Remove register-sized CUDA kernel args check, unnecessary since rust-…

…lang/rust#94703
juntyr · Sep 14, 2023 · 0bf3e0d · 0bf3e0d
1 parent d7f1e70
commit 0bf3e0d
Show file tree

Hide file tree

Showing 5 changed files with 7 additions and 59 deletions.
diff --git a/examples/single-source/src/main.rs b/examples/single-source/src/main.rs
@@ -43,6 +43,11 @@ pub struct Empty([u8; 0]);
 #[layout(crate = "rc::const_type_layout")]
 pub struct Tuple(u32, i32);
 
+#[repr(C)]
+#[derive(rc::const_type_layout::TypeLayout)]
+#[layout(crate = "rc::const_type_layout")]
+pub struct Triple(i32, i32, i32);
+
 #[rc::common::kernel(use link_kernel! as impl Kernel<KernelArgs, KernelPtx> for Launcher)]
 #[kernel(crate = "rc")]
 #[kernel(
@@ -56,6 +61,7 @@ pub fn kernel<'a, T: rc::common::RustToCuda>(
     #[kernel(pass = SafeDeviceCopy, jit)] _v @ _w: &'a core::sync::atomic::AtomicU64,
     #[kernel(pass = LendRustToCuda)] _: Wrapper<T>,
     #[kernel(pass = SafeDeviceCopy)] Tuple(s, mut __t): Tuple,
+    #[kernel(pass = SafeDeviceCopy)] q: Triple,
     // #[kernel(pass = SafeDeviceCopy)] shared3: ThreadBlockShared<u32>,
 ) where
     T: rc::safety::StackOnly + rc::safety::NoAliasing,
@@ -70,7 +76,7 @@ pub fn kernel<'a, T: rc::common::RustToCuda>(
         (*shared.index_mut_unchecked(1)).0 = (f64::from(s) * 2.0) as u32;
     }
     unsafe {
-        (*shared2.index_mut_unchecked(2)).1 = 24;
+        (*shared2.index_mut_unchecked(2)).1 = q.0 + q.1 + q.2;
     }
     // unsafe { core::arch::asm!("hi") }
     // unsafe {

diff --git a/rust-cuda-derive/src/kernel/wrapper/generate/cpu_wrapper/kernel_func_async/mod.rs b/rust-cuda-derive/src/kernel/wrapper/generate/cpu_wrapper/kernel_func_async/mod.rs
@@ -102,14 +102,8 @@ pub(super) fn quote_kernel_func_async(
                     #[allow(dead_code)]
                     fn assert_impl_no_aliasing<T: #crate_path::safety::NoAliasing>() {}
 
-                    #[allow(dead_code)]
-                    fn assert_impl_fits_into_device_register<
-                        T: #crate_path::safety::FitsIntoDeviceRegister,
-                    >(_val: &T) {}
-
                     #(assert_impl_devicecopy(&#func_params);)*
                     #(assert_impl_no_aliasing::<#cpu_func_unboxed_types>();)*
-                    #(assert_impl_fits_into_device_register(&#func_params);)*
                 }
 
                 let #crate_path::host::LaunchConfig {

diff --git a/rust-cuda-derive/src/kernel/wrapper/generate/cuda_wrapper.rs b/rust-cuda-derive/src/kernel/wrapper/generate/cuda_wrapper.rs
@@ -136,14 +136,8 @@ pub(in super::super) fn quote_cuda_wrapper(
                 #[allow(dead_code)]
                 fn assert_impl_no_aliasing<T: #crate_path::safety::NoAliasing>() {}
 
-                #[allow(dead_code)]
-                fn assert_impl_fits_into_device_register<
-                    T: #crate_path::safety::FitsIntoDeviceRegister,
-                >(_val: &T) {}
-
                 #(assert_impl_devicecopy(&#func_params);)*
                 #(assert_impl_no_aliasing::<#ptx_func_unboxed_types>();)*
-                #(assert_impl_fits_into_device_register(&#func_params);)*
             }
 
             #ptx_func_input_unwrap

diff --git a/src/safety/mod.rs b/src/safety/mod.rs
@@ -1,7 +1,6 @@
 mod arch;
 mod device_copy;
 mod no_aliasing;
-mod register_fit;
 mod stack_only;
 #[cfg(any(feature = "alloc", doc))]
 mod unified_heap;
@@ -13,7 +12,6 @@ pub mod type_layout;
 
 pub use device_copy::SafeDeviceCopy;
 pub use no_aliasing::NoAliasing;
-pub use register_fit::FitsIntoDeviceRegister;
 pub use stack_only::StackOnly;
 #[cfg(any(feature = "alloc", doc))]
 pub use unified_heap::UnifiedHeapOnly;
diff --git a/src/safety/register_fit.rs b/src/safety/register_fit.rs