BugFix with ParseInputDataType from LodTensorArray

[Aurelius84]

PR types

Bug fixes

PR changes

Others

Describe

1. PR内容

修复LoDTensorArray作为Op的输入时，在GetExepectedKernelType函数中ParseInputDataType的极其隐藏bug。

2. BUG原因

报错信息：

2021-05-13 14:16:32 ----------------------
2021-05-13 14:16:32 Error Message Summary:
2021-05-13 14:16:32 ----------------------
2021-05-13 14:16:32 NotFoundError: Operator (slice) does not have kernel for data_type[bool]:data_layout[ANY_LAYOUT]:place[CUDAPlace(0)]:library_type[PLAIN].
2021-05-13 14:16:32   [Hint: Expected kernel_iter != kernels.end(), but received kernel_iter == kernels.end().] (at /paddle/paddle/fluid/framework/operator.cc:1278)
2021-05-13 14:16:32   [operator < slice > error]  [operator < run_program > error]

问题代码：
https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/fluid/framework/operator.cc#L1541-L1572

     // ...(省略)
      } else if (var->IsType<LoDTensorArray>()) {
        auto t_arr = var->Get<LoDTensorArray>();   // 此处为赋值构造函数
        // VLOG(1) << "t_arr :" << &t_arr;
        // auto t_arr1 = var->Get<LoDTensorArray>();   // 执行两次，是连个不同的copy
       // VLOG(1) << "t_arr1 :" << &t_arr1;
        for (size_t j = 0; j < t_arr.size(); j++) {
          if (t_arr[j].IsInitialized()) {
            t = &(t_arr[j]);
          }
        }  // 出了else if 的作用域，则t指向了一个已析构了的Tensor对象
      }
      if (t != nullptr) {
        // VLOG(1) << "tensor ptr:" << t << "holder: " << t->Holder() << " , " <<  t->Holder()->Ptr();
        proto::VarType::Type tmp = t->type();  // 存在一定的风险。因为t是个野指针

3. 解决方案

将t_arr改为引用，因为此处每次返回指向的Tensor都是内在的Tensor地址。

PR之前：

I0514 10:06:15.922755  3916 device_context.cc:94] DeviceContextPool Get: CUDAPlace(0)
I0514 10:06:15.922760  3916 device_context.cc:94] DeviceContextPool Get: CUDAPlace(0)
I0514 10:06:15.922771  3916 variable.h:40] a ptr: 0x3ec8d908 holder->Ptr(): 0x3ec8d908
I0514 10:06:15.922776  3916 operator.cc:1555] t_arr: 0x7fdd850267b0     # 第一次copy
I0514 10:06:15.922781  3916 variable.h:40] a ptr: 0x3ec8d908 holder->Ptr(): 0x3ec8d908
I0514 10:06:15.922785  3916 operator.cc:1557] t_arr1: 0x7fdd850267d0    # 第二次copy （不同变量）
I0514 10:06:15.922791  3916 operator.cc:1566] tensor array[0] ptr: 0x7fdcdc00efe0 holder: 0x7fdcdc0092f0 , 0x7fdd1f8c8b00
# 以下4次的 t 的指针都是一致的，但是后两行为出了if block后，t->Holder()就瞎指了
I0514 10:06:15.922794  3916 operator.cc:1568]  tensor ptr: 0x7fdcdc00efe0 holder: 0x7fdcdc0092f0 , 0x7fdd1f8c8b00 tmp dtype int  
I0514 10:06:15.922799  3916 operator.cc:1570]  tensor ptr: 0x7fdcdc00efe0 holder: 0x7fdcdc0092f0 , 0x7fdd1f8c8b00 tmp dtype int
I0514 10:06:15.922804  3916 operator.cc:1572]  tensor ptr: 0x7fdcdc00efe0 holder: 0x7fdcdc000118 , 0x7fdcdc00b0a0 tmp dtype int
I0514 10:06:15.922809  3916 operator.cc:1580]  tensor ptr: 0x7fdcdc00efe0 holder: 0x7fdcdc000118 , 0x7fdcdc00b0a0 tmp dtype int

修复之后：

I0514 08:16:24.946592 17077 device_context.cc:94] DeviceContextPool Get: CUDAPlace(0)
I0514 08:16:24.946602 17077 device_context.cc:94] DeviceContextPool Get: CUDAPlace(0)0x7fc3ef8c8700
# 以下4次的 t 的指针都是一致的，后两行为出了if block后，t->Holder()也是一致的
I0514 08:16:24.946631 17077 operator.cc:1556] tensor array[0] ptr: 0x7fc3a4013df8 holder: 0x7fc3a400de40 , 0x7fc3ef8c8600
I0514 08:16:24.946635 17077 operator.cc:1557] tensor array[0] ptr: 0x7fc3a4013df8 holder: 0x7fc3a400de40 , 0x7fc3ef8c8600
I0514 08:16:24.946640 17077 operator.cc:1561] tmp dtype int tensor ptr: 0x7fc3a4013df8 holder: 0x7fc3a400de40 , 0x7fc3ef8c8600
I0514 08:16:24.946645 17077 operator.cc:1569] tmp dtype int tensor ptr: 0x7fc3a4013df8 holder: 0x7fc3a400de40 , 0x7fc3ef8c8600

4. 之前为啥没有触发

1. LoDTensorArray的dtype作为OpKernel的选择，此类分支被触发的比较少
2. 个人觉得，虽然t是个野指针，但其为 undefined behavior，如果block外的访问的内存还未被释放或未被重新重新分配和修改，可能误打误撞返回有效值。
3. 从多次执行可以稳定复现，log中也出现了有时会返回无效值

[paddle-bot-old]

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

[zhhsplendid]

LGTM

[zhiqiu]

LGTM