[cherry-pick] Optimize performance of dygraph (#42093, #42103, #42137) (#42171)

* optimiaze performance of PreparePhiData (#42093)

* Dygraph performance optimization (v2) (#42103)

* optimiaze performance of PreparePhiData

* dygraph performance optimization

* optimize performance of dygraph (#42137)

paddle/fluid/framework/infershape_utils.cc

@@ -402,21 +402,20 @@ std::vector<phi::MetaTensor*> CompatInferMetaContext::MutableOutputBetween(
 CompatInferMetaContext BuildInferMetaContext(InferShapeContext* ctx,
                                              const std::string& op_type) {
   // 1. get kernel args
-  auto arg_map_fn = phi::OpUtilsMap::Instance().GetArgumentMappingFn(op_type);
-  PADDLE_ENFORCE_NOT_NULL(
-      arg_map_fn, platform::errors::NotFound(
-                      "The ArgumentMappingFn of %s op is not found.", op_type));
+  auto* arg_map_fn = phi::OpUtilsMap::Instance().GetArgumentMappingFn(op_type);
   InferShapeArgumentMappingContext arg_map_context(*ctx);
-  auto signature = arg_map_fn(arg_map_context);
+  KernelSignature signature =
+      arg_map_fn ? (*arg_map_fn)(arg_map_context)
+                 : phi::DefaultKernelSignatureMap::Instance().Get(op_type);
   VLOG(3) << "BuildInferMetaContext: op kernel signature - " << signature;
 
   // 2. build infermeta context
   CompatInferMetaContext infer_meta_context(
       {ctx->IsRuntime(), ctx->IsRunMKLDNNKernel()});
 
-  auto& input_names = std::get<0>(signature.args);
-  auto& attr_names = std::get<1>(signature.args);
-  auto& output_names = std::get<2>(signature.args);
+  const auto& input_names = signature.input_names;
+  const auto& attr_names = signature.attr_names;
+  const auto& output_names = signature.output_names;
 
   const auto& args_def =
       phi::KernelFactory::Instance().GetFirstKernelArgsDef(signature.name);

paddle/fluid/framework/operator.cc

@@ -1200,8 +1200,10 @@ bool OperatorWithKernel::SupportsMKLDNN(
 
 bool OperatorWithKernel::CanMKLDNNBeUsed(const framework::ExecutionContext& ctx,
                                          proto::VarType::Type data_type) const {
-  bool use_mkldnn_ctx = ctx.HasAttr("use_mkldnn") &&
-                        ctx.Attr<bool>("use_mkldnn") &&
+  const auto& attrs_map = ctx.Attrs();
+  auto iter = attrs_map.find("use_mkldnn");
+  bool use_mkldnn_ctx = iter != attrs_map.end() &&
+                        BOOST_GET_CONST(bool, iter->second) &&
                         platform::is_cpu_place(ctx.GetPlace());
   return use_mkldnn_ctx && this->SupportsMKLDNN(data_type);
 }
@@ -2117,8 +2119,16 @@ KernelSignature OperatorWithKernel::GetExpectedPhiKernelArgs(
     const ExecutionContext& ctx) const {
   ExecutionArgumentMappingContext arg_mapping_ctx(ctx);
   if (arg_map_fn_ == nullptr) {
-    arg_map_fn_.reset(new phi::ArgumentMappingFn(
-        phi::OpUtilsMap::Instance().GetArgumentMappingFn(Type())));
+    auto* arg_map_fn = phi::OpUtilsMap::Instance().GetArgumentMappingFn(type_);
+    if (arg_map_fn) {
+      arg_map_fn_.reset(new phi::ArgumentMappingFn(*arg_map_fn));
+    } else {
+      auto func =
+          [this](const phi::ArgumentMappingContext& ctx) -> KernelSignature {
+        return phi::DefaultKernelSignatureMap::Instance().Get(type_);
+      };
+      arg_map_fn_.reset(new phi::ArgumentMappingFn(func));
+    }
   }
   return (*arg_map_fn_)(arg_mapping_ctx);
 }
@@ -2126,7 +2136,7 @@ KernelSignature OperatorWithKernel::GetExpectedPhiKernelArgs(
 Scope* OperatorWithKernel::PreparePhiData(
     const Scope& scope, const phi::Kernel& pt_kernel,
     const KernelSignature& pt_kernel_signature, RuntimeContext* ctx) const {
-  auto& input_names = std::get<0>(pt_kernel_signature.args);
+  const auto& input_names = pt_kernel_signature.input_names;
   auto input_defs = pt_kernel.args_def().input_defs();
   PADDLE_ENFORCE_EQ(input_names.size(), input_defs.size(),
                     platform::errors::InvalidArgument(
@@ -2178,11 +2188,15 @@ Scope* OperatorWithKernel::PreparePhiData(
       if (in_def.backend == phi::Backend::ALL_BACKEND) {
         continue;
       }
-      auto expected_place = phi::TransToPhiPlace(in_def.backend);
-      if (platform::is_same_place(tensor_in->place(), expected_place)) {
+
+      auto tensor_backend = phi::TransToPhiBackend(tensor_in->place());
+      if (in_def.backend == tensor_backend ||
+          (in_def.backend == phi::Backend::GPUDNN &&
+           tensor_backend == phi::Backend::GPU)) {
         continue;
       }
 
+      auto expected_place = phi::TransToPhiPlace(in_def.backend);
       VLOG(3) << "phi Transform Variable " << input_names[i] << " from "
               << tensor_in->place() << " to " << expected_place;
 
@@ -2219,9 +2233,9 @@ void OperatorWithKernel::BuildPhiKernelContext(
     phi::KernelContext* pt_kernel_context) const {
   pt_kernel_context->SetDeviceContext(dev_ctx);
 
-  auto& input_names = std::get<0>(pt_kernel_signature_->args);
-  auto& attr_names = std::get<1>(pt_kernel_signature_->args);
-  auto& output_names = std::get<2>(pt_kernel_signature_->args);
+  auto& input_names = pt_kernel_signature_->input_names;
+  auto& attr_names = pt_kernel_signature_->attr_names;
+  auto& output_names = pt_kernel_signature_->output_names;
 
   auto input_defs = pt_kernel_->args_def().input_defs();
   auto attr_defs = pt_kernel_->args_def().attribute_defs();

paddle/fluid/imperative/prepared_operator.cc

@@ -37,6 +37,8 @@ namespace paddle {
 namespace imperative {
 
 static const phi::Kernel empty_kernel;
+static const framework::RuntimeContext empty_ctx({}, {});
+static const framework::Scope empty_scope;
 
 const std::shared_ptr<VariableWrapper>& GetVariableWrapper(
     const std::shared_ptr<paddle::imperative::VarBase>& var) {
@@ -138,8 +140,6 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
   platform::DeviceContextPool& pool = platform::DeviceContextPool::Instance();
   auto* dev_ctx = pool.Get(place);
 
-  framework::RuntimeContext ctx({}, {});
-
 #ifdef PADDLE_WITH_MKLDNN
   // MKLDNN variant of code reads attributes in some of GetKernelTypeForVar and
   // GetKernelType functions, so we need to copy the attributes there.
@@ -158,7 +158,7 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
 
   // 1. get expected kernel key
   auto dygraph_exe_ctx = DygraphExecutionContext<VarType>(
-      op, framework::Scope(), *dev_ctx, ctx, ins, outs, attrs, default_attrs);
+      op, empty_scope, *dev_ctx, empty_ctx, ins, outs, attrs, default_attrs);
   auto expected_kernel_key = op.GetExpectedKernelType(dygraph_exe_ctx);
 
   framework::KernelSignature pt_kernel_signature;
@@ -172,11 +172,26 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
       paddle::platform::is_in_xpu_black_list(op.Type());
 
 #endif
-  if (phi::KernelFactory::Instance().HasCompatiblePhiKernel(op.Type())) {
-    pt_kernel_signature =
-        std::move(op.GetExpectedPhiKernelArgs(dygraph_exe_ctx));
-    VLOG(6) << pt_kernel_signature;
 
+  bool has_phi_kernel = false;
+
+  const auto* arg_map_fn =
+      phi::OpUtilsMap::Instance().GetArgumentMappingFn(op.Type());
+  if (arg_map_fn) {
+    has_phi_kernel = true;
+    pt_kernel_signature = (*arg_map_fn)(
+        framework::ExecutionArgumentMappingContext(dygraph_exe_ctx));
+  } else {
+    const auto* kernel_sig =
+        phi::DefaultKernelSignatureMap::Instance().GetNullable(op.Type());
+    if (kernel_sig) {
+      has_phi_kernel = true;
+      pt_kernel_signature = *kernel_sig;
+    }
+  }
+
+  if (has_phi_kernel) {
+    VLOG(6) << pt_kernel_signature;
     pt_kernel_name = pt_kernel_signature.name;
 // NOTE(Liu-xiandong): The register kernel used KP have library_type[KP],
 // But the default library_type is Plain, so we need to modify the
@@ -231,7 +246,7 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
         dev_ctx = pool.Get(expected_kernel_key.place_);
       }
 
-      return PreparedOp(op, ctx, expected_kernel_key,
+      return PreparedOp(op, empty_ctx, expected_kernel_key,
                         std::move(pt_kernel_signature), pt_kernel, dev_ctx);
     } else {
       VLOG(6) << "Dynamic mode ChoosePhiKernel - kernel `" << pt_kernel_name
@@ -280,7 +295,7 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
                 << " | kernel key: " << pt_cpu_kernel_key
                 << " | kernel: " << pt_cpu_kernel;
         auto* cpu_ctx = pool.Get(paddle::platform::CPUPlace());
-        return PreparedOp(op, ctx, expected_kernel_key,
+        return PreparedOp(op, empty_ctx, expected_kernel_key,
                           std::move(pt_kernel_signature), pt_cpu_kernel,
                           cpu_ctx);
       }
@@ -373,7 +388,8 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
     dev_ctx = pool.Get(expected_kernel_key.place_);
   }
 
-  return PreparedOp(op, ctx, expected_kernel_key, kernel_iter->second, dev_ctx);
+  return PreparedOp(op, empty_ctx, expected_kernel_key, kernel_iter->second,
+                    dev_ctx);
 }
 
 PreparedOp PreparedOp::Prepare(const NameVarMap<VarBase>& ins,

paddle/fluid/imperative/prepared_operator.h

@@ -233,9 +233,9 @@ void BuildDygraphPhiKernelContext(
     platform::DeviceContext* dev_ctx, phi::KernelContext* kernel_ctx) {
   kernel_ctx->SetDeviceContext(dev_ctx);
 
-  auto& input_names = std::get<0>(pt_kernel_signature.args);
-  auto& attr_names = std::get<1>(pt_kernel_signature.args);
-  auto& output_names = std::get<2>(pt_kernel_signature.args);
+  const auto& input_names = pt_kernel_signature.input_names;
+  const auto& attr_names = pt_kernel_signature.attr_names;
+  const auto& output_names = pt_kernel_signature.output_names;
 
   auto& input_defs = pt_kernel.args_def().input_defs();
   auto& output_defs = pt_kernel.args_def().output_defs();
@@ -570,7 +570,7 @@ template <typename VarType>
 void PreparePhiData(const phi::Kernel& pt_kernel,
                     const framework::KernelSignature& pt_kernel_signature,
                     const NameVarMap<VarType>& ins) {
-  auto& input_names = std::get<0>(pt_kernel_signature.args);
+  const auto& input_names = pt_kernel_signature.input_names;
   auto& input_defs = pt_kernel.args_def().input_defs();
 
   PADDLE_ENFORCE_EQ(input_names.size(), input_defs.size(),
@@ -581,10 +581,11 @@ void PreparePhiData(const phi::Kernel& pt_kernel,
 
   for (size_t i = 0; i < input_names.size(); ++i) {
     auto& in_def = input_defs.at(i);
-    if (ins.find(input_names[i]) == ins.end()) {
+    auto iter = ins.find(input_names[i]);
+    if (iter == ins.end()) {
       continue;
     }
-    auto& ins_vector = ins.at(input_names[i]);
+    auto& ins_vector = iter->second;
 
     for (size_t offset = 0; offset < ins_vector.size(); ++offset) {
       auto& var = ins_vector[offset];
@@ -593,11 +594,15 @@ void PreparePhiData(const phi::Kernel& pt_kernel,
         if (in_def.backend == phi::Backend::ALL_BACKEND) {
           continue;
         }
-        auto expected_place = phi::TransToPhiPlace(in_def.backend);
-        if (platform::is_same_place(tensor_in->place(), expected_place)) {
+        auto tensor_backend = phi::TransToPhiBackend(tensor_in->place());
+        if (in_def.backend == tensor_backend ||
+            (in_def.backend == phi::Backend::GPUDNN &&
+             tensor_backend == phi::Backend::GPU)) {
           continue;
         }
 
+        auto expected_place = phi::TransToPhiPlace(in_def.backend);
+
         VLOG(3) << "Phi Transform Variable " << input_names[i] << " from "
                 << tensor_in->place() << " to " << expected_place;
 

paddle/fluid/pybind/imperative.cc

@@ -2050,9 +2050,9 @@ void BindImperative(py::module *m_ptr) {
               };
               auto ret = self.GetExpectedKernelSignature(type, ins_map,
                                                          outs_map, attrs);
-              auto kernelsig_ins = input_to_vector(std::get<0>(ret.args));
-              auto kernelsig_attrs = attr_to_vector(std::get<1>(ret.args));
-              auto kernelsig_outs = output_to_vector(std::get<2>(ret.args));
+              auto kernelsig_ins = input_to_vector(ret.input_names);
+              auto kernelsig_attrs = attr_to_vector(ret.attr_names);
+              auto kernelsig_outs = output_to_vector(ret.output_names);
               return std::make_tuple(kernelsig_ins, kernelsig_attrs,
                                      kernelsig_outs);
             }

paddle/fluid/pybind/kernel_signature_generator.cc

@@ -58,10 +58,10 @@ int main(int argc, char **argv) {
     if (kernel_signature_map.Has(op_name)) {
       kernel_signature_map_str =
           kernel_signature_map_str + "\"" + op_kernel_pair.first + "\":{";
-      auto &args = kernel_signature_map.Get(op_name).args;
+      const auto &args = kernel_signature_map.Get(op_name);
 
       kernel_signature_map_str += "\"inputs\":[";
-      auto inputs_ = std::get<0>(args);
+      auto inputs_ = args.input_names;
       for (size_t i = 0; i < inputs_.size(); i++) {
         kernel_signature_map_str =
             kernel_signature_map_str + "\"" + inputs_[i] + "\",";
@@ -69,14 +69,14 @@ int main(int argc, char **argv) {
       if (inputs_.size()) kernel_signature_map_str.pop_back();
 
       kernel_signature_map_str += "],\"attrs\":[";
-      auto attrs_ = std::get<1>(args);
+      auto attrs_ = args.attr_names;
       for (size_t i = 0; i < attrs_.size(); i++) {
         kernel_signature_map_str =
             kernel_signature_map_str + "\"" + attrs_[i] + "\",";
       }
       if (attrs_.size()) kernel_signature_map_str.pop_back();
       kernel_signature_map_str += "],\"outputs\":[";
-      auto outputs_ = std::get<2>(args);
+      auto outputs_ = args.output_names;
       for (size_t i = 0; i < outputs_.size(); i++) {
         kernel_signature_map_str =
             kernel_signature_map_str + "\"" + outputs_[i] + "\",";

paddle/infrt/dialect/phi/pass/phi_op_convert_pass.cc

@@ -108,14 +108,14 @@ void PhiOpConvertPass::convertStage() {
       op->replaceAllUsesWith(kernel_op.getResults());
     } else {
       ::phi::KernelSignature kernel_sign =
-          ::phi::OpUtilsMap::Instance().GetArgumentMappingFn(op_name)(
+          (*::phi::OpUtilsMap::Instance().GetArgumentMappingFn(op_name))(
               infrt::ProtoArgumentMappingContext(op));
       VLOG(3) << "IncompatiblePhiKernel: op(" << op_name << "), kernel("
               << kernel_sign.name << ")";
       // resort input&output according to kernel_sign
       ::llvm::SmallVector<mlir::Value, 4> inputs, ori_output;
       ::llvm::SmallVector<mlir::Type, 4> output_types;
-      for (const std::string &str : std::get<0>(kernel_sign.args)) {
+      for (const std::string &str : kernel_sign.input_names) {
         if (pd_dialect_inputs_info_map_.at(op_name).count(str) == 0) {
           LOG(ERROR) << "No input info for Op " << op_name << " and argument "
                      << str;
@@ -125,7 +125,7 @@ void PhiOpConvertPass::convertStage() {
         inputs.push_back(op->getOperands()[index]);
       }
 
-      for (const std::string &str : std::get<2>(kernel_sign.args)) {
+      for (const std::string &str : kernel_sign.output_names) {
         if (pd_dialect_outputs_info_map_.at(op_name).count(str) == 0) {
           LOG(ERROR) << "No output info for Op " << op_name << " and argument "
                      << str;

paddle/phi/core/compat/arg_map_context.cc

@@ -20,11 +20,11 @@ limitations under the License. */
 namespace phi {
 std::ostream& operator<<(std::ostream& os, KernelSignature signature) {
   os << "Kernel Signature - name: " << signature.name << "; inputs: "
-     << paddle::string::join_strings(std::get<0>(signature.args), ", ")
+     << paddle::string::join_strings(signature.input_names, ", ")
      << "; attributes: "
-     << paddle::string::join_strings(std::get<1>(signature.args), ", ")
+     << paddle::string::join_strings(signature.attr_names, ", ")
      << "; outputs: "
-     << paddle::string::join_strings(std::get<2>(signature.args), ", ");
+     << paddle::string::join_strings(signature.output_names, ", ");
   return os;
 }
 

paddle/phi/core/compat/arg_map_context.h

@@ -33,7 +33,9 @@ using KernelArgsTuple = std::tuple<paddle::SmallVector<const char*>,
 
 struct KernelSignature {
   const char* name;
-  KernelArgsTuple args;
+  paddle::SmallVector<const char*> input_names;
+  paddle::SmallVector<const char*> attr_names;
+  paddle::SmallVector<const char*> output_names;
 
   KernelSignature() = default;
 
@@ -41,18 +43,26 @@ struct KernelSignature {
                   paddle::SmallVector<const char*>&& inputs,
                   paddle::SmallVector<const char*>&& attrs,
                   paddle::SmallVector<const char*>&& outputs)
-      : name(kernel_name), args(std::make_tuple(inputs, attrs, outputs)) {}
+      : name(kernel_name),
+        input_names(std::move(inputs)),
+        attr_names(std::move(attrs)),
+        output_names(std::move(outputs)) {}
   KernelSignature(const char* kernel_name,
                   const paddle::SmallVector<const char*>& inputs,
                   const paddle::SmallVector<const char*>& attrs,
                   const paddle::SmallVector<const char*>& outputs)
-      : name(kernel_name), args(std::make_tuple(inputs, attrs, outputs)) {}
+      : name(kernel_name),
+        input_names(inputs),
+        attr_names(attrs),
+        output_names(outputs) {}
 
   // TODO(chenweihang): add assign constructor to solve windows compile
   // problem, remove it later
   KernelSignature& operator=(const KernelSignature& other) {
     name = other.name;
-    args = other.args;
+    input_names = other.input_names;
+    attr_names = other.attr_names;
+    output_names = other.output_names;
     return *this;
   }
 };

paddle/phi/core/compat/op_utils.h

@@ -86,6 +86,14 @@ class DefaultKernelSignatureMap {
     return it->second;
   }
 
+  const KernelSignature* GetNullable(const std::string& op_type) const {
+    auto it = map_.find(op_type);
+    if (it != map_.end()) {
+      return &it->second;
+    }
+    return nullptr;
+  }
+
   void Insert(std::string op_type, KernelSignature signature) {
     PADDLE_ENFORCE_NE(
         Has(op_type),
@@ -148,16 +156,13 @@ class OpUtilsMap {
     }
   }
 
-  ArgumentMappingFn GetArgumentMappingFn(const std::string& op_type) const {
+  const ArgumentMappingFn* GetArgumentMappingFn(
+      const std::string& op_type) const {
     auto it = arg_mapping_fn_map_.find(op_type);
     if (it == arg_mapping_fn_map_.end()) {
-      auto func =
-          [&op_type](const ArgumentMappingContext& ctx) -> KernelSignature {
-        return DefaultKernelSignatureMap::Instance().Get(op_type);
-      };
-      return func;
+      return nullptr;
     } else {
-      return it->second;
+      return &it->second;
     }
   }