Module ifds_gpu

Expand description

GPU-native IFDS/IDE driver (G3).

Build a BFS step over the exploded supergraph:

use weir::ifds_gpu::{ifds_gpu_step, IfdsShape};

let shape = IfdsShape {
    num_procs: 1, blocks_per_proc: 4, facts_per_proc: 4, edge_count: 4,
};
let step = ifds_gpu_step(shape, "fin", "fout").unwrap();

GPU-native IFDS/IDE driver (G3).

§What this does

IFDS / IDE reframes interprocedural dataflow as reachability on the exploded supergraph: each (proc, block, fact) triple is a graph vertex; the analysis reduces to BFS + bitset-fixpoint

primitives vyre already owns.

The pieces live in vyre_primitives::graph::exploded (node encoding + CSR builder), vyre_primitives::graph::csr_forward_traverse (BFS step), and vyre_primitives::fixpoint::bitset_fixpoint (convergence loop). This module composes them.

§Entry points

solve_cpu - in-process CPU reference. Conformance tests run this against the GPU output bit-for-bit.
[solve_borrowed_via] - production solver route: builds the exploded CSR through GPU dispatch and reuses borrowed ProgramGraph buffers through the fixpoint loop.
[ifds_gpu_step] - one BFS step over the exploded supergraph as a GPU [Program]. Caller dispatches this in a loop over (frontier_in, frontier_out) until the frontier stops growing (classic BFS-to-fixpoint). Allocates the exploded supergraph’s ProgramGraph buffers internally - the caller only provides the two frontier buffer names.

Structs§

IfdsBorrowedSolveScratch: Caller-owned host scratch for repeated one-shot IFDS solves.
IfdsGpu: Marker type for the GPU-native IFDS dataflow primitive.
IfdsPrepareScratch: Caller-owned host scratch for repeated IFDS CSR preparation.
IfdsShape: Dispatch geometry for one IFDS BFS step on the GPU.
IfdsSolveScratch: Caller-owned host scratch for repeated prepared IFDS solves.
PreparedIfdsCsr: Reusable IFDS exploded-supergraph CSR prepared once for many seed queries.
ResidentIfdsHostScratch: Caller-owned host scratch for repeated resident IFDS single-query solves.
ResidentIfdsParallelHostScratch: Caller-owned host scratch for repeated resident IFDS parallel solves.
ResidentIfdsParallelScratch: Reusable resident scratch buffers for parallel IFDS query batches.
ResidentIfdsScratch: Reusable resident IFDS scratch buffers.
ResidentPreparedIfdsCsr: IFDS CSR and invariant ProgramGraph buffers uploaded once to a backend.

Constants§

OP_ID: Stable primitive id for Weir’s GPU-native IFDS fixpoint route.

Traits§

IfdsResidentDispatch: Backend-neutral resident dispatch hooks for IFDS.

Functions§

allocate_resident_ifds_parallel_scratch: Allocate reusable resident scratch buffers for parallel IFDS batch solves.
allocate_resident_ifds_parallel_scratch_with_capacities: Allocate reusable resident scratch buffers for parallel IFDS batch solves with explicit query, iteration, and seed staging capacities.
allocate_resident_ifds_parallel_scratch_with_iterations: Allocate reusable resident scratch buffers for parallel IFDS batch solves with an explicit convergence iteration budget.
allocate_resident_ifds_scratch: Allocate reusable resident scratch buffers for IFDS solves.
allocate_resident_ifds_scratch_with_seed_capacity: Allocate reusable resident scratch buffers for IFDS solves with an explicit per-query seed fact staging capacity.
free_resident_ifds_parallel_scratch: Free reusable parallel IFDS scratch buffers.
free_resident_ifds_scratch: Free reusable IFDS scratch buffers created by allocate_resident_ifds_scratch.
free_resident_prepared_ifds_csr: Free invariant resident buffers created by upload_prepared_ifds_csr_resident or prepare_ifds_csr_resident_via.
ifds_gpu: Backwards-compatible three-string shim over ifds_gpu_step. The _exploded_adj argument is the ProgramGraph’s pg_edge_targets buffer; it is ignored here because the step kernel binds pg_edge_* at the canonical names. Kept so older call sites compile unchanged.
ifds_gpu_step: Emit one GPU BFS step over the exploded supergraph.
prepare_ifds_csr_borrowed_via: Build and pack the invariant IFDS CSR once for repeated GPU solves.
prepare_ifds_csr_borrowed_with_scratch_via: Prepare IFDS CSR buffers while reusing caller-owned dispatch output storage.
prepare_ifds_csr_resident_via: Build, pack, and upload IFDS invariant CSR buffers once for resident solves.
solve_borrowed_many_via: Solve several IFDS seed sets through one borrowed-buffer GPU preparation.
solve_borrowed_many_with_scratch_via: Solve several IFDS seed sets through one borrowed-buffer GPU preparation, reusing all host-side preparation and fixpoint scratch across queries.
solve_borrowed_via: Solve IFDS reachability through a caller-provided borrowed-buffer GPU dispatch.
solve_borrowed_with_scratch_via: Solve IFDS reachability through a borrowed-buffer GPU dispatch while reusing all host-side preparation and fixpoint scratch across calls.
solve_prepared_borrowed_into_via: Solve IFDS reachability over a prepared CSR while reusing caller-owned dispatch output storage across fixpoint iterations.
solve_prepared_borrowed_many_with_scratch_into_via: Solve several IFDS seed sets over one prepared CSR into caller-owned result rows while reusing all host-side solve scratch between queries.
solve_prepared_borrowed_many_with_scratch_via: Solve several IFDS seed sets over one prepared CSR while reusing all host-side solve scratch between queries.
solve_prepared_borrowed_via: Solve IFDS reachability over a CSR prepared by [prepare_ifds_csr_borrowed_via].
solve_prepared_borrowed_with_adapter_scratch_via: Solve IFDS reachability over a prepared CSR while adapting a borrowed Vec-returning dispatch into caller-owned solve scratch.
solve_prepared_borrowed_with_scratch_into_via: Solve IFDS reachability over a prepared CSR into caller-owned result storage while reusing all host-side solve scratch across calls.
solve_prepared_borrowed_with_scratch_via: Solve IFDS reachability over a prepared CSR while reusing all host-side frontier, decode, and backend-output scratch across calls.
solve_resident_prepared_many_parallel_via: Solve several IFDS seed queries against one resident CSR in parallel.
solve_resident_prepared_many_parallel_with_scratch_and_host_into_via: Solve several IFDS seed queries in parallel using caller-owned reusable resident buffers, host packing/readback scratch, and result storage.
solve_resident_prepared_many_parallel_with_scratch_and_host_via: Solve several IFDS seed queries in parallel using caller-owned reusable resident buffers and caller-owned host packing/readback scratch.
solve_resident_prepared_many_parallel_with_scratch_via: Solve several IFDS seed queries in parallel using caller-owned reusable parallel-batch scratch buffers.
solve_resident_prepared_many_via: Solve several IFDS seed queries against one resident CSR with one scratch frontier allocation and one scratch changed-flag allocation.
solve_resident_prepared_many_with_scratch_and_host_into_via: Solve several IFDS seed queries against one resident CSR using caller-owned reusable resident buffers, host packing/readback scratch, and result storage.
solve_resident_prepared_many_with_scratch_into_via: Solve several IFDS seed queries against one resident CSR using caller-owned reusable scratch buffers and caller-owned result storage.
solve_resident_prepared_many_with_scratch_via: Solve several IFDS seed queries against one resident CSR using caller-owned reusable scratch buffers.
solve_resident_prepared_via: Solve IFDS reachability using resident CSR/frontier buffers.
solve_via: Solve IFDS reachability through a caller-provided owned-buffer GPU dispatch.
upload_prepared_ifds_csr_resident: Upload an already prepared IFDS CSR into backend-resident buffers.
validate_ifds_problem

Module ifds_gpu

Module ifds_gpu Copy item path

§What this does

§Entry points

Structs§

Constants§

Traits§

Functions§

Module ifds_gpu