From a mailbox and two spinning cores to a char device.

softgpu is a software model of a GPU. The device is threads and memcpy on a 1 GHz clock. The stack above it is the real subject: rings, fences, interrupts, a page table, a runlist, and the user/kernel boundary those things sit on.

Submit stores into the mapped ring and rings the doorbell. It does not ioctl. The two engines are threads on a 1 GHz clock.

2.00 → 0.03engine cores while idle
~48 µsto wake a gated engine
40 → 9 GB/spinned 1 MiB copy, after the page table

Ten stages, in the order they were built.

Each one started from a number in the previous record. The stage ids follow the design notes, which is why 7 landed before 5.

What idle costs.

One 64-byte fill after a quiet gap. spin keeps both engines hot. sleep gates immediately and pays the wake on every command. Hybrid 50 µs is the default: gaps inside the budget are free, and a quiet device drops to a few hundredths of a core.

Latency of the command that breaks the gap. Log scale. A 10 µs budget gates on the 30 µs row. 50 µs and 200 µs both wait until 300 µs; the longer budget does not make that command faster. Source: gate-spin.json, gated.json, gate-sleep.json, vm7-idle10k.json, vm7-idle200k.json.

Both engines together. 2.0 is the bill this project paid from stage 0 until stage 7, without printing it.

A few rows that decided something.

All of these are from the VM, canary-gated, best pass per row. The fast cluster itself drifted from about 208 ns to about 375 ns across the project, so a later round trip is not automatically a regression.

Pinned host-to-device, 1 MiB. The page table is the break.
Stage	GB/s
3a pinned	65.6
4 no global lock	52.0
7 idle gating	39.5
5 virtual memory	9.0

Managed sequential copy, 1 MiB. Prefetch is 16 pages.
Mode	MB/s	µs/page
Prefetch on	469	8.7
Prefetch off	29	140
Engines spinning	883	4.6

Stage 6, `fair`. A higher priority does not wait out the slice. A 5 µs slice spends the engine on switching.
Run	First probe	Pair GB/s	Preempts / op
No timeslice, equal priority	210 µs	39.2	0
50 µs slice, equal	78 µs	39.2	0.09
5 µs slice, equal pair	2.7 µs	24.9	0.94
Priority, no slice	28 µs	—	0.10