simvx.graphics.gpu_gate

GPU capability gating for the test suite: a single 3-state policy.

The engine’s tests fall into three buckets when a Vulkan GPU is involved, and collapsing “could not run” into “passed” is the one outcome we never allow (a green-but-empty GPU lane silently hides regressions). The three states:

• AVAILABLE – a usable Vulkan device is present; the test runs.

• ABSENT – no vulkan binding, or no Vulkan-capable device. By default the test skips (so contributors and CI hosts without a GPU stay green), unless --require-gpu is passed – then absence is a failure, because that flag means the caller (the GPU CI lane, tools/run_benchmarks.py) asked to exercise the GPU and an empty run must not look green.

• BROKEN – a device is present but initialisation raises. This is a real defect, not an absence, so it is always a failure, regardless of --require-gpu.

A test that needs a GPU is identified by the vulkan marker. The graphics conftest auto-applies that marker to any test requesting a GPU fixture (:data:GPU_FIXTURES), so “uses a GPU fixture” and “is a GPU test” are the same thing and nothing has to be remembered per test. The runtime guard (:func:ungated_gpu_error_hint) catches the residual case – a test that builds its own device without gating – by recognising the absent-device error.

This module holds the pure decision logic so it is unit-testable and monkeypatchable without a GPU; the conftest hooks are thin wrappers over it.

Module Contents

Classes

GpuStatus
Decision

Functions

reset_cache	Forget the cached probe result (test helper).
probe_gpu	Detect GPU status once per process. Returns (status, human_reason).
needs_gpu	True if the test requests a GPU fixture (so it should be vulkan-marked).
decide	Map (probe status, --require-gpu) to a run/skip/fail decision + message.
apply_gate	Resolve the run/skip/fail decision for one test (probes the GPU on demand).
ungated_gpu_error_hint	Return an actionable hint if a non-GPU-gated test errored from a missing GPU.

Data

__all__
GPU_FIXTURES
VULKAN_MARKER

API

simvx.graphics.gpu_gate.__all__

[‘GPU_FIXTURES’, ‘VULKAN_MARKER’, ‘GpuStatus’, ‘Decision’, ‘probe_gpu’, ‘reset_cache’, ‘needs_gpu’, …

simvx.graphics.gpu_gate.GPU_FIXTURES

‘frozenset(…)’

simvx.graphics.gpu_gate.VULKAN_MARKER

‘vulkan’

class simvx.graphics.gpu_gate.GpuStatus

Bases: enum.Enum

AVAILABLE

‘available’

ABSENT

‘absent’

BROKEN

‘broken’

__new__(value)

__repr__()

__str__()

__dir__()

__format__(format_spec)

__hash__()

__reduce_ex__(proto)

__deepcopy__(memo)

__copy__()

name()

value()

class simvx.graphics.gpu_gate.Decision

Bases: enum.Enum

RUN

‘run’

SKIP

‘skip’

FAIL

‘fail’

__new__(value)

__repr__()

__str__()

__dir__()

__format__(format_spec)

__hash__()

__reduce_ex__(proto)

__deepcopy__(memo)

__copy__()

name()

value()

simvx.graphics.gpu_gate.reset_cache() → None

Forget the cached probe result (test helper).

simvx.graphics.gpu_gate.probe_gpu() → tuple[simvx.graphics.gpu_gate.GpuStatus, str]

Detect GPU status once per process. Returns (status, human_reason).

simvx.graphics.gpu_gate.needs_gpu(fixturenames) → bool

True if the test requests a GPU fixture (so it should be vulkan-marked).

simvx.graphics.gpu_gate.decide(status: simvx.graphics.gpu_gate.GpuStatus, reason: str, require_gpu: bool) → tuple[simvx.graphics.gpu_gate.Decision, str]

Map (probe status, --require-gpu) to a run/skip/fail decision + message.

simvx.graphics.gpu_gate.apply_gate(is_vulkan_marked: bool, require_gpu: bool) → tuple[simvx.graphics.gpu_gate.Decision, str]

Resolve the run/skip/fail decision for one test (probes the GPU on demand).

This is the exact decision the conftest pytest_runtest_setup hook applies; the hook only maps the result onto pytest.skip / pytest.fail. Non-GPU tests (not vulkan-marked) always run.

simvx.graphics.gpu_gate.ungated_gpu_error_hint(report_text: str, is_vulkan_marked: bool) → str | None

Return an actionable hint if a non-GPU-gated test errored from a missing GPU.

The auto-marker covers every test that uses a GPU fixture; this is the safety net for a test that builds its own device without gating. On a GPU-less host such a test errors with the absent-device message instead of skipping, which reads as a real bug. We recognise that and tell the author how to gate it. Returns None for a correctly-gated test or an unrelated failure.