A custom, lock-free C++17 fiber scheduler with topological DAG logic gates and rigid 64 byte task layouts optimized for L1 cache boundaries. Solves starvation/deadlock via age-based promotion and priority inheritance.
currently working on a windows/directx ecosystem of game libs and this is the core of the engine
my other repos have stuff in them like the renderer, and other libs and some games (although the platformer demo wont run because i need to re upload it with the assets in the distribution and i dont have time before work today)
either way this is a really cool scheduler i built with a lot of features and options, a fast path if you dont need fiber features, fork join support, topology-aware stealing, built in parallelfor, a built in task-dag, and much more, very fast!
currently windows-only but highly feature rich and efficient
Last benchmarks:
Throughput: 0.82M tasks/sec — this is actually measuring the producer side, not the pool: one thread (main) creating+pushing 200k tasks serially at ~1.2µs each while workers drain faster than main can feed them. So it's your task-creation cost, which is the number the 64-byte change helps.
Latency: 6.9µs push→run→wait round trip. That's the floor for "schedule one thing and wait for it" — solid, and it bounds how fine-grained a task can be before overhead eats it: anything under ~10µs of real work isn't worth its own task.
ParallelFor: 3.4x speedup — don't read this as "scheduler only scales 3.4x on 32 threads." The kernel is one multiply-add per float over 64MB, which is memory-bandwidth-bound; 3.4x means you saturated DRAM with ~4 cores and the other 28 had nothing to feed on. A compute-heavy kernel (e.g. sqrtf chains) would show the real scaling ceiling.
Frame DAG: 31.9µs per 6-node graph — Building, submitting, running, and tearing down a Game01-shaped frame graph costs ~32µs, i.e. 0.2% of a 60fps frame budget.
fork-join : completed run 0: 12.99 ms (pushed=254, executed=128)
fork-join : 1M recursive (10k-elem leaf) best-of-1 -> 12.99 ms
A custom, lock-free C++17 fiber scheduler with topological DAG logic gates and rigid 64 byte task layouts optimized for L1 cache boundaries. Solves starvation/deadlock via age-based promotion and priority inheritance.
currently working on a windows/directx ecosystem of game libs and this is the core of the engine
its running my platformer and a parallel commandlist renderer shown here https://streamable.com/qakm4h
my other repos have stuff in them like the renderer, and other libs and some games (although the platformer demo wont run because i need to re upload it with the assets in the distribution and i dont have time before work today)
either way this is a really cool scheduler i built with a lot of features and options, a fast path if you dont need fiber features, fork join support, topology-aware stealing, built in parallelfor, a built in task-dag, and much more, very fast!
currently windows-only but highly feature rich and efficient
Last benchmarks:
Throughput: 0.82M tasks/sec — this is actually measuring the producer side, not the pool: one thread (main) creating+pushing 200k tasks serially at ~1.2µs each while workers drain faster than main can feed them. So it's your task-creation cost, which is the number the 64-byte change helps. Latency: 6.9µs push→run→wait round trip. That's the floor for "schedule one thing and wait for it" — solid, and it bounds how fine-grained a task can be before overhead eats it: anything under ~10µs of real work isn't worth its own task. ParallelFor: 3.4x speedup — don't read this as "scheduler only scales 3.4x on 32 threads." The kernel is one multiply-add per float over 64MB, which is memory-bandwidth-bound; 3.4x means you saturated DRAM with ~4 cores and the other 28 had nothing to feed on. A compute-heavy kernel (e.g. sqrtf chains) would show the real scaling ceiling. Frame DAG: 31.9µs per 6-node graph — Building, submitting, running, and tearing down a Game01-shaped frame graph costs ~32µs, i.e. 0.2% of a 60fps frame budget.
fork-join : completed run 0: 12.99 ms (pushed=254, executed=128) fork-join : 1M recursive (10k-elem leaf) best-of-1 -> 12.99 ms
very fast