[BUG] Kernel caching false positives

[gdaviet]

Bug Description

Constructs involving function pointers that depend on a type or constant value may lead to caching false positives (functions being omitted from the hash computation). See examples of such constructs below.

This seems to be stemming from a difference between how the hash computation Adjoint.get_references resolve static expressions compared to the actual codegen (e.g. emit_Call); the hash computation uses a more lightweight approach.

Repro:

Run once, change message and re-run; kernels will be loaded from cache and output won't have changed. Manually clear cache, re-run; output is properly updated.

import warp as wp

# wp.clear_kernel_cache()


@wp.func
def message():
    wp.print("A")  # Run once, then change this and rerun


# Repro 1 (type-dependent)


@wp.struct
class A:
    fn = message


@wp.kernel
def kn1(a: A):
    a.fn()


@wp.kernel
def kn2(a: A):
    type(a).fn()


wp.launch(kn1, dim=1, inputs=[A()])
wp.launch(kn2, dim=1, inputs=[A()])

wp.synchronize()


# Repro 2 (constant-dependent)

funcs = [message]


@wp.kernel
def kn3():
    for fn in range(wp.static(len(funcs))):
        wp.static(funcs[fn])()


@wp.kernel
def kn4():
    FN = 0
    wp.static(funcs[FN])()


wp.launch(kn3, dim=1, inputs=[])
wp.launch(kn4, dim=1, inputs=[])
wp.synchronize()

### System Information

_No response_

[shi-eric]

Second case has been fixed by @eric-heiden in 3c6388890e7a18c0164504b68b53ccb60859cbc0