[IR] refine program translator (#54719)

* refine program translator

* fix warning: not override

* fix bug

* merge new modifications

* modify by reviews

* resolve conflicts

* resolve conflicts

* fix

* fix

* fix conflicts

* add unittest for special op transcriber

* set cpu as default backend

* modify by reviews

paddle/fluid/eager/auto_code_generator/generator/eager_gen.py

@@ -1013,12 +1013,9 @@ class DygraphFunctionGeneratorBase(FunctionGeneratorBase):
                         name in forward_outputs_position_map.keys()
                     ), AssertMessage(name, forward_outputs_position_map.keys())
 
-                if is_optional:
-                    set_tensor_wrappers = f"{indent}if({name}) grad_node->SetTensorWrapper{name}(*{name});"
-                else:
-                    set_tensor_wrappers = (
-                        f"{indent}grad_node->SetTensorWrapper{name}({name});"
-                    )
+                set_tensor_wrappers = (
+                    f"{indent}grad_node->SetTensorWrapper{name}({name});"
+                )
                 set_output_tensor_wrappers_list.append(set_tensor_wrappers)
         set_input_tensor_wrappers_str = "\n".join(
             set_input_tensor_wrappers_list

paddle/fluid/ir/dialect/pd_op.yaml

@@ -19,7 +19,7 @@
   attrs:
   - {typename: str, name: name}
   outputs:
-  - {typename: Tensor, name: out, optional: false, intermediate: false}
+    - {typename: Tensor, name: out, optional: false, intermediate: false}
   no_need_buffer: null
   data_transform: null
   infer_meta:
@@ -35,3 +35,128 @@
     force_backend: null
   inplace: null
   backward: null
+
+- name: share_buffer_
+  inputs:
+  - typename: Tensor[]
+    name: x
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs:
+  - {typename: 'bool[]', name: share_dims_and_dtype, default_value: '{}'}
+  outputs:
+  - {typename: 'Tensor[]', name: out, size: x.size(), optional: false, intermediate: false}
+  - {typename: 'Tensor[]', name: xout, size: x.size(), optional: false, intermediate: false}
+  no_need_buffer: null
+  data_transform: null
+  inplace: null
+  backward: null
+
+- name: assert
+  inputs:
+  - typename: Tensor
+    name: cond
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  - typename: Tensor[]
+    name: data
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs:
+  - {typename: 'int64_t', name: summarize, default_value: '-1'}
+  outputs: []
+  no_need_buffer: null
+  data_transform: null
+  inplace: null
+  backward: null
+
+- name: print
+  inputs:
+  - typename: Tensor
+    name: in
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs:
+  - {typename: 'int', name: first_n}
+  - {typename: 'str', name: message}
+  - {typename: 'int', name: summarize}
+  - {typename: 'bool', name: print_tensor_name, default_value: 'true'}
+  - {typename: 'bool', name: print_tensor_type, default_value: 'true'}
+  - {typename: 'bool', name: print_tensor_shape, default_value: 'true'}
+  - {typename: 'bool', name: print_tensor_layout, default_value: 'true'}
+  - {typename: 'bool', name: print_tensor_lod, default_value: 'true'}
+  - {typename: 'str', name: print_phase, default_value: 'BOTH'}
+  - {typename: 'bool', name: is_forward, default_value: 'true'}
+  outputs:
+    - typename: Tensor
+      name: out
+      optional: false
+      no_need_buffer: false
+      data_transform: {}
+  no_need_buffer: null
+  data_transform: null
+  inplace: null
+  backward: null
+
+- name: add_n_
+  inputs:
+  - typename: Tensor[]
+    name: inputs
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs: []
+  outputs:
+  - {typename: Tensor, name: out, optional: false, intermediate: false}
+  no_need_buffer: null
+  data_transform: null
+  invoke: {func: add_n_impl, args: inputs}
+  backward: add_n_grad
+
+- name: write_to_array
+  inputs:
+  - typename: Tensor
+    name: i
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  - typename: Tensor
+    name: x
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs: []
+  outputs:
+  - {typename: 'Tensor[]', name: out, optional: false, intermediate: false}
+  no_need_buffer: null
+  data_transform: null
+  backward: write_to_array_grad
+
+- name: lod_array_length
+  inputs:
+  - typename: Tensor[]
+    name: x
+    optional: false
+    no_need_buffer: false
+    data_transform: {}
+  attrs: []
+  outputs:
+  - {typename: 'Tensor', name: out, optional: false, intermediate: false}
+  no_need_buffer: null
+  data_transform: null
+
+- name: py_func_
+  inputs:
+  -  {typename: 'Tensor', name: x, optional: false, no_need_buffer: false, data_transform: {}}
+  attrs:
+  - {typename: 'int', name: forward_callable_id, default_value: '0'}
+  - {typename: 'int', name: backward_callable_id, default_value: '-1'}
+  - {typename: 'str[]', name: backward_skip_vars, default_value: '{}'}
+  outputs:
+  - {typename: 'Tensor', name: out, optional: false, intermediate: false}
+  no_need_buffer: null
+  data_transform: null

paddle/fluid/ir/transforms/pd_op_to_kernel_pass.cc

@@ -173,6 +173,10 @@ phi::KernelKey GetKernelKey(
     }
   }
 
+  if (kernel_backend == phi::Backend::UNDEFINED) {
+    kernel_backend = paddle::experimental::ParseBackend(place);
+  }
+
   phi::KernelKey res(kernel_backend, kernel_layout, kernel_data_type);
   return res;
 }

paddle/fluid/ir_adaptor/translator/attribute_translator.cc

@@ -42,6 +42,11 @@ class AttributeVisitor {
     return ir::Int32Attribute::get(ctx, i);
   }
 
+  virtual ir::Attribute operator()(int64_t i) {
+    VLOG(10) << "translating int";
+    return ir::Int64Attribute::get(ctx, i);
+  }
+
   virtual ir::Attribute operator()(float f) {
     VLOG(10) << "translating float";
     return ir::FloatAttribute::get(ctx, f);
@@ -146,6 +151,21 @@ class AttributeVisitor {
   }
 };
 
+class Int64ArrayAttributeVisitor : public AttributeVisitor {
+ public:
+  using AttributeVisitor::AttributeVisitor;
+
+  ir::Attribute operator()(const std::vector<int>& is) override {
+    VLOG(10) << "translating vector<int64>";
+    std::vector<ir::Attribute> attrs;
+    attrs.reserve(is.size());
+    for (const auto& v : is) {
+      attrs.push_back(ir::Int64Attribute::get(ctx, v));
+    }
+    return ir::ArrayAttribute::get(ctx, attrs);
+  }
+};
+
 class IntArrayAttributeVisitor : public AttributeVisitor {
  public:
   using AttributeVisitor::AttributeVisitor;
@@ -171,6 +191,11 @@ class DataTypeAttributeVisitor : public AttributeVisitor {
     auto phi_dtype = phi::TransToPhiDataType(i);
     return paddle::dialect::DataTypeAttribute::get(ctx, phi_dtype);
   }
+
+  ir::Attribute operator()(const paddle::blank& blank) override {
+    VLOG(10) << "translating paddle::blank to DataType::UNDEFINED";
+    return paddle::dialect::DataTypeAttribute::get(ctx, phi::DataType());
+  }
 };
 
 class PlaceAttributeVisitor : public AttributeVisitor {
@@ -178,8 +203,8 @@ class PlaceAttributeVisitor : public AttributeVisitor {
   using AttributeVisitor::AttributeVisitor;
 
   ir::Attribute operator()(const paddle::blank& blank) override {
-    VLOG(10) << "translating paddle::blank";
-    phi::Place data(phi::AllocationType::CPU);
+    VLOG(10) << "translating paddle::blank to Place::UNDEFINED";
+    phi::Place data(phi::AllocationType::UNDEFINED);
     return paddle::dialect::PlaceAttribute::get(ctx, data);
   }
 };
@@ -192,6 +217,8 @@ AttributeTranslator::AttributeTranslator() {
       new DataTypeAttributeVisitor();
   special_visitors["paddle::dialect::PlaceAttribute"] =
       new PlaceAttributeVisitor();
+  special_visitors["ir::ArrayAttribute<ir::Int64Attribute>"] =
+      new Int64ArrayAttributeVisitor();
 }
 
 ir::Attribute AttributeTranslator::operator()(

paddle/fluid/ir_adaptor/translator/op_translator.h

@@ -36,7 +36,7 @@ class OpTranslator {
   using BlockDesc = paddle::framework::BlockDesc;
   using VarDesc = paddle::framework::VarDesc;
   using OpTranslateFn = std::function<ir::Operation*(
-      ir::IrContext*, TranslationContext*, ir::Program*, const OpDesc&)>;
+      ir::IrContext*, TranslationContext*, const OpDesc&, ir::Program*)>;
 
  private:
   OpTranslator();  // Disallow instantiation outside of the class.

paddle/fluid/ir_adaptor/translator/program_translator.cc

@@ -111,25 +111,46 @@ void ProgramTranslator::GetParameterForSingleBlock(const BlockDesc& block) {
     parameter_name_mappings_[var->Name()] = var;
   }
 
+  std::unordered_set<std::string> inner_defining_variables;
+
   for (auto op_desc : block.AllOps()) {
     for (const auto& n : op_desc->Inputs()) {
       const auto& input_var_names = n.second;
       for (const auto& var_name : input_var_names) {
-        bool need_get_parameter_op = (parameter_name_mappings_.find(var_name) !=
-                                      parameter_name_mappings_.end());
-        need_get_parameter_op &= (parameter_visited_.count(var_name) == 0);
+        if (no_cast_var_names.count(var_name) != 0) continue;
+        VarDesc* var_desc = nullptr;
+
+        bool is_parameter = (parameter_name_mappings_.find(var_name) !=
+                             parameter_name_mappings_.end());
+        is_parameter &= (parameter_visited_.count(var_name) == 0);
+        if (is_parameter) {
+          var_desc = parameter_name_mappings_[var_name];
+        }
+        bool is_unseen_variable =
+            (inner_defining_variables.count(var_name) == 0);
+        if (is_unseen_variable) {
+          var_desc = block.FindVarRecursive(var_name);
+        }
+
+        bool need_get_parameter_op = is_parameter || is_unseen_variable;
         if (need_get_parameter_op) {
-          ir::Operation* op =
-              InsertGetParamaterOp(ctx_, parameter_name_mappings_[var_name]);
+          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;
 
           program_->SetParameter(var_name, nullptr);
           parameter_visited_.insert(var_name);
+          inner_defining_variables.insert(var_name);
         }
       }
     }
+    for (const auto& n : op_desc->Outputs()) {
+      const auto& output_var_names = n.second;
+      for (const auto& var_name : output_var_names) {
+        inner_defining_variables.insert(var_name);
+      }
+    }
   }
 }
 
@@ -137,7 +158,7 @@ void ProgramTranslator::InsertOperationToSingleBlock(const BlockDesc& block) {
   auto& op_translator = OpTranslator::instance();
   for (auto op : block.AllOps()) {
     OpTranslateFn& fn = op_translator[op->Type()];
-    ir::Operation* operation = fn(ctx_, &param_map_, program_, *op);
+    ir::Operation* operation = fn(ctx_, &param_map_, *op, program_);
     VLOG(10) << "[op translated][special]" << operation;
   }
 }

paddle/fluid/ir_adaptor/translator/translate.cc

@@ -19,6 +19,7 @@
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/ir/dialect/pd_dialect.h"
 #include "paddle/fluid/ir_adaptor/translator/program_translator.h"
+#include "paddle/ir/core/builtin_dialect.h"
 #include "paddle/ir/core/program.h"
 
 namespace paddle {
@@ -28,7 +29,9 @@ using Program = ::ir::Program;
 
 std::unique_ptr<Program> TranslateLegacyProgramToProgram(
     const LegacyProgramDesc& legacy_program) {
-  auto program = std::make_unique<Program>(ir::IrContext::Instance());
+  ir::IrContext* ctx = ir::IrContext::Instance();
+  ctx->GetOrRegisterDialect<dialect::PaddleDialect>();
+  auto program = std::make_unique<Program>(ctx);
 
   translator::ProgramTranslator program_translator(&legacy_program,
                                                    program.get());

paddle/fluid/ir_adaptor/translator/type_translator.cc

@@ -93,6 +93,20 @@ TypeTranslator::TypeTranslator() {
          size_t offset = 0;
          return DenseTensorType::get(ctx, dtype, dim, layout, lod, offset);
        }},
+      {VarType::LOD_TENSOR_ARRAY,
+       [&](ir::IrContext* ctx, const VarDesc& var_desc) -> ir::Type {
+         VLOG(10) << "[vartype translating]"
+                  << "[" << var_desc.Name() << "] from LOD_TENSOR_ARRAY";
+
+         return ir::VectorType::get(ctx, std::vector<ir::Type>{});
+       }},
+      {VarType::SELECTED_ROWS,
+       [&](ir::IrContext* ctx, const VarDesc& var_desc) -> ir::Type {
+         VLOG(10) << "[vartype translating]"
+                  << "[" << var_desc.Name() << "] from SELECTED_ROWS";
+
+         return this->operator[](VarType::LOD_TENSOR)(ctx, var_desc);
+       }},
   };
 }
 

paddle/phi/api/yaml/legacy_ops.yaml

@@ -84,6 +84,7 @@
   kernel :
     func : assign
   backward : assign_grad
+  inplace : (x -> out)
 
 - op : assign_out_
   args : (Tensor x, Tensor output)
@@ -120,6 +121,7 @@
     data_type : x
   view : (mean -> mean_out), (variance -> variance_out)
   backward : batch_norm_grad
+  optional : reserve_space
 
 - op : cast
   args : (Tensor x, DataType dtype)

paddle/phi/api/yaml/op_compat.yaml

@@ -769,7 +769,7 @@
     attrs : ['int[] slots = {}']
 
 - op : divide (elementwise_div)
-  backward : divide_grad (elementwise_div)
+  backward : divide_grad (elementwise_div_grad)
   inputs :
     {x: X, y : Y}
   outputs :
@@ -1776,6 +1776,8 @@
 
 - op : mish
   backward : mish_grad
+  inputs:
+    lambda:  threshold
   extra :
     attrs : [bool use_mkldnn = false]
 
@@ -2839,6 +2841,8 @@
     yolo_loss : GetYoloLossExpectedKernelType
     yolo_loss_grad : GetYoloLossExpectedKernelType
 
+- op: full_batch_size_like (fill_constant_batch_size_like)
+
 - op: lu
   backward: lu_grad
   inputs:

test/ir/new_ir/test_special_op_translator.py

+# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import unittest
+
+import numpy as np
+
+import paddle
+from paddle.fluid import core
+
+paddle.enable_static()
+
+
+class TestCastOpTranscriber(unittest.TestCase):
+    def test_op(self):
+        place = core.Place()
+        place.set_place(paddle.CPUPlace())
+        new_scope = paddle.static.Scope()
+        main_program = paddle.static.Program()
+        with paddle.static.scope_guard(new_scope):
+            with paddle.static.program_guard(main_program):
+                x = paddle.to_tensor([2, 3, 4], 'float64')
+                y = paddle.cast(x, 'uint8')
+
+                default_job = core.Job("default")
+                type_to_program = {"default": main_program.desc}
+                plan = core.Plan([default_job], type_to_program)
+                new_exe = core.StandaloneExecutor(place, plan, new_scope)
+
+
+class TestEmbeddingOpTranscriber(unittest.TestCase):
+    def test_op(self):
+        place = core.Place()
+        place.set_place(paddle.CPUPlace())
+        new_scope = paddle.static.Scope()
+        main_program = paddle.static.Program()
+        with paddle.static.scope_guard(new_scope):
+            with paddle.static.program_guard(main_program):
+                x = paddle.static.data(name="x", shape=[2, 4], dtype=np.int64)
+                embedding = paddle.nn.Embedding(
+                    10, 3, weight_attr=paddle.nn.initializer.Constant(value=1.0)
+                )
+                output = embedding(x)
+
+                default_job = core.Job("default")
+                type_to_program = {"default": main_program.desc}
+                plan = core.Plan([default_job], type_to_program)
+                new_exe = core.StandaloneExecutor(place, plan, new_scope)
+
+
+class TestIncrementOpTranscriber(unittest.TestCase):
+    def test_op(self):
+        place = core.Place()
+        place.set_place(paddle.CPUPlace())
+        new_scope = paddle.static.Scope()
+        main_program = paddle.static.Program()
+        with paddle.static.scope_guard(new_scope):
+            with paddle.static.program_guard(main_program):
+                data = paddle.zeros(shape=[1], dtype='float32')
+                counter = paddle.increment(data)
+
+                default_job = core.Job("default")
+                type_to_program = {"default": main_program.desc}
+                plan = core.Plan([default_job], type_to_program)
+                new_exe = core.StandaloneExecutor(place, plan, new_scope)
+
+
+class TestAssignValueOpTranscriber(unittest.TestCase):
+    def test_op(self):
+        place = core.Place()
+        place.set_place(paddle.CPUPlace())
+        new_scope = paddle.static.Scope()
+        main_program = paddle.static.Program()
+        with paddle.static.scope_guard(new_scope):
+            with paddle.static.program_guard(main_program):
+                x = paddle.to_tensor(
+                    [[0.1, 0.2], [0.3, 0.4]],
+                    place=paddle.CPUPlace(),
+                    stop_gradient=False,
+                )
+
+                default_job = core.Job("default")
+                type_to_program = {"default": main_program.desc}
+                plan = core.Plan([default_job], type_to_program)
+                new_exe = core.StandaloneExecutor(place, plan, new_scope)
+
+
+class TestRnnOpTranscriber(unittest.TestCase):
+    def test_op(self):
+        place = core.Place()
+        place.set_place(paddle.CPUPlace())
+        new_scope = paddle.static.Scope()
+        main_program = paddle.static.Program()
+        with paddle.static.scope_guard(new_scope):
+            with paddle.static.program_guard(main_program):
+                x = paddle.randn((4, 16))
+                prev_h = paddle.randn((4, 32))
+
+                cell = paddle.nn.SimpleRNNCell(16, 32)
+                y, h = cell(x, prev_h)
+
+                default_job = core.Job("default")
+                type_to_program = {"default": main_program.desc}
+                plan = core.Plan([default_job], type_to_program)
+                new_exe = core.StandaloneExecutor(place, plan, new_scope)
+
+
+if __name__ == "__main__":
+    unittest.main()