[libc] Allow benchmarking several implementations at the same time.

Next step is to generate an archive with all implementations and a header listing them all.

Differential Revision: https://reviews.llvm.org/D107336

libc/benchmarks/LibcMemoryBenchmark.cpp

@@ -104,11 +104,11 @@ void ParameterBatch::checkValid(const ParameterType &P) const {
             .concat(llvm::Twine(BufferSize)));
 }
 
-CopyHarness::CopyHarness()
+CopySetup::CopySetup()
     : ParameterBatch(2), SrcBuffer(ParameterBatch::BufferSize),
       DstBuffer(ParameterBatch::BufferSize) {}
 
-ComparisonHarness::ComparisonHarness()
+ComparisonSetup::ComparisonSetup()
     : ParameterBatch(2), LhsBuffer(ParameterBatch::BufferSize),
       RhsBuffer(ParameterBatch::BufferSize) {
   // The memcmp buffers always compare equal.
@@ -116,7 +116,7 @@ ComparisonHarness::ComparisonHarness()
   memset(RhsBuffer.begin(), 0xF, BufferSize);
 }
 
-SetHarness::SetHarness()
+SetSetup::SetSetup()
     : ParameterBatch(1), DstBuffer(ParameterBatch::BufferSize) {}
 
 } // namespace libc_benchmarks

libc/benchmarks/LibcMemoryBenchmark.h

@@ -185,19 +185,43 @@ struct ParameterBatch {
   std::vector<ParameterType> Parameters;
 };
 
+/// Memory function prototype and configuration.
+using MemcpyFunction = void *(*)(void *__restrict, const void *__restrict,
+                                 size_t);
+struct MemcpyConfiguration {
+  MemcpyFunction Function;
+  llvm::StringRef Name;
+};
+
+using MemsetFunction = void *(*)(void *, int, size_t);
+struct MemsetConfiguration {
+  MemsetFunction Function;
+  llvm::StringRef Name;
+};
+
+using BzeroFunction = void (*)(void *, size_t);
+struct BzeroConfiguration {
+  BzeroFunction Function;
+  llvm::StringRef Name;
+};
+
+using MemcmpFunction = int (*)(const void *, const void *, size_t);
+struct MemcmpConfiguration {
+  MemcmpFunction Function;
+  llvm::StringRef Name;
+};
+
 /// Provides source and destination buffers for the Copy operation as well as
 /// the associated size distributions.
-struct CopyHarness : public ParameterBatch {
+struct CopySetup : public ParameterBatch {
-  CopyHarness();
+  CopySetup();
 
   inline static const ArrayRef<MemorySizeDistribution> getDistributions() {
     return getMemcpySizeDistributions();
   }
 
-  inline void *Call(ParameterType Parameter,
+  inline void *Call(ParameterType Parameter, MemcpyFunction Memcpy) {
-                    void *(*memcpy)(void *__restrict, const void *__restrict,
+    return Memcpy(DstBuffer + Parameter.OffsetBytes,
-                                    size_t)) {
-    return memcpy(DstBuffer + Parameter.OffsetBytes,
                   SrcBuffer + Parameter.OffsetBytes, Parameter.SizeBytes);
   }
 
@@ -208,21 +232,20 @@ private:
 
 /// Provides destination buffer for the Set operation as well as the associated
 /// size distributions.
-struct SetHarness : public ParameterBatch {
+struct SetSetup : public ParameterBatch {
-  SetHarness();
+  SetSetup();
 
   inline static const ArrayRef<MemorySizeDistribution> getDistributions() {
     return getMemsetSizeDistributions();
   }
 
-  inline void *Call(ParameterType Parameter,
+  inline void *Call(ParameterType Parameter, MemsetFunction Memset) {
-                    void *(*memset)(void *, int, size_t)) {
+    return Memset(DstBuffer + Parameter.OffsetBytes,
-    return memset(DstBuffer + Parameter.OffsetBytes,
                   Parameter.OffsetBytes % 0xFF, Parameter.SizeBytes);
   }
 
-  inline void *Call(ParameterType Parameter, void (*bzero)(void *, size_t)) {
+  inline void *Call(ParameterType Parameter, BzeroFunction Bzero) {
-    bzero(DstBuffer + Parameter.OffsetBytes, Parameter.SizeBytes);
+    Bzero(DstBuffer + Parameter.OffsetBytes, Parameter.SizeBytes);
     return DstBuffer.begin();
   }
 
@@ -232,16 +255,15 @@ private:
 
 /// Provides left and right buffers for the Comparison operation as well as the
 /// associated size distributions.
-struct ComparisonHarness : public ParameterBatch {
+struct ComparisonSetup : public ParameterBatch {
-  ComparisonHarness();
+  ComparisonSetup();
 
   inline static const ArrayRef<MemorySizeDistribution> getDistributions() {
     return getMemcmpSizeDistributions();
   }
 
-  inline int Call(ParameterType Parameter,
+  inline int Call(ParameterType Parameter, MemcmpFunction Memcmp) {
-                  int (*memcmp)(const void *, const void *, size_t)) {
+    return Memcmp(LhsBuffer + Parameter.OffsetBytes,
-    return memcmp(LhsBuffer + Parameter.OffsetBytes,
                   RhsBuffer + Parameter.OffsetBytes, Parameter.SizeBytes);
   }
 

libc/benchmarks/LibcMemoryGoogleBenchmarkMain.cpp

@@ -2,10 +2,26 @@
 #include "LibcMemoryBenchmark.h"
 #include "MemorySizeDistributions.h"
 #include "benchmark/benchmark.h"
+#include "llvm/ADT/ArrayRef.h"
+#include "llvm/ADT/Twine.h"
+#include <chrono>
 #include <cstdint>
 #include <random>
 #include <vector>
 
+using llvm::Align;
+using llvm::ArrayRef;
+using llvm::Twine;
+using llvm::libc_benchmarks::BzeroConfiguration;
+using llvm::libc_benchmarks::ComparisonSetup;
+using llvm::libc_benchmarks::CopySetup;
+using llvm::libc_benchmarks::MemcmpConfiguration;
+using llvm::libc_benchmarks::MemcpyConfiguration;
+using llvm::libc_benchmarks::MemorySizeDistribution;
+using llvm::libc_benchmarks::MemsetConfiguration;
+using llvm::libc_benchmarks::OffsetDistribution;
+using llvm::libc_benchmarks::SetSetup;
+
 namespace __llvm_libc {
 
 extern void *memcpy(void *__restrict, const void *__restrict, size_t);
@@ -15,84 +31,92 @@ extern int memcmp(const void *, const void *, size_t);
 
 } // namespace __llvm_libc
 
-using llvm::Align;
+// List of implementations to test.
-using llvm::ArrayRef;
+static constexpr MemcpyConfiguration kMemcpyConfigurations[] = {
-using llvm::libc_benchmarks::ComparisonHarness;
+    {__llvm_libc::memcpy, "__llvm_libc::memcpy"}};
-using llvm::libc_benchmarks::CopyHarness;
-using llvm::libc_benchmarks::MemorySizeDistribution;
-using llvm::libc_benchmarks::OffsetDistribution;
-using llvm::libc_benchmarks::SetHarness;
 
+static constexpr MemcmpConfiguration kMemcmpConfigurations[] = {
+    {__llvm_libc::memcmp, "__llvm_libc::memcmp"}};
+
+static constexpr MemsetConfiguration kMemsetConfigurations[] = {
+    {__llvm_libc::memset, "__llvm_libc::memset"}};
+
+static constexpr BzeroConfiguration kBzeroConfigurations[] = {
+    {__llvm_libc::bzero, "__llvm_libc::bzero"}};
+
+// Alignment to use for when accessing the buffers.
 static constexpr Align kBenchmarkAlignment = Align::Constant<1>();
 
-template <typename Harness> struct Randomized : public Harness {
+static std::mt19937_64 &getGenerator() {
-  Randomized(benchmark::State &State)
+  static std::mt19937_64 Generator(
-      : State(State), Distribution(Harness::getDistributions()[State.range(0)]),
+      std::chrono::system_clock::now().time_since_epoch().count());
+  return Generator;
+}
+
+template <typename SetupType, typename ConfigurationType> struct Runner {
+  Runner(benchmark::State &S, llvm::ArrayRef<ConfigurationType> Configurations)
+      : State(S), Distribution(SetupType::getDistributions()[State.range(0)]),
         Probabilities(Distribution.Probabilities),
         SizeSampler(Probabilities.begin(), Probabilities.end()),
-        OffsetSampler(Harness::BufferSize, Probabilities.size() - 1,
+        OffsetSampler(Setup.BufferSize, Probabilities.size() - 1,
-                      kBenchmarkAlignment) {
+                      kBenchmarkAlignment),
-    for (auto &P : Harness::Parameters) {
+        Configuration(Configurations[State.range(1)]) {
-      P.OffsetBytes = OffsetSampler(Gen);
+    for (auto &P : Setup.Parameters) {
-      P.SizeBytes = SizeSampler(Gen);
+      P.OffsetBytes = OffsetSampler(getGenerator());
-      Harness::checkValid(P);
+      P.SizeBytes = SizeSampler(getGenerator());
+      Setup.checkValid(P);
     }
   }
 
-  ~Randomized() {
+  ~Runner() {
-    const size_t AvgBytesPerIteration =
+    const size_t AvgBytesPerIteration = Setup.getBatchBytes() / Setup.BatchSize;
-        Harness::getBatchBytes() / Harness::BatchSize;
     const size_t TotalBytes = State.iterations() * AvgBytesPerIteration;
     State.SetBytesProcessed(TotalBytes);
-    State.SetLabel(Distribution.Name.str());
+    State.SetItemsProcessed(State.iterations());
+    State.SetLabel((Twine(Configuration.Name) + "," + Distribution.Name).str());
     State.counters["bytes_per_cycle"] = benchmark::Counter(
         TotalBytes / benchmark::CPUInfo::Get().cycles_per_second,
         benchmark::Counter::kIsRate);
   }
 
-  template <typename Function> inline void runBatch(Function foo) {
+  inline void runBatch() {
-    for (const auto &P : Harness::Parameters)
+    for (const auto &P : Setup.Parameters)
-      benchmark::DoNotOptimize(Harness::Call(P, foo));
+      benchmark::DoNotOptimize(Setup.Call(P, Configuration.Function));
   }
 
+  size_t getBatchSize() const { return Setup.BatchSize; }
+
 private:
+  SetupType Setup;
   benchmark::State &State;
-  Harness UP;
   MemorySizeDistribution Distribution;
   ArrayRef<double> Probabilities;
   std::discrete_distribution<unsigned> SizeSampler;
   OffsetDistribution OffsetSampler;
-  std::mt19937_64 Gen;
+  ConfigurationType Configuration;
 };
 
-template <typename Harness> static int64_t getMaxIndex() {
+#define BENCHMARK_MEMORY_FUNCTION(BM_NAME, SETUP, CONFIGURATION_TYPE,          \
-  return Harness::getDistributions().size() - 1;
+                                  CONFIGURATION_ARRAY_REF)                     \
-}
+  void BM_NAME(benchmark::State &State) {                                      \
+    Runner<SETUP, CONFIGURATION_TYPE> Setup(State, CONFIGURATION_ARRAY_REF);   \
+    const size_t BatchSize = Setup.getBatchSize();                             \
+    while (State.KeepRunningBatch(BatchSize))                                  \
+      Setup.runBatch();                                                        \
+  }                                                                            \
+  BENCHMARK(BM_NAME)->Apply([](benchmark::internal::Benchmark *benchmark) {    \
+    const int64_t DistributionSize = SETUP::getDistributions().size();         \
+    const int64_t ConfigurationSize = CONFIGURATION_ARRAY_REF.size();          \
+    for (int64_t DistIndex = 0; DistIndex < DistributionSize; ++DistIndex)     \
+      for (int64_t ConfIndex = 0; ConfIndex < ConfigurationSize; ++ConfIndex)  \
+        benchmark->Args({DistIndex, ConfIndex});                               \
+  })
 
-void BM_Memcpy(benchmark::State &State) {
+BENCHMARK_MEMORY_FUNCTION(BM_Memcpy, CopySetup, MemcpyConfiguration,
-  Randomized<CopyHarness> Harness(State);
+                          llvm::makeArrayRef(kMemcpyConfigurations));
-  while (State.KeepRunningBatch(Harness.BatchSize))
+BENCHMARK_MEMORY_FUNCTION(BM_Memcmp, ComparisonSetup, MemcmpConfiguration,
-    Harness.runBatch(__llvm_libc::memcpy);
+                          llvm::makeArrayRef(kMemcmpConfigurations));
-}
+BENCHMARK_MEMORY_FUNCTION(BM_Memset, SetSetup, MemsetConfiguration,
-BENCHMARK(BM_Memcpy)->DenseRange(0, getMaxIndex<CopyHarness>());
+                          llvm::makeArrayRef(kMemsetConfigurations));
-
+BENCHMARK_MEMORY_FUNCTION(BM_Bzero, SetSetup, BzeroConfiguration,
-void BM_Memcmp(benchmark::State &State) {
+                          llvm::makeArrayRef(kBzeroConfigurations));
-  Randomized<ComparisonHarness> Harness(State);
-  while (State.KeepRunningBatch(Harness.BatchSize))
-    Harness.runBatch(__llvm_libc::memcmp);
-}
-BENCHMARK(BM_Memcmp)->DenseRange(0, getMaxIndex<ComparisonHarness>());
-
-void BM_Memset(benchmark::State &State) {
-  Randomized<SetHarness> Harness(State);
-  while (State.KeepRunningBatch(Harness.BatchSize))
-    Harness.runBatch(__llvm_libc::memset);
-}
-BENCHMARK(BM_Memset)->DenseRange(0, getMaxIndex<SetHarness>());
-
-void BM_Bzero(benchmark::State &State) {
-  Randomized<SetHarness> Harness(State);
-  while (State.KeepRunningBatch(Harness.BatchSize))
-    Harness.runBatch(__llvm_libc::bzero);
-}
-BENCHMARK(BM_Bzero)->DenseRange(0, getMaxIndex<SetHarness>());