[PROTON] Reduce msgpack serialization overhead-1

third_party/proton/csrc/include/Data/Metric.h

@@ -13,6 +13,7 @@
 #include <set>
 #include <shared_mutex>
 #include <stdexcept>
+#include <string_view>
 #include <type_traits>
 #include <utility>
 #include <variant>
@@ -65,7 +66,7 @@ class Metric {
 
   virtual const std::string &getName() const = 0;
 
-  virtual const std::string &getValueName(int valueId) const = 0;
+  virtual std::string_view getValueName(int valueId) const = 0;
 
   virtual bool isProperty(int valueId) const = 0;
 
@@ -81,8 +82,8 @@ class Metric {
     if (values[valueId].index() != value.index()) {
       throw makeInvalidArgument(
           std::string("Metric value type mismatch for valueId ") +
-          std::to_string(valueId) + " (" + getValueName(valueId) + ")" +
-          ": current=" + getTypeNameForIndex(values[valueId].index()) +
+          std::to_string(valueId) + " (" + std::string(getValueName(valueId)) +
+          ")" + ": current=" + getTypeNameForIndex(values[valueId].index()) +
           ", new=" + getTypeNameForIndex(value.index()));
     }
     // Handle string and other values separately
@@ -105,7 +106,8 @@ class Metric {
                   throw makeInvalidArgument(
                       std::string("Vector metric size mismatch for "
                                   "valueId ") +
-                      std::to_string(valueId) + " (" + getValueName(valueId) +
+                      std::to_string(valueId) + " (" +
+                      std::string(getValueName(valueId)) +
                       "): current=" + std::to_string(currentValue.size()) +
                       ", new=" + std::to_string(otherValue.size()));
                 }
@@ -116,7 +118,8 @@ class Metric {
                 throw makeLogicError(
                     std::string("Metric aggregation not supported for "
                                 "valueId ") +
-                    std::to_string(valueId) + " (" + getValueName(valueId) +
+                    std::to_string(valueId) + " (" +
+                    std::string(getValueName(valueId)) +
                     "): type=" + getTypeNameForIndex(values[valueId].index()));
               }
             }
@@ -172,7 +175,8 @@ class FlexibleMetric : public Metric {
 
   const std::string &getName() const override { return name; }
 
-  const std::string &getValueName(int valueId) const override {
+  // Flexible metrics carry their name as per-instance state.
+  std::string_view getValueName(int valueId) const override {
     return valueName;
   }
 
@@ -219,7 +223,13 @@ class KernelMetric : public Metric {
 
   const std::string &getName() const override { return name; }
 
-  const std::string &getValueName(int valueId) const override {
+  // Fast path for callers that already know they are working with KernelMetric.
+  static constexpr std::string_view getValueName(kernelMetricKind valueId) {
+    return VALUE_NAMES[valueId];
+  }
+
+  // Virtual access used through the Metric interface.
+  std::string_view getValueName(int valueId) const override {
     return VALUE_NAMES[valueId];
   }
 
@@ -232,7 +242,7 @@ class KernelMetric : public Metric {
       true, true, false, false, true, true, true, true};
   const static inline bool EXCLUSIVE[kernelMetricKind::Count] = {
       false, false, false, false, true, true, true, true};
-  const static inline std::string VALUE_NAMES[kernelMetricKind::Count] = {
+  static constexpr std::string_view VALUE_NAMES[kernelMetricKind::Count] = {
       "start_time (ns)", "end_time (ns)", "count",     "time (ns)",
       "device_id",       "device_type",   "stream_id", "is_metric_kernel",
   };
@@ -278,14 +288,21 @@ class PCSamplingMetric : public Metric {
 
   const std::string &getName() const override { return name; }
 
-  const std::string &getValueName(int valueId) const override {
+  // Fast path for callers that already know they are working with
+  // PCSamplingMetric.
+  static constexpr std::string_view getValueName(PCSamplingMetricKind valueId) {
+    return VALUE_NAMES[valueId];
+  }
+
+  // Virtual access used through the Metric interface.
+  std::string_view getValueName(int valueId) const override {
     return VALUE_NAMES[valueId];
   }
 
   bool isProperty(int valueId) const override { return false; }
   bool isExclusive(int valueId) const override { return false; }
 
-  const static inline std::string VALUE_NAMES[PCSamplingMetricKind::Count] = {
+  static constexpr std::string_view VALUE_NAMES[PCSamplingMetricKind::Count] = {
       "num_samples",
       "num_stalled_samples",
       "stalled_branch_resolving",
@@ -359,7 +376,13 @@ class CycleMetric : public Metric {
 
   const std::string &getName() const override { return name; }
 
-  const std::string &getValueName(int valueId) const override {
+  // Fast path for callers that already know they are working with CycleMetric.
+  static constexpr std::string_view getValueName(CycleMetricKind valueId) {
+    return VALUE_NAMES[valueId];
+  }
+
+  // Virtual access used through the Metric interface.
+  std::string_view getValueName(int valueId) const override {
     return VALUE_NAMES[valueId];
   }
 
@@ -374,7 +397,7 @@ class CycleMetric : public Metric {
   const static inline bool EXCLUSIVE[CycleMetricKind::Count] = {
       false, false, true, true, true,  true,  true, true,
       true,  true,  true, true, false, false, false};
-  const static inline std::string VALUE_NAMES[CycleMetricKind::Count] = {
+  static constexpr std::string_view VALUE_NAMES[CycleMetricKind::Count] = {
       "start_cycle", "end_cycle",      "cycles",         "normalized_cycles",
       "kernel_id",   "kernel_name",    "block_id",       "processor_id",
       "unit_id",     "device_id",      "device_type",    "time_shift_cost",

third_party/proton/csrc/include/Profiler/Graph.h

@@ -9,7 +9,6 @@
 #include <map>
 #include <memory>
 #include <mutex>
-#include <optional>
 #include <set>
 #include <shared_mutex>
 #include <unordered_map>
@@ -45,6 +44,7 @@ struct GraphState {
   // Capture tag to identify captured call paths
   static constexpr const char *captureTag = "<captured_at>";
   static constexpr const char *metricTag = "<metric>";
+
   struct NodeState {
     // The graph node id for this node
     uint64_t nodeId{};

third_party/proton/csrc/include/Utility/MsgPackWriter.h

@@ -3,6 +3,7 @@
 
 #include <cstddef>
 #include <cstdint>
+#include <cstring>
 #include <string_view>
 #include <vector>
 
@@ -21,6 +22,17 @@ class MsgPackWriter {
   void packInt(int64_t value);
   void packDouble(double value);
   void packStr(std::string_view value);
+  template <size_t N> void packFixStrLiteral(const char (&value)[N]) {
+    static_assert(N > 0);
+    constexpr uint32_t size = static_cast<uint32_t>(N - 1);
+    // MsgPack fixstr stores the string length in 5 bits, so literals must fit
+    // in the 0..31 byte range.
+    static_assert(size <= 31);
+    out.push_back(static_cast<uint8_t>(0xa0 | size));
+    const auto offset = out.size();
+    out.resize(offset + size);
+    std::memcpy(out.data() + offset, value, size);
+  }
   void packArray(uint32_t size);
   void packMap(uint32_t size);