c/parallel should be built with CUB_DISABLE_CDP (NVIDIA#4422)

* c/parallel should be built with CUB_DISABLE_CDP

Library functions launch kernels from the host only.
Setting this variable causes CUB_RUNTIME_FUNCTION to
be _CCCL_HOST, and hence compiler won't generate CUB
dispatch functions for the device, speeding up compilation.

* Add CUB_DISABLE_CDP to jit compiler arguments

* [pre-commit.ci] auto code formatting

---------

Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>

Author: oleksandr-pavlyk

c/parallel/CMakeLists.txt

@@ -41,7 +41,7 @@ target_link_libraries(cccl.c.parallel PRIVATE
   Thrust::Thrust
 )
 target_compile_definitions(cccl.c.parallel PUBLIC CCCL_C_EXPERIMENTAL=1 _CUB_HAS_TRANSFORM_UBLKCP=0)
-target_compile_definitions(cccl.c.parallel PRIVATE NVRTC_GET_TYPE_NAME=1)
+target_compile_definitions(cccl.c.parallel PRIVATE NVRTC_GET_TYPE_NAME=1 CUB_DISABLE_CDP=1)
 target_compile_options(cccl.c.parallel PRIVATE $<$<COMPILE_LANG_AND_ID:CUDA,NVIDIA>:--extended-lambda>)
 
 target_include_directories(cccl.c.parallel PUBLIC "include")

c/parallel/src/for.cu

@@ -94,8 +94,9 @@ CUresult cccl_device_for_build(
 
     const std::string arch = std::format("-arch=sm_{0}{1}", cc_major, cc_minor);
 
-    constexpr size_t num_args  = 7;
-    const char* args[num_args] = {arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto"};
+    constexpr size_t num_args  = 8;
+    const char* args[num_args] = {
+      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};

c/parallel/src/merge_sort.cu

@@ -408,8 +408,9 @@ struct device_merge_sort_vsmem_helper {{
 
     const std::string arch = std::format("-arch=sm_{0}{1}", cc_major, cc_minor);
 
-    constexpr size_t num_args  = 7;
-    const char* args[num_args] = {arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto"};
+    constexpr size_t num_args  = 8;
+    const char* args[num_args] = {
+      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};

c/parallel/src/reduce.cu

@@ -332,8 +332,9 @@ struct device_reduce_policy {{
 
     const std::string arch = std::format("-arch=sm_{0}{1}", cc_major, cc_minor);
 
-    constexpr size_t num_args  = 7;
-    const char* args[num_args] = {arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto"};
+    constexpr size_t num_args  = 8;
+    const char* args[num_args] = {
+      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};

c/parallel/src/scan.cu

@@ -295,8 +295,9 @@ struct device_scan_policy {{
 
     const std::string arch = std::format("-arch=sm_{0}{1}", cc_major, cc_minor);
 
-    constexpr size_t num_args  = 7;
-    const char* args[num_args] = {arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto"};
+    constexpr size_t num_args  = 8;
+    const char* args[num_args] = {
+      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};

c/parallel/src/segmented_reduce.cu

@@ -332,8 +332,9 @@ struct device_segmented_reduce_policy {{
 
     const std::string arch = std::format("-arch=sm_{0}{1}", cc_major, cc_minor);
 
-    constexpr size_t num_args  = 7;
-    const char* args[num_args] = {arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto"};
+    constexpr size_t num_args  = 8;
+    const char* args[num_args] = {
+      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};

c/parallel/src/transform.cu

@@ -270,9 +270,17 @@ struct device_transform_policy {{
     // Note: `-default-device` is needed because of the use of lambdas
     // in the transform kernel code. Qualifying those explicitly with
     // `__device__` seems not to be supported by NVRTC.
-    constexpr size_t num_args  = 8;
+    constexpr size_t num_args  = 9;
     const char* args[num_args] = {
-      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-default-device"};
+      arch.c_str(),
+      cub_path,
+      thrust_path,
+      libcudacxx_path,
+      ctk_path,
+      "-rdc=true",
+      "-dlto",
+      "-default-device",
+      "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};

c/parallel/src/unique_by_key.cu

@@ -363,8 +363,9 @@ struct device_unique_by_key_vsmem_helper {{
 
     const std::string arch = std::format("-arch=sm_{0}{1}", cc_major, cc_minor);
 
-    constexpr size_t num_args  = 7;
-    const char* args[num_args] = {arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto"};
+    constexpr size_t num_args  = 8;
+    const char* args[num_args] = {
+      arch.c_str(), cub_path, thrust_path, libcudacxx_path, ctk_path, "-rdc=true", "-dlto", "-DCUB_DISABLE_CDP"};
 
     constexpr size_t num_lto_args   = 2;
     const char* lopts[num_lto_args] = {"-lto", arch.c_str()};