NextCore GPU

nextcore-gpu provides graphics resource contracts, a software rasterizer, and an optional host Vulkan compute backend. This module is developed in its own repository and included by the 26x86 integration repository as a Git submodule. It has no sibling NextCore crate dependencies.

Recorded compute commands now execute through an explicitly supplied ComputePipelineManager: each submission waits for Vulkan completion and commits readback before the next command. An unavailable backend or failed command returns an error; completed earlier commands remain committed. The Vulkan backend reuses one native pipeline for identical shader code, entry point, and ordered buffer bindings, and recompiles when that key changes.

The product target is an x86 EFI host executing an ARM64E macOS guest. The Vulkan backend runs under a host operating system as development and reference code. The EFI GOP scanout implementation belongs to Nextcore-ISE, with its caller in Nextcore-EFI.

Build and verify

From this standalone repository, with Rust stable installed:

cargo test --no-default-features
cargo test --features vulkan
cargo check --features vulkan --example vulkan_compute

These commands do not require a working hardware GPU. The ignored Linux Vulkan lifecycle test requires a Vulkan loader, the Mesa llvmpipe ICD, and Khronos spirv-val. On the validated Ubuntu 24.04 setup:

VK_DRIVER_FILES=/usr/share/vulkan/icd.d/lvp_icd.json \
  timeout --kill-after=2s 30s \
  cargo test --features vulkan --lib \
  software_vulkan_pipeline_reuse_and_replacement -- --ignored --nocapture

The 2026-09-08 WSL2 validation used Rust 1.98.1 and Mesa 25.2.8 llvmpipe (LLVM 20.1.2). The default suite passed 99 tests; the Vulkan suite passed 102 tests with one ignored test. Running that ignored test explicitly passed five actual software Vulkan dispatches, independently checked 1,280 output values, and observed three native pipeline compilations.

A subsequent run on the same date cross-built the unchanged hardware acceptance example from WSL2 and executed it natively on Windows with an AMD Radeon RX 6800 XT (1002:73bf, driver 32.0.21043.10005). Six actual hardware dispatches passed 1,280 independent readback comparisons and guard/fence/failure checks. Three native pipeline compilations established shader reuse and replacement; cpu_fallback=false. The hardware validation evidence includes the selected device, stdout arrays, source hashes, and reproduction instructions. No driver replacement or reboot was needed.

See VULKAN_COMPUTE.md for the supported shader/buffer contract, recorded command API, hardware acceptance example, and failure semantics.

The SGPU submission adapter now shares APLS's existing LE command codec, rejects malformed counts and ranges before allocation, and maps explicitly registered resources/kernels into the same compute executor. Physical wire-path validation passed on the RX 6800 XT: one buffer copy, two GPU dispatches, 256 checked final values, guards, rejection and resource-lifetime checks. That milestone passed 109 GPU default tests, 112 GPU Vulkan tests (one ignored), and 48 APLS tests. Actual guest driver submission and EFI GPU command submission remain unverified.

Current limits

• The default compute manager has no execution backend. Public Vulkan construction requires an exact integrated/discrete device match and rejects CPU/virtual GPUs.
• Guest Metal registration, driver transport, Metal shader translation, and a physical PCI GPU command backend for EFI are not implemented by this module.
• VirtualMetalDevice supports software buffer copies only and reports Metal support as false. Its command API rejects unsupported compute, clear, and present operations.
• The hardware result establishes Windows-host Vulkan compute on the recorded GPU and driver. macOS boot, guest Metal execution, and EFI GPU command submission remain unverified.

Provenance

The historical 26x86-Nextcore-GPU-v0.1.1 snapshot was exported from integration commit dcc90013109eac694ccbf997b1e44a7018480f78. The current development tree includes subsequent command execution and pipeline lifecycle changes; it is not an unchanged copy of that release. repository.json preserves the original release metadata and describes the development state. repository-files.json records current public file sizes and SHA256 hashes, excluding itself, Git metadata, ignored files, build outputs, and private _isolated paths.

This repository contains public source and an independently authored shader test fixture. It includes no Apple firmware, operating-system binaries, or private research inputs. The repository license applies.

BP28 portable host GPU selection

The Vulkan backend now exposes actual queue/feature/compute-limit/memory inventory and accepts an explicit device UUID to distinguish identical GPU models while preserving the existing two-ID constructor. Compatible noncoherent host-visible memory uses explicit flush/invalidate; memory types requiring unenabled optional features are excluded. Device selection and memory decisions use actual Vulkan capabilities without a vendor allowlist or automatic fallback.

The BP28 hardware receipt records successful legacy-ID and observed-UUID runs on the RX 6800 XT: eight GPU dispatches each, checked readback and shader-cache/SGPU behavior. Vulkan tests passed 119, default tests 109 and native Windows library tests 10. NVIDIA/Intel integrated hardware and a real noncoherent allocation remain unverified, as do physical EFI GPU submission and guest macOS Metal. See the implementation contract and Vulkan usage.