Struct FileBatch

Source

pub struct FileBatch { /* private fields */ }

Expand description

Persistent device-owned batch buffers.

Implementations§

Source §

impl FileBatch

Source

pub fn upload( device_queue: Arc<(Device, Queue)>, files: &[BatchFile], rule_count: u32, hit_capacity: u32, ) -> Result<Self, PipelineError>

Upload a new multi-file batch into persistent GPU buffers.

§Errors

Returns PipelineError::QueueFull when the batch exceeds the current u32 table limits or the work queue would overflow.

Source

pub fn refresh( &mut self, files: &[BatchFile], rule_count: u32, hit_capacity: u32, ) -> Result<(), PipelineError>

Refresh this batch in place, reusing host staging vectors and resident GPU buffers whenever the new batch fits existing allocations.

§Errors

Returns PipelineError::QueueFull before mutating the batch when the requested file/rule fanout cannot fit the megakernel batch protocol.

Source

pub fn refresh_with_report( &mut self, files: &[BatchFile], rule_count: u32, hit_capacity: u32, ) -> Result<FileBatchRefreshReport, PipelineError>

Refresh this batch in place and return allocation/transfer telemetry.

§Errors

Returns PipelineError::QueueFull before mutating the batch when the requested file/rule fanout cannot fit the megakernel batch protocol.

Source

pub fn reset_queue_state(&self) -> Result<(), PipelineError>

Reset the persistent queue indices before another dispatch.

§Errors

Returns PipelineError::Backend when the queue-state upload fails.

Source

pub fn set_segmentation( &mut self, seg_len: u32, overlap: u32, ) -> Result<(), PipelineError>

Re-tile this batch at a new window geometry: rebuild the segments table (each file split into ceil(len / seg_len) windows, files shorter than seg_len staying whole), grow/rewrite the resident segments buffer, and update queue_len = segment_count * rule_count + the queue-state header so the next dispatch claims the new work space. seg_len = u32::MAX restores one segment per file (the legacy whole-file scan).

SOUNDNESS CONTRACT: overlap MUST be at least segmentation::catalog_sync_overlap over the rules that will scan this batch. With a shorter warm-up a window can reconstruct the wrong DFA state at emit_start and drop or fabricate matches. The geometry is the caller’s decision (it owns the rule catalog); this method enforces only the structural invariant seg_len > 0.

§Errors

Returns PipelineError::QueueFull when seg_len == 0 or the new work queue overflows the device claim protocol, or PipelineError::Backend when the buffer upload fails.

Source

pub fn file_count(&self) -> usize

Number of files in the batch.

Source

pub const fn queue_len(&self) -> u32

Number of queued (file, rule, layer) items.

Source

pub const fn hit_capacity(&self) -> u32

Sparse-hit capacity.

Source

pub fn device_queue(&self) -> Arc<(Device, Queue)>

Device queue used for every buffer in this batch.

Source

pub const fn haystack(&self) -> &GpuBufferHandle

Packed haystack buffer.

Source

pub const fn offsets(&self) -> &GpuBufferHandle

Prefix-sum offset table. Length = file_count + 1.

Source

pub const fn metadata(&self) -> &GpuBufferHandle

Per-file metadata table.

Source

pub const fn segments(&self) -> &GpuBufferHandle

Flat segment table (segment_count * SEGMENT_WORDS u32s). The device claim decode reads row seg_idx = claim / rule_count to derive the window (file_idx, scan_start, emit_start, emit_end).

Source