Fix gemm example on CUDA 13.0.

`#[address_space(shared)] static mut` is used to model GPU shared
memory. It's a bit weird. In particular, GPU shared memory is
uninitialized, but `static mut` requires an initializer in Rust. `gemm`
uses a zero initialize, but this initializer is ignored by NVVM. At
least, it was in CUDA 12.x, but in CUDA 13.0 the `gemm` example fails
with this error:
```
thread 'rustc' panicked at crates/rustc_codegen_nvvm/src/nvvm.rs:120:9:
Malformed NVVM IR program rejected by libnvvm, dumping verifier log:

error: Error: : Global Variable `_ZN12gemm_kernels10gemm_tiled10gemm_tiled6TILE_A17hc9c66e758c373a7eE':
  context: @_ZN12gemm_kernels10gemm_tiled10gemm_tiled6TILE_A17hc9c66e758c373a7eE = internal unnamed_addr addrspace(3) global <{ [1024 x i8] }> zeroinitializer, align 4
  Shared variables can't be initialized
```
This memory looks like it's initialized to zero but isn't, and then is
written and read normally. This is incredibly dodgy and very likely UB.
The proper way to deal with uninitialized memory in Rust is with
`MaybeUninit`, and there are strict rules around its used, e.g. writes
must be done with `write` and `assume_init` must be used values after
they are written.

This commit changes `gemm` to use `MaybeUninit` for the shared memory.
This fixes the error on CUDA 13.0 and the example runs correctly.

(This is the only executed use of GPU shared memory in rust-cuda. There
is a `shared_array!` macro defined but it's only used in a compiletest
where it is compiled but not run. That macro is extremely dubious but I
will deal with it in a separate PR because it's not necessary to get
CUDA 13.0 working.)

Author: nnethercote

Cargo.lock

@@ -1,3 +1,4 @@
+use core::mem::MaybeUninit;
 use cuda_std::address_space;
 use cuda_std::kernel;
 use cuda_std::thread;
@@ -38,11 +39,19 @@ pub unsafe fn gemm_tiled(
     beta: f32,
 ) {
     const TILE_SIZE: usize = 16;
+    const TILE_SIZE_2D: usize = TILE_SIZE * TILE_SIZE;
 
+    // Shared GPU memory is modelled with `#[address_space(shared)] static mut`. Unlike normal
+    // `static mut`, it is not initialized, and only exists for the duration of the kernel's
+    // (multi-)execution. Because it is not initialized, it must be marked with `MaybeUninit`,
+    // written with `write` (in unsafe blocks because writing a `static mut` is unsafe), and
+    // subsequently read with `assume_init`.
     #[address_space(shared)]
-    static mut TILE_A: [f32; TILE_SIZE * TILE_SIZE] = [0.; TILE_SIZE * TILE_SIZE];
+    static mut TILE_A: [MaybeUninit<f32>; TILE_SIZE_2D] =
+        [const { MaybeUninit::uninit() }; TILE_SIZE_2D];
     #[address_space(shared)]
-    static mut TILE_B: [f32; TILE_SIZE * TILE_SIZE] = [0.; TILE_SIZE * TILE_SIZE];
+    static mut TILE_B: [MaybeUninit<f32>; TILE_SIZE_2D] =
+        [const { MaybeUninit::uninit() }; TILE_SIZE_2D];
 
     // Thread indices within the block.
     let tx = thread::thread_idx_x() as usize;
@@ -57,20 +66,22 @@ pub unsafe fn gemm_tiled(
     for kk in (0..k).step_by(TILE_SIZE) {
         // Collaborative loading of tiles into shared memory.
         if row < m && (kk + tx) < k {
-            unsafe { TILE_A[ty * TILE_SIZE + tx] = mat_a[row * k + (kk + tx)] };
+            unsafe { TILE_A[ty * TILE_SIZE + tx].write(mat_a[row * k + (kk + tx)]); }
         } else {
-            unsafe { TILE_A[ty * TILE_SIZE + tx] = 0.0f32 };
+            unsafe { TILE_A[ty * TILE_SIZE + tx].write(0.0f32); }
         }
         if col < n && (kk + ty) < k {
-            unsafe { TILE_B[ty * TILE_SIZE + tx] = mat_b[(kk + ty) * n + col] };
+            unsafe { TILE_B[ty * TILE_SIZE + tx].write(mat_b[(kk + ty) * n + col]); }
         } else {
-            unsafe { TILE_B[ty * TILE_SIZE + tx] = 0.0f32 };
+            unsafe { TILE_B[ty * TILE_SIZE + tx].write(0.0f32); }
         }
         thread::sync_threads();
 
         // Perform the computation on the tile.
         for i in 0..TILE_SIZE {
-            sum += unsafe { TILE_A[ty * TILE_SIZE + i] * TILE_B[i * TILE_SIZE + tx] };
+            sum += unsafe {
+                TILE_A[ty * TILE_SIZE + i].assume_init() * TILE_B[i * TILE_SIZE + tx].assume_init()
+            };
         }
         thread::sync_threads();
     }