render-test: capture shader-abort message to stdout + abort message tests

[nv-slang-bot]

Motivation

Follow-up to #11790 / #11792. PR #11792 bumps external/slang-rhi to the head of
slang-rhi#782, which adds the Vulkan
Feature::ShaderAbort ("shader-abort") path that recovers a shader abort()'s message after the
device-fault and forwards it to the host debug callback. To actually exercise that message
round-trip from a .slang test, render-test needs to surface the captured message where a
slang-test directive can assert it.

Per maintainer guidance on #11792: "capture the handleMessage output automatically when
-render-features shader-abort is set, and let a filecheck= line assert it." This PR implements
exactly that, plus the requested %f / multiple-argument / mixed-type abort cases.

Proposed solution

When (and only when) -render-features shader-abort is requested, render-test enters a
device-loss-tolerant capture mode: it scopes a local debug callback over the dispatch, then prints
any "Shader abort: ..." message it received to stdout. A COMPARE_COMPUTE(filecheck=...)
line FileChecks render-test's stdout, so the formatted abort text can be asserted directly. The
mode is strictly gated, so every other render test is byte-for-byte unchanged.

Change summary

File	What
tools/render-test/render-test-main.cpp	_isShaderAbortRequested + _printCapturedShaderAbortMessages helpers; in _innerMain, a gated branch that scopes a capture callback over app.update() and prints the abort text to stdout. Non-abort path unchanged.
tests/spirv/abort-message.slang (new)	Four aborting COMPARE_COMPUTE(filecheck=...) -vk cases: plain string, %f, multi-arg, mixed %d %f. Each is its own entry point + //TEST: line (a fired abort loses the device, terminal for that invocation).
tests/spirv/abort-message.slang{,.1,.2,.3}.expected.txt (new, empty)	Satisfy runComputeComparisonImpl's mandatory buffer comparison: a lost device writes no buffer, so the empty pre-cleared .actual.txt matches an empty .expected.txt.

The emit-side coverage (-target spirv -capability abort → OpAbortKHR) and the device-keep-alive
buffer-compare line continue to live in tests/spirv/abort-runtime.slang (PR #11792, unchanged).

Concepts and vocabulary

• filecheck= vs filecheck-buffer= — in a render-test COMPARE_COMPUTE, filecheck= runs
  FileCheck against render-test's stdout/stderr (_validateOutput), while filecheck-buffer=
  runs against the readback buffer file. The abort message arrives on stdout, so these tests use
  filecheck=.
• device-fault path — on a shader abort the Vulkan device is lost; slang-rhi reads the abort
  message back via vkGetDeviceFaultDebugInfoKHR and delivers it as a host debug message prefixed
  "Shader abort: ".
• ScopedCoreDebugCallback / CoreDebugCallback — the existing render-test debug-callback
  bridge/buffer (tools/render-test/slang-support.h); the capture mode reuses them rather than
  adding any new interface or ABI.

Process report

• Strictly additive + gated. The capture branch is under shaderAbortMode = _isShaderAbortRequested(options); the else-branch is the identical app.update();. Zero behavior
  change for any test that does not request shader-abort. _isShaderAbortRequested reuses the
  canonical _getFeatureFromName name→feature map (one source of truth), which returns _Count
  for unknown names with no side effects, so failure timing for other tests is unchanged.
• Device-loss tolerance. A fired abort loses the device, so app.update()'s result is swallowed
  in this mode (it was already unchecked on the normal path) so we reach the print; filecheck=
  ignores the process result code. The capture callback is scoped to just the dispatch.
• Buffer-comparison branch. Rather than fabricate a readback after device loss, the empty
  .expected.txt siblings match the empty (never-written) output buffer — the cleanest neutralizer
  for the always-on buffer comparison.
• Verification / honesty. Verified here: the C++ compiles; the existing abort-runtime.slang
  SIMPLE emit line still PASSES; all four new -vk -render-features shader-abort lines SKIP cleanly
  with no GPU (SLANG_E_NOT_AVAILABLE → ignored, not failed). Not verifiable here (no GPU): the
  abort actually firing → device-fault → "Shader abort: ..." round-trip → FileCheck match. That
  end-to-end is a maintainer hardware run.

Dependency / merge status

Draft, and cannot merge before #11792. This branch is based on #11792's fix/issue-11790
branch because rhi::Feature::ShaderAbort only exists in the bumped submodule — the capture code
does not compile against master. #11792 itself is gated on
slang-rhi#782 merging. Kept separate (rather
than folded into #11792) so #11792 stays a minimal submodule-bump + gated-test that can come out of
draft the moment slang-rhi#782 lands; this can fold into #11792 instead if preferred.

Related to #11790. Does not auto-close it (the GPU round-trip is verified separately).

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

Remove all of "*.expected.txt" files.
We prefer to use inlined //CHK: sytax.

[nv-slang-bot]

Done — pushed fa1a35d976. Removed all four *.expected.txt files and moved the assertions inline.

Each case now combines two inline FileCheck prefixes in one directive:

• filecheck=<PREFIX> asserts the captured Shader abort: … message on render-test's stdout (what the capture mode prints), and
• filecheck-buffer=EMPTY asserts the readback buffer stays empty — a fired abort loses the device, so nothing is written back; the shared // EMPTY-NOT: {{.}} satisfies COMPARE_COMPUTE's mandatory buffer comparison without a sidecar file.

I kept filecheck= for the message per your earlier note. One choice worth your call: since the abort loses the device there's no readback, so the buffer side is an emptiness assert rather than a positive value check. If you'd rather have render-test write the captured Shader abort: … text into the output buffer so a single positive filecheck-buffer= checks it (dropping the stdout path and the empty-assert entirely), I'm happy to switch — just say the word.

Verified locally: the combined directives parse and the four -vk lines skip cleanly on a GPU-less runner, and the existing -target spirv emit case still passes. As before, the abort-fires → message round-trip → FileCheck match is the on-hardware run.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}

[jkwak-work]

Looks good to me.

But blocked until the following PR gets merged on slang-rhi:

• Vulkan: support VK_KHR_shader_abort + VK_KHR_device_fault slang-rhi#782

[nv-slang-bot]

Thanks! Agreed on the gating.

I'll hold this draft — and the underlying submodule bump it stacks on (#11792) — until slang-rhi#782 merges. At that point I'll move the external/slang-rhi pin from the PR-branch commit to the merged slang-rhi main SHA and take both out of draft. No further changes needed here in the meantime.

_{🤖 Generated by an automated Slang coworker — may be inaccurate. A human maintainer should verify.}