Remove unnecessary uses of thrust::pair (pytorch#168941)

This PR replaces unnecessary uses of thrust::pair with std::pair.

Pull Request resolved: pytorch#168941
Approved by: https://github.com/albanD

Author: cyyever

aten/src/ATen/native/cuda/ReflectionPad.cu

@@ -23,15 +23,14 @@
 #include <ATen/ops/reflection_pad3d_backward_native.h>
 #endif
 
-#include <thrust/pair.h>
 
 namespace at::native {
 namespace {
 
 using at::cuda::detail::canUse32BitIndexMath;
 
 __device__
-inline thrust::pair<int64_t, int64_t> get_index_mapping1d(
+inline std::pair<int64_t, int64_t> get_index_mapping1d(
     int64_t input_w, int64_t output_w,
     int64_t output_x,
     int64_t pad_l) {
@@ -50,13 +49,13 @@ inline thrust::pair<int64_t, int64_t> get_index_mapping1d(
                     + 2 * pad_l + input_w - 1
                     - o_start_x + i_start_x;
 
-  return thrust::make_pair<int64_t, int64_t>(
+  return std::make_pair<int64_t, int64_t>(
     input_offset + input_x, output_offset + output_x);
 }
 
 
 __device__
-inline thrust::pair<int64_t, int64_t>  get_index_mapping2d(
+inline std::pair<int64_t, int64_t>  get_index_mapping2d(
     int64_t input_dim_x, int64_t input_dim_y,
     int64_t output_dim_x, int64_t output_dim_y,
     int64_t pad_l, int64_t pad_t,
@@ -87,7 +86,7 @@ inline thrust::pair<int64_t, int64_t>  get_index_mapping2d(
                  + 2 * pad_t + input_dim_y - 1
                  - o_start_y + i_start_y;
 
-  return thrust::make_pair<int64_t, int64_t>(
+  return std::make_pair<int64_t, int64_t>(
     input_offset + input_y * input_dim_x + input_x,
     output_offset + output_y * output_dim_x + output_x);
 }

aten/src/ATen/native/cuda/group_norm_kernel.cu

@@ -38,7 +38,7 @@ __global__ void RowwiseMomentsCUDAKernel(
   using T_ACC = acc_type<T, true>;
   using WelfordType = WelfordData<T_ACC, int64_t>;
   using WelfordOp =
-      WelfordOps<T_ACC, T_ACC, int64_t, thrust::pair<T_ACC, T_ACC>>;
+      WelfordOps<T_ACC, T_ACC, int64_t, std::pair<T_ACC, T_ACC>>;
 
   const int64_t i = blockIdx.x;
   WelfordOp welford_op = {/*correction=*/0, /*take_sqrt=*/false};

aten/src/ATen/native/cuda/layer_norm_kernel.cu

@@ -64,7 +64,7 @@ __global__ void RowwiseMomentsCUDAKernel(
     T_ACC* rstd) {
   using WelfordType = WelfordData<T_ACC, int64_t>;
   using WelfordOp =
-      WelfordOps<T_ACC, T_ACC, int64_t, thrust::pair<T_ACC, T_ACC>>;
+      WelfordOps<T_ACC, T_ACC, int64_t, std::pair<T_ACC, T_ACC>>;
 
   __shared__
       typename std::aligned_storage<sizeof(WelfordType), alignof(WelfordType)>::

aten/src/ATen/native/sparse/cuda/SparseCUDATensor.cu

@@ -88,7 +88,7 @@ SparseTensor _coalesce_sparse_cuda(const SparseTensor& self) {
   );
 
   // this forces device-host synchronization!
-  thrust::pair<thrust_ptr, thrust_ptr> newEnd = thrust::unique_by_key(policy,
+  auto newEnd = thrust::unique_by_key(policy,
     indicesIter, indicesIter + nnz,
     uniqueOffsetsIter
   );