Struct TestContext 

pub struct TestContext {
    pub provider: CudaKernelProvider,
    pub device: Arc<CudaDevice>,
    pub memory: Arc<GpuMemoryManager>,
    /* private fields */
}

Test context providing CUDA resources for certification tests.

Fields

provider: CudaKernelProvider

device: Arc<CudaDevice>

memory: Arc<GpuMemoryManager>

Implementations

impl TestContext

pub fn with_budget(budget_bytes: usize) -> Result<Self>

Create test context with specific memory budget in bytes. Backend is chosen by XLOG_USE_DEVICE_RUNTIME (default legacy).

When XLOG_REQUIRE_CUDA=1, any initialization failure panics via enforce_cuda_required so callers’ skip-on-error paths cannot turn a missing GPU into a vacuous pass.

pub fn uses_device_runtime(&self) -> bool

Return whether this context was built against the runtime-backed allocator stack (vs the legacy cudarc-only stack). Cert categories can use this to gate behavior or diagnostics.

pub fn reap_pending(&self)

Drain pending async frees on the runtime allocator, if any. No-op for the legacy backend. Cert harnesses call this between categories so the GlobalDeviceBudget reservation bookkeeping releases bytes that have been freed via cuMemFreeAsync but whose owning stream has not yet been synchronized — without this, a long sequence of small allocate-then-drop tests piles up pending frees and exhausts the reservation pool even though the GPU has plenty of free memory.

pub fn new() -> Result<Self>

Create test context with default 1GB memory budget.

pub fn large() -> Result<Self>

Create test context with large 4GB memory budget for stress tests.

pub fn sync_and_check(&self) -> Result<()>

Force device synchronization and check for async errors.

pub fn memory_used(&self) -> usize

Get current memory usage in bytes.

pub fn memory_budget(&self) -> usize

Get memory budget in bytes.

pub fn reset_transfer_counters(&self)

Reset host/device transfer counters (in bytes).

pub fn transfer_counters(&self) -> (u64, u64)

Snapshot transfer counters (htod_bytes, dtoh_bytes).

pub fn htod_sync_copy_into<T: DeviceRepr, Dst: DevicePtrMut<T>>( &self, src: &[T], dst: &mut Dst, ) -> Result<()>

Track a synchronous host-to-device copy.

pub fn dtoh_sync_copy<T: DeviceRepr, Src: DevicePtr<T>>( &self, src: &Src, ) -> Result<Vec<T>>

Track a synchronous device-to-host copy returning a Vec.

pub fn dtoh_sync_copy_into<T: DeviceRepr, Src: DevicePtr<T>>( &self, src: &Src, dst: &mut [T], ) -> Result<()>

Track a synchronous device-to-host copy into a slice.

pub fn measure_gpu_ms<F>(&self, f: F) -> Result<f32>

where F: FnOnce() -> Result<()>,

Measure GPU time (in milliseconds) for work enqueued on the device stream.

pub fn multi_gpu_available(&self) -> bool

Check if multi-GPU is available.

pub fn compute_capability(&self) -> Result<(u32, u32)>

Get compute capability of current device.

pub fn device_row_count(&self, buffer: &CudaBuffer) -> u64

Read device-resident row count for a buffer.

Panics if the count cannot be read; certification tests treat this as fatal.