1# @ajax/mesh-rt-sim
2
3mesh-rt's device primitives on the CPU, and a launcher that runs a kernel's nodes as POSIX threads, so a
4mesh GPU program written in C can be tested without a GPU.
5
6- A node is `warps` x 32 threads with its own zeroed shared memory (`RT_SHARED_BYTES` + the program's).
7- The block barrier is a `pthread_barrier` over the node's threads. Every warp collective (shuffles,
8  ballot, `bar.warp.sync`) is an exchange through a per-warp array between two barriers over its 32
9  lanes, so programs that follow mesh-rt's rules (all lanes call warp collectives, all threads call block
10  ones) behave as on a GPU.
11- Loads, stores, atomics and fences are C11 atomics with the same orders; the global timer is
12  `CLOCK_MONOTONIC`; `trap` aborts with the node and thread.
13- Each simulated thread runs with f32 flush-to-zero (x86 MXCSR FTZ|DAZ, AArch64 FPCR.FZ), the PTX `.ftz`
14  semantics of the cluster float mode.
15
16```c
17mesh_sim_launch *l = mesh_sim_start(mesh_main, mailbox, slot, nodes, warps, shared_bytes);
18/* publish tickets, wait for marks (@ajax/mesh-rt-sim-mailbox) */
19mesh_sim_join(l);
20```
21
22- **Ported from:** the semantics of alpha-rust-mesh `crates/mesh-rt/src/ptx.rs` (the Rust has no CPU twin of
23  the GPU runtime; its native workers, `crates/mesh-worker`, run a different SDK).
24- **Tests:** every program under `programs/` runs its tasks on it and compares against the CPU references.
25- **Differences:** x86's FTZ/DAZ also applies to f64, which PTX never flushes: a program whose f64 values
26  become subnormal would differ here (none of the ported programs does). Divergent warp collectives (not
27  allowed by mesh-rt's rules) deadlock here rather than misbehave.