Graphics on Windows¶

Apple's GPU has no Windows driver and no public documentation. The stack here is built from Mesa's reverse-engineered AGX support, ported to Windows and driven by a Windows kernel driver.

1.4.354

Vulkan advertised

11.0

D3D11 feature level

1,200

frames presented

0

leaked objects after teardown

Measured on a MacBook Pro 14-inch M2 Pro (J414s, Apple T6020, G14S B1) running Windows on ARM64 under the m1n1 hypervisor.

The stack¶

    Direct3D 11 app          Vulkan app
          │                      │
          ▼                      │
     DXVK (D3D9-11)              │
          └──────────┬───────────┘
                     ▼
         Honeykrisp (Vulkan ICD, ported)
                     │
                     ▼
            AppleAgxGpu.sys
        KMDF render node, ACPI\NTAS0023\0
                     │
                     ▼
                   AGX

Direct3D 12 is served by a port of vkd3d-proton on the same Vulkan base.

What works¶

Capability	State	Evidence
Vulkan API	Working	Honeykrisp advertises Vulkan 1.4.354 on hardware
Compute, buffer and image paths	Passing	Empty, fill, constants, shader verification, meta-clear, IMG2BUF, offscreen render and CPU readback
Window-system presentation	Passing	20 fresh one-frame processes, 120 frames, and 1,200 frames on the physical console
Repeated process teardown	Passing	Firmware contexts, heap garbage, sync objects, VMs, BOs, queues and bindings all return to zero
Direct3D 11 via DXVK	Working	Feature level 11.0; compute 64/64 and raster/readback 8/8
dxgkrnl / WDDM miniport	In progress	Bring-up transaction and memory/UAT backend build and test clean; controlled hardware start is next
Accelerated DWM desktop	Not yet	Needs the WDDM submission/fence/present path and a native D3D11 UMD DDI

Working is not conformant

The Vulkan result is a runtime milestone, not a Khronos conformance claim. No dEQP/CTS binary is staged on the current machine.

Do not describe the ICD as conformant until the applicable Vulkan CTS profile actually passes. "Advertises 1.4.354 and runs real workloads" and "is a conformant Vulkan implementation" are different statements, and only the first is currently supported by evidence.

Components¶

Path	Role
`drivers/AppleAgxGpu`	KMDF render node — command streams, memory, sync objects, event completion
`drivers/AppleAgxWddm`	WDDM display miniport, in bring-up
`gpu/honeykrisp-port`	Mesa's AGX Vulkan driver, cross-built for aarch64-windows
`gpu/dxvk-port`	DXVK — Direct3D 9/10/11 on Vulkan
`gpu/vkd3d-port`	vkd3d-proton — Direct3D 12 on Vulkan
`gpu/d3d10umd-port`	Mesa's `d3d10umd` Gallium frontend as a D3D10 UMD
`gpu/vktests`, `gpu/vkprobe`	Validation harnesses

How results are evidenced¶

GPU claims here carry package hashes, driver versions and log locations, so a result can be tied to the exact binary that produced it. A recent validation run recorded:

submit_ioctl = materialize = prepare = 2690
prepare_failures  = 0
sync_wait_pends   = sync_wait_completions = 1348
wait_eval_error   = 0
VM/BO/queue/binding objects = 0
render submissions  = 1343
compute submissions = 1347

The zeros matter as much as the totals: every allocated object was returned, and no wait was left unevaluated. A GPU stack that renders correctly but leaks contexts fails differently — later, and harder to diagnose.

Relationship to d3d12agx¶

d3d12agx is a separate research effort on Linux: a native Direct3D 12 driver for AGX with no Vulkan layer at all.

If it matures and is ported to a Windows UMD, it would replace the vkd3d-port → Honeykrisp half of the chain for Direct3D 12 — one translation layer instead of two. That is a long way off; today the shipping path for D3D12 is vkd3d-proton on the Vulkan stack described above.