diff --git a/mace/core/testing/test_benchmark.cc b/mace/core/testing/test_benchmark.cc
index 4d158b25d99a84cc9efd071a3094adcb2ec8ca45..17f56d0cf811d91ea99576f9580789612c103a44 100644
--- a/mace/core/testing/test_benchmark.cc
+++ b/mace/core/testing/test_benchmark.cc
@@ -37,6 +37,18 @@ Benchmark::Benchmark(const char* name, void (*fn)(int, int, int))
   Register();
 }
 
+Benchmark* Benchmark::Arg(int x) {
+  CHECK(num_args_ == 1);
+  args_.push_back(std::make_pair(x, -1));
+  return this;
+}
+
+Benchmark* Benchmark::ArgPair(int x, int y) {
+  CHECK(num_args_ == 2);
+  args_.push_back(std::make_pair(x, y));
+  return this;
+}
+
 
 // Run all benchmarks
 void Benchmark::Run() {
@@ -44,32 +56,31 @@ void Benchmark::Run() {
 
   // Compute name width.
   int width = 10;
-  string name;
+  char name[100];
   for (auto b : *all_benchmarks) {
-    name = b->name_;
     for (auto arg : b->args_) {
-      name.resize(b->name_.size());
+      strcpy(name, b->name_.c_str());
       if (arg.first >= 0) {
-        name += "/" + arg.first;
+        sprintf(name, "%s/%d", name, arg.first);
         if (arg.second >= 0) {
-          name += "/" + arg.second;
+          sprintf(name, "%s/%d", name, arg.second);
         }
       }
 
-      width = std::max<int>(width, name.size());
+      width = std::max<int>(width, strlen(name));
     }
   }
 
   printf("%-*s %10s %10s\n", width, "Benchmark", "Time(ns)", "Iterations");
   printf("%s\n", string(width + 22, '-').c_str());
   for (auto b : *all_benchmarks) {
-    name = b->name_;
+
     for (auto arg : b->args_) {
-      name.resize(b->name_.size());
+      strcpy(name, b->name_.c_str());
       if (arg.first >= 0) {
-        name += "/" + arg.first;
+        sprintf(name, "%s/%d", name, arg.first);
         if (arg.second >= 0) {
-          name += "/" + arg.second;
+          sprintf(name, "%s/%d", name, arg.second);
         }
       }
 
@@ -89,7 +100,7 @@ void Benchmark::Run() {
                  (items_processed * 1e-6) / seconds);
         full_label += buf;
       }
-      printf("%-*s %10.0f %10d\t%s\n", width, name.c_str(),
+      printf("%-*s %10.0f %10d\t%s\n", width, name,
              seconds * 1e9 / iters, iters, full_label.c_str());
     }
   }
diff --git a/mace/core/testing/test_benchmark.h b/mace/core/testing/test_benchmark.h
index 896399415bfd3372df4ef7241a64ce8346fbad2e..44a352f54df4cae609b3955eb5343c9b78d34126 100644
--- a/mace/core/testing/test_benchmark.h
+++ b/mace/core/testing/test_benchmark.h
@@ -24,6 +24,8 @@ class Benchmark {
   Benchmark(const char* name, void (*fn)(int));
   Benchmark(const char* name, void (*fn)(int, int));
   Benchmark(const char* name, void (*fn)(int, int, int));
+  Benchmark* Arg(int x);
+  Benchmark* ArgPair(int x, int y);
 
   static void Run();
 
@@ -40,12 +42,10 @@ class Benchmark {
 };
 
 void RunBenchmarks();
-void SetLabel(const std::string& label);
 void BytesProcessed(int64);
 void ItemsProcessed(int64);
 void StartTiming();
 void StopTiming();
-void UseRealTime();
 
 }  // namespace testing
 }  // namespace mace
diff --git a/mace/examples/benchmark_example.cc b/mace/examples/benchmark_example.cc
index dc3b9c4a2f6e7423b0d27446832e2a3729d89159..106c6c3c4ccdf7dcc091a6c9f9bbc8c0c15d2611 100644
--- a/mace/examples/benchmark_example.cc
+++ b/mace/examples/benchmark_example.cc
@@ -25,3 +25,21 @@ static void foo(int iters) {
 BENCHMARK(foo);
 
 
+static void bar(int iters, int n) {
+  const int64 tot = static_cast<int64>(iters) * n;
+  mace::testing::ItemsProcessed(tot);
+  mace::testing::BytesProcessed(tot * (sizeof(float)));
+
+  float* inp = new float[n];
+  float* out = new float[n];
+
+  while (iters--) {
+    for (int i=0; i < n; i++) {
+      out[i] = inp[i] * 2.0;
+    }
+  }
+  delete[] inp;
+  delete[] out;
+}
+
+BENCHMARK(bar)->Arg(32)->Arg(64)->Arg(128);