Struct CachedKernel 

pub struct CachedKernel {
    pub program: Box<dyn Program>,
    pub device: String,
    pub code: String,
    pub entry_point: String,
    pub var_names: Vec<String>,
    pub globals: Vec<usize>,
    pub outs: Vec<usize>,
    pub ins: Vec<usize>,
    pub host_parallel_safe: bool,
    pub global_size: [Arc<UOp>; 3],
    pub local_size: Option<[Arc<UOp>; 3]>,
}

Cached kernel that can be reused across tensors.

Note: This struct does not implement Clone because Box<dyn Program> is not Clone. Use Arc<CachedKernel> for sharing.

Fields

program: Box<dyn Program>

The compiled, executable program.

device: String

Device string (e.g., “CPU”, “CUDA:0”).

code: String

Generated source code (for debugging/profiling).

entry_point: String

Entry point name.

var_names: Vec<String>

Variable names in order for converting HashMap to positional vals. Matches the order expected by the compiled program.

globals: Vec<usize>

Global buffer slots in kernel argument order. Matches Tinygrad’s ProgramSpec.globals semantics.

outs: Vec<usize>

Output buffer slots written by STORE operations. Matches Tinygrad’s ProgramSpec.outs semantics.

ins: Vec<usize>

Input buffer slots read by LOAD operations. Matches Tinygrad’s ProgramSpec.ins semantics.

host_parallel_safe: bool

Whether host-level scheduling may overlap this program with other kernels.

Thread-safety is required by the Program trait. This flag is about backend/kernel semantics, not Rust synchronization safety.

global_size: [Arc<UOp>; 3]

Symbolic global work size evaluated with runtime vars before dispatch.

local_size: Option<[Arc<UOp>; 3]>

Symbolic local work size. None means direct global-id execution.