algo::TimingsFormat: Improve handling of subtimers.

algo/base/util/TimingsFormat.cxx

@@ -21,7 +21,11 @@ namespace cbm::algo
       fSS    = std::stringstream();
     }
 
-    void Title(std::string_view title) { fSS << fmt::format("{:<{}}\n", title, fAlign); }
+    void Title(std::string_view title)
+    {
+      Indent();
+      fSS << fmt::format("{:<{}}\n", title, fAlign);
+    }
 
     std::string Finalize() { return fSS.str(); }
 
@@ -35,13 +39,20 @@ namespace cbm::algo
       Measurement("Memset", t.memset(), t.throughput_memset());
       NewLine();
 
+      // Merge subtimers with identical names
+      // Useful eg in unpacking, where unpacker might be called multiple times per TS
+      std::unordered_map<std::string, xpu::timings> subtimers;
       for (xpu::timings& st : t.children()) {
-        if (st.kernels().empty()) {
-          Measurement(st.name(), st.wall(), st.throughput());
+        subtimers[std::string(st.name())].merge(st);
+      }
+
+      for (auto& [name, st] : subtimers) {
+        if (st.kernels().empty() && st.children().empty()) {
+          Measurement(name, st.wall(), st.throughput());
           NewLine();
         }
         else {
-          Title(st.name());
+          Title(name);
           Fmt(st);
           NewLine();
         }
@@ -52,8 +63,10 @@ namespace cbm::algo
         NewLine();
       }
 
-      Measurement("Kernel time", t.kernel_time(), t.throughput_kernels());
-      NewLine();
+      if (!t.kernels().empty()) {
+        Measurement("Kernel time", t.kernel_time(), t.throughput_kernels());
+        NewLine();
+      }
       Measurement("Wall time", t.wall(), t.throughput());
       fIndent -= 2;
     }
@@ -88,7 +101,7 @@ namespace cbm::algo
   private:
     void Measurement(std::string_view name, f64 time, f64 throughput)
     {
-      fSS << std::setw(fIndent) << std::setfill(' ') << std::right << "";
+      Indent();
       fSS << std::setw(fAlign) << std::setfill(' ') << std::left << fmt::format("{}:", name);
       Real(time, 10, 3, "ms");
       if (std::isnormal(throughput)) {
@@ -113,6 +126,8 @@ namespace cbm::algo
       }
     }
 
+    void Indent() { fSS << std::setw(fIndent) << std::setfill(' ') << std::left << ""; }
+
     size_t fAlign  = 0;
     size_t fIndent = 0;
     std::stringstream fSS;