Module compressed_buffer 

Hardware-accelerated lossless memory compression bookkeeping.

On Ampere (compute capability 8.0) and newer GPUs the virtual-memory management API (cuMemCreate) accepts a CU_MEM_ALLOCATION_COMPRESSION_TYPE_GENERIC allocation flag that enables the memory controller’s lossless compression engine. Compressible allocations transparently reduce the number of bytes that travel across the L2 ↔ DRAM interface, raising effective bandwidth for compressible data (e.g. activation tensors with large runs of zeros).

This module models the host-side bookkeeping required to drive that feature: capability gating, compression-granularity alignment, an effective-bandwidth model, and a CompressedDeviceBuffer descriptor that tracks the logical / physical footprint of a compressed allocation. The actual cuMemCreate call requires a GPU and lives behind the device-gated path; everything here is deterministic and unit-testable on the host.

Example

// Ampere supports generic compression; the granularity is 2 MiB.
let support = CompressionSupport::for_compute_capability(8, 0);
assert!(support.is_supported());

// A 5 MiB request is rounded up to a multiple of the granularity.
let plan = CompressionPlan::new(5 * 1024 * 1024, support)?;
assert_eq!(plan.physical_bytes() % support.granularity(), 0);

Structs

CompressedDeviceBuffer

Host-side descriptor for a compressed device allocation.

CompressionPlan

A validated, granularity-aligned plan for a compressed allocation.

CompressionSupport

Describes whether and how a given device supports generic compression.

Enums

CompressionType

The kind of memory compression requested for an allocation.

Constants

DEFAULT_COMPRESSION_GRANULARITY

The default compression page granularity on supported hardware (2 MiB).