Repository navigation
Releases
mac_linuxgpu 0.1.163
- The HSA runtime adds the entry points HRX System main uses. LemonSeed Engine 0.5.8 and later need them; without them libhsa failed to load.
hsa_amd_queue_createwith version-1 queue descriptors. A compute descriptor creates an AQL queue ofqueue_size_bytes / 64packets. Device-memory placement, SDMA and AIE engines, cooperative queues and non-normal priorities fail that descriptor withHSA_STATUS_ERROR_INVALID_QUEUE_CREATION. In a batch, a failed descriptor's queue is NULL and the first error is returned.hsa_amd_queue_cu_set_maskaccepts a mask that enables every CU the device reports. A mask that leaves a CU out still fails withHSA_STATUS_ERROR_INVALID_QUEUE.hsa_amd_vmem_get_accessreports a mapping's permissions.- Target ids with an unknown OS field are accepted, and the
MEMORY_PROPERTIESandPM4_EMULATIONagent queries are answered. - The vendored HSA headers move to hsa-runtime-headers 4285513, the revision HRX System pins.
- On an R9700, LemonSeed Engine on HRX main loads
/Library/MacAMDGPU/runtime/libhsa-runtime64.0.1.0.dylibwith no runtime override and serves plain, MTP=3 and DFlash2 requests.
- The display agent closes a mirroring process's output pipe at end of file. It used to spin on the closed pipe and keep its descriptor open after every mirroring process exit: after 29 hours, 170–260% CPU and 2,556 open pipes. Now: 0.0–0.3% CPU and 2 pipes over 9 minutes.
- A mirroring process no longer ends when its STATUS poll is answered
kIOReturnBusy. The daemon's polls and the mirroring process's IMPORT share the driver's display gate, and each collision ended the mirror for 10 s, about once a minute. IMPORT is asked again for up to 2 s onkIOReturnBusy. The daemon logs each mirroring process's exit status.
Upgrading from an earlier build: the old build does not yet leave on its own. Turn the display off in the menu bar, then use Disconnect GPU (or power-cycle the GPU), and the new build takes over.
mac_linuxgpu 0.1.162
- Code objects are sub-allocated from a code heap. The HSA runtime carves each program's code objects out of shared 4 MiB VRAM chunks, stages them in host memory and uploads them in one SDMA copy per batch, followed by one code-cache sync, instead of one allocation, one copy and several driver calls per object.
- Loading a code object: 0.437 ms and 6.2 driver calls → 0.009 ms and 0.02 calls (98% time saved).
- Destroying 1,000 code objects: 438 ms → 1.2 ms.
- LemonSeed Engine's 262 code objects load in 2.3–3.4 ms instead of 130 ms, and its live driver buffers drop from 624 to 366.
- Code objects no longer take entries in the driver's 8,192-entry buffer table, which every program on the GPU shares. 20,000 code objects load in one process.
LemonSeed Engine's time to first token and decode speed are unchanged.
Upgrading from an earlier build: the old build does not yet leave on its own. Turn the display off in the menu bar, then use Disconnect GPU (or power-cycle the GPU), and the new build takes over.
mac_linuxgpu 0.1.161
- Kernel arguments in VRAM for programs that opt in. The HSA runtime reports the HDP flush registers and lets the CPU map VRAM allocations it is granted, as ROCr does on a large-BAR card. The HDP flush register page is exported in the 4 KiB hole next to it, so 16 KiB host pages can map it. With LemonSeed Engine's HRX on an R9700: DFlash2 decode 53.0 → 54.2 tok/s (+2.3%), MTP=3 53.6 → 55.4 / 54.0 tok/s, plain 30.7 → 31.8 tok/s (+3.6%); time to first token unchanged, output identical.
- Every CPU store a client makes into the GPU on its submission path (kernel arguments, HDP flush, doorbell) happens inside one bracket on the client's gate. The driver closes the gate and waits for open brackets before a power change, device reset, session close, quarantine or fault. Once the device is gone, the runtime and libmlg_drm replace the process's GPU mappings with ordinary memory.
- Writes into VRAM use aligned 64-bit stores, since Apple silicon maps the BAR as device memory.
Upgrading from an earlier build: the old build does not yet leave on its own. Turn the display off in the menu bar, then use Disconnect GPU (or power-cycle the GPU), and the new build takes over.
mac_linuxgpu 0.1.160
- Programs ring their GPU queues' doorbells themselves, through the doorbell page KFD gives each process, as on Linux. A doorbell store takes 0.38 µs instead of 14.7 µs, and an empty dispatch round trip 32.4 µs instead of 43.9 µs (Linux on the same card: 31.3 µs). Doorbells stop before any power change, device reset, session close or removal.
- GPU recovery: a hang that a queue reset can't end can reset the whole device, behind a gate and with a time bound, after which the BARs, MSI-X and AQL queues are restored. Device resets are off unless the driver's MacLinuxGPUDeviceReset setting is on; without it a failed queue reset marks the GPU wedged, as before.
Upgrading from an earlier build: the old build does not yet leave on its own. Turn the display off in the menu bar, then use Disconnect GPU (or power-cycle the GPU), and the new build takes over.
mac_linuxgpu 0.1.159
Includes everything since 0.1.156 (builds 261–263; 261 and 262 were not released).
- Upgrades: the installer shuts the old build's GPU session down when it is idle. The old build does not yet leave on its own, so after installing, turn the display off in the menu bar and use Disconnect GPU (or power-cycle the GPU) for the new build to take over.
- Faster HSA runtime: a signal store is a plain CPU store (0.04 µs), doorbells are not waited for, blocked waits poll 200 µs before sleeping, async copies run on kept workers, and the shipped library is built optimized.
- Copies between system memory and VRAM run on SDMA through TTM's GART windows (6.3–6.5 GB/s). Copies below 256 KiB take the staging path.
- The dext is built with Linux's code-generation flags and copies normal memory with NEON.
- AQL queue kicks run on the delivery thread.
- A KFD queue's context-save area lives in VRAM.
- A dext that ends itself closes its PCI session first, and driver events (stop, power changes) are written to the unified log and to a trail on disk.
mac_linuxgpu 0.1.156
Includes everything since 0.1.151 (builds 256–260; 256–259 were not released).
- Faster first request: loading GPU code no longer launches a job per kernel. One cache sync runs per batch of loads, before the next submission. 60 kernel loads take 36 ms instead of 600 ms; a cold 137-token LemonSeed Engine request drops from 4.5 s to 1.6 s.
- The ~50 µs stall on dispatches that exactly fill a shader engine is down to ~17 µs: compute queue descriptors (MQDs) live in CPU-visible VRAM instead of host memory behind Thunderbolt.
- A program's GPU memory fault stays that program's: its queues are evicted and it gets a clear error, while other programs and the display keep running. A faulted queue left on the command processor is reset instead of wedging the GPU.
- HSA runtime: no more silent signal failures after long transfers, much cheaper signal create/destroy, and memory faults reported the way ROCr reports them.
- Performance controls in the menu bar (performance level, power profile, live clocks/temperature/power) and on the command line, kept and reapplied whenever the GPU comes up.
- Safety: every PCI call from the driver is validated before it reaches the kernel, and no kernel mapping is made for VRAM outside the CPU-visible window.
- GPU firmware updated to linux-firmware 20260916.
mac_linuxgpu 0.1.151
Includes everything since 0.1.143 (builds 248–255; 248–254 were not released).
- Faster PCIe link over Thunderbolt: amdgpu now sees the enclosure's port above the GPU, so the link trains at 16 GT/s instead of staying at Gen2. Host↔VRAM copies go from 1.65 GB/s to 6.9 GB/s; LemonSeed Engine loads Qwen3.8-27B Q4 + DFlash2 in 3.9 s.
- The GPU stays up across programs: the first program brings it up, later ones join the running device, and several programs (display, Vulkan, HSA, observers) run at once. "Disconnect GPU" closes the device for every program on it and makes it safe to unplug.
- Fixed the GPU dropping off the bus when the display reconnected while a compute program held VRAM: process memory written by upstream code (copy-job page tables) is no longer mistaken for a register access.
- CPU-visible VRAM stays free: compute buffers are placed above the 256 MB BAR window again (gpu_buddy top-down fallback), display framebuffers live outside it, and display/Vulkan clients no longer open a KFD process.
- A driver instance always exits: Stop and Retire never wait forever behind a stuck call, and a lost device ends its waits, so upgrades and power cycles no longer leave a stuck driver behind. Client records no longer leak (compute stopped working after 64 programs).
mac_linuxgpu 0.1.143
Includes everything since 0.1.138 (builds 243–247; 243–246 were not released).
- A stuck or slow GPU request can no longer freeze the driver for every app: every driver call that can wait runs off the driver's call-delivery thread (display, Linux-file/Vulkan and HSA session calls), and the driver's own stop and power messages have their own queue.
- GPU recovery as on Linux: a hung job resets its queue (MES/SDMA queue reset), the guilty context reports a reset, other work continues; a reset that can't complete marks the GPU wedged with a clear power-cycle message instead of waiting forever.
- Fixes found on hardware: Vulkan apps no longer hang at their first buffer map or fall back to the CPU on a cold GPU; HSA/LemonSeed Engine gets the GPU again (owner calls carry 16 scalars) and no longer crashes at exit; a leak that quarantined the driver after ~128 sessions is fixed; displays mirror again and unplugging with a live display is safe.
- Clients refuse drivers older than build 243 immediately instead of hanging; every async call has a bound and names the selector if it never completes.
mac_linuxgpu 0.1.138
- A stuck GPU can no longer freeze the driver for every app: display operations and Linux-file requests that can wait now run on worker threads, and unplug and shutdown are handled even while a request is blocked.
- Unplugging or powering off the GPU with a display live is safe: every driver access to VRAM goes through the kernel, and the menu bar has Disconnect GPU / Reconnect GPU.
- The installer puts libmlg_drm and libdrm_mlg in /usr/local/lib next to the HSA runtime, so local apps use the library that matches the installed driver.
- Fixes: the first syncobj of a render-node file no longer gets handle 0; important driver events go to the unified log (
log show --predicate 'eventMessage BEGINSWITH "mac.linuxgpu: EVENT"'); CI runs the offline tests and unsigned builds on every push.
Known issue: monitors on the GPU do not mirror in this build (the host passes no output buffer to the async display status calls). Fixed in 0.1.139.
mac_linuxgpu 0.1.136
- Display output on the GPU's monitors as macOS displays, started automatically with the driver and controlled from a menu bar item. Only changed pixels are copied, scrolls move in VRAM, and typed text reaches the screen in about 20 ms.
- Vulkan through RADV: Apple-silicon memory placement, faster driver requests (lock-free RCU, persistent wait workers, constant-time allocation tracking compiled into debug builds only), no fixed DMA budget, and a llama.cpp patch that queues graph inputs in the GPU stream (Qwen3.6-27B decode 12.96 → 22.9 tok/s).
- Games scan out directly from VRAM: Quake3e (Vulkan) runs full screen or windowed on the GPU's monitor through
mlg-run, with the window followed and cropped as it moves. - Safer lifecycle: GPU power-off with a live display no longer touches the device, quarantines after client exits and after removal are fixed, and the installer names apps that block an upgrade.