[LatencyPredictor] add hardware (#1089)

* Add rk3288 predictor

* fix some bugs for sparse conv2d.

paddleslim/analysis/extract_features.py

@@ -61,7 +61,7 @@ def get_features_from_paramkey(param_key, op_type, data_type):
         if quant_bits not in param_key:
             return None
 
-        weight = re.search(r'weight=(\(\d*, \d*, \d*, \d*\))',
+        weight = re.search(r'weight=(\(\d*, -?\d*, \d*, \d*\))',
                            param_key).group().split('=')[-1].strip(
                                '('
                                ')').split(', ')

paddleslim/analysis/latency_predictor.py

@@ -68,7 +68,7 @@ class TableLatencyPredictor(LatencyPredictor):
     Args:
         table_file(str): The path of file that records the device latency of operators.
     """
-    hardware_list = ['SD625', 'SD710']
+    hardware_list = ['SD625', 'SD710', 'RK3288']
 
     def __init__(self, table_file='SD710'):
         self.table_file = table_file

paddleslim/analysis/parse_ops.py

@@ -24,7 +24,10 @@ def get_key_from_op(op):
     if op_type == 'sparse_conv2d':
         out_shape = op.all_outputs()[0].shape()
         in_shape = op.inputs('Input')[0].shape()
-        weight_shape = (out_shape[1], in_shape[1], 1, 1)
+        if in_shape:
+            weight_shape = (out_shape[1], in_shape[1], 1, 1)
+        else:
+            weight_shape = (out_shape[1], -1, 1, 1)
         NonZeroWeights = op.inputs('NonZeroWeights')[0].shape()[0]
 
         stride = op.attr('strides')[1]
@@ -147,14 +150,13 @@ def get_key_from_op(op):
 
     elif op_type == 'stack':
         data = op.all_inputs()
-        X = "["
+        X = ""
         for x in data:
             X += f"{x.shape()}"
-        X += "]"
         axis = op.attr('axis')
         out_shape = op.all_outputs()[0].shape()
 
-        param_key = f'{op_type} X={X} axis={axis} out={out_shape}'
+        param_key = f'{op_type} in={X} axis={axis} out={out_shape}'
 
     elif op_type == 'exp':
         in_shape = op.all_inputs()[-1].shape()
@@ -219,7 +221,7 @@ def get_key_from_op(op):
         in_shape2 = op.all_inputs()[1].shape()
         out_shape = op.all_outputs()[0].shape()
 
-        param_key = f'{op_type} in={in_shape1} in={in_shape2}  out={out_shape}'
+        param_key = f'{op_type} X={in_shape1} Y={in_shape2}  out={out_shape}'
 
     elif op_type in ['calib', 'floor']:
         in_shape = op.all_inputs()[-1].shape()