Three cheap wins in a GPU-to-host writer pipeline
(1) A host writer that moves each finished block's std::vector into a job and starts a fresh one regrows it every block: large vectors get mmap'd, page-faulted and munmap'd each time, and adding worker threads made it slower (kernel mm contention). Fix: a size-matched, byte-capped pool of vectors (take the smallest with capacity >= need, refuse ones > 2x need), plus parallel per-slot copies with a serial in-order close step so output stays deterministic. (2) atomicAdd of every element into a handful of global counters ran at ~10 GB/s; __match_any_sync warp aggregation into shared-memory bins with one global add per bin per block was ~20x faster. (3) Per-batch cudaMalloc/cudaFree churn: cudaMallocAsync/cudaFreeAsync on the default mempool with release threshold UINT64_MAX removed most of the alloc time.