CONTROL_MEM      = 0x14408;
+
 }
 
 #endif

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

@@ -117,7 +117,8 @@ namespace xdp {
         && shimTileMetricsSettings.empty() && shimGraphMetricsSettings.empty()) {
       isValidMetrics = false;
     } else {
-      getConfigMetricsForTiles(aieTileMetricsSettings, aieGraphMetricsSettings, module_type::core);
+      // Use DMA type here to include both core-active tiles and DMA-only tiles
+      getConfigMetricsForTiles(aieTileMetricsSettings, aieGraphMetricsSettings, module_type::dma);
       getConfigMetricsForTiles(memTileMetricsSettings, memGraphMetricsSettings, module_type::mem_tile);
       getConfigMetricsForInterfaceTiles(shimTileMetricsSettings, shimGraphMetricsSettings);
       setTraceStartControl(compilerOptions.graph_iterator_event);
@@ -314,8 +315,8 @@ namespace xdp {
 
     std::set<tile_type> allValidTiles;
     auto validTilesVec = metadataReader->getTiles("all", type, "all");
-    std::unique_copy(validTilesVec.begin(), validTilesVec.end(), std::inserter(allValidTiles, allValidTiles.end()), 
-                     tileCompare);
+    std::unique_copy(validTilesVec.begin(), validTilesVec.end(), 
+                     std::inserter(allValidTiles, allValidTiles.end()), tileCompare);
 
     // STEP 1 : Parse per-graph and/or per-kernel settings
 
@@ -606,6 +607,8 @@ namespace xdp {
 
     // Set default, check validity, and remove "off" tiles
     bool showWarning = true;
+    bool showWarning2 = true;
+    bool showWarning3 = true;
     std::vector<tile_type> offTiles;
     auto defaultSet = defaultSets[type];
     auto coreSets = metricSets[module_type::core];
@@ -635,15 +638,30 @@ namespace xdp {
         }
         tileMetric.second = defaultSet;
       }
+
+      // Check for deprecated metric set names
+      if (tileMetric.second == "functions_partial_stalls") {
+        if (showWarning2) {
+          xrt_core::message::send(severity_level::warning, "XRT", 
+              "The metric set functions_partial_stalls is being renamed to partial_stalls. "
+              "Please use the new set name starting in 2024.2.");
+          showWarning2 = false;
+        }
+      }
+      if (tileMetric.second == "functions_all_stalls") {
+        if (showWarning3) {
+          xrt_core::message::send(severity_level::warning, "XRT", 
+              "The metric set functions_all_stalls is being renamed to all_stalls. "
+              "Please use the new set name starting in 2024.2.");
+          showWarning3 = false;
+        }
+      }
     }
 
     // Remove all the "off" tiles
     for (auto &t : offTiles) {
       configMetrics.erase(t);
     }
-
-    // If needed, turn on debug fal messages
-    // xaiefal::Logger::get().setLogLevel(xaiefal::LogLevel::DEBUG);
   }
 
   // Resolve metrics for interface tiles

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

@@ -138,13 +138,16 @@ class AieTraceMetadata {
 
     std::map<module_type, std::string> defaultSets {
       { module_type::core,     "functions"},
+      { module_type::dma,      "functions"},
       { module_type::mem_tile, "input_channels"},
       { module_type::shim,     "input_ports"}
     };
 
     std::map <module_type, std::vector<std::string>> metricSets {
       { module_type::core,     {"functions", "functions_partial_stalls", 
-                                "functions_all_stalls", "all"} },
+                                "functions_all_stalls", "partial_stalls",
+                                "all_stalls", "all_dma", "all_stalls_dma", 
+                                "s2mm_channels_stalls", "mm2s_channels_stalls"} },
       { module_type::mem_tile, {"input_channels", "input_channels_stalls", 
                                 "output_channels", "output_channels_stalls",
                                 "s2mm_channels", "s2mm_channels_stalls",