[NewIR] Insert `get_parameter` only for paramters (#56325)

* fix inset get_parameter op bug

* fix bug: insert  only for parameters

* fix bug: wrong idx in vector

---------
Co-authored-by: Nzhangbo9674 <zhangbo54@baidu.com>

paddle/fluid/framework/executor_cache.cc

@@ -367,7 +367,7 @@ std::unique_ptr<::ir::Program> ConstructFowardIrProgram(
     }
   }
 
-  // add fetch with place op to program
+  // add data op to program
   auto *block = local_program.MutableBlock(0);
   for (auto &in_t : x) {
     auto name = in_t.name();

paddle/fluid/framework/new_executor/new_ir_interpreter.cc

@@ -478,9 +478,10 @@ std::string NewIRInterpreter::DebugValueInfo() {
                    platform::errors::PreconditionNotMet(
                        "var(%s) should exist in var_name_2_id_", kv.second));
     auto* var = InnerScope()->FindVar(kv.second);
-    PADDLE_ENFORCE(var != nullptr,
+    PADDLE_ENFORCE(
+        var != nullptr,
         platform::errors::PreconditionNotMet(
-                       "var(%s) should exist in var_name_2_id_", kv.second));
+            "var(%s) should exist in scope (%p)", kv.second, InnerScope()));
     os << kv.first.impl() << " -> " << kv.second << " -> "
        << var_name_2_id_.at(kv.second) << " -> " << var << "\n";
   }

paddle/fluid/ir_adaptor/translator/op_translator.cc

@@ -49,7 +49,9 @@ namespace translator {
 
 namespace {
 
-using ResultIdx = size_t;
+using IdxInOp = size_t;
+using IdxInVector = size_t;
+using ResultIdx = std::tuple<IdxInOp, IdxInVector>;
 using OpDesc = paddle::framework::OpDesc;
 using BlockDesc = paddle::framework::BlockDesc;
 using VarDesc = paddle::framework::VarDesc;
@@ -325,18 +327,15 @@ void OpTranscriber::InsertSliceOperationForInput(
   // scan all inputs to see if any of them is generated as a vector<Tensor>
   // so need an additional `SliceOp` to take it out.
   for (const auto& n : op_desc.Inputs()) {
-    auto& name = n.first;
     auto& args = n.second;
 
     for (const auto& arg_name : args) {
       bool check =
-          param_map->count(arg_name) != 0 || !yaml_input_set.count(arg_name);
-      IR_ENFORCE(check,
-                 "arg %s.%s as input should be exists before prasing %s",
-                 name,
-                 arg_name,
-                 op_desc.Type());
-      auto defining_info = (*param_map)[arg_name];
+          param_map->count(arg_name) != 0 && !yaml_input_set.count(arg_name);
+      if (!check) {
+        continue;
+      }
+      auto defining_info = param_map->at(arg_name);
       if (defining_info.generated_by_vector) {
         InsertSliceOperationForTarget(
             ctx, param_map, program, defining_info, arg_name);
@@ -391,7 +390,8 @@ std::vector<ir::OpResult> OpTranscriber::GenerateOperationInput(
       }
     }
     VLOG(10) << "[op:" << op_desc.Type() << "][input]" << info.name << " "
-             << legacy_input_name << " " << legacy_input_vars.size();
+             << legacy_input_name << " " << legacy_input_vars.size() << "["
+             << legacy_input_vars << "]";
 
     if (legacy_input_vars.empty() && mutable_attributes != nullptr &&
         mutable_attributes->count(info.name) != 0) {
@@ -507,7 +507,7 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
         ir::Type translated_var_type =
             type_translator[var->GetType()](ctx, *var);
         op_output_types.push_back(translated_var_type);
-        arg_to_idx[var->Name()] = cur_output_idx;
+        arg_to_idx[var->Name()] = {cur_output_idx, 0};
         continue;
       }
     }
@@ -535,7 +535,7 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
 
       ir::Type translated_var_type = type_translator[var->GetType()](ctx, *var);
 
-      arg_to_idx[var_name] = cur_output_idx;
+      arg_to_idx[var_name] = {cur_output_idx, 0};
       op_output_types.push_back(translated_var_type);
 
       // if src type is Vector<Tesnor>
@@ -544,10 +544,12 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
                << "[" << op_desc.Type() << "]" << info.name << " :"
                << info.type_name << " var: " << legacy_output_name;
       std::vector<ir::Type> types;
-      for (const auto& var_name : legacy_output_vars) {
+      for (IdxInVector idx_in_vec = 0; idx_in_vec < legacy_output_vars.size();
+           idx_in_vec++) {
+        const auto& var_name = legacy_output_vars[idx_in_vec];
         if (var_name == kEmptyVarName) {
           types.emplace_back(nullptr);
-          arg_to_idx[var_name] = cur_output_idx;
+          arg_to_idx[var_name] = {cur_output_idx, idx_in_vec};
           continue;
         }
         VarDesc* var = block->FindVarRecursive(var_name);
@@ -557,7 +559,7 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
         ir::Type translated_var_type =
             type_translator[var->GetType()](ctx, *var);
         types.push_back(translated_var_type);
-        arg_to_idx[var_name] = cur_output_idx;
+        arg_to_idx[var_name] = {cur_output_idx, idx_in_vec};
       }
       ir::Type vec_type = ir::VectorType::get(ctx, types);
       op_output_types.push_back(vec_type);
@@ -613,45 +615,16 @@ void OpTranscriber::RecordOpResultMapping(ir::IrContext* ctx,
                                           const OpDesc& op_desc,
                                           ir::Operation* operation,
                                           const OpOutputMapping& arg_to_idx) {
-  for (const auto& n : op_desc.Outputs()) {
-    auto& name = n.first;
-    VLOG(10) << "[output recording]"
-             << "[" << op_desc.Type() << "]" << name;
-    const auto& args = n.second;
-    size_t idx_in_vector = 0;
-    for (const auto& arg_name : args) {
-      if (arg_name == kEmptyVarName) {
-        idx_in_vector++;
-        continue;
-      }
-      auto idx_iter = arg_to_idx.find(arg_name);
-      if (idx_iter == arg_to_idx.end()) {
-        VLOG(4) << "[output recording]"
-                << "[" << op_desc.Type() << "][skip]" << arg_name;
-        continue;
-      }
-      auto idx = idx_iter->second;
+  for (const auto& [arg_name, idx] : arg_to_idx) {
+    const auto& [idx_in_op, idx_in_vec] = idx;
     VLOG(10) << "[output recording]"
-               << "[" << op_desc.Type() << "]" << arg_name << " " << idx;
-
-      ir::OpResult value = operation->result(idx);
+             << "[" << op_desc.Type() << "]" << arg_name << " " << idx_in_op
+             << " " << idx_in_vec;
+    ir::OpResult value = operation->result(idx_in_op);
     bool generated_by_vector = value.type().isa<ir::VectorType>();
 
-      // Specially process TensorArray, this because we cannot distinguish it
-      // with Vector<DenseTensor> by other conditions but we cannot support it
-      // like Vector<DenseTensor>
-      if (args.size() == 1) {
-        VarDesc* var = op_desc.Block()->FindVarRecursive(args[0]);
-        if (var->GetType() ==
-            paddle::framework::proto::VarType::LOD_TENSOR_ARRAY) {
-          generated_by_vector = false;
-        }
-      }
-
     (*param_map)[arg_name] = VariableDefiningInfo(
-          value, generated_by_vector, generated_by_vector ? idx_in_vector : -1);
-      idx_in_vector++;
-    }
+        value, generated_by_vector, generated_by_vector ? idx_in_vec : -1);
   }
 }
 
@@ -1439,9 +1412,10 @@ struct ElementwiseGradTranscriber : public OpTranscriber {
     if (idx_iter == arg_to_idx.end()) {
       IR_THROW("op[%s] should have got its y_grad", op_desc.Type());
     }
-    auto idx = idx_iter->second;
+    auto [idx_in_op, idx_in_vec] = idx_iter->second;
     VLOG(10) << "[output recording]"
-             << "[" << op_desc.Type() << "]" << y_grad_var_name << " " << idx;
+             << "[" << op_desc.Type() << "]" << y_grad_var_name << " "
+             << idx_in_op << " " << idx_in_vec;
 
     auto y_names = op_desc.Input("Y", true);
     auto y_name = y_names[0];
@@ -1465,7 +1439,7 @@ struct ElementwiseGradTranscriber : public OpTranscriber {
         y_type.dyn_cast<dialect::DenseTensorType>();
     std::vector<int64_t> y_shape = phi::vectorize(y_tensor_type.dims());
 
-    ir::OpResult value = operation->result(idx);
+    ir::OpResult value = operation->result(idx_in_op);
     ir::Builder builder(ctx, operation->GetParent());
     auto reshape_op = builder.Build<dialect::ReshapeOp>(value, y_shape);
     (*param_map)[y_grad_var_name] =

paddle/fluid/ir_adaptor/translator/program_translator.cc

@@ -143,12 +143,12 @@ void ProgramTranslator::GetParameterForSingleBlock(const BlockDesc& block) {
           var_desc = block.FindVarRecursive(var_name);
         }
 
-        bool need_get_parameter_op = is_parameter || is_unseen_variable;
+        bool need_get_parameter_op = is_parameter && is_unseen_variable;
         if (need_get_parameter_op) {
           ir::Operation* op = InsertGetParamaterOp(ctx_, var_desc);
           program_->block()->push_back(op);
           param_map_[var_name] = VariableDefiningInfo(op->result(0));
-          VLOG(10) << "[op translated][get parameter]" << op;
+          VLOG(10) << "[op translated][get parameter]" << var_name;
 
           program_->SetParameter(var_name, nullptr);
           parameter_visited_.insert(var_name);
@@ -224,7 +224,7 @@ void ProgramTranslator::SetParameterFromSingleBlock(const BlockDesc& block) {
           insert_pos++;
 
           block->insert(insert_pos, op);
-          VLOG(10) << "[op translated][set parameter]" << op;
+          VLOG(10) << "[op translated][set parameter]" << var_name;
 
           program_->SetParameter(var_name, nullptr);
           parameter_visited_.insert(var_name);

paddle/fluid/ir_adaptor/translator/utils.cc

@@ -18,6 +18,7 @@
 
 #include "paddle/ir/core/builtin_attribute.h"
 #include "paddle/ir/core/builtin_type.h"
+#include "paddle/ir/core/utils.h"
 
 namespace paddle {
 namespace translator {
@@ -46,5 +47,15 @@ ir::Operation* InsertSliceOperationForTarget(
   return operation;
 }
 
+std::ostream& operator<<(std::ostream& os,
+                         const std::vector<std::string>& vec_str) {
+  ir::PrintInterleave(
+      vec_str.begin(),
+      vec_str.end(),
+      [&os](std::string s) { os << s; },
+      [&os]() { os << ", "; });
+  return os;
+}
+
 }  // namespace translator
 }  // namespace paddle

paddle/fluid/ir_adaptor/translator/utils.h

@@ -31,5 +31,8 @@ ir::Operation* InsertSliceOperationForTarget(
     const VariableDefiningInfo& defining_info,
     const std::string& arg_name);
 
+std::ostream& operator<<(std::ostream& os,
+                         const std::vector<std::string>& vec_str);
+
 }  // namespace translator
 }  // namespace paddle

python/paddle/static/input.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import os
+
 import paddle
 from paddle.fluid import Variable, core
 from paddle.fluid.data_feeder import check_type
@@ -96,7 +98,7 @@ def data(name, shape, dtype=None, lod_level=0):
             shape[i] = -1
 
     if dtype:
-        return helper.create_global_variable(
+        out = helper.create_global_variable(
             name=name,
             shape=shape,
             dtype=dtype,
@@ -107,7 +109,7 @@ def data(name, shape, dtype=None, lod_level=0):
             need_check_feed=True,
         )
     else:
-        return helper.create_global_variable(
+        out = helper.create_global_variable(
             name=name,
             shape=shape,
             dtype=paddle.get_default_dtype(),
@@ -118,6 +120,21 @@ def data(name, shape, dtype=None, lod_level=0):
             need_check_feed=True,
         )
 
+    if os.environ.get("FLAGS_enable_new_ir_in_executor", None):
+        helper = LayerHelper('data', **locals())
+        helper.append_op(
+            type='data',
+            inputs={},
+            outputs={'out': out},
+            attrs={
+                'index': 0,
+                'dtype': 0,
+                'place': 0,
+                'name': name,
+            },
+        )
+    return out
+
 
 class InputSpec:
     """

test/ir/new_ir/test_ir_vjp.py

@@ -106,7 +106,7 @@ class TestMeanVjp(unittest.TestCase):
                 .source()
                 .get_defining_op()
                 .name(),
-                "builtin.get_parameter",
+                "pd.data",
             )
             self.assertEqual(
                 grad_outs[0][0]

test/prim/new_ir_prim/test_decomp_op.py

@@ -47,7 +47,7 @@ class TestBuildOp(unittest.TestCase):
         self.assertEqual(
             op_name_list,
             [
-                'builtin.get_parameter',
+                'pd.data',
                 'pd.matmul',
                 'pd.add',
                 'pd.full_int_array',