Module kernels 

Built-in SoftGPU Functional IR kernels (Rust builders + JSON twins).

Functions

atomic_inc_reduce

SoftGPU Phase 7: each lane atomic_add(counter, 1); write previous value to out[i].

cross_workgroup_store_zero

SoftGPU Phase 8 (negative): every workgroup stores to the same global cell.

group_exchange

SoftGPU Phase 7: group[lid]=lid; barrier; out[gid]=group[(lid+1)%wg].

group_exchange_missing_barrier

SoftGPU Phase 8 (negative): group exchange without barrier (MissingBarrier).

infinite_loop_watchdog

SoftGPU Phase 7: busy loop until step budget trips (watchdog).

kernarg_one_ptr

kernarg_two_ptrs

Pack two u64 pointers into a 16-byte kernarg.

oob_store_past_end

SoftGPU Phase 8 (negative): store at base + (n * 4) — one past last element.

predicated_inc

SoftGPU Phase 7: even lanes add 1 to buf[i]; odd lanes leave value unchanged.

race_all_store_global_zero

SoftGPU Phase 8 (negative): every workitem stores to global offset 0 (race).

sanitize_clean_index

SoftGPU Phase 8 (positive): private per-lane global store (no race).

tiny_add

b[i] = a[i] + 1 for i = global_id_x. Pointers in kernarg at 0 and 8.

tiny_copy

b[i] = a[i] copy.

tiny_index

out[i] = global_id_x (index write).

uninit_group_read

SoftGPU Phase 8 (negative): read group memory before any store (uninit).

wave_lane_ids

SoftGPU Phase 7: write (wave_id * 16) + lane_id to out[global_id].