Skip to content

A8 GPU firmware in the A8 boards' firmware recipes - #203

Open
brkosebay wants to merge 10 commits into
OrchardLinux:mainfrom
brkosebay:a8-gpu-firmware
Open

brkosebay wants to merge 10 commits into
OrchardLinux:mainfrom
brkosebay:a8-gpu-firmware

Conversation

@brkosebay

Copy link
Copy Markdown
Contributor

Depends on #196 (the A8 GPU driver). This branch carries #196's commits, so review only the last 2 commits.

powervr's A8 variant in #196 requests apple/gfx-t7000.bin (the GPU firmware) and apple/gfx-t7000-border.bin (its border colour table), and refuses to probe without them. No A8 board's recipe installs them yet, so an installed system has no GPU firmware. This PR adds them to the five A8 boards: N102, N56, N61, J96 and J97.

Commits

  • firmware: A8 GPU firmware steps; inputs from another restore image
    • Two new steps:
      • agx-g4p-ldr-from-kernelcache@1: AGXFirmwareKextG4P's META LDR, byte for byte.
      • agx-g4p-border-from-kernelcache@1: the first 0x800 bytes of AGXG4P's border colour table.
    • The iOS 12 kernelcaches have no per-kext load addresses in their prelink metadata, so the A9's section lookup can't be used here. Both steps find their data by structure and return the candidate whose hash the output pins, as bluetool-hcd@1 does.
    • An ipsw input can now read a board's kernelcache from another pinned restore image: {"kernelcache": "n102ap", "restore": {version, build, url, bytes, sha256}}. Such an input must pin its own sha256 (checked in recipe.supported() and in the schema). _produce opens each restore image only when an input needs it.
    • Schema, docs/firmware-recipes.md and tests (synthetic bytes only).
  • devices: the A8 GPU firmware on every A8 board
    • Each of the five recipes reads the N102 kernelcache of the iPod touch 6's iOS 12.4 (16G77) restore image. All five pin the same two hashes:
      • gfx-t7000.bin: 115534 bytes, 0b60def6…;
      • gfx-t7000-border.bin: 2048 bytes, 233ffc1d….
    • The device.toml comments say where the GPU firmware comes from.
    • test_n61_data picks the touch output by destination, now that N61 has more than one output.

Why the iPod's 16G77 image on every board

The driver was written and tested against the 16G77 image, and relies on its segment layout and object offsets. None of the boards' own pinned restore images has that exact image:

Board Own restore image Its G4P firmware
N102, N56 iOS 12.5.8 (16H88) same as 16G77 except the build-date string and the 16-bit check word of the block that holds it
N61 iOS 12.5.1 (16H22) the same, with another date
J96, J97 iPadOS 15.8.8 (19H422) another firmware build (114794 bytes), never run by the driver

All of them hold the same border colour table. Reading the tested image everywhere means every A8 board gets the exact bytes that ran, and no board's own pin changes. Each board's other inputs (Wi-Fi, touch) still come from its own restore image.

Tested

Notes

Proposed changelog line: "A8 boards: the GPU firmware is installed with the other firmware (iPod touch 6 iOS 12.4 image)."

🤖 Generated with Claude Code

The A8's GX6450 runs the same family of firmware as the A9's, so it
becomes a variant of the S8003 chip file rather than a new driver
(drivers/gpu/0063-0073):

- 0063-0064: the S8003 chip described in a variant table, and chip
  hooks for its own GPU ID read and heap table. These touch the A9 and
  A10 paths; their values and functions are unchanged.
- 0065: dt-bindings for apple,t7000-gpu.
- 0066: a GPU node without power domains, and the failing probe step
  named (shared powervr code; nodes with power domains are unchanged).
- 0067-0069: the A8's firmware boot and power up and down in Apple's
  order, compute jobs, and TA/3D render jobs with the parameter-manager
  pool (a8_pm_blocks, a8_pm_grow) and the region-header ring kept in
  the HWRT data set.
- 0070: fault and hang reports and recovery.
- 0071: completion checks in the firmware interrupt's thread, the
  re-check after an early interrupt, a publish that cannot leave a
  failed job in the firmware's ring, Apple's doorbell.
- 0072: a fragment job submitted behind its geometry job, with two
  alternating 3D records (a8_joint, off by default).
- 0073: statistics and the A8's job timeline in debugfs apple_jobs,
  last. It also points the firmware's TA and 3D time words at the
  channel's times; the series without it has not run on the device.

Config: CONFIG_DRM_POWERVR_APPLE goes from unset to =y on apple-4k, as
on apple-16k, so the A8 gets powervr's Apple glue. It builds into the
powervr module on every A7, A8 and A8X kernel; only boards whose device
tree enables an apple,t7000-gpu node bind it.

Assisted-by: LLM
The iPod touch 6 RAM-boots through checkm8, PongoOS and m1n1 to the USB
serial shell. Its descriptor, boot assets and firmware recipe pin the
public iOS 12.5.8 IPSW.

devices/t7000/0011 describes the A8's GPU in t7000.dtsi (the Rogue core,
Fender, the power manager and GPU bridge registers, five interrupts,
the firmware name) and enables it on the N102, where powervr's A8
variant runs Vulkan and OpenGL ES (through Zink) jobs.

Orchard's feature list has no GPU entry and the chip's [gpu] table
records dEQP results, which the A8 does not have yet, so the GPU state
is in devices/n102/README.md. docs/compatibility.md is regenerated.

Assisted-by: LLM
The iPad mini 4 (J96, J97) and the iPhone 6 and 6 Plus (N61, N56) have
the iPod touch 6's A8 GPU. devices/t7000/0012 enables the GPU node in
their shared board files (t7000-mini4.dtsi, t7000-6.dtsi) so people with
these boards can test it; the driver has run only on the iPod touch 6.
Each board's README asks for the probe and ladder results. The A8X
(T7001) has a different GPU and is unchanged.

Assisted-by: LLM
minimal-boot with a constant USB serial number, the kernel's modules,
the GPU firmware and a8-gpu-drm-test, which loads powervr and checks the
probe, Mesa's Vulkan path and two DRM-level replays, each step under a
hardware watchdog with its kernel lines saved. It builds for every A8
board (n102, j96, j97, n61, n56).

tools/extract-fw.py writes the GPU firmware and its border colour table
from Apple's public iPod touch 6 iOS 12.4 IPSW, read over HTTP ranges
with orchard's own IPSW and kernelcache code, or from another IPSW or
kernelcache. Nothing from Apple is in the repository, and everything
build.sh stages is git-ignored.

Assisted-by: LLM
a8-vk-smoke (Vulkan probe, compute, a triangle) and a8-vk-ladder, a set
of Vulkan tests from one clear to vkmark's scenes with every pixel or
value checked on the CPU, both MIT. vk/make-bundle.sh builds them with a
Mesa build of the A8 branch into the self-contained /opt/a8vk bundle
that build.sh's A8_VK_DIR stages, with vkmark when it is installed.

Assisted-by: LLM
- agx-g4p-ldr-from-kernelcache@1 and agx-g4p-border-from-kernelcache@1:
  the A8 GPU firmware (AGXFirmwareKextG4P's META LDR, byte for byte) and
  the border colour table powervr's A8 variant loads beside it (0x800
  bytes of AGXG4P's data). The iOS 12 kernelcaches that carry them have
  no per-kext load addresses in their prelink metadata, so both are found
  by structure, and the output's pinned hash selects the one installed.
- An ipsw input may read a board's kernelcache from another pinned
  restore image: {"kernelcache": "<board>ap", "restore": {...}}, with
  the same fields as the recipe's own ipsw. Such an input must pin its
  sha256. Each restore image is opened only when an input needs it.

Tests use synthetic bytes only.

Assisted-by: LLM
N102, N56, N61, J96 and J97 install apple/gfx-t7000.bin and
apple/gfx-t7000-border.bin, the two files powervr's A8 variant requests.
Both come from the N102 kernelcache of the iPod touch 6's iOS 12.4
(16G77), the image the driver was written and tested against, so every
board pins the same two hashes.

The boards' own restore images don't have that image: iOS 12.5.8 and
12.5.1 (N102, N56, N61) differ from it in the firmware's build-date
string and the check word of the block that holds it; iPadOS 15.8.8 (J96,
J97) has another firmware build, which the driver has not run. Every one
holds the same border colour table.

Checked against Apple's server with firmware.ensure() for all five
boards, into an empty cache.

Assisted-by: LLM
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant