Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/akhildevelops/cudaz/llms.txt

Use this file to discover all available pages before exploring further.

cudaz provides two complementary GPU buffer types. CudaSlice(T) is a generic, comptime-typed buffer that gives you full type safety and lets the Zig compiler verify element types at build time. CudaSliceR is its runtime-typed counterpart that carries a DType tag instead of a comptime type parameter, making it useful for dynamic dispatch and generic pipelines where the element type is not known until runtime. Both types are thin wrappers around a CUdeviceptr and expose clone and free methods.

​
CudaSlice(T)

CudaSlice(T) is a comptime-generic struct parameterized by the element type T. It is the return type of Device.alloc, Device.allocZeros, Device.htodCopy, and Rng.genrandom.

​
Fields

FieldTypeDescription
device_ptrcuda.CUdeviceptrOpaque handle to the GPU memory region
lenusizeNumber of elements (not bytes) in the buffer
deviceDeviceThe Device on which this memory was allocated

​
Comptime Constants

ConstantTypeDescription
element_typetypeComptime constant equal to T; accessible as CudaSlice(T).element_type via pub const element_type: type = T

​
clone

pub fn clone(self: @This()) CudaError.Error!@This()
Performs a device-to-device memory copy using cuMemcpyDtoD_v2, allocating a new CudaSlice(T) of the same length on the same device. The original slice is not freed. The caller owns the returned slice and must call free() on it. Returns: CudaError.Error!CudaSlice(T)

​
free

pub fn free(self: @This()) void
Frees the GPU memory held by this slice via cuMemFree_v2. Panics if the underlying CUDA call fails. Always call free (or pair it with defer) when you are done with a slice.
free panics on error rather than returning one. If you need error handling on deallocation, use Device.free(slice.device_ptr) directly.

​
Example

const device = try Cuda.Device.default();
defer device.deinit();

const slice = try device.alloc(f32, 512);
defer slice.free();

const copy = try slice.clone();
defer copy.free();

​
CudaSliceR

CudaSliceR is the runtime-typed GPU slice. Instead of a comptime T, it stores a DType value in its element_type field. This makes it suitable for scenarios where you do not know the element type at compile time, or where you want a single variable that can hold slices of different types.

​
Fields

FieldTypeDescription
device_ptrcuda.CUdeviceptrOpaque handle to the GPU memory region
lenusizeNumber of elements (not bytes) in the buffer
deviceDeviceThe Device on which this memory was allocated
element_typeDTypeRuntime element type tag (.f16 or .f32)

​
clone

pub fn clone(self: *const CudaSliceR) CudaError.Error!CudaSliceR
Performs a device-to-device copy using cuMemcpyDtoD_v2. The element_type field is preserved in the returned slice. The byte count is computed as element_type.size() * len. Returns: CudaError.Error!CudaSliceR

​
free

pub fn free(self: CudaSliceR) void
Frees GPU memory via cuMemFree_v2. Panics on error.

​
Example

const device = try Cuda.Device.default();
defer device.deinit();

const slice = try device.allocR(.f32, 512);
defer slice.free();

const copy = try slice.clone();
defer copy.free();
To bring data from a CudaSliceR back to the host, use Device.syncReclaimR. You must supply the concrete comptime type at the call site — it must match slice.element_type.

​
DType

DType is an enum that represents a GPU element type at runtime. It is used by CudaSliceR, Device.allocR, and Device.allocZerosR. 
pub const DType = enum(u8) {
    f16 = 0,
    f32 = 1,
    // ...
};

​
Variants

VariantBacking valueElement size
DType.f1602 bytes (@sizeOf(f16))
DType.f3214 bytes (@sizeOf(f32))

​
size

pub fn size(self: DType) usize
Returns the element size in bytes for the given DType.
​
self
DType
required
The runtime element type whose size to query.
Returns: usize — 2 for .f16, 4 for .f32 
const bytes_per_elem = DType.f32.size(); // 4
const total_bytes = bytes_per_elem * 1024; // 4096

Build docs developers (and LLMs) love

Get started for freeTalk to us