fix

Signed-off-by: intwanghao <hao3.wang@intel.com>

Author: intwanghao

clang/examples/DPCT/CUB/cub$$BlockReduce$$Reduce.cu

@@ -0,0 +1,10 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+
+__device__ void test(int data) {
+  // Start
+  __shared__ typename cub::BlockReduce<int, 4>::TempStorage temp_storage;
+  cub::BlockReduce<int, 4>(temp_storage).Reduce(data/*int*/, cub::Sum()/*ReduceOp*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockReduce$$Sum.cu

@@ -0,0 +1,10 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+
+__device__ void test(int data) {
+  // Start
+  __shared__ typename cub::BlockReduce<int, 4>::TempStorage temp_storage;
+  cub::BlockReduce<int, 4>(temp_storage).Sum(data/*int*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockScan$$ExclusiveScan.cu

@@ -0,0 +1,10 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+
+__device__ void test(int input, int output, int init) {
+  // Start
+  __shared__ typename cub::BlockScan<int, 4>::TempStorage temp_storage;
+  cub::BlockScan<int, 4>(temp_storage).ExclusiveScan(input/*int*/, output/*int &*/, init/*int*/, cub::Sum()/*ScanOp*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockScan$$ExclusiveSum.cu

@@ -0,0 +1,10 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+
+__device__ void test(int input, int output) {
+  // Start
+  __shared__ typename cub::BlockScan<int, 4>::TempStorage temp_storage;
+  cub::BlockScan<int, 4>(temp_storage).ExclusiveSum(input/*int*/, output/*int &*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockScan$$InclusiveScan.cu

@@ -0,0 +1,10 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+
+__device__ void test(int input, int output) {
+  // Start
+  __shared__ typename cub::BlockScan<int, 4>::TempStorage temp_storage;
+  cub::BlockScan<int, 4>(temp_storage).InclusiveScan(input/*int*/, output/*int &*/, cub::Sum()/*ScanOp*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockScan$$InclusiveSum.cu

@@ -0,0 +1,10 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+
+__device__ void test(int input, int output) {
+  // Start
+  __shared__ typename cub::BlockScan<int, 4>::TempStorage temp_storage;
+  cub::BlockScan<int, 4>(temp_storage).InclusiveSum(input/*int*/, output/*int &*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockShuffle$$Down.cu

@@ -0,0 +1,11 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+#include <cub/block/block_shuffle.cuh>
+
+__device__ void test(int (&input)[4], int (&output)[4]) {
+  // Start
+  __shared__ typename cub::BlockShuffle<int, 128>::TempStorage temp_storage;
+  cub::BlockShuffle<int, 128>(temp_storage).Down(input/*int*/, output/*int &*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockShuffle$$Offset.cu

@@ -0,0 +1,11 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+#include <cub/block/block_shuffle.cuh>
+
+__device__ void test(int input, int output, int distance) {
+  // Start
+  __shared__ typename cub::BlockShuffle<int, 128>::TempStorage temp_storage;
+  cub::BlockShuffle<int, 128>(temp_storage).Offset(input/*int*/, output/*int &*/, distance/*int*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockShuffle$$Rotate.cu

@@ -0,0 +1,11 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+#include <cub/block/block_shuffle.cuh>
+
+__device__ void test(int input, int output, unsigned int distance) {
+  // Start
+  __shared__ typename cub::BlockShuffle<int, 128>::TempStorage temp_storage;
+  cub::BlockShuffle<int, 128>(temp_storage).Rotate(input/*int*/, output/*int &*/, distance/*unsigned int*/);
+  // End
+}

clang/examples/DPCT/CUB/cub$$BlockShuffle$$Up.cu

@@ -0,0 +1,11 @@
+// clang-format off
+#include <cstddef>
+#include <cub/cub.cuh>
+#include <cub/block/block_shuffle.cuh>
+
+__device__ void test(int (&input)[4], int (&output)[4]) {
+  // Start
+  __shared__ typename cub::BlockShuffle<int, 128>::TempStorage temp_storage;
+  cub::BlockShuffle<int, 128>(temp_storage).Up(input/*int*/, output/*int &*/);
+  // End
+}