Some updates to C++ benchmark (#708)

- Fix whitespace in usage text
- Fix memory usage output on MacOS
- Add readme

benchmark/c/options.cpp

@@ -23,7 +23,7 @@ namespace {
   s << "Usage: " << program_name << " -i <model path> <other options>\n"
     << "  Options:\n"
     << "    -i,--input_folder <path>\n"
-    << "      Path to the ONNX model directory to benchmark, compatible with onnxruntime-genai.\n "
+    << "      Path to the ONNX model directory to benchmark, compatible with onnxruntime-genai.\n"
     << "    -b,--batch_size <number>\n"
     << "      Number of sequences to generate in parallel. Default: " << defaults.batch_size << "\n"
     << "    -l,--prompt_length <number>\n"

benchmark/c/posix/resource_utils.cpp

@@ -17,7 +17,13 @@ size_t GetPeakWorkingSetSizeInBytes() {
     throw std::runtime_error("getrusage failed with error code " + std::to_string(errno));
   }
 
-  return static_cast<size_t>(rusage.ru_maxrss * 1024L);
+#if defined(__APPLE__)
+  constexpr size_t kBytesPerMaxRssUnit = 1;
+#else
+  constexpr size_t kBytesPerMaxRssUnit = 1024;
+#endif
+
+  return static_cast<size_t>(rusage.ru_maxrss) * kBytesPerMaxRssUnit;
 }
 
 }  // namespace benchmark::utils

benchmark/c/readme.md

@@ -0,0 +1,15 @@
+# model_benchmark
+
+`model_benchmark` is an end-to-end benchmark program for ONNX Runtime GenAI models.
+It is written in C++ and built as part of the ONNX Runtime GenAI build (e.g., via [build.py](../../build.py)).
+
+It is an alternative to the [Python benchmark script](../python/benchmark_e2e.py) that can be run in environments where Python is not available.
+
+Example usage:
+```
+model_benchmark -i <path to model directory>
+```
+
+Run with `--help` to see information about additional options.
+
+Note: On some platforms, such as Android, you may need to set the environment variable `LD_LIBRARY_PATH` to the directory containing the onnxruntime shared library for `model_benchmark` to be able to run.