Feature: Add capability information for mace_run

1. add capability show information for mace_run
2. remove flatten_atrous_conv of tensorflow transform_graph

mace/libmace/mace_version_script.lds

@@ -7,6 +7,7 @@ mace {
     *CreateMaceEngineFromProto*;
     *GetBigLittleCoreIDs*;
     *MaceVersion*;
+    *GetCapability*;
 
     # api for static library of models
     *mace*logging*LogMessage*;

mace/public/mace.h

@@ -102,7 +102,7 @@ class RunMetadata {
 
 /// Consistent with Android NNAPI
 struct PerformanceInfo {
-  // Time of executing some workload.
+  // Time of executing some workload(millisecond).
   // negative value for unsupported.
   float exec_time;
 };

mace/python/tools/converter_tool/tensorflow_converter.py

@@ -123,39 +123,16 @@ TFOpType = Enum('TFOpType', [(op, op) for op in TFSupportedOps], type=str)
 
 TFSupportedOps = [six.b(op) for op in TFSupportedOps]
 
-TFTransformGraphOptions = {
-    base_converter.DeviceType.CPU.value: [
-        'strip_unused_nodes',
-        'remove_nodes(op=Identity, op=CheckNumerics)',
-        'fold_constants(ignore_errors=true)',
-        'fold_batch_norms',
-        'fold_old_batch_norms',
-        'remove_control_dependencies',
-        'strip_unused_nodes',
-        'sort_by_execution_order'
-    ],
-    base_converter.DeviceType.GPU.value: [
-        'strip_unused_nodes',
-        'remove_nodes(op=Identity, op=CheckNumerics)',
-        'fold_constants(ignore_errors=true)',
-        'flatten_atrous_conv',
-        'fold_batch_norms',
-        'fold_old_batch_norms',
-        'remove_control_dependencies',
-        'strip_unused_nodes',
-        'sort_by_execution_order'
-    ],
-    base_converter.DeviceType.HEXAGON.value: [
-        'strip_unused_nodes',
-        'remove_nodes(op=Identity, op=CheckNumerics)',
-        'fold_constants(ignore_errors=true)',
-        'fold_batch_norms',
-        'fold_old_batch_norms',
-        'remove_control_dependencies',
-        'strip_unused_nodes',
-        'sort_by_execution_order'
-    ]
-}
+TFTransformGraphOptions = [
+    'strip_unused_nodes',
+    'remove_nodes(op=Identity, op=CheckNumerics)',
+    'fold_constants(ignore_errors=true)',
+    'fold_batch_norms',
+    'fold_old_batch_norms',
+    'remove_control_dependencies',
+    'strip_unused_nodes',
+    'sort_by_execution_order'
+]
 
 
 class TensorflowConverter(base_converter.ConverterInterface):
@@ -289,15 +266,13 @@ class TensorflowConverter(base_converter.ConverterInterface):
 
         self._placeholders = {}
 
-        print("Run transform_graph: %s" % TFTransformGraphOptions[
-            option.device])
+        print("Run transform_graph: %s" % TFTransformGraphOptions)
         try:
             print("output keys: ", option.output_nodes.keys())
             transformed_graph_def = TransformGraph(tf_graph_def,
                                                    option.input_nodes.keys(),
                                                    option.output_nodes.keys(),
-                                                   TFTransformGraphOptions[
-                                                       option.device])
+                                                   TFTransformGraphOptions)
         except Exception as ex:
             print("Failed to transform graph using tf tool: %s" % ex)
             transformed_graph_def = tf_graph_def

mace/tools/validation/mace_run.cc

@@ -225,7 +225,8 @@ bool RunModel(const std::string &model_name,
               const std::vector<DataFormat> &input_data_formats,
               const std::vector<std::string> &output_names,
               const std::vector<std::vector<int64_t>> &output_shapes,
-              const std::vector<DataFormat> &output_data_formats) {
+              const std::vector<DataFormat> &output_data_formats,
+              float cpu_capability) {
   DeviceType device_type = ParseDeviceType(FLAGS_device);
 
   int64_t t0 = NowMicros();
@@ -446,11 +447,11 @@ bool RunModel(const std::string &model_name,
   }
 
   // Metrics reporting tools depends on the format, keep in consistent
-  printf("========================================\n");
-  printf("            init      warmup     run_avg\n");
-  printf("========================================\n");
-  printf("time %11.3f %11.3f %11.3f\n",
-         init_millis, warmup_millis, model_run_millis);
+  printf("========================================================\n");
+  printf("     capability(CPU)        init      warmup     run_avg\n");
+  printf("========================================================\n");
+  printf("time %15.3f %11.3f %11.3f %11.3f\n",
+         cpu_capability, init_millis, warmup_millis, model_run_millis);
 
 
   for (size_t i = 0; i < output_count; ++i) {
@@ -532,13 +533,16 @@ int Main(int argc, char **argv) {
   }
 
 
+  // get cpu capability
+  Capability cpu_capability = GetCapability(DeviceType::CPU);
+
   bool ret = false;
   for (int i = 0; i < FLAGS_restart_round; ++i) {
     VLOG(0) << "restart round " << i;
-    ret =
-        RunModel(FLAGS_model_name,
-                 input_names, input_shape_vec, input_data_formats,
-                 output_names, output_shape_vec, output_data_formats);
+    ret = RunModel(FLAGS_model_name,
+        input_names, input_shape_vec, input_data_formats,
+        output_names, output_shape_vec, output_data_formats,
+        cpu_capability.float32_performance.exec_time);
   }
   if (ret) {
     return 0;