add int8 benchmark, test=develop (#2338)

lite/api/benchmark.cc

@@ -32,7 +32,9 @@ DEFINE_string(input_shape,
 DEFINE_string(result_filename, "", "save test result");
 DEFINE_bool(run_model_optimize,
             false,
-            "apply model_optimize_tool to model, use optimized model to test");
+            "if set true, apply model_optimize_tool to model, use optimized "
+            "model to test");
+DEFINE_bool(is_quantized_model, false, "if set true, test the quantized model");
 
 namespace paddle {
 namespace lite_api {
@@ -42,12 +44,14 @@ void OutputOptModel(const std::string& load_model_dir,
                     const std::vector<std::vector<int64_t>>& input_shapes) {
   lite_api::CxxConfig config;
   config.set_model_dir(load_model_dir);
-  config.set_valid_places({
-      Place{TARGET(kARM), PRECISION(kFloat)},
-      Place{TARGET(kARM), PRECISION(kInt8)},
-      Place{TARGET(kX86), PRECISION(kFloat)},
-      Place{TARGET(kOpenCL), PRECISION(kFloat)},
-  });
+  std::vector<Place> vaild_places = {Place{TARGET(kARM), PRECISION(kFloat)},
+                                     Place{TARGET(kX86), PRECISION(kFloat)},
+                                     Place{TARGET(kOpenCL), PRECISION(kFloat)}};
+  if (FLAGS_is_quantized_model) {
+    vaild_places.insert(vaild_places.begin(),
+                        Place{TARGET(kARM), PRECISION(kInt8)});
+  }
+  config.set_valid_places(vaild_places);
   auto predictor = lite_api::CreatePaddlePredictor(config);
 
   int ret = system(

lite/tools/benchmark.sh

@@ -8,6 +8,7 @@ then
     echo "Usage:"
     echo "  sh benchmark.sh <benchmark_bin_path> <benchmark_models_path> <result_filename>"
     echo "  sh benchmark.sh <benchmark_bin_path> <benchmark_models_path> <result_filename> <is_run_model_optimize: [true|false]>"
+    echo "  sh benchmark.sh <benchmark_bin_path> <benchmark_models_path> <result_filename> <is_run_model_optimize: [true|false]> <is_run_quantized_model: [trur|false]>"
     exit
 fi
 
@@ -20,6 +21,7 @@ RESULT_FILENAME=$3
 WARMUP=10
 REPEATS=30
 IS_RUN_MODEL_OPTIMIZE=false
+IS_RUN_QUANTIZED_MODEL=false
 NUM_THREADS_LIST=(1 2 4)
 MODELS_LIST=$(ls $MODELS_DIR)
 
@@ -28,6 +30,10 @@ if [ $# -gt  3 ];
 then
     IS_RUN_MODEL_OPTIMIZE=$4
 fi
+if [ $# -gt  4 ];
+then
+    IS_RUN_QUANTIZED_MODEL=$5
+fi
 
 # Adb push benchmark_bin, models
 adb push $BENCHMARK_BIN $ANDROID_DIR/benchmark_bin
@@ -46,7 +52,8 @@ for threads in ${NUM_THREADS_LIST[@]}; do
                    --repeats=$REPEATS \
                    --threads=$threads \
                    --result_filename=$ANDROID_DIR/$RESULT_FILENAME \
-                   --run_model_optimize=$IS_RUN_MODEL_OPTIMIZE"
+                   --run_model_optimize=$IS_RUN_MODEL_OPTIMIZE \
+                   --is_quantized_model=$IS_RUN_QUANTIZED_MODEL"
     done
     adb shell "echo >> $ANDROID_DIR/$RESULT_FILENAME"
 done