fix and open 3 tests with backward (#59947)

chen2016013 · web-flow · commit a8eba167d63a · 2023-12-13T10:56:07.000+08:00
diff --git a/paddle/fluid/framework/new_executor/instruction/if_instruction.cc b/paddle/fluid/framework/new_executor/instruction/if_instruction.cc
@@ -75,6 +75,21 @@ IfInstruction::IfInstruction(size_t id,
   auto& false_branch_block = if_op.false_block();
   false_outside_inputs =
       GetExternalInputs(&false_branch_block, *value_exec_info, &inputs);
+  // NOTE(chenxi67): the variable corresponding to container value if a
+  // <VariableRefArray> Type. It will recursively get the ID of internal
+  // variables when use GetValueId() method. However, the copy_var pushed into
+  // the tuple does not have a corresponding ID, and will insert a -1. Here we
+  // remove the value of -1.
+  for (auto& item : inputs) {
+    auto& var_vec = item.second;
+    for (auto it = var_vec.begin(); it != var_vec.end();) {
+      if (*it == -1) {
+        it = var_vec.erase(it);
+      } else {
+        ++it;
+      }
+    }
+  }
   SetInputs(inputs);
 
   std::unordered_map<pir::Value, std::vector<int>> outputs;
@@ -95,7 +110,16 @@ IfInstruction::IfInstruction(size_t id,
 
   InsertTuplePushContinerToOuts(
       &if_op.false_block(), *value_exec_info, &outputs);
-
+  for (auto& item : outputs) {
+    auto& var_vec = item.second;
+    for (auto it = var_vec.begin(); it != var_vec.end();) {
+      if (*it == -1) {
+        it = var_vec.erase(it);
+      } else {
+        ++it;
+      }
+    }
+  }
   SetOutputs(outputs);
   VLOG(6) << "finish process inputs outputs index";
 
diff --git a/paddle/fluid/framework/new_executor/instruction/instruction_util.cc b/paddle/fluid/framework/new_executor/instruction/instruction_util.cc
@@ -229,7 +229,14 @@ std::unordered_set<pir::Value> GetInternalOutputs(pir::Block* block) {
     inner_outputs.insert(block->arg(arg_id));
   }
   for (auto& op : *block) {
-    VLOG(8) << "GetInternalOutputs of " << op.name();
+    std::string op_name = op.name();
+    if (op.attributes().count("op_name")) {
+      op_name = op.attributes()
+                    .at("op_name")
+                    .dyn_cast<pir::StrAttribute>()
+                    .AsString();
+    }
+    VLOG(8) << "GetInternalOutputs of " << op_name;
     if (op.num_regions()) {
       for (size_t i = 0; i < op.num_regions(); ++i) {
         for (auto& sub_block : op.region(i)) {
@@ -239,8 +246,10 @@ std::unordered_set<pir::Value> GetInternalOutputs(pir::Block* block) {
         }
       }
     }
+
     for (size_t i = 0; i < op.num_results(); ++i) {
       inner_outputs.insert(op.result(i));
+      VLOG(10) << op_name << "'s inner_output: " << op.result(i).impl();
     }
   }
   return inner_outputs;
@@ -249,7 +258,14 @@ std::unordered_set<pir::Value> GetInternalOutputs(pir::Block* block) {
 std::unordered_set<pir::Value> GetInternalInputs(pir::Block* block) {
   std::unordered_set<pir::Value> inner_inputs;
   for (auto& op : *block) {
-    VLOG(8) << "GetInternalInputs of " << op.name();
+    std::string op_name = op.name();
+    if (op.attributes().count("op_name")) {
+      op_name = op.attributes()
+                    .at("op_name")
+                    .dyn_cast<pir::StrAttribute>()
+                    .AsString();
+    }
+    VLOG(8) << "GetInternalInputs of " << op_name;
     if (op.num_regions()) {
       for (size_t i = 0; i < op.num_regions(); ++i) {
         for (auto& sub_block : op.region(i)) {
@@ -265,8 +281,7 @@ std::unordered_set<pir::Value> GetInternalInputs(pir::Block* block) {
     }
     for (size_t i = 0; i < op.num_operands(); ++i) {
       inner_inputs.insert(op.operand_source(i));
-      VLOG(10) << op.name()
-               << "'s inner_input: " << op.operand_source(i).impl();
+      VLOG(10) << op_name << "'s inner_input: " << op.operand_source(i).impl();
     }
   }
   return inner_inputs;
diff --git a/paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc b/paddle/fluid/framework/new_executor/interpreter/interpreter_util.cc
@@ -1264,6 +1264,7 @@ const paddle::framework::Variable* GetVariableByName(
       return kv.first;
     }
   }
+  VLOG(8) << "cannot find var: " << var_name;
   return nullptr;
 }
 
diff --git a/paddle/fluid/framework/new_executor/pir_adaptor/pir_adaptor_util.cc b/paddle/fluid/framework/new_executor/pir_adaptor/pir_adaptor_util.cc
@@ -405,7 +405,7 @@ void HandleForSpecialOp(pir::Operation* op,
     value_exe_info->Add(value, fetch_var_name);
   } else if (op_name == paddle::dialect::FeedOp::name() ||
              op_name == paddle::dialect::DataOp::name()) {
-    VLOG(6) << "Handle for" << op_name;
+    VLOG(6) << "Handle for " << op_name;
     auto value = op->result(0);
     VLOG(6) << "link feed output to feed in variable"
             << value_exe_info->GetScope();
@@ -415,6 +415,7 @@ void HandleForSpecialOp(pir::Operation* op,
     Variable* var = value_exe_info->GetScope()->FindVar(name);
     if (var == nullptr) {
       var = value_exe_info->GetScope()->Var(name);
+      var->GetMutable<phi::DenseTensor>();
     }
     PADDLE_ENFORCE(var,
                    paddle::platform::errors::InvalidArgument(
diff --git a/paddle/fluid/framework/new_executor/pir_adaptor/pir_adaptor_util.h b/paddle/fluid/framework/new_executor/pir_adaptor/pir_adaptor_util.h
@@ -228,7 +228,7 @@ void BuildPhiContext(pir::Operation* op,
                                               var->Type()));
     }
   }
-
+  VLOG(8) << "EmplaceBackInput done";
   // EmplaceBackAttributes
   auto& vec_kernel_fn_attr_params = op_yaml_info.AttrParams(is_kernel);
   for (auto& t : vec_kernel_fn_attr_params) {
@@ -416,6 +416,7 @@ void BuildPhiContext(pir::Operation* op,
     }
     VLOG(6) << "ctx->EmplaceBackAttr: " << t;
   }
+  VLOG(8) << "EmplaceBackBackAttributes done";
 
   // EmplaceBackOutputs
   VLOG(8) << "ctx->EmplaceBackOutput: ";
@@ -482,6 +483,7 @@ void BuildPhiContext(pir::Operation* op,
           phi::errors::Unimplemented("only support DenseTensor and vector "));
     }
   }
+  VLOG(8) << "EmplaceBackOutputs done";
 
   VLOG(6) << "Done build phi context";
 }
diff --git a/paddle/fluid/framework/new_executor/pir_interpreter.cc b/paddle/fluid/framework/new_executor/pir_interpreter.cc
@@ -1048,6 +1048,8 @@ void PirInterpreter::CalculateLastLiveOps() {
       ins_and_outs.insert(outs.begin(), outs.end());
     }
 
+    VLOG(4) << "get gc check vars for: " << instr->Name();
+
     for (auto& item : ins_and_outs) {
       for (auto var_id : item.second) {
         // skip no_need_buffer input vars
@@ -1057,7 +1059,6 @@ void PirInterpreter::CalculateLastLiveOps() {
         gc_check_vars.insert(var_id);
       }
     }
-    VLOG(4) << "get gc check vars for: " << instr->Name();
 
     for (auto var_id : gc_check_vars) {
       Scope* inner_scope = InnerScope();
diff --git a/python/paddle/autograd/ir_backward.py b/python/paddle/autograd/ir_backward.py
@@ -55,11 +55,11 @@ def check_all_puts(block, inputs, outputs):
             )
 
 
-def append_full_like(float_value, value, state, backward_ops):
+def append_full_like(float_value, copy_value, value, state, backward_ops):
     value_grad = paddle.full_like(
-        value,
+        copy_value,
         float_value,
-        dtype=value.dtype,
+        dtype=copy_value.dtype,
     )
     full_like_op = value_grad.get_defining_op()
     full_op = full_like_op.operand_source(1).get_defining_op()
@@ -118,7 +118,7 @@ def prepare_grad_outputs(grad_outputs, outputs, state):
         # fwd : op1 -> op2 -> op3 -> output
         # bwd : op1G <- op2G <- op3G <- outputG <- full_likeop/feedop
         if grad is None:
-            append_full_like(1.0, output, state, backward_ops)
+            append_full_like(1.0, output, output, state, backward_ops)
         else:
             if output.shape != grad.shape:
                 raise ValueError(
@@ -156,7 +156,7 @@ def prepare_grad_outputs(grad_outputs, outputs, state):
                     ]
                 else:
                     grad_value = append_full_like(
-                        0.0, opresult, state, backward_ops
+                        0.0, opresult, opresult, state, backward_ops
                     )
                     visited_output.add(opresult)
 
@@ -441,7 +441,9 @@ def make_output_with_output_grad(op):
                     # last bwd_op return None because input in no_grad_set,
                     # but this bwd_op need a input.
 
-                    append_full_like(0.0, value, state, backward_ops)
+                    append_full_like(
+                        0.0, new_value[0], value, state, backward_ops
+                    )
                     zero_flag[i] = True
 
             outputs.append(new_value)
@@ -562,7 +564,7 @@ def append_yield(block, base_inputs, base_inputs_grad):
                     inputs_grad.append(state.value_to_valuegrad[value][0][0])
                 else:
                     value_grad = append_full_like(
-                        0.0, value, state, backward_ops
+                        0.0, value, value, state, backward_ops
                     )
                     inputs_grad.append(value_grad)
             paddle.base.libpaddle.pir.cf_yield(inputs_grad)
diff --git a/python/paddle/base/executor.py b/python/paddle/base/executor.py
@@ -1111,7 +1111,17 @@ def _get_pir_program_and_executor(self, cached_data):
         plan = core.Plan([default_job], type_to_program)
 
         new_exe = _StandaloneExecutor(place, plan, scope)
-        return program, new_exe
+
+        data_op_infos = []
+        global_block = program.global_block()
+        for op in global_block.ops:
+            if op.name() == 'pd_op.data':
+                feed_target_name = op.attrs()["name"]
+                var_type = paddle_type_to_proto_type[op.attrs()["dtype"]]
+                var_shape = op.attrs()["shape"]
+                tup = (feed_target_name, var_type, var_shape)
+                data_op_infos.append(tup)
+        return program, new_exe, data_op_infos
 
 
 class Executor:
@@ -1328,26 +1338,23 @@ def _feed_data(self, program, feed, feed_var_name, scope):
             else:
                 break
 
-    def _pir_feed_data(self, program, feed, scope):
+    def _pir_feed_data(self, program, feed, scope, data_op_infos):
         # feed var to framework
         feed_target_names = set()
-        global_block = program.global_block()
-        for op in global_block.ops:
-            if op.name() == 'pd_op.data':
-                feed_target_name = op.attrs()["name"]
-                feed_target_names.add(feed_target_name)
-                var_type = paddle_type_to_proto_type[op.attrs()["dtype"]]
-                var_shape = op.attrs()["shape"]
-                cur_feed = feed[feed_target_name]
-                if not isinstance(cur_feed, core.LoDTensor):
-                    cur_feed = _as_lodtensor(cur_feed, self.place, var_type)
-                pir_check_feed_shape_type(
-                    cur_feed, feed_target_name, var_shape, var_type
-                )
-                # the last arg of set_feed_variable has no effect in pir, we pass 0 by default.
-                core.set_feed_variable(scope, cur_feed, feed_target_name, 0)
-            else:
-                break
+        for data_op_info in data_op_infos:
+            feed_target_name = data_op_info[0]
+            feed_target_names.add(feed_target_name)
+            var_type = data_op_info[1]
+            var_shape = data_op_info[2]
+
+            cur_feed = feed[feed_target_name]
+            if not isinstance(cur_feed, core.LoDTensor):
+                cur_feed = _as_lodtensor(cur_feed, self.place, var_type)
+            pir_check_feed_shape_type(
+                cur_feed, feed_target_name, var_shape, var_type
+            )
+            # the last arg of set_feed_variable has no effect in pir, we pass 0 by default.
+            core.set_feed_variable(scope, cur_feed, feed_target_name, 0)
 
         # pop variable which is not found in program
         for feed_name in list(feed.keys()):
@@ -2034,7 +2041,11 @@ def _run_pir_impl(
                 % (type(feed))
             )
 
-        program, new_exe = self._executor_cache.get_pir_program_and_executor(
+        (
+            program,
+            new_exe,
+            data_op_infos,
+        ) = self._executor_cache.get_pir_program_and_executor(
             program,
             feed,
             fetch_list,
@@ -2044,7 +2055,7 @@ def _run_pir_impl(
             scope,
         )
 
-        self._pir_feed_data(program, feed, scope)
+        self._pir_feed_data(program, feed, scope, data_op_infos)
 
         if hasattr(program, 'lr_scheduler'):
             from paddle.optimizer.lr import LRScheduler
diff --git a/test/legacy_test/test_cond.py b/test/legacy_test/test_cond.py
@@ -480,7 +480,7 @@ def test_extremely_simple_net_with_op_in_condition(self):
 
 
 class TestCondNestedControlFlow(unittest.TestCase):
-    # @test_with_pir_api
+    @test_with_pir_api
     def test_cond_inside_cond(self):
         """
         pseudocode:
@@ -548,7 +548,7 @@ def greater_equal_branch(i, a):
                         da = g
                 ret = exe.run(
                     main_program,
-                    feed={'i': np.full((1), feed_i)},
+                    feed={'i': np.full((1), feed_i, np.float32)},
                     fetch_list=[out, da],
                 )
             else:
@@ -560,7 +560,7 @@ def greater_equal_branch(i, a):
             self.assertEqual(ret[0][0], expected_ret)
             self.assertEqual(ret[1][0], expected_a_grad)
 
-    # @test_with_pir_api
+    @test_with_pir_api
     def test_cond_inside_cond_0d_tensor(self):
         """
         pseudocode:
@@ -635,7 +635,7 @@ def greater_equal_branch(i, a):
         )
         self.assertEqual(ret[1].shape, ())
 
-    # @test_with_pir_api
+    @test_with_pir_api
     def test_cond_op_in_condition(self):
         paddle.enable_static()
         main_program = paddle.static.Program()

Original file line number	Diff line number	Diff line change
`@@ -1264,6 +1264,7 @@ const paddle::framework::Variable* GetVariableByName(`
`1264`	`1264`	`return kv.first;`
`1265`	`1265`	`}`
`1266`	`1266`	`}`
	`1267`	`+ VLOG(8) << "cannot find var: " << var_name;`
`1267`	`1268`	`return nullptr;`
`1268`	`1269`	`}`
`1269`	`1270`
Original file line number	Diff line number	Diff line change
`@@ -228,7 +228,7 @@ void BuildPhiContext(pir::Operation* op,`
`228`	`228`	`var->Type()));`
`229`	`229`	`}`
`230`	`230`	`}`
`231`		`-`
	`231`	`+ VLOG(8) << "EmplaceBackInput done";`
`232`	`232`	`// EmplaceBackAttributes`
`233`	`233`	`auto& vec_kernel_fn_attr_params = op_yaml_info.AttrParams(is_kernel);`
`234`	`234`	`for (auto& t : vec_kernel_fn_attr_params) {`
`@@ -416,6 +416,7 @@ void BuildPhiContext(pir::Operation* op,`
`416`	`416`	`}`
`417`	`417`	`VLOG(6) << "ctx->EmplaceBackAttr: " << t;`
`418`	`418`	`}`
	`419`	`+ VLOG(8) << "EmplaceBackBackAttributes done";`
`419`	`420`
`420`	`421`	`// EmplaceBackOutputs`
`421`	`422`	`VLOG(8) << "ctx->EmplaceBackOutput: ";`
`@@ -482,6 +483,7 @@ void BuildPhiContext(pir::Operation* op,`
`482`	`483`	`phi::errors::Unimplemented("only support DenseTensor and vector "));`
`483`	`484`	`}`
`484`	`485`	`}`
	`486`	`+ VLOG(8) << "EmplaceBackOutputs done";`
`485`	`487`
`486`	`488`	`VLOG(6) << "Done build phi context";`
`487`	`489`	`}`
Original file line number	Diff line number	Diff line change
`@@ -1048,6 +1048,8 @@ void PirInterpreter::CalculateLastLiveOps() {`
`1048`	`1048`	`ins_and_outs.insert(outs.begin(), outs.end());`
`1049`	`1049`	`}`
`1050`	`1050`
	`1051`	`+ VLOG(4) << "get gc check vars for: " << instr->Name();`
	`1052`	`+`
`1051`	`1053`	`for (auto& item : ins_and_outs) {`
`1052`	`1054`	`for (auto var_id : item.second) {`
`1053`	`1055`	`// skip no_need_buffer input vars`
`@@ -1057,7 +1059,6 @@ void PirInterpreter::CalculateLastLiveOps() {`
`1057`	`1059`	`gc_check_vars.insert(var_id);`
`1058`	`1060`	`}`
`1059`	`1061`	`}`
`1060`		`- VLOG(4) << "get gc check vars for: " << instr->Name();`
`1061`	`1062`
`1062`	`1063`	`for (auto var_id : gc_check_vars) {`
`1063`	`1064`	`Scope* inner_scope = InnerScope();`