CUDA Allocator

cuda.CudaAllocator implements the std.mem.Allocator interface, enabling device memory to be used anywhere Zig expects a standard allocator.

Initialisation

var cuda_alloc = cuda.CudaAllocator.init();
const allocator = cuda_alloc.allocator();

Usage

const slice = try allocator.alloc(f32, 1024);
defer allocator.free(slice);

// Pass to any allocator-aware code
var list = try std.ArrayListUnmanaged(f32).initCapacity(allocator, 512);
defer list.deinit(allocator);

WARNING

Slices allocated by CudaAllocator live in device (GPU) memory. Attempting to read or write them from the host causes undefined behaviour. Always use cudaMemcpy or the DeviceBuffer copy helpers to move data between host and device.

Resize and Free

The allocator supports resize (no-op that always fails, forcing reallocation) and free via cudaFree.

CPU Fallback Mode

In fallback mode CudaAllocator falls back to the system heap (std.heap.c_allocator), so code written against the allocator interface continues to work transparently.

When to Use CudaAllocator vs DeviceBuffer

Use DeviceBuffer(T) when…	Use CudaAllocator when…
You know the element type statically	You need a generic Allocator interface
You want copy helpers (H2D, D2H)	You're integrating with allocator-aware abstractions
You manage one homogeneous array	You need multiple typed allocations