Struct NvrtcCache 

pub struct NvrtcCache { /* private fields */ }

Read-through disk cache for compiled NVRTC kernels.

Cross-process safe: writes go to a temp file and are atomically renamed into place. Reads tolerate concurrent writers (a partial .tmp is invisible to readers; a corrupt .bin returns None).

Implementations

impl NvrtcCache

pub fn new() -> Result<Self, GpuError>

Construct a cache rooted at the OS-default location.

Probe order:

1. $XDG_CACHE_HOME/atomr-accel/nvrtc/
2. $HOME/.cache/atomr-accel/nvrtc/ (via [dirs::cache_dir])
3. <temp_dir>/atomr-accel/nvrtc/

pub fn with_dir(path: PathBuf) -> Result<Self, GpuError>

Construct a cache rooted at an explicit directory. Creates the directory (recursively) if it does not exist.

pub fn dir(&self) -> &Path

Cache root directory.

pub fn get(&self, key: NvrtcCacheKey) -> Option<Arc<CachedKernel>>

Look up a cached kernel.

Returns None if the entry is absent, unreadable, mis-encoded, or stamped with a non-matching atomr_accel_version. A read hit on disk but cold in memory promotes the entry into the in-memory map for subsequent lookups.

pub fn insert( &self, key: NvrtcCacheKey, value: CachedKernel, ) -> Result<(), GpuError>

Store a kernel.

The on-disk write is atomic (write to <name>.tmp then rename). Concurrent writers race the rename; the loser’s bytes are silently overwritten — bincode payloads of the same key are expected to be identical so this is benign.

pub fn clear_in_memory(&self)

Drop every in-memory entry. Disk contents are left untouched — subsequent get calls will re-populate from disk.

Trait Implementations

impl Debug for NvrtcCache

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.