Amend tests after reverting arguments order

cub/test/catch2_test_device_scan.cu

@@ -38,24 +38,7 @@
 #include "catch2_test_helper.h"
 #include "catch2_test_launch_helper.h"
 
-// Circumvate DECLARE_LAUNCH_WRAPPER appending stream as the last argument
-// by default since init_value needs to be the last argument.
-template <typename InputIteratorT, typename OutputIteratorT, typename ScanOpT, typename InitT>
-CUB_RUNTIME_FUNCTION static cudaError_t inclusive_scan_init_wrapper(
-  void* d_temp_storage,
-  size_t& temp_storage_bytes,
-  InputIteratorT d_in,
-  OutputIteratorT d_out,
-  ScanOpT scan_op,
-  InitT init,
-  int num_items,
-  cudaStream_t stream = 0)
-{
-  return cub::DeviceScan::InclusiveScanInit(
-    d_temp_storage, temp_storage_bytes, d_in, d_out, scan_op, init, num_items, stream);
-}
-
-DECLARE_LAUNCH_WRAPPER(inclusive_scan_init_wrapper, device_inclusive_scan_with_init);
+DECLARE_LAUNCH_WRAPPER(cub::DeviceScan::InclusiveScanInit, device_inclusive_scan_with_init);
 DECLARE_LAUNCH_WRAPPER(cub::DeviceScan::ExclusiveSum, device_exclusive_sum);
 DECLARE_LAUNCH_WRAPPER(cub::DeviceScan::ExclusiveScan, device_exclusive_scan);
 DECLARE_LAUNCH_WRAPPER(cub::DeviceScan::InclusiveSum, device_inclusive_sum);

cub/test/catch2_test_device_scan_api.cu

@@ -34,28 +34,28 @@
 
 CUB_TEST("Device inclusive scan works", "[scan][device]")
 {
-  // Determine temporary device storage requirements for inclusive scan
-  cudaStream_t stream{};
-  REQUIRE(cudaSuccess == cudaStreamCreate(&stream));
   // example-begin device-inclusive-scan
   thrust::device_vector<int> input{0, -1, 2, -3, 4, -5, 6};
   thrust::device_vector<int> out(input.size());
 
-  void* d_temp_storage{};
-  size_t temp_storage_bytes{};
-
   int init = 1;
+  size_t temp_storage_bytes{};
 
   cub::DeviceScan::InclusiveScanInit(
-    d_temp_storage, temp_storage_bytes, input.begin(), out.begin(), cub::Max{}, init, static_cast<int>(input.size()));
+    nullptr, temp_storage_bytes, input.begin(), out.begin(), cub::Max{}, init, static_cast<int>(input.size()));
 
   // Allocate temporary storage for inclusive scan
   thrust::device_vector<std::uint8_t> temp_storage(temp_storage_bytes);
-  d_temp_storage = thrust::raw_pointer_cast(temp_storage.data());
 
   // Run inclusive prefix sum
   cub::DeviceScan::InclusiveScanInit(
-    d_temp_storage, temp_storage_bytes, input.begin(), out.begin(), cub::Max{}, init, static_cast<int>(input.size()));
+    thrust::raw_pointer_cast(temp_storage.data()),
+    temp_storage_bytes,
+    input.begin(),
+    out.begin(),
+    cub::Max{},
+    init,
+    static_cast<int>(input.size()));
 
   thrust::host_vector<int> expected{1, 1, 2, 2, 4, 4, 6};
   // example-end device-inclusive-scan