[PTen] Update all forward argument maping fns

[chenwhql]

PR types

Function optimization

PR changes

Others

Describe

[PTen] Update all forward argument maping fns

为满足infrt需求，将用于Op和Kernel参数匹配的函数，lower到pten中作为兼容组件，且不依赖fluid，本PR迁移了前向kernel相关的匹配函数

由于infrt不仅需要在运行时动态调用此参数映射函数，还需要在没有任何信息的情况下，提前将该函数所有的可能结果在运行之前提前静态化声明如下定义：

//为op到每个kernel的映射生成一个规则描述
def PDKEL_Reshape_to_CPU : Pat<
          (PD_ReshapeOp $x, $shape_tensor， $shape_attr), // OpMaker参数列表
          (PDKEL_ReshapeKernelAttr $x, fn($shape_attr, (INFRT_createBoolAttr<"false">))>; // Kernel需要的参数列表
def PDKEL_Reshape_to_CPU : Pat<
          (PD_ReshapeOp $x, $shape_tensor， $shape_attr),
          (PDKEL_ReshapeKernelAttr $x, fn($shape_tensor, (INFRT_createBoolAttr<"false">))>; // use_mkldnn = false

因此，我们需要将参数映射函数的每个结果都具体写出来，infrt通过正则抽取return KernelSignature("full", {}, {"ShapeTensor", "value"}, {"Out"});，在运行之前拿到所有的结果用于上述静态定义的生成，这导致一些函数会写得比较冗长，复杂的可能会有几百行，对于编程体验有一定副作用，示例如下

fill_constant_op映射函数原写法：

framework::KernelSignature GetExpectedPtenKernelArgs(
      const framework::ExecutionContext& ctx) const override {
    std::string shape;
    if (ctx.HasInput("ShapeTensor")) {
      shape = "ShapeTensor";
    } else if (ctx.MultiInput<framework::Tensor>("ShapeTensorList").size()) {
      shape = "ShapeTensorList";
    } else {
      shape = "shape";
    }
    std::string value;
    if (ctx.HasInput("ValueTensor")) {
      value = "ValueTensor";
    } else {
      const auto& str_value = ctx.Attr<std::string>("str_value");
      value = str_value.empty() ? "value" : "str_value";
    }
    if (!ctx.OutputVar("Out")->IsType<pten::SelectedRows>()) {
      return framework::KernelSignature("full", {}, {shape, value}, {"Out"});
    }
    return framework::KernelSignature("fill_constant.unregistered", {}, {}, {});
  }

为适配infrt更改后，fill_constant_op映射函数写法：

KernelSignature FillConstantOpArgumentMapping(
    const ArgumentMappingContext& ctx) {
  if (ctx.IsDenseTensorOutput("Out")) {
    if (ctx.HasInput("ShapeTensor")) {
      if (ctx.HasInput("ValueTensor")) {
        return KernelSignature(
            "full", {}, {"ShapeTensor", "ValueTensor"}, {"Out"});
      } else {
        const auto& str_value =
            paddle::any_cast<std::string>(ctx.Attr("str_value"));
        if (str_value.empty()) {
          return KernelSignature("full", {}, {"ShapeTensor", "value"}, {"Out"});
        } else {
          return KernelSignature(
              "full", {}, {"ShapeTensor", "str_value"}, {"Out"});
        }
      }
    } else if (ctx.InputSize("ShapeTensorList") > 0) {
      if (ctx.HasInput("ValueTensor")) {
        return KernelSignature(
            "full", {}, {"ShapeTensorList", "ValueTensor"}, {"Out"});
      } else {
        const auto& str_value =
            paddle::any_cast<std::string>(ctx.Attr("str_value"));
        if (str_value.empty()) {
          return KernelSignature(
              "full", {}, {"ShapeTensorList", "value"}, {"Out"});
        } else {
          return KernelSignature(
              "full", {}, {"ShapeTensorList", "str_value"}, {"Out"});
        }
      }
    } else {
      if (ctx.HasInput("ValueTensor")) {
        return KernelSignature("full", {}, {"shape", "ValueTensor"}, {"Out"});
      } else {
        const auto& str_value =
            paddle::any_cast<std::string>(ctx.Attr("str_value"));
        if (str_value.empty()) {
          return KernelSignature("full", {}, {"shape", "value"}, {"Out"});
        } else {
          return KernelSignature("full", {}, {"shape", "str_value"}, {"Out"});
        }
      }
    }
  }
  return KernelSignature("unregistered", {}, {}, {});
}

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[XiaoguangHu01]

LGTM