Skip to content

Apple's open-source XNU kernel on Android phones: its own loader and drivers, Metal on the Adreno. Runs from RAM; a reboot gives the phone back to Android.

C

docs: the UNVERIFIED pair is settled, so stop calling the whole file unverified for that reason

vcvkkbb2f5c19c657
diff --git a/docs/air.md b/docs/air.md
index 38fde65..858ecf5 100644
--- a/docs/air.md
+++ b/docs/air.md
@@ -287,9 +287,10 @@ Three references, each independent of the translator:
and the instruction translation — and no statement at all about `air_intrinsics.c`. The
shaders that do call `air.*` are exactly the ones this check cannot judge: the texture ones
have a varying for a parameter, the math-heavy ones are compute kernels or `air.visible`
- library functions. So **all ~50 operations in `air_intrinsics.c` are unverified**, including
- two whose meaning is a guess and marked `UNVERIFIED` in the code (`air.fast_exp`: natural
- `exp` or `exp2`? and `air.fast_round`: away from zero or to even?). What would settle them is
+ library functions. So **no operation in `air_intrinsics.c` is numerically verified**, even the
+ two whose *identity* section 3.2 has since settled against Apple's own compiler: knowing that
+ `air.fast_round` is `round` and not `rint` says nothing about whether this file computes
+ `round` correctly. What would settle that is
a shader whose intrinsic result is observable through a device buffer, compared the same way
— which needs the buffer-readback half of this tool, i.e. M4's acceptance test. Writing C
implementations of the `air.*` functions for the CPU side would *not* settle them: those