Fix SIR cache when call func many times (#21)

* Fix SIR cache when call func many times

* Rename ut

* Polish code

* Remove bind

* Rename var and fun

Author: 0x45f

symbolic_trace/trace_cache_entrance.py

@@ -9,16 +9,42 @@
 
 
 def trace_cache(func):
+    @no_eval_frame
     def call_with_cache(*args, **kwargs):
-        args, kwargs = convert_arguments(args), convert_arguments(kwargs) 
+        args, kwargs = convert_arguments(args), convert_arguments(kwargs)
+        args, kwargs, outter_names = construct_inner_proxy_tensor(func.__name__, *args, **kwargs)
+
         if frame_enter(func.__name__, args):
-            return cache_and_return(func.__name__, args)
+            return cache_and_return(func.__name__, outter_names)
         ret = func(*args)
-        frame_leave(func.__name__, ret)
+        frame_leave(func.__name__, outter_names, ret)
         return ret
     return call_with_cache
 
 
+def construct_inner_proxy_tensor(func_name, *args, **kwargs):
+    flat_args = paddle.utils.flatten(args)
+    flat_kwargs = paddle.utils.flatten(kwargs)
+    outter_names = []
+    name_i = 0
+    for i, v in enumerate(flat_args):
+        if isinstance(v, ProxyTensor):
+            name = '{}_input_{}'.format(func_name, name_i)
+            outter_names.append(v.name)
+            flat_args[i] = ProxyTensor(name, v.meta)
+            name_i = name_i + 1
+    for i, v in enumerate(flat_kwargs):
+        if isinstance(v, ProxyTensor):
+            name = '{}_input_{}'.format(func_name, name_i)
+            outter_names.append(v.name)
+            flat_kwargs[i] = ProxyTensor(name, v.meta)
+            name_i = name_i + 1
+
+    args = paddle.utils.pack_sequence_as(args, flat_args)
+    kwargs = paddle.utils.pack_sequence_as(kwargs, flat_kwargs)
+
+    return args, kwargs, outter_names
+
 @no_eval_frame
 # should generate a unique name for every function
 def frame_enter(name, inputs):
@@ -52,7 +78,7 @@ def frame_enter(name, inputs):
 
 
 @no_eval_frame
-def frame_leave(name, outputs):
+def frame_leave(name, outter_names, outputs):
     key_name = SymbolicTraceContext().sir_key_stack[-1]
     SymbolicTraceContext().sir_key_stack.pop()
 
@@ -88,24 +114,20 @@ def frame_leave(name, outputs):
             return
 
     # at the first time, the inputs and outputs need not change
-    SymbolicTraceContext().call_SIR(cur_sir.name, cur_sir.inputs, cur_sir.outputs)
+    SymbolicTraceContext().call_SIR(cur_sir.name, [Symbol(name) for name in outter_names], cur_sir.outputs)
     log(1, cur_sir, "\n")
     return
 
 
 @no_eval_frame
-def cache_and_return(name, inputs):
+def cache_and_return(name, outter_names):
     key_name = SymbolicTraceContext().sir_key_stack[-1]
     SymbolicTraceContext().sir_key_stack.pop()
 
     # find sir and it's origin_outputs
     cached_sir = SymbolicTraceContext().statement_factory[key_name]
     origin_outputs = SIRRuntimeCache().get_origin_outputs(key_name)
 
-    # gen call_SIR inputs
-    flat_inputs = paddle.utils.flatten(inputs)
-    symbol_inputs = [Symbol(x.name) for x in flat_inputs if isinstance(x, ProxyTensor)]
-
     # create return value
     outputs = gen_new_proxy_tensor_output(origin_outputs)
 
@@ -114,7 +136,7 @@ def cache_and_return(name, inputs):
     symbol_outputs = [Symbol(x.name) for x in flat_outputs if isinstance(x, ProxyTensor)]
 
     # add call_SIR
-    SymbolicTraceContext().call_SIR(cached_sir.name, symbol_inputs, symbol_outputs)
+    SymbolicTraceContext().call_SIR(cached_sir.name, [Symbol(name) for name in outter_names], symbol_outputs)
     return outputs
 
 

tests/error_test_sir_call.py

@@ -1,14 +1,12 @@
 import unittest
 import paddle
 from symbolic_trace import symbolic_trace
-from symbolic_trace.trace_cache_entrance import frame_enter, frame_leave, cache_and_return
+from symbolic_trace.trace_cache_entrance import trace_cache
 
 
+@trace_cache
 def sum(x, y):
-    if frame_enter("sum", (x, y)):
-        return cache_and_return("sum", (x, y))
     ret = x + y
-    frame_leave("sum", (ret))
     return ret
 
 def main(x, y):
@@ -21,7 +19,7 @@ def test_return_callable(self):
         x = paddle.to_tensor([1.0])
         y = paddle.to_tensor([2.0])
         ret = symbolic_trace(main)(x, y)
-        assert (ret.item() == 3.0), "Should be 4.0"
+        assert (ret.item() == 3.0), "Should be 3.0"
 
 if __name__ == "__main__":
     unittest.main()