Struct Renderer 

pub struct Renderer {
    pub device: String,
    pub has_local: bool,
    pub has_shared: bool,
    pub has_threads: bool,
    pub shared_max: usize,
    pub global_max: Option<Vec<usize>>,
    pub local_max: Option<usize>,
    pub upcast_max: usize,
    pub buffer_max: Option<usize>,
    pub tensor_cores: Vec<TensorCore>,
}

Backend renderer capabilities.

Describes what features and optimizations a particular backend supports. Used by the optimizer to determine valid transformations and enforce device limits.

Fields

device: String

Backend device identifier (e.g., “CUDA”, “Metal”, “CPU”).

has_local: bool

Whether the backend supports local/shared memory (GPU workgroups).

has_shared: bool

Whether the backend supports shared memory across threads in a workgroup.

has_threads: bool

Whether the backend supports CPU-style threading (not GPU threads).

shared_max: usize

Maximum shared memory size in bytes.

Used to validate GROUP/GROUPTOP optimizations that allocate shared memory. Typical values: 48KB-96KB for modern GPUs.

global_max: Option<Vec<usize>>

Maximum global work dimensions [x, y, z].

Maximum size for each global thread dimension. Used to validate thread count in THREAD optimization. None if unlimited or not applicable.

local_max: Option<usize>

Maximum local work group size.

Maximum number of threads in a workgroup (product of local dimensions). Typical values: 256-1024 for GPUs.

upcast_max: usize

Maximum vectorization width (upcast limit).

Maximum number of elements that can be processed as a vector. Typical values: 8-16 for SIMD, 4 for GPU float4.

buffer_max: Option<usize>

Maximum number of buffers/arguments per kernel.

Some backends have limits on kernel arguments. Metal: 31, WebGPU: 8, CUDA: typically unlimited.

tensor_cores: Vec<TensorCore>

Available tensor core configurations.

Hardware-accelerated matrix multiplication units with specific size constraints. Empty if tensor cores not available.

Implementations

impl Renderer

pub fn cpu() -> Self

Create a CPU renderer configuration.

pub fn cuda() -> Self

Create a CUDA GPU renderer configuration (SM80/Ampere by default).

For specific architectures, use cuda_sm75(), cuda_sm80(), or cuda_sm89().

pub fn cuda_sm75() -> Self

Create a CUDA GPU renderer for SM75 (Turing - RTX 20xx, T4).

pub fn cuda_sm80(allow_tf32: bool) -> Self

Create a CUDA GPU renderer for SM80 (Ampere - A100, RTX 30xx).

pub fn cuda_sm89(allow_tf32: bool) -> Self

Create a CUDA GPU renderer for SM89 (Hopper - H100).

pub fn metal() -> Self

Create a Metal GPU renderer configuration (Apple M1/M2/M3).

pub fn apple_amx() -> Self

Create an Apple AMX renderer configuration (M1/M2/M3 matrix coprocessor).

pub fn is_amx(&self) -> bool

Whether this renderer is for Apple AMX (CPU matrix coprocessor).

pub fn amd_rdna3() -> Self

Create an AMD RDNA3 GPU renderer (RX 7000 series).

pub fn amd_rdna4() -> Self

Create an AMD RDNA4 GPU renderer.

pub fn amd_cdna3() -> Self

Create an AMD CDNA3 GPU renderer (MI300 series).

pub fn amd_cdna4() -> Self

Create an AMD CDNA4 GPU renderer.

pub fn intel_xe() -> Self

Create an Intel Xe GPU renderer.

pub fn webgpu() -> Self

Create a WebGPU renderer configuration.

Trait Implementations

impl Clone for Renderer

fn clone(&self) -> Renderer

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for Renderer

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.