[Inference]Fix analyzer_capi_exp_tester and analyzer_capi_exp_ner_tester in PIR mode (#67248)

YuanRisheng · web-flow · commit b8a44dabe612 · 2024-08-16T17:21:50.000+08:00
* fix analyzer_capi_exp_tester and analyzer_capi_exp_ner_tester

* fix download model

* fix random
diff --git a/paddle/fluid/ir_adaptor/translator/op_compat_gen.py b/paddle/fluid/ir_adaptor/translator/op_compat_gen.py
@@ -140,6 +140,13 @@ def insert_new_mutable_attributes(
         "out": "Output",
     }
 
+    op_name_mappings["lookup_table"] = "embedding"
+    op_arg_name_mappings["lookup_table"] = {
+        "x": "Ids",
+        "weight": "W",
+        "out": "Out",
+    }
+
     op_arg_name_mappings["set_value_grad"]["values_grad"] = "ValueTensor@GRAD"
     op_arg_name_mappings["fetch"] = {"x": "X"}
     op_arg_name_mappings["elementwise_add_grad_grad"] = {
diff --git a/paddle/fluid/ir_adaptor/translator/op_translator.cc b/paddle/fluid/ir_adaptor/translator/op_translator.cc
@@ -1085,7 +1085,105 @@ struct Conv2dOpTranscriber : public OpTranscriber {
   }
 };
 
+using ValueInfo =
+    std::tuple<std::vector<int64_t>, dialect::DenseTensorType, pir::Value>;
+
+ValueInfo GetTensorInfoByVarName(const OpDesc& op_desc,
+                                 const std::vector<std::string>& names,
+                                 TranslationContext* param_map,
+                                 const std::string& var_name) {
+  PADDLE_ENFORCE_EQ(
+      names.size(),
+      1UL,
+      common::errors::InvalidArgument(
+          "Expected op[%s]'s input %s has only 1 variable, but got %d",
+          op_desc.Type(),
+          var_name,
+          names.size()));
+  const auto& name = names[0];
+  PADDLE_ENFORCE_GT(
+      param_map->count(name),
+      0UL,
+      common::errors::InvalidArgument(
+          "Expected op[%s]'s input %s has been parsed", op_desc.Type(), name));
+  const auto& defining_info = param_map->at(name);
+
+  pir::Value value = defining_info.value;
+  PADDLE_ENFORCE_NE(
+      value,
+      nullptr,
+      common::errors::PreconditionNotMet(
+          "Expected op[%s]'s input %s is not null", op_desc.Type(), name));
+  const pir::Type& type = value.type();
+  PADDLE_ENFORCE_EQ(type.isa<dialect::DenseTensorType>(),
+                    true,
+                    common::errors::InvalidArgument(
+                        "Expected op[%s]'s input %s is DenseTensor but got %s",
+                        op_desc.Type(),
+                        name,
+                        type));
+  dialect::DenseTensorType tensor_type =
+      type.dyn_cast<dialect::DenseTensorType>();
+
+  std::vector<int64_t> shape = common::vectorize(tensor_type.dims());
+
+  return std::make_tuple(shape, tensor_type, value);
+}
+
 struct EmbeddingOpTranscriber : public OpTranscriber {
+  void RecordOpResultMapping(pir::IrContext* ctx,
+                             TranslationContext* param_map,
+                             const OpDesc& op_desc,
+                             pir::Operation* operation,
+                             const OpOutputMapping& arg_to_idx) override {
+    OpTranscriber::RecordOpResultMapping(
+        ctx, param_map, op_desc, operation, arg_to_idx);
+    if (op_desc.Type() == "lookup_table") {
+      ValueInfo out_info = GetTensorInfoByVarName(
+          op_desc, op_desc.Output("Out"), param_map, "Out");
+      const auto& output_vars = op_desc.Output("Out");
+      const auto& output_name = output_vars[0];
+
+      const dialect::DenseTensorType& out_tensor_type = std::get<1>(out_info);
+      pir::Value& out_value = std::get<2>(out_info);
+
+      ValueInfo ids_info = GetTensorInfoByVarName(
+          op_desc, op_desc.Input("Ids", true), param_map, "Ids");
+      const std::vector<int64_t>& ids_shape = std::get<0>(ids_info);
+
+      ValueInfo w_info = GetTensorInfoByVarName(
+          op_desc, op_desc.Input("W", true), param_map, "W");
+
+      const std::vector<int64_t>& w_shape = std::get<0>(w_info);
+
+      std::vector<int64_t> out_new_shape(
+          ids_shape.begin(), ids_shape.begin() + ids_shape.size() - 1);
+      out_new_shape.insert(out_new_shape.end(), w_shape[1]);
+
+      pir::Builder builder(ctx, operation->GetParent());
+      dialect::ReshapeOp reshape_op_out =
+          builder.Build<dialect::ReshapeOp>(out_value, out_new_shape);
+      pir::Value out_new = reshape_op_out.out();
+      VLOG(6) << "[" << op_desc.Type() << "] out_shape change from "
+              << out_tensor_type.dims() << " to "
+              << common::make_ddim(out_new_shape);
+
+      param_map->PushValue(output_name,
+                           VariableDefiningInfo(out_new, false, -1));
+    }
+  }
+
+  pir::OpInfo LookUpOpInfo(pir::IrContext* ctx,
+                           const OpDesc& op_desc) override {
+    auto op_info = ctx->GetRegisteredOpInfo("pd_op.embedding");
+    if (!op_info) {
+      IR_THROW("Op %d should have corresponding OpInfo %d",
+               op_desc.Type(),
+               "pd_op.embedding");
+    }
+    return op_info;
+  }
+
   void HandleNonexistentAttribute(pir::IrContext* ctx,
                                   pir::AttributeMap* attribute_map,
                                   const OpAttributeInfo& info) override {
@@ -1393,6 +1491,20 @@ struct CrossEntropyWithSoftmaxOpTranscriber : public OpTranscriber {
   }
 };
 
+struct BoxCoderOpTranscriber : public OpTranscriber {
+  void HandleNonexistentAttribute(pir::IrContext* ctx,
+                                  pir::AttributeMap* attribute_map,
+                                  const OpAttributeInfo& info) override {
+    if (info.name == "axis") {
+      (*attribute_map)[info.name] = pir::Int32Attribute::get(ctx, 0);
+    }
+    if (info.name == "variance") {
+      std::vector<pir::Attribute> variance;
+      (*attribute_map)[info.name] = pir::ArrayAttribute::get(ctx, variance);
+    }
+  }
+};
+
 struct DepthwiseConv2dOpTranscriber : public OpTranscriber {
   void HandleNonexistentAttribute(pir::IrContext* ctx,
                                   pir::AttributeMap* attribute_map,
@@ -1657,51 +1769,6 @@ struct TrilAndTriuGradOpTranscriber : public OpTranscriber {
   }
 };
 
-using ValueInfo =
-    std::tuple<std::vector<int64_t>, dialect::DenseTensorType, pir::Value>;
-
-ValueInfo GetTensorInfoByVarName(const OpDesc& op_desc,
-                                 const std::vector<std::string>& names,
-                                 TranslationContext* param_map,
-                                 const std::string& var_name) {
-  PADDLE_ENFORCE_EQ(
-      names.size(),
-      1UL,
-      common::errors::InvalidArgument(
-          "Expected op[%s]'s input %s has only 1 variable, but got %d",
-          op_desc.Type(),
-          var_name,
-          names.size()));
-  const auto& name = names[0];
-  PADDLE_ENFORCE_GT(
-      param_map->count(name),
-      0UL,
-      common::errors::InvalidArgument(
-          "Expected op[%s]'s input %s has been parsed", op_desc.Type(), name));
-  const auto& defining_info = param_map->at(name);
-
-  pir::Value value = defining_info.value;
-  PADDLE_ENFORCE_NE(
-      value,
-      nullptr,
-      common::errors::PreconditionNotMet(
-          "Expected op[%s]'s input %s is not null", op_desc.Type(), name));
-  const pir::Type& type = value.type();
-  PADDLE_ENFORCE_EQ(type.isa<dialect::DenseTensorType>(),
-                    true,
-                    common::errors::InvalidArgument(
-                        "Expected op[%s]'s input %s is DenseTensor but got %s",
-                        op_desc.Type(),
-                        name,
-                        type));
-  dialect::DenseTensorType tensor_type =
-      type.dyn_cast<dialect::DenseTensorType>();
-
-  std::vector<int64_t> shape = common::vectorize(tensor_type.dims());
-
-  return std::make_tuple(shape, tensor_type, value);
-}
-
 struct MulOpTranscriber : public OpTranscriber {
   pir::Operation* operator()(pir::IrContext* ctx,
                              TranslationContext* param_map,
@@ -3566,6 +3633,37 @@ struct CEmbeddingOpTranscriber : public OpTranscriber {
   }
 };
 
+struct GatherOpTranscriber : public OpTranscriber {
+  pir::Value GetAttributeAsInput(pir::IrContext* ctx,
+                                 pir::Block* block,
+                                 const OpDesc& op_desc,
+                                 const OpInputInfo& input_info) override {
+    auto& attribute_translator = AttributeTranslator::instance();
+    auto& op_normalizer = OpNameNormalizer::instance();
+
+    auto legacy_attr_name =
+        op_normalizer.GetLegacyAttrName(op_desc.Type(), input_info.name);
+
+    if (!op_desc.HasAttr(legacy_attr_name)) {
+      VLOG(10) << "[" << op_desc.Type() << "][attribute]"
+               << " name: " << legacy_attr_name << " not found and fill 0.";
+      pir::Attribute new_attr = pir::Int64Attribute::get(ctx, 0);
+      pir::Operation* defining_op =
+          InsertFullOperationForAttributeInput(ctx, block, new_attr);
+      return defining_op->result(0);
+    } else {
+      paddle::framework::Attribute legacy_attr =
+          op_desc.GetAttr(legacy_attr_name);
+      VLOG(10) << "[" << op_desc.Type() << "][attribute]"
+               << " name: " << legacy_attr_name << " " << legacy_attr.index();
+      pir::Attribute new_attr = attribute_translator(legacy_attr);
+      pir::Operation* defining_op =
+          InsertFullOperationForAttributeInput(ctx, block, new_attr);
+      return defining_op->result(0);
+    }
+  }
+};
+
 struct QuantizeLinearOpTranscriber : public OpTranscriber {
   void HandleNonexistentAttribute(pir::IrContext* ctx,
                                   pir::AttributeMap* attribute_map,
@@ -3756,6 +3854,7 @@ OpTranslator::OpTranslator() {
   special_handlers["increment"] = IncrementOpTranscriber();
   special_handlers["lookup_table_v2"] = EmbeddingOpTranscriber();
   special_handlers["lookup_table_v2_grad"] = EmbeddingGradOpTranscriber();
+  special_handlers["lookup_table"] = EmbeddingOpTranscriber();
   special_handlers["one_hot_v2"] = OneHotTranscriber();
   special_handlers["pool2d"] = Pool2dOpTranscriber();
   special_handlers["randint"] = RandIntOpTranscriber();
@@ -3784,6 +3883,8 @@ OpTranslator::OpTranslator() {
   special_handlers["softmax"] = SoftmaxOpTranscriber();
   special_handlers["softmax_with_cross_entropy"] =
       SoftmaxWithCrossEntropyOpTranscriber();
+  special_handlers["gather"] = GatherOpTranscriber();
+  special_handlers["box_coder"] = BoxCoderOpTranscriber();
 
   // To adapt LodTensorArray
   special_handlers["lod_array_length"] = LodArrayLengthOpTranscriber();
diff --git a/paddle/phi/infermeta/binary.cc b/paddle/phi/infermeta/binary.cc
@@ -2904,6 +2904,7 @@ void MatmulInferMeta(const MetaTensor& x,
     out->set_dtype(x.dtype());
   }
   out->set_layout(x.layout());
+  out->share_lod(x);
 }
 
 void MatmulWithFlattenInferMeta(const MetaTensor& x,
diff --git a/test/cpp/inference/api/CMakeLists.txt b/test/cpp/inference/api/CMakeLists.txt
@@ -1007,9 +1007,31 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST)
       --infer_model=${INT8_DATA_DIR}/resnet50/model)
   endif()
 
+  # chinese_ner
+  set(CHINESE_NER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/chinese_ner")
+  download_model_and_data_without_verify(
+    ${CHINESE_NER_INSTALL_DIR} "chinese_ner_model.tar.gz"
+    "chinese_ner-data.txt.tar.gz")
+  inference_analysis_test(
+    test_analyzer_capi_exp_ner
+    SRCS
+    analyzer_capi_exp_ner_tester.cc
+    EXTRA_DEPS
+    paddle_inference_c_shared
+    ARGS
+    --infer_model=${CHINESE_NER_INSTALL_DIR}/model)
+
   # resnet50
   set(RESNET50_MODEL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/resnet50")
   download_data_without_verify(${RESNET50_MODEL_DIR} "resnet50_model.tar.gz")
+  inference_analysis_test(
+    test_analyzer_capi_exp
+    SRCS
+    analyzer_capi_exp_tester.cc
+    EXTRA_DEPS
+    paddle_inference_c_shared
+    ARGS
+    --infer_model=${RESNET50_MODEL_DIR}/model)
 
   if(WITH_GPU)
     inference_analysis_test(
diff --git a/test/cpp/inference/api/analyzer_capi_exp_ner_tester.cc b/test/cpp/inference/api/analyzer_capi_exp_ner_tester.cc
diff --git a/test/cpp/inference/api/analyzer_capi_exp_tester.cc b/test/cpp/inference/api/analyzer_capi_exp_tester.cc
diff --git a/test/deprecated/cpp/inference/api/CMakeLists.txt b/test/deprecated/cpp/inference/api/CMakeLists.txt
@@ -500,15 +500,6 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST)
     test_analyzer_multi_model_prediction ${MMP_INSTALL_DIR}
     analyzer_mmp_tester.cc EXTRA_DEPS common)
 
-  inference_analysis_test(
-    test_analyzer_capi_exp
-    SRCS
-    analyzer_capi_exp_tester.cc
-    EXTRA_DEPS
-    paddle_inference_c_shared
-    ARGS
-    --infer_model=${RESNET50_MODEL_DIR}/model)
-
   inference_analysis_test(
     test_analyzer_capi_exp_pd_tensor
     SRCS
@@ -563,15 +554,6 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST)
     --infer_data=${OCR_INSTALL_DIR}/data.txt
     --refer_result=${OCR_INSTALL_DIR}/result.txt)
 
-  inference_analysis_test(
-    test_analyzer_capi_exp_ner
-    SRCS
-    analyzer_capi_exp_ner_tester.cc
-    EXTRA_DEPS
-    paddle_inference_c_shared
-    ARGS
-    --infer_model=${CHINESE_NER_INSTALL_DIR}/model)
-
   set_tests_properties(test_analyzer_mobilenet_transpose PROPERTIES TIMEOUT 120)
   set_tests_properties(test_analyzer_resnet50 PROPERTIES TIMEOUT 120)
   set_tests_properties(test_analyzer_ner PROPERTIES TIMEOUT 120)
diff --git a/test/legacy_test/test_uniform_random_op.py b/test/legacy_test/test_uniform_random_op.py
@@ -340,14 +340,12 @@ def test_api(self):
             y = linear(x)
 
             place = base.CPUPlace()
-            x_tensor = base.create_lod_tensor(
-                np.random.rand(3, 16).astype("float32"), [[1, 2]], place
-            )
+            x_data = np.random.rand(3, 16).astype("float32")
             exe = base.Executor(place)
             exe.run(paddle.static.default_startup_program())
             ret = exe.run(
                 paddle.static.default_main_program(),
-                feed={'x': x_tensor},
+                feed={'x': x_data},
                 fetch_list=[y],
                 return_numpy=False,
             )

Original file line number	Diff line number	Diff line change
`@@ -2904,6 +2904,7 @@ void MatmulInferMeta(const MetaTensor& x,`
`2904`	`2904`	`out->set_dtype(x.dtype());`
`2905`	`2905`	`}`
`2906`	`2906`	`out->set_layout(x.layout());`
	`2907`	`+ out->share_lod(x);`
`2907`	`2908`	`}`
`2908`	`2909`
`2909`	`2910`	`void MatmulWithFlattenInferMeta(const MetaTensor& x,`