Update toolchain to nightly-03-02.

Cargo.toml

@@ -1,7 +1,5 @@
 [workspace]
-
 resolver = "2"
-
 members = [
   "crates/*",
   "crates/optix/examples/ex*",

crates/cuda_builder/src/lib.rs

@@ -376,10 +376,13 @@ fn invoke_rustc(builder: &CudaBuilder) -> Result<PathBuf, CudaBuilderError> {
 
     let new_path = get_new_path_var();
 
-    let mut rustflags = vec![format!(
-        "-Zcodegen-backend={}",
-        rustc_codegen_nvvm.display(),
-    )];
+    let mut rustflags = vec![
+        format!("-Zcodegen-backend={}", rustc_codegen_nvvm.display()),
+        "-Zcrate-attr=feature(register_tool)".into(),
+        "-Zcrate-attr=register_tool(nvvm_internal)".into(),
+        "-Zcrate-attr=no_std".into(),
+        "-Zsaturating_float_casts=false".into(),
+    ];
 
     if let Some(emit) = &builder.emit {
         let string = match emit {
@@ -437,8 +440,7 @@ fn invoke_rustc(builder: &CudaBuilder) -> Result<PathBuf, CudaBuilderError> {
         "--lib",
         "--message-format=json-render-diagnostics",
         "-Zbuild-std=core,alloc",
-        "--target",
-        "nvptx64-nvidia-cuda",
+        "--target=nvptx64-nvidia-cuda",
     ]);
 
     cargo.args(&builder.build_args);

crates/cuda_std/src/atomic/intrinsics.rs

@@ -1,6 +1,7 @@
 //! Raw CUDA-specific atomic functions that map to PTX instructions.
 
 use crate::gpu_only;
+use core::arch::asm;
 use core::concat;
 use paste::paste;