Dtrace json output of control instrumentation to reduce overhead at hwcontext destructor

profile/plugin/aie_dtrace/aie_dtrace_plugin.cpp

@@ -4,6 +4,7 @@
 #define XDP_PLUGIN_SOURCE
 
 #include "xdp/profile/plugin/aie_dtrace/aie_dtrace_plugin.h"
+#include "xdp/profile/plugin/aie_dtrace/util/aie_dtrace_util.h"
 
 #include "core/common/api/hw_context_int.h"
 #include "core/common/config_reader.h"
@@ -31,6 +32,9 @@ namespace xdp {
   {
     AieDtracePlugin::live = true;
 
+    if (xrt_core::config::get_aie_dtrace())
+      aie::dtrace::initDtraceOutputConfig();
+
     db->registerPlugin(this);
     db->registerInfo(info::aie_dtrace);
     db->getStaticInfo().setAieApplication();
@@ -68,6 +72,8 @@ namespace xdp {
     if (!xrt_core::config::get_aie_dtrace())
       return;
 
+    aie::dtrace::initDtraceOutputConfig();
+
     if (!handle)
       return;
 

profile/plugin/aie_dtrace/util/aie_dtrace_util.cpp

@@ -5,8 +5,43 @@
 
 #include "xdp/profile/plugin/aie_dtrace/util/aie_dtrace_util.h"
 
+#include "core/common/config_reader.h"
+#include "core/common/message.h"
+
+#include <mutex>
+
 namespace xdp::aie::dtrace {
 
+  namespace {
+    using severity_level = xrt_core::message::severity_level;
+
+    static constexpr unsigned int DEFAULT_COALESCE_RESULT_MEMORY_MB = 256;
+  } // anonymous namespace
+
+  void
+  initDtraceOutputConfig()
+  {
+    static std::once_flag once;
+    std::call_once(once, []() {
+      try {
+        xrt_core::config::detail::set("Debug.dtrace_output_json_format", "true");
+        xrt_core::config::detail::set("Debug.dtrace_coalesce_result", "true");
+        xrt_core::config::detail::set("Debug.dtrace_coalesce_result_memory_mb",
+                                      std::to_string(DEFAULT_COALESCE_RESULT_MEMORY_MB));
+      }
+      catch (const std::exception& e) {
+        xrt_core::message::send(severity_level::warning, "XRT",
+            std::string("AIE dtrace: could not apply default dtrace output settings: ")
+            + e.what());
+        return;
+      }
+
+      xrt_core::message::send(severity_level::info, "XRT",
+          "AIE dtrace: enabled JSON dtrace_dump output with coalesced results "
+          "(dtrace_dump_ctx_<slot>_<timestamp>.json on hw context teardown)");
+    });
+  }
+
   std::map<std::string, std::vector<XAie_Events>>
   getBandwidthInterfaceTileEventSets(int hwGen)
   {

profile/plugin/aie_dtrace/util/aie_dtrace_util.h

@@ -17,6 +17,10 @@ namespace xdp::aie::dtrace {
   // Shim bandwidth metric sets used for Debug.aie_dtrace (not part of standard aie_profile ini).
   std::map<std::string, std::vector<XAie_Events>> getBandwidthInterfaceTileEventSets(int hwGen);
 
+  // Enable JSON dtrace_dump output with coalesced results by default.
+  // Must run before XRT creates the first dtrace module (config keys lock on first read).
+  void initDtraceOutputConfig();
+
 } // namespace xdp::aie::dtrace
 
 #endif

profile/plugin/aie_dtrace/ve2/aie_dtrace_ct_writer.cpp

@@ -636,9 +636,12 @@ bool AieDtraceCTWriter::writeCTFile(const std::vector<ASMFileInfo>& asmFileInfoL
     ctFile << "{\n";
     ctFile << "    ts_" << asmFileInfo.asmId << " = timestamp32()\n";
 
-    // Write counter reads using _ as throwaway variable
+    // Write counter reads. Each counter is assigned a sequential result
+    // variable (_0, _1, _2, ...) so the dtrace JSON dump emits one key per
+    // counter ("_0", "_1", ...) in probe order; counter semantics are provided
+    // by the COUNTER_METADATA block above.
     for (size_t i = 0; i < asmFileInfo.counters.size(); i++) {
-      ctFile << "    _ = read_reg("
+      ctFile << "    _" << i << " = read_reg("
              << formatAddress(asmFileInfo.counters[i].address) << ")\n";
     }
 
@@ -1156,9 +1159,12 @@ bool AieDtraceCTWriter::writeBandwidthCTFile(
     ctFile << "{\n";
     ctFile << "    ts_" << asmFileInfo.asmId << " = timestamp32()\n";
 
+    // Each counter is assigned a sequential result variable (_0, _1, _2, ...)
+    // so the dtrace JSON dump emits one key per counter ("_0", "_1", ...) in
+    // probe order; counter semantics live in the COUNTER_METADATA block above.
     for (size_t i = 0; i < asmFileInfo.counters.size(); i++) {
       const auto& ctr = asmFileInfo.counters[i];
-      ctFile << "    _ = read_reg(" << formatAddress(ctr.address) << ")\n";
+      ctFile << "    _" << i << " = read_reg(" << formatAddress(ctr.address) << ")\n";
     }
 
     ctFile << "}\n\n";

profile/plugin/aie_dtrace/ve2/aie_dtrace_ve2.cpp

@@ -6,6 +6,7 @@
 #include "xdp/profile/plugin/aie_dtrace/ve2/aie_dtrace_ve2.h"
 #include "xdp/profile/plugin/aie_dtrace/ve2/aie_dtrace_ct_writer.h"
 #include "xdp/profile/plugin/aie_dtrace/ve2/elf_helper.h"
+#include "xdp/profile/plugin/aie_dtrace/util/aie_dtrace_util.h"
 
 #include "core/common/api/hw_context_int.h"
 #include "core/common/api/kernel_int.h"
@@ -134,6 +135,8 @@ namespace xdp {
     if (!ctWriter.generateBandwidthCT(outputPath, hwctx, it->second, bandwidthMetricSet, bandwidthChannel))
       return;
 
+    aie::dtrace::initDtraceOutputConfig();
+
     xrt_core::message::send(severity_level::debug, "XRT",
         "AIE dtrace: Bandwidth CT generated for kernel '" + kernel_name
         + "' with metric set '" + bandwidthMetricSet + "'");