Fix/implement checked_{add,sub,mul} and saturating_{add,sub} (using SPIR-V instructions were possible).

[eddyb]

After this PR:

• signed checked_{add,sub} still have no direct SPIR-V equivalents
  • our custom implementation used to have bugs around e.g. 0 + 0 / 0 - 0
    (fixed in this PR, and verified via Alive2: broken vs correct)
• unsigned checked_{add,sub} and signed/unsigned checked_mul map to SPIR-V instructions
  (OpIAddCarry, OpISubBorrow, OpSMulExtended/OpUMulExtended)
• saturating_{add,sub} use an Intel extension instead of plainly erroring
  (OpIAddSatINTEL/OpISubSatINTEL)
  • in time, we might choose to emulate these instructions (later in the pipeline)

---

TODO: add tests for the edge cases, consider the need for OpUAddSatINTEL/OpUSubSatINTEL.

[nazar-pc]

Could you also comment on #403? Seems directly applicable here.

[Firestar99]

The checked functions look good, the saturating ones emitting Intel-extension code can't really be used / tested anywhere but not like you've been able to do that in the first place either.

And then I read this:

TODO: add tests for the edge cases, consider the need for OpUAddSatINTEL/OpUSubSatINTEL.

And thought I'd make you a quick difftest between CPU and spirv to see if it actually resolved the issue at checked-saturating-spv-difftest, and you can see it fail:

    ┌────────┬─────────────────────────────┬──────────────────────────────────┐
    │ Offset │ math-checked-saturating-cpu │ math-checked-saturating-rust-gpu │
    ├────────┼─────────┬────────┬──────────┼──────────┬──────────┬────────────┤
    │        │   Hex   │  Dec   │  ASCII   │   Hex    │   Dec    │   ASCII    │
    ├────────┼─────────┼────────┼──────────┼──────────┼──────────┼────────────┤
    │ 0x0084 │    00   │    0   │   \0     │    01    │     1    │            │
    ├────────┼─────────┼────────┼──────────┼──────────┼──────────┼────────────┤
    │ 0x008c │    00   │    0   │   \0     │    01    │     1    │            │
    ├────────┼─────────┼────────┼──────────┼──────────┼──────────┼────────────┤
    │ 0x00b4 │    00   │    0   │   \0     │    01    │     1    │            │
    ├────────┼─────────┼────────┼──────────┼──────────┼──────────┼────────────┤
    │ 0x00bc │    00   │    0   │   \0     │    01    │     1    │            │
    └────────┴─────────┴────────┴──────────┴──────────┴──────────┴────────────┘

Whereas with this PR:

thread 'main' (564342) panicked at /home/firestar99/.cargo/registry/src/index.crates.io-1949cf8c6b5b557f/wgpu-27.0.1/src/backend/wgpu_core.rs:1058:
wgpu error: Validation Error

Caused by:
  In Device::create_shader_module, label = 'Compute Shader'
        
Shader 'Compute Shader' parsing error: UnsupportedInstruction(Function, IAddCarry)

naga doesn't support IAddCarry...

So what about using shader passthrough on our difftest? Well, it exists with -1 and no error... caused by a segfault... caused by shader passthrough requiring you to pass explicit PipelineLayout since wgpu's autogenerated one is just empty... which required refactoring the wgpu runner... and there's 3 different copies of a wgpu runner, thanks AI, so this may just be a rewrite of our difftest wgpu runner.

So... the difftest only works if rebased on #334 :D

[eddyb]

consider the need for OpUAddSatINTEL/OpUSubSatINTEL.

Definitely needed, on top of @Firestar99's difftest commit, this tells signedness apart:

0i32.saturating_sub(2), // -2
0u32.saturating_sub(2), // 0

(and even though the driver has the extension, there's a few other changes needed, will push ASAP)