ML timeline buffer needs to be 12KB aligned now

src/runtime_src/xdp/profile/plugin/ml_timeline/clientDev/ml_timeline.cpp

@@ -60,8 +60,8 @@ namespace xdp {
       }
   };
 
-  MLTimelineClientDevImpl::MLTimelineClientDevImpl(VPDatabase*dB)
-    : MLTimelineImpl(dB)
+  MLTimelineClientDevImpl::MLTimelineClientDevImpl(VPDatabase*dB, uint32_t sz)
+    : MLTimelineImpl(dB, sz)
   {
     xrt_core::message::send(xrt_core::message::severity_level::debug, "XRT", 
               "Created ML Timeline Plugin for Client Device.");
@@ -134,16 +134,16 @@ namespace xdp {
     // Record Timer TS in JSON
     // Assuming correct Stub has been called and Write Buffer contains valid data
 
-    uint32_t max_count = mBufSz / (3*sizeof(uint32_t));
+    uint32_t maxCount = mBufSz / RECORD_TIMER_ENTRY_SZ_IN_BYTES;
     // Each record timer entry has 32bit ID and 32bit AIE High Timer + 32bit AIE Low Timer value.
 
-    uint32_t numEntries = max_count;
+    uint32_t numEntries = maxCount;
     std::stringstream msg;
     msg << "A maximum of " << numEntries << " record can be accommodated in given buffer of bytes size 0x"
         << std::hex << mBufSz << std::dec << std::endl;
     xrt_core::message::send(xrt_core::message::severity_level::debug, "XRT", msg.str());
 
-    if (numEntries <= max_count) {
+    if (numEntries <= maxCount) {
       for (uint32_t i = 0 ; i < numEntries; i++) {
         boost::property_tree::ptree ptIdTS;
         uint32_t id = *ptr;

src/runtime_src/xdp/profile/plugin/ml_timeline/clientDev/ml_timeline.h

@@ -27,7 +27,7 @@ namespace xdp {
   {
     std::unique_ptr<ResultBOContainer> mResultBOHolder;
     public :
-      MLTimelineClientDevImpl(VPDatabase* dB);
+      MLTimelineClientDevImpl(VPDatabase* dB, uint32_t sz);
 
       ~MLTimelineClientDevImpl();
 

src/runtime_src/xdp/profile/plugin/ml_timeline/ml_timeline_impl.h

@@ -21,6 +21,9 @@
 
 namespace xdp {
 
+  // Each record timer entry has 32bit ID and 32bit AIE High Timer + 32bit AIE Low Timer value.
+  constexpr uint32_t RECORD_TIMER_ENTRY_SZ_IN_BYTES = 3*sizeof(uint32_t);
+
   class VPDatabase;
 
   class MLTimelineImpl
@@ -30,8 +33,9 @@ namespace xdp {
       uint32_t mBufSz;
 
     public:
-      MLTimelineImpl(VPDatabase* dB)
-        : db(dB)
+      MLTimelineImpl(VPDatabase* dB, uint32_t sz)
+        : db(dB),
+          mBufSz(sz)
       {}
 
       MLTimelineImpl() = delete;
@@ -40,11 +44,6 @@ namespace xdp {
 
       virtual void updateDevice(void*) = 0;
       virtual void finishflushDevice(void*, uint64_t) = 0;
-
-      void setBufSize(uint32_t sz)
-      {
-        mBufSz = sz;
-      }
   };
 
 }

src/runtime_src/xdp/profile/plugin/ml_timeline/ml_timeline_plugin.cpp

@@ -38,30 +38,47 @@ namespace xdp {
 
   uint32_t ParseMLTimelineBufferSizeConfig()
   {
-    uint32_t bufSz = 0x20000;
+    uint32_t bufSz = 0;
     std::string szCfgStr = xrt_core::config::get_ml_timeline_buffer_size();
     std::smatch subStr;
 
+    std::stringstream msg;
+
     const std::regex validSzRegEx("\\s*([0-9]+)\\s*(K|k|M|m|)\\s*");
     if (std::regex_match(szCfgStr, subStr, validSzRegEx)) {
+      uint32_t szKB = 0;
       try {
         if ("K" == subStr[2] || "k" == subStr[2]) {
-          bufSz = (uint32_t)std::stoull(subStr[1]) * uint_constants::one_kb;
+          szKB = (uint32_t)std::stoull(subStr[1]);
         } else if ("M" == subStr[2] || "m" == subStr[2]) {
-          bufSz = (uint32_t)std::stoull(subStr[1]) * uint_constants::one_mb;
+          // Convert to KB now
+          szKB = (uint32_t)std::stoull(subStr[1]) * uint_constants::one_kb;
+        }
+        if (0 != (szKB % RECORD_TIMER_ENTRY_SZ_IN_BYTES)) {
+          /* Adjust given ML Timeline Buffer Size for alignment to avoid incorrect reads when Host Buffer
+           * gets overwritten with excess record timer data.
+           */
+          uint32_t q = szKB / RECORD_TIMER_ENTRY_SZ_IN_BYTES;
+          // Round up to next 12KB aligned size 
+          szKB = (q + 1) * RECORD_TIMER_ENTRY_SZ_IN_BYTES;
+          bufSz = szKB * uint_constants::one_kb;
+          std::stringstream amsg;
+          amsg << "Adjusting given ML Timeline Buffer Size " << szCfgStr
+               << " to 0x" << std::hex << bufSz << std::dec << " (in bytes) for alignment." << std::endl;
+          xrt_core::message::send(xrt_core::message::severity_level::debug, "XRT", amsg.str());
+        } else if (szKB) {
+          bufSz = szKB * uint_constants::one_kb;
         }
-
       } catch (const std::exception &e) {
-        std::stringstream msg;
-        msg << "Invalid string specified for ML Timeline Buffer Size. Using default size of 128KB."
-            << e.what() << std::endl;
-        xrt_core::message::send(xrt_core::message::severity_level::warning, "XRT", msg.str());
+        msg << "Hit exception " << e.what() << ". ";
       }
-
-    } else {
-      xrt_core::message::send(xrt_core::message::severity_level::warning, "XRT",
-                "Invalid string specified for ML Timeline Buffer Size. Using default size of 128KB.");
     }
+    if (0 == bufSz) {
+      bufSz = 0x30000;
+      msg << "Invalid string " << szCfgStr << " specified for ML Timeline Buffer Size. Using default size of 192KB." << std::endl;
+      xrt_core::message::send(xrt_core::message::severity_level::warning, "XRT", msg.str());
+    }
+
     return bufSz;
   }
 
@@ -117,9 +134,8 @@ namespace xdp {
     (db->getStaticInfo()).updateDeviceClient(deviceId, coreDevice, false);
     (db->getStaticInfo()).setDeviceName(deviceId, winDeviceName);
 
-    mMultiImpl[hwCtxImpl] = std::make_pair(implId, std::make_unique<MLTimelineClientDevImpl>(db));
+    mMultiImpl[hwCtxImpl] = std::make_pair(implId, std::make_unique<MLTimelineClientDevImpl>(db, mBufSz));
     auto mlImpl = mMultiImpl[hwCtxImpl].second.get();
-    mlImpl->setBufSize(mBufSz);
     mlImpl->updateDevice(hwCtxImpl);
 
 #endif