From b5d6ea22691feeec62e28a64ac6d634209ab76d7 Mon Sep 17 00:00:00 2001
From: ReversedGravity <rg@youxplode.com>
Date: Sat, 23 Sep 2023 17:54:57 +0100
Subject: [PATCH 1/2] Added Deref on VarProxy.

---
 luisa_compute/src/lang/ops.rs   | 4 +---
 luisa_compute/src/lang/types.rs | 2 +-
 2 files changed, 2 insertions(+), 4 deletions(-)

diff --git a/luisa_compute/src/lang/ops.rs b/luisa_compute/src/lang/ops.rs
index 287e9855..d9af8b10 100644
--- a/luisa_compute/src/lang/ops.rs
+++ b/luisa_compute/src/lang/ops.rs
@@ -4,10 +4,10 @@ use std::ops::*;
 use super::types::core::{Floating, Integral, Numeric, Primitive, Signed};
 use super::types::vector::{VectorAlign, VectorElement};
 
+mod cast_impls;
 mod impls;
 mod spread;
 mod traits;
-mod cast_impls;
 
 pub use spread::*;
 pub use traits::*;
@@ -24,8 +24,6 @@ pub trait Linear: Value {
     const N: usize;
     type Scalar: VectorElement;
     type WithScalar<S: VectorElement>: Linear<Scalar = S>;
-    // We don't actually know that the vector has equivalent vectors of every
-    // primitive type.
 }
 impl<T: VectorElement> Linear for T {
     const N: usize = 1;
diff --git a/luisa_compute/src/lang/types.rs b/luisa_compute/src/lang/types.rs
index 45f5abcc..11bad6b0 100644
--- a/luisa_compute/src/lang/types.rs
+++ b/luisa_compute/src/lang/types.rs
@@ -62,7 +62,7 @@ pub trait ExprProxy: Copy + 'static {
 /// For example, `Var<[f32; 4]>` dereferences to `ArrayVar<f32, 4>`, which
 /// exposes [`Index`](std::ops::Index) and [`IndexMut`](std::ops::IndexMut)
 /// impls.
-pub trait VarProxy: Copy + 'static {
+pub trait VarProxy: Copy + 'static + Deref<Target = Expr<Self::Value>> {
     type Value: Value<Var = Self>;
     fn as_var_from_proxy(&self) -> &Var<Self::Value>;
 

From 898009a0c46650478660cb30817ddb4e6203a844 Mon Sep 17 00:00:00 2001
From: ReversedGravity <rg@youxplode.com>
Date: Sat, 23 Sep 2023 17:55:13 +0100
Subject: [PATCH 2/2] Kernel dispatch fix.

---
 luisa_compute/src/runtime.rs | 30 ++++++++++++++++++------------
 1 file changed, 18 insertions(+), 12 deletions(-)

diff --git a/luisa_compute/src/runtime.rs b/luisa_compute/src/runtime.rs
index 6c1fb73f..6d1ae634 100644
--- a/luisa_compute/src/runtime.rs
+++ b/luisa_compute/src/runtime.rs
@@ -401,12 +401,14 @@ impl Device {
         }
     }
 
-    /// Compile a [`KernelDef`] into a [`Kernel`]. See [`Kernel`] for more details on kernel creation
+    /// Compile a [`KernelDef`] into a [`Kernel`]. See [`Kernel`] for more
+    /// details on kernel creation
     pub fn compile_kernel<S: KernelSignature>(&self, k: &KernelDef<S>) -> Kernel<S> {
         self.compile_kernel_with_options(k, KernelBuildOptions::default())
     }
 
-    /// Compile a [`KernelDef`] into a [`Kernel`] asynchronously. See [`Kernel`] for more details on kernel creation
+    /// Compile a [`KernelDef`] into a [`Kernel`] asynchronously. See [`Kernel`]
+    /// for more details on kernel creation
     pub fn compile_kernel_async<S: KernelSignature>(&self, k: &KernelDef<S>) -> Kernel<S> {
         self.compile_kernel_with_options(
             k,
@@ -417,7 +419,8 @@ impl Device {
         )
     }
 
-    /// Compile a [`KernelDef`] into a [`Kernel`] with options. See [`Kernel`] for more details on kernel creation
+    /// Compile a [`KernelDef`] into a [`Kernel`] with options. See [`Kernel`]
+    /// for more details on kernel creation
     pub fn compile_kernel_with_options<S: KernelSignature>(
         &self,
         k: &KernelDef<S>,
@@ -1084,7 +1087,11 @@ impl RawKernel {
         }
     }
 
-    pub fn dispatch_async(&self, args: KernelArgEncoder, dispatch_size: [u32; 3]) -> Command {
+    pub fn dispatch_async(
+        &self,
+        args: KernelArgEncoder,
+        dispatch_size: [u32; 3],
+    ) -> Command<'static> {
         let mut rt = ResourceTracker::new();
         rt.add(Arc::new(args.uniform_data));
         let args = args.args;
@@ -1218,8 +1225,8 @@ pub struct KernelDef<T: KernelSignature> {
 /// use luisa_compute::prelude::*;
 /// let ctx = Context::new(std::env::current_exe().unwrap());
 /// let device = ctx.create_device("cpu");
-/// let kernel = KernelDef::<fn(Buffer<f32>, Buffer<f32>, Buffer<f32>)>::new(&device, track!(|a,b,c|{  }));
-/// // Compilation:
+/// let kernel = KernelDef::<fn(Buffer<f32>, Buffer<f32>,
+/// Buffer<f32>)>::new(&device, track!(|a,b,c|{  })); // Compilation:
 /// let kernel = device.compile_kernel(&kernel);
 /// ```
 /// - Recording and compilation in one step:
@@ -1227,11 +1234,10 @@ pub struct KernelDef<T: KernelSignature> {
 /// use luisa_compute::prelude::*;
 /// let ctx = Context::new(std::env::current_exe().unwrap());
 /// let device = ctx.create_device("cpu");
-/// let kernel = Kernel::<fn(Buffer<f32>, Buffer<f32>, Buffer<f32>)>::new(&device, track!(|a,b,c|{ }));
-/// ```
+/// let kernel = Kernel::<fn(Buffer<f32>, Buffer<f32>,
+/// Buffer<f32>)>::new(&device, track!(|a,b,c|{ })); ```
 /// - Asynchronous compilation use [`Kernel::<T>::new_async`]
 /// - Custom build options using [`Kernel::<T>::new_with_options`]
-///
 pub struct Kernel<T: KernelSignature> {
     pub(crate) inner: RawKernel,
     pub(crate) _marker: PhantomData<T>,
@@ -1335,10 +1341,10 @@ macro_rules! impl_dispatch_for_kernel {
             }
             #[allow(non_snake_case)]
             #[allow(unused_mut)]
-            pub fn dispatch_async<'a>(
-                &'a self,
+            pub fn dispatch_async(
+                &self,
                 dispatch_size: [u32; 3], $($Ts:&impl AsKernelArg<$Ts>),*
-            ) -> Command<'a> {
+            ) -> Command<'static> {
                 let mut encoder = KernelArgEncoder::new();
                 $($Ts.encode(&mut encoder);)*
                 self.inner.dispatch_async(encoder, dispatch_size)