Function exec_kernel

tinygrad/engine/realize.py:170–180

(ctx:ExecContext, call:UOp, ast:UOp)

  return None

def exec_kernel(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
  et = None
  for bufs, device_vars in unwrap_multi(call, resolve_params(call, ctx.input_uops)):
    var_vals = {**ctx.var_vals, **device_vars}
    prg_bufs = [bufs[i].ensure_allocated() for i in ast.arg.globals]
    rt = get_runtime(device:=bufs[0].device, ast, cache=ctx.cache)
    global_size, local_size = ast.arg.launch_dims(var_vals)
    with track_stats(ctx, call, device, prg_bufs, var_vals) as tm:
      et = tm[0] = rt(*[b._buf for b in prg_bufs], global_size=global_size, local_size=local_size, vals=ast.arg.vals(var_vals),
                       wait=ctx.wait, timeout=ctx.timeout)
  return et

def exec_validate(ctx:ExecContext, call:UOp, ast:UOp) -> float|None:
  import numpy as np