Add sum aggregation feature for cycles event.

0003-Add-sum-aggregation-feature-for-cycles-event.patch

@@ -0,0 +1,96 @@
+From 4a60eb0b1b90825e874c5fb60bd4b2e8f8285dea Mon Sep 17 00:00:00 2001
+From: l00822163 <liufeiyang6@huawei.com>
+Date: Mon, 11 Mar 2024 16:37:14 +0800
+Subject: [PATCH] Add sum aggregation feature for cycles event. Previously, PMU
+ event counts per line/function are aggregated by MAX operation since some PMU
+ events such as cache miss count requires an estimation of operand count for
+ memory access while SUM operation without the number of operands on that line
+ cannot render an average expression for operand count. However, when we use
+ the number of cycles to evaluate the importance of a function, we can
+ directly add up cycles within the function.
+
+---
+ instruction_map.cc | 19 +++++++++++++++----
+ profile.cc         | 10 +++++++++-
+ symbol_map.cc      |  3 +++
+ 3 files changed, 27 insertions(+), 5 deletions(-)
+
+diff --git a/instruction_map.cc b/instruction_map.cc
+index f460d67..768ab35 100644
+--- a/instruction_map.cc
++++ b/instruction_map.cc
+@@ -38,11 +38,22 @@ void InstructionMap::BuildPerFunctionInstructionMap(
+     addr2line_->GetInlineStack(addr, &info->source_stack);
+     inst_map_.insert(InstMap::value_type(addr, info));
+     if (info->source_stack.size() > 0) {
+-      symbol_map_->AddSourceCount(name, info->source_stack, 0, 1,
+-                                  SymbolMap::MAX, std::string(), is_repeat);
+-      for (const auto& event_name : symbol_map_->GetProcessingEventName()) {
++      if (symbol_map_->GetProcessingEventName().empty()) {
++        // For perf sampling with single PMU event, even if we collect "cycles"
++        // info, we will not change operation from "MAX" to "SUM".
+         symbol_map_->AddSourceCount(name, info->source_stack, 0, 1,
+-                                    SymbolMap::MAX, event_name, is_repeat);
++                                    SymbolMap::MAX, std::string(), is_repeat);
++      } else {
++        for (const auto& event_name : symbol_map_->GetProcessingEventName()) {
++          SymbolMap::Operation op = SymbolMap::MAX;
++          if (event_name.find("cycles") == 0) {
++            // When we collect "cycles" info by perf, switch aggregation
++            // operation from "MAX" to "SUM".
++            op = SymbolMap::SUM;
++          }
++          symbol_map_->AddSourceCount(name, info->source_stack, 0, 1,
++                                      op, event_name, is_repeat);
++        }
+       }
+     }
+   }
+diff --git a/profile.cc b/profile.cc
+index a306562..e563c5e 100644
+--- a/profile.cc
++++ b/profile.cc
+@@ -146,10 +146,16 @@ void Profile::ProcessPerFunctionProfile(string func_name,
+           continue;
+         }
+         if (info->source_stack.size() > 0) {
++          SymbolMap::Operation op = SymbolMap::MAX;
++          if (event_address.first.find("cycles") == 0) {
++            // When we collect "cycles" info by perf, switch aggregation
++            // operation from "MAX" to "SUM".
++            op = SymbolMap::SUM;
++          }
+           symbol_map_->AddSourceCount(
+               func_name, info->source_stack,
+               address_count.second * info->source_stack[0].DuplicationFactor(), 0,
+-              SymbolMap::MAX, event_address.first, is_repeat);
++              op, event_address.first, is_repeat);
+         }
+       }
+     }
+@@ -165,6 +171,8 @@ void Profile::ProcessPerFunctionProfile(string func_name,
+         continue;
+       }
+       if (info->source_stack.size() > 0) {
++        // For perf sampling with single PMU event, even if we collect "cycles"
++        // info, we will not change operation from "MAX" to "SUM".
+         symbol_map_->AddSourceCount(
+             func_name, info->source_stack,
+             address_count.second * info->source_stack[0].DuplicationFactor(), 0,
+diff --git a/symbol_map.cc b/symbol_map.cc
+index 233bbb2..322439f 100644
+--- a/symbol_map.cc
++++ b/symbol_map.cc
+@@ -398,6 +398,9 @@ void SymbolMap::AddSourceCount(const string &symbol_name,
+     }
+   } else if (op == SUM) {
+     symbol->pos_counts[offset].count += count;
++    if (!event_name.empty()) {
++      symbol->event_pos_counts[event_name][offset].count += count;
++    }
+   } else {
+     LOG(FATAL) << "op not supported.";
+   }
+-- 
+2.33.0
+

autofdo.spec

@@ -1,6 +1,6 @@
 Name:           autofdo
 Version:        0.19
-Release:        2
+Release:        3
 Summary:        A tool to convert perf.data profile to AutoFDO profile
 License:        Apache-2.0
 URL:            https://github.com/google/autofdo
@@ -10,6 +10,7 @@ URL:            https://github.com/google/autofdo
 Source:         %{name}-%{version}.tar.xz
 Patch0:         0001-Multi-event-processing-support.patch
 Patch1:         0002-Arm-spe-parser-support.patch
+Patch2:		0003-Add-sum-aggregation-feature-for-cycles-event.patch
 
 BuildRequires:  gcc gcc-c++ libtool autoconf automake git elfutils-libelf-devel openssl-devel pkg-config
 Requires:	glibc openssl-libs elfutils libgcc libstdc++ zlib
@@ -42,6 +43,12 @@ make -j 1
 %exclude %{_bindir}/create_llvm_prof
 
 %changelog
+* Wed May 29 2024 yuzeyang <yuzeyang4@huawei.com> - 0.19-3
+- Type:enhancement
+- ID:NA
+- SUG:NA
+- DESC:Add sum aggregation feature for cycles event
+
 * Sun Feb 13 2022 liyancheng <412998149@qq.com> - 0.19-2
 - Type:enhancement
 - ID:NA