Run each model benchmark at least 10x to amortize setup noise.

PiperOrigin-RevId: 708329672
google · Dec 27, 2024 · 72e4d09 · 72e4d09
1 parent c2bddb4
commit 72e4d09
Show file tree

Hide file tree

Showing 3 changed files with 80 additions and 62 deletions.
diff --git a/bench/models/benchmark.cc b/bench/models/benchmark.cc
@@ -5,14 +5,12 @@
 
 #include <benchmark/benchmark.h>
 
-#include <algorithm>
 #include <cassert>
 #include <cstddef>
 #include <cstdint>
 #include <cstdlib>
 #include <cstring>
 #include <functional>
-#include <iostream>
 #include <memory>
 #include <vector>
 
@@ -23,8 +21,9 @@
 #include "xnnpack/subgraph.h"
 #include "pthreadpool.h"
 
-int FLAGS_num_threads = 1;
-uint32_t FLAGS_xnn_runtime_flags = 0;
+namespace {
+static const size_t kMinIterations = 10;
+}  // namespace
 
 struct ModelRuntime {
   std::unique_ptr<xnn_subgraph, decltype(&xnn_delete_subgraph)> model;
@@ -116,11 +115,15 @@ static void BenchmarkInvoke(benchmark::State& state,
     return;
   }
 
-  for (auto _ : state) {
-    benchmark::utils::WipePthreadpoolL2Caches(state, model_runtime.threadpool);
-    if (!model_runtime.Invoke()) {
-      state.SkipWithError("failed to invoke runtime");
-      return;
+  while (state.KeepRunningBatch(kMinIterations)) {
+    for (int iter = 0; iter < kMinIterations; iter++) {
+      benchmark::utils::WipePthreadpoolL2Caches(state,
+                                                model_runtime.threadpool);
+
+      if (!model_runtime.Invoke()) {
+        state.SkipWithError("failed to invoke runtime");
+        return;
+      }
     }
   }
 
@@ -239,46 +242,4 @@ BENCHMARK(QD8Attention)
 
 BENCHMARK(QS8MobileNetV2)->Unit(benchmark::kMicrosecond)->UseRealTime();
 
-int ProcessArgs(int& argc, char**& argv) {
-  for (int i = 1; i < argc;) {
-    if (strncmp(argv[i], "--num_threads=", 14) == 0) {
-      FLAGS_num_threads = atoi(argv[i] + 14);
-      if (FLAGS_num_threads <= 0) {
-        std::cerr << "Invalid --num_threads: " << FLAGS_num_threads << "\n";
-        return 1;
-      }
-      std::copy(argv + i + 1, argv + argc, argv + i);
-      argc -= 1;
-    } else if (strncmp(argv[i], "--xnn_runtime_flags=", 20) == 0) {
-      const char* v = argv[i] + 20;
-      if (strlen(v) > 2 && strncmp(v, "0x", 2) == 0) {
-        FLAGS_xnn_runtime_flags = strtoul(v + 2, nullptr, 16);
-      } else {
-        FLAGS_xnn_runtime_flags = strtoul(v, nullptr, 10);
-      }
-      std::copy(argv + i + 1, argv + argc, argv + i);
-      argc -= 1;
-    } else {
-      ++i;
-    }
-  }
-  return 0;
-}
-
-#ifdef BENCHMARK_ARGS_BOTTLENECK
-// We are provided with a main that will call this function
-extern "C" {
-int BenchmarkArgBottleneck(int& argc, char**& argv) {
-  return ProcessArgs(argc, argv);
-}
-}
-#else
-int main(int argc, char** argv) {
-  ::benchmark::Initialize(&argc, argv);
-  int status = ProcessArgs(argc, argv);
-  if (status != 0) return status;
-  if (::benchmark::ReportUnrecognizedArguments(argc, argv)) return 1;
-  ::benchmark::RunSpecifiedBenchmarks();
-}
-#endif
-
+XNN_BENCHMARK_MAIN();
diff --git a/bench/utils.cc b/bench/utils.cc
@@ -3,33 +3,38 @@
 // This source code is licensed under the BSD-style license found in the
 // LICENSE file in the root directory of this source tree.
 
+#include "utils.h"
+
+#include <algorithm>
 #include <cstdint>
 #include <cstdio>
 #include <cstdlib>
 #include <cstring>
+#include <iostream>
 #include <mutex>
 
-#include "xnnpack/common.h"
-#include <benchmark/benchmark.h>
-#include "pthreadpool.h"
-
 #ifdef __linux__
-  #include <sched.h>
+#include <sched.h>
 #endif
 #if defined(__ANDROID__) || defined(_WIN32) || defined(__CYGWIN__)
-  #include <malloc.h>
+#include <malloc.h>
 #endif
 #if defined(__SSE__) || defined(__x86_64__)
-  #include <xmmintrin.h>
+#include <xmmintrin.h>
 #endif
 
 #if XNN_ENABLE_CPUINFO
-  #include <cpuinfo.h>
+#include <cpuinfo.h>
 #endif  // XNN_ENABLE_CPUINFO
 
+#include "xnnpack/common.h"
 #include "xnnpack/hardware-config.h"
+#include <benchmark/benchmark.h>
+#include "pthreadpool.h"
 
-#include "utils.h"
+// Common flags for all benchmarks.
+int FLAGS_num_threads = 1;
+uint32_t FLAGS_xnn_runtime_flags = 0;
 
 namespace benchmark {
 namespace utils {
@@ -88,6 +93,33 @@ void PthreadpoolClearL2Cache(void* context, size_t id) {
 
 };  // namespace
 
+int ProcessArgs(int& argc, char**& argv) {
+  for (int i = 1; i < argc;) {
+    if (strncmp(argv[i], "--num_threads=", 14) == 0) {
+      FLAGS_num_threads = atoi(argv[i] + 14);
+      if (FLAGS_num_threads <= 0) {
+        std::cerr << "Invalid --num_threads: " << FLAGS_num_threads << "\n";
+        return 1;
+      }
+      std::copy(argv + i + 1, argv + argc, argv + i);
+      argc -= 1;
+    } else if (strncmp(argv[i], "--xnn_runtime_flags=", 20) == 0) {
+      const char* v = argv[i] + 20;
+      if (strlen(v) > 2 && strncmp(v, "0x", 2) == 0) {
+        FLAGS_xnn_runtime_flags = strtoul(v + 2, nullptr, 16);
+      } else {
+        FLAGS_xnn_runtime_flags = strtoul(v, nullptr, 10);
+      }
+      std::copy(argv + i + 1, argv + argc, argv + i);
+      argc -= 1;
+    } else {
+      ++i;
+    }
+  }
+  // InitGoogle(...);
+  return 0;
+}
+
 uint32_t PrefetchToL1(const void* ptr, size_t size) {
   uint32_t step = 16;
   #if XNN_ENABLE_CPUINFO
@@ -154,7 +186,7 @@ void DisableDenormals() {
 #endif
 }
 
-// Return clockrate in Hz
+// Return clock rate in Hz.
 uint64_t GetCurrentCpuFrequency() {
 #ifdef __linux__
   int freq = 0;

diff --git a/bench/utils.h b/bench/utils.h
@@ -13,9 +13,34 @@
 #include <benchmark/benchmark.h>
 #include "pthreadpool.h"
 
+#ifdef BENCHMARK_ARGS_BOTTLENECK
+#define XNN_BENCHMARK_MAIN()                            \
+  extern "C" {                                          \
+  int BenchmarkArgBottleneck(int& argc, char**& argv) { \
+    return benchmark::utils::ProcessArgs(argc, argv);   \
+  }                                                     \
+  }
+#else
+#define XNN_BENCHMARK_MAIN()                                            \
+  int main(int argc, char** argv) {                                     \
+    ::benchmark::Initialize(&argc, argv);                               \
+    int status = benchmark::utils::ProcessArgs(argc, argv);             \
+    if (status != 0) return status;                                     \
+    if (::benchmark::ReportUnrecognizedArguments(argc, argv)) return 1; \
+    ::benchmark::RunSpecifiedBenchmarks();                              \
+  }                                                                     \
+  int main(int, char**)
+#endif  // BENCHMARK_ARGS_BOTTLENECK
+
+// Common flags for all benchmarks.
+extern int FLAGS_num_threads;
+extern uint32_t FLAGS_xnn_runtime_flags;
+
 namespace benchmark {
 namespace utils {
 
+int ProcessArgs(int& argc, char**& argv);
+
 uint32_t WipeCache();
 uint32_t PrefetchToL1(const void* ptr, size_t size);