Improve CPU Cache info reporting -- Add Windows support. (#486)

* Improve CPU Cache info reporting -- Add Windows support.

This patch does a couple of thing regarding CPU Cache reporting.

First, it adds an implementation on Windows. Second it fixes
the JSONReporter to correctly (and actually) output the CPU
configuration information.

And finally, third, it detects and reports the number of
physical CPU's that share the same cache.

include/benchmark/benchmark.h

@@ -1159,6 +1159,7 @@ struct CPUInfo {
     std::string type;
     int level;
     int size;
+    int num_sharing;
   };
 
   int num_cpus;

src/json_reporter.cc

@@ -87,6 +87,27 @@ bool JSONReporter::ReportContext(const Context& context) {
   out << indent << FormatKV("cpu_scaling_enabled", info.scaling_enabled)
       << ",\n";
 
+  out << indent << "\"caches\": [\n";
+  indent = std::string(6, ' ');
+  std::string cache_indent(8, ' ');
+  for (size_t i = 0; i < info.caches.size(); ++i) {
+    auto& CI = info.caches[i];
+    out << indent << "{\n";
+    out << cache_indent << FormatKV("type", CI.type) << ",\n";
+    out << cache_indent << FormatKV("level", static_cast<int64_t>(CI.level))
+        << ",\n";
+    out << cache_indent
+        << FormatKV("size", static_cast<int64_t>(CI.size) * 1000u) << ",\n";
+    out << cache_indent
+        << FormatKV("num_sharing", static_cast<int64_t>(CI.num_sharing))
+        << "\n";
+    out << indent << "}";
+    if (i != info.caches.size() - 1) out << ",";
+    out << "\n";
+  }
+  indent = std::string(4, ' ');
+  out << indent << "],\n";
+
 #if defined(NDEBUG)
   const char build_type[] = "release";
 #else

src/reporter.cc

@@ -45,7 +45,10 @@ void BenchmarkReporter::PrintBasicContext(std::ostream *out,
     Out << "CPU Caches:\n";
     for (auto &CInfo : info.caches) {
       Out << "  L" << CInfo.level << " " << CInfo.type << " "
-          << (CInfo.size / 1000) << "K\n";
+          << (CInfo.size / 1000) << "K";
+      if (CInfo.num_sharing != 0)
+        Out << " (x" << (info.num_cpus / CInfo.num_sharing) << ")";
+      Out << "\n";
     }
   }
 

src/sysinfo.cc

@@ -32,7 +32,10 @@
 #endif
 
 #include <algorithm>
+#include <array>
+#include <bitset>
 #include <cerrno>
+#include <climits>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
@@ -123,6 +126,15 @@ struct ValueUnion {
       return Buff->uint64_value;
     BENCHMARK_UNREACHABLE();
   }
+
+  template <class T, int N>
+  std::array<T, N> GetAsArray() {
+    const int ArrSize = sizeof(T) * N;
+    CHECK_LE(ArrSize, Size);
+    std::array<T, N> Arr;
+    std::memcpy(Arr.data(), data(), ArrSize);
+    return Arr;
+  }
 };
 
 #ifdef __GNUC__
@@ -158,6 +170,14 @@ bool GetSysctl(std::string const& Name, Tp* Out) {
   *Out = static_cast<Tp>(Buff.GetAsUnsigned());
   return true;
 }
+
+template <class Tp, size_t N>
+bool GetSysctl(std::string const& Name, std::array<Tp, N>* Out) {
+  auto Buff = GetSysctlImp(Name);
+  if (!Buff) return false;
+  *Out = Buff.GetAsArray<Tp, N>();
+  return true;
+};
 #endif
 
 template <class ArgT>
@@ -186,6 +206,25 @@ bool CpuScalingEnabled(int num_cpus) {
   return false;
 }
 
+int CountSetBitsInCPUMap(std::string Val) {
+  auto CountBits = [](std::string Part) {
+    using CPUMask = std::bitset<sizeof(std::uintptr_t) * CHAR_BIT>;
+    Part = "0x" + Part;
+    CPUMask Mask(std::stoul(Part, nullptr, 16));
+    return static_cast<int>(Mask.count());
+  };
+  size_t Pos;
+  int total = 0;
+  while ((Pos = Val.find(',')) != std::string::npos) {
+    total += CountBits(Val.substr(0, Pos));
+    Val = Val.substr(Pos + 1);
+  }
+  if (!Val.empty()) {
+    total += CountBits(Val);
+  }
+  return total;
+}
+
 BENCHMARK_MAYBE_UNUSED
 std::vector<CPUInfo::CacheInfo> GetCacheSizesFromKVFS() {
   std::vector<CPUInfo::CacheInfo> res;
@@ -214,6 +253,10 @@ std::vector<CPUInfo::CacheInfo> GetCacheSizesFromKVFS() {
       PrintErrorAndDie("Failed to read from file ", FPath, "type");
     if (!ReadFromFile(StrCat(FPath, "level"), &info.level))
       PrintErrorAndDie("Failed to read from file ", FPath, "level");
+    std::string map_str;
+    if (!ReadFromFile(StrCat(FPath, "shared_cpu_map"), &map_str))
+      PrintErrorAndDie("Failed to read from file ", FPath, "shared_cpu_map");
+    info.num_sharing = CountSetBitsInCPUMap(map_str);
     res.push_back(info);
   }
 
@@ -223,14 +266,18 @@ std::vector<CPUInfo::CacheInfo> GetCacheSizesFromKVFS() {
 #ifdef BENCHMARK_OS_MACOSX
 std::vector<CPUInfo::CacheInfo> GetCacheSizesMacOSX() {
   std::vector<CPUInfo::CacheInfo> res;
+  std::array<uint64_t, 4> CacheCounts{{0, 0, 0, 0}};
+  GetSysctl("hw.cacheconfig", &CacheCounts);
+
   struct {
     std::string name;
     std::string type;
     int level;
-  } Cases[] = {{"hw.l1dcachesize", "Data", 1},
-               {"hw.l1icachesize", "Instruction", 1},
-               {"hw.l2cachesize", "Unified", 2},
-               {"hw.l3cachesize", "Unified", 3}};
+    size_t num_sharing;
+  } Cases[] = {{"hw.l1dcachesize", "Data", 1, CacheCounts[1]},
+               {"hw.l1icachesize", "Instruction", 1, CacheCounts[1]},
+               {"hw.l2cachesize", "Unified", 2, CacheCounts[2]},
+               {"hw.l3cachesize", "Unified", 3, CacheCounts[3]}};
   for (auto& C : Cases) {
     int val;
     if (!GetSysctl(C.name, &val)) continue;
@@ -238,15 +285,67 @@ std::vector<CPUInfo::CacheInfo> GetCacheSizesMacOSX() {
     info.type = C.type;
     info.level = C.level;
     info.size = val;
+    info.num_sharing = static_cast<int>(C.num_sharing);
     res.push_back(std::move(info));
   }
   return res;
 }
+#elif defined(BENCHMARK_OS_WINDOWS)
+std::vector<CPUInfo::CacheInfo> GetCacheSizesWindows() {
+  std::vector<CPUInfo::CacheInfo> res;
+  DWORD buffer_size = 0;
+  using PInfo = SYSTEM_LOGICAL_PROCESSOR_INFORMATION;
+  using CInfo = CACHE_DESCRIPTOR;
+
+  using UPtr = std::unique_ptr<PInfo, decltype(&std::free)>;
+  GetLogicalProcessorInformation(nullptr, &buffer_size);
+  UPtr buff((PInfo*)malloc(buffer_size), &std::free);
+  if (!GetLogicalProcessorInformation(buff.get(), &buffer_size))
+    PrintErrorAndDie("Failed during call to GetLogicalProcessorInformation: ",
+                     GetLastError());
+
+  PInfo* it = buff.get();
+  PInfo* end = buff.get() + (buffer_size / sizeof(PInfo));
+
+  for (; it != end; ++it) {
+    if (it->Relationship != RelationCache) continue;
+    using BitSet = std::bitset<sizeof(ULONG_PTR) * CHAR_BIT>;
+    BitSet B(it->ProcessorMask);
+    // To prevent duplicates, only consider caches where CPU 0 is specified
+    if (!B.test(0)) continue;
+    CInfo* Cache = &it->Cache;
+    CPUInfo::CacheInfo C;
+    C.num_sharing = B.count();
+    C.level = Cache->Level;
+    C.size = Cache->Size;
+    switch (Cache->Type) {
+      case CacheUnified:
+        C.type = "Unified";
+        break;
+      case CacheInstruction:
+        C.type = "Instruction";
+        break;
+      case CacheData:
+        C.type = "Data";
+        break;
+      case CacheTrace:
+        C.type = "Trace";
+        break;
+      default:
+        C.type = "Unknown";
+        break;
+    }
+    res.push_back(C);
+  }
+  return res;
+}
 #endif
 
 std::vector<CPUInfo::CacheInfo> GetCacheSizes() {
 #ifdef BENCHMARK_OS_MACOSX
   return GetCacheSizesMacOSX();
+#elif defined(BENCHMARK_OS_WINDOWS)
+  return GetCacheSizesWindows();
 #else
   return GetCacheSizesFromKVFS();
 #endif

test/reporter_output_test.cc

@@ -13,6 +13,41 @@ ADD_CASES(TC_ConsoleOut,
           {{"^[-]+$", MR_Next},
            {"^Benchmark %s Time %s CPU %s Iterations$", MR_Next},
            {"^[-]+$", MR_Next}});
+static int AddContextCases() {
+  AddCases(TC_ConsoleErr,
+           {
+               {"%int[-/]%int[-/]%int %int:%int:%int$", MR_Default},
+               {"Run on \\(%int X %float MHz CPU s\\)", MR_Next},
+           });
+  AddCases(TC_JSONOut, {{"^\\{", MR_Default},
+                        {"\"context\":", MR_Next},
+                        {"\"date\": \"", MR_Next},
+                        {"\"num_cpus\": %int,$", MR_Next},
+                        {"\"mhz_per_cpu\": %float,$", MR_Next},
+                        {"\"cpu_scaling_enabled\": ", MR_Next},
+                        {"\"caches\": \\[$", MR_Next}});
+  auto const& Caches = benchmark::CPUInfo::Get().caches;
+  if (!Caches.empty()) {
+    AddCases(TC_ConsoleErr, {{"CPU Caches:$", MR_Next}});
+  }
+  for (size_t I = 0; I < Caches.size(); ++I) {
+    std::string num_caches_str =
+        Caches[I].num_sharing != 0 ? " \\(x%int\\)$" : "$";
+    AddCases(
+        TC_ConsoleErr,
+        {{"L%int (Data|Instruction|Unified) %intK" + num_caches_str, MR_Next}});
+    AddCases(TC_JSONOut, {{"\\{$", MR_Next},
+                          {"\"type\": \"", MR_Next},
+                          {"\"level\": %int,$", MR_Next},
+                          {"\"size\": %int,$", MR_Next},
+                          {"\"num_sharing\": %int$", MR_Next},
+                          {"}[,]{0,1}$", MR_Next}});
+  }
+
+  AddCases(TC_JSONOut, {{"],$"}});
+  return 0;
+}
+int dummy_register = AddContextCases();
 ADD_CASES(TC_CSVOut, {{"%csv_header"}});
 
 // ========================================================================= //