Output Marker

Differential programming in warp requires you place all input parameters ahead of output parameters in kernel function, and provide inputs and outputs accordingly on lauching kernel, which bring implicit constaint and add mental burden on lauching kernel.

@warp.kernel
def func(read: wp.array(...), write: wp.array(...)): ...

warp.launch(func, shape, inputs = (read,), outputs = (write,))

In this example, you have to place read ahead of write in parameters. When defining func, you have no way to tell the type system read is input and write is output, which you have to figure out every time launching func, providing parameters in inputs and outputs correctly.

from warp_hint import Out, kernel

@kernel
def func(write: Out[wp.array(...)], read: wp.array(...)): ...

func(shape, write, read)

Thanks to alias resolve, warp_hint let you use a generic alias to mark a parameter is output, and place it anywhere in the type annotations of kernel function parameters.

@kernel
def func(a: Out[int], b: int, c: Out[int], d: int, ...): ...

func(a, b, c, d, ...)

You can very arbitarily mark the output parameters, not affecting the kernel call.