Skip to content

arm64: dts: apple: t6050-j775d: Give the Mac Studio (M5 Ultra) a working GPU, and up to 1.5x faster light GPU submits - #1973

Open
funkzy wants to merge 3 commits into
aurora-silicon:integration/aurorafrom
funkzy:m5-ultra
Open

funkzy wants to merge 3 commits into
aurora-silicon:integration/aurorafrom
funkzy:m5-ultra

Conversation

@funkzy

@funkzy funkzy commented Nov 15, 2026 •

Copy link
Copy Markdown
Contributor

What is missing

  • No GPU. The GPU driver has the M5 Ultra's configuration, but the J775D device tree leaves the GPU and both
    GPU coprocessor nodes disabled, with no mailboxes and no reservations. The Mac Studio boots without a render
    node and its desktop session cannot start.
  • No CPU clusters. t6050-ultra.dtsi has no cpu-map and names no cache, so Linux sees thirty CPUs in one
    package with one L2 shared by all thirty. parse_dt_topology() reads capacity-dmips-mhz only while it walks
    cpu-map, so the M and P capacities from arm64: dts: apple: t6050: Give the M and P cores their capacities #1399 never reach the scheduler (every CPU reports 1024), and a wakeup
    may pick any idle CPU of the thirty, on either die. The cpu-map commit for the one-die parts (68c10dd) says
    "The two-die description keeps its flat layout for now"; this fills it in.

What this adds

Three commits, J775D only:

  1. The CPU clusters and their L2s (t6050-ultra.dtsi). A cpu-map in the shape of the one-die description's:
    six clusters of five, MPIDR Aff1 0, 1, 2 for die 0's M0, M1 and P and 0x10, 0x11, 0x12 for die 1's, as the
    ADT's cluster-id and die-id give them, listed flat as t6002.dtsi and t6022.dtsi list theirs. One L2 node a
    cluster and each CPU's next-level-cache: 8 MiB in the M clusters, 16 MiB in the P clusters, the ADT's
    l2-cache-size. One commit because the gain needs both: with the cpu-map alone every CPU still reports its L2
    as shared by all thirty and the rows below do not move.
  2. Binding: apple,agx-g17d with the apple,agx-g17s fallback, added to the entry ewb's 4c7fdf1 made for
    the M5 Max's apple,agx-g17c: his entry becomes a list of two names, the form the G13 and G14 parts use. The
    driver matches the family name and tells the parts apart by the GPU's ID registers; the name is there so the
    device tree says which part it is. The Ultra's GPU node has the M5 Pro's hand-off shape (one sgx range, two
    firmware mailboxes, six reserved regions, no operating-point tables), so the M5 Pro hand-off entry, which since
    sithglan's 8f418a4 applies to a node whose only compatible is apple,agx-g17s, applies to a node with
    apple,agx-g17d as well.
  3. t6050-j775d-gpu.dtsi, in the shape of dj's t6050-j714s-gpu-panel.dtsi (32a31cf): the gpu alias,
    the six reservations the boot loader fills, the two firmware mailboxes in place of the inventory's coprocessor
    nodes, GPX and AFR held on, and the GPU node's handoff properties. The J775D ADT gives the same addresses and
    interrupts as J714s's: both firmware roles and their mailboxes are on die 0. It differs from the J714s file in
    two things: no panel or display part (the Studio's display chain is described already), and the part's name.
    Everything stays disabled in the source; the boot loader enables the nodes after checking the ADT and the GPU
    firmware build.

The GPU part now reads much the same in four board files (J714s, J716c, K114c and this one). It could be one shared
file that they include; this leaves the others as they are. dj, say if you would rather have that here.

Who runs the changed code

  • Only t6050-j775d.dts includes t6050-ultra.dtsi and the new file. The ten other T6050 board device trees
    (J714c, J714s, J716c, J716s, J775c, J873s, K114c, K114s, K116c, K116s) and T8142's J704, J813 and J815 are
    byte-identical with and without the three commits.
  • The binding allows one more name and applies the M5 Pro hand-off entry to it. A node whose only compatible is
    apple,agx-g17s and an apple,agx-g17c node take the entries they took before; no existing device tree uses the
    new name.
  • No driver change. The one-die description (t6050-max.dtsi) has a cpu-map and no cache nodes and is left as
    it is. dj, J714s may show the same thing (one L2 "shared" by every CPU in
    /sys/devices/system/cpu/cpu0/cache/index2/shared_cpu_list); if it does, the same L2 nodes would be a small
    change there. satchlj's open arm64: dts: apple: port the boot-tested J775c compatibility layout #1619 gives the J775c (Mac Studio, M5 Max) per-cluster L2 nodes in its own files;
    the node names here follow t6002.dtsi and t6022.dtsi, and can follow his instead if reviewers prefer one
    scheme.

Depends on

  • Nothing to build or boot: a loader that does not know the Ultra's GPU firmware build leaves the GPU off, as
    today, and the cluster commit needs nothing.
  • For the GPU to come up: the loader pull request that admits the Ultra's build in the T6050 GPU handoff
    (aura, "kboot: T6050: hand the M5 Ultra's GPU to Linux instead of refusing its firmware build").

Overlaps

dj's draft #2102 (J775c support) touches the same binding and t6050-j775d.dts; it merges with this branch without
a conflict.

Validation

Rebased onto integration/aurora at 59b02d5 (from 1f38ff8, and that from fc05050) without a conflict;
the three diffs are unchanged but for the binding's context (the new apple,agx-t8142 entry). On 59b02d5:

  • dtschema 2026.9: dt_binding_check DT_SCHEMA_FILES=gpu/apple,agx.yaml is clean on the head and with this pull
    request. dtbs_check on all eleven T6050 boards reports the same 18 distinct lines with and without it (all
    already on the head), none new and none on J775D, whose GPU node also validates against the binding with its
    status set to "okay".

Earlier, on integration/aurora at 7c71d6d (the merge of #1901):

  • Each commit builds with integration's own configuration, in order, with 0 errors and no dtc warning; the full
    build has the same warnings with and without the series (this pull request and that day's other M5 Ultra
    commits, among them a compute-timestamp change that is no longer in this set).
  • scripts/checkpatch.pl --strict: clean on the three, but for the generic new-file note on commit 3 (MAINTAINERS
    already covers arch/arm64/boot/dts/apple/).
  • tools/asahi/run-host-tests.py on a Linux host, on integration/aurora at fc05050: 321 of 321 runs pass, on
    the head and on the head plus this pull request and the other linux pull request below; no run differs.
  • dtschema 2026.9: dt_binding_check DT_SCHEMA_FILES=gpu/apple,agx.yaml is clean on the head and with this pull
    request. dtbs_check on all eleven T6050 boards reports the same 55 lines with and without it (all already on
    the head: the display mux lists of J716c and J714s, J716c's operating points, J714s's DART reg-names, the PCIe
    port0 nodes), none on J775D. The J775D GPU node also validates against the binding with its status set to
    "okay", as the boot loader leaves it.

Hardware (Mac Studio, M5 Ultra, J775D)

integration/aurora at 7c71d6d with this pull request:

  • Through aura main plus the loader pull request above, the kernel uploaded over USB (that boot's kernel also
    carried a compute-timestamp change that is no longer in this set): the loader reports "T6050 G17D 26A428
    enabled", the driver probes 2300000000.gpu with both firmware roles' initdata acknowledged, and a render node
    appears, where integration's head has none. The GPU's compute and chained-submit checks, 46 concurrent clients and
    a 628-case render CTS subset (571 pass, 57 not supported) pass, with no GPU recovery, fault or timeout in the
    kernel log. The same in each of five further boots through a loader that also hands over the display, four of
    them without that change.
  • With and without the cluster commit, two boots each, three passes of each row a boot:
    • cpu_capacity reads 754 on the twenty M cores and 1024 on the ten P cores (1024 on all thirty without);
      cluster_cpus_list reads six lists of five (thirty lists of one without); each CPU's L2 reads 8 MiB or 16 MiB
      shared by its own five (no size, shared by all thirty, without).
    • A light compute submit with 8 in flight takes 16.8-17.6 us against 20.1-23.5, and with 64 in flight 13.7-14.5
      us against 16.1-24.0. A 720p fill with 4 in flight takes 43.3-45.4 us against 50.5-70.3. By the median of the six
      passes these are 1.31x, 1.52x and 1.40x.
    • A single-triangle render with 8 in flight takes 31.9-32.5 us against 30.9-37.6, and with 32 in flight
      29.1-29.9 us against 28.2-38.0: steadier rather than faster.
    • Single submits and the heavy compute rows are unchanged; a semaphore hop between two queues takes 42.3-46.5 us
      against 45.0-49.1.
  • Rebased onto integration/aurora at fc05050 (the merge of drm/asahi: g17: let a refused render take back idle render pools #1948), with only the other linux pull request
    below, through aura main plus the pack's aura pull requests, one boot: the driver probes the GPU with both
    firmware roles' initdata acknowledged and a render node appears; the compute and chained-submit checks, 46
    concurrent clients and the 628-case subset (571 pass, 57 not supported, two runs) pass, with no GPU recovery,
    fault or timeout in the kernel log; all 30 CPUs are up; the desktop runs on the GPU on HDMI and DisplayPort.
  • On integration/aurora at 1f38ff8 with this pull request and sithglan's 27.2 dfb_ready_query commit (merged
    since, in drm/apple: iomfb: firmware 27.2: the AP link's release is A550 #2063), through aura main at 03d120dc plus the pack's aura pull requests:

Not tested

  • Any board other than J775D (none includes the changed files).
  • Suspend, and a long soak.

Part of the M5 Ultra platform pack

Each pull request below stands on its own; together they bring the Mac Studio (M5 Ultra, J775D) up from public
branches to the desktop.

…2s, for up to 1.5x faster and steadier light GPU submits on M5 Ultra

The M5 Ultra's CPU nodes have no cpu-map and name no cache, so Linux
sees thirty CPUs in one package with no clusters, and one L2 shared by
all thirty. Two things follow on the J775D (Mac Studio):

- parse_dt_topology() reads capacity-dmips-mhz only while it walks
  cpu-map, so the M and P core capacities that t6050-base.dtsi gives
  every T6050 CPU (dj's aurora-silicon#1399) never reach the scheduler: every CPU
  reports cpu_capacity 1024.
- With no last-level cache to group the CPUs by, the scheduler's
  cache-sharing domain is the whole package: a wakeup may pick any idle
  CPU of the thirty, on either die.

Describe the six clusters the way the one-die description does, commit
68c10dd ("arm64: dts: apple: t6050: Describe the CPU clusters"):
MPIDR Aff1 0, 1 and 2 are die 0's M0, M1 and P clusters and 0x10, 0x11
and 0x12 are die 1's, five cores each, as the ADT's cluster-id and
die-id give them. The clusters are listed flat, cluster0-2 on die 0 and
cluster3-5 on die 1, as the M1 and M2 Ultra descriptions (t6002.dtsi,
t6022.dtsi) list theirs; the topology parser does not support clusters
of clusters. The boot loader prunes the entries of cores it did not
start. The capacities in t6050-base.dtsi are kept as they are.

Add one L2 cache node per cluster and point each CPU's next-level-cache
at its cluster's: the ADT gives each CPU an l2-cache-size of 8 MiB in
the M clusters and 16 MiB in the P clusters. It gives no L1 sizes and no
L2 geometry, so the nodes carry the level and the size only. They sit in
/cpus as l2-cache-N, as in t6002.dtsi and t6022.dtsi.

Both are in one commit because the gain needs both. With the cpu-map
alone the clusters and capacities appear, but every CPU still reports
its L2 as shared by all thirty, and the rows below read as without it
(18.4-23.0 us at depth 8, on integration's head of 2026-11-10).

Only t6050-j775d.dts includes t6050-ultra.dtsi, so the J775D is the only
board whose device tree changes: the ten other T6050 board device trees
(J714c, J714s, J716c, J716s, J775c, J873s, K114c, K114s, K116c, K116s)
are byte-identical with and without the series. The one-die description
(t6050-max.dtsi) has a cpu-map and no cache nodes; this commit leaves
it unchanged.

Testing, on J775D, integration/aurora's head of 2026-11-14 (the merge
of aurora-silicon#1901) with the J775D GPU described (its own pull request), two boots
with and two without this commit, three passes of each row a boot:
- cpu_capacity reads 754 on the twenty M cores and 1024 on the ten P
  cores (1024 on all thirty without); cluster_cpus_list reads six lists
  of five (thirty lists of one without); each CPU's L2 reads 8 MiB or 16
  MiB shared by its own five (no size, shared by all thirty, without).
- A light compute submit with 8 in flight takes 16.8-17.6 us against
  20.1-23.5, and with 64 in flight 13.7-14.5 us against 16.1-24.0. A
  720p fill with 4 in flight takes 43.3-45.4 us against 50.5-70.3. By
  the median of the six passes these are 1.31x, 1.52x and 1.40x. A
  single-triangle render with 8 in flight takes 31.9-32.5 us against
  30.9-37.6, and with 32 in flight 29.1-29.9 us against 28.2-38.0:
  steadier rather than faster. Single submits and the heavy compute rows
  are unchanged, and a semaphore hop between two queues takes 42.3-46.5
  us against 45.0-49.1.
- The GPU's compute and chained-submit checks, 46 concurrent clients and
  a 628-case render CTS subset (571 pass, 57 not supported) read the
  same either way, with no GPU recovery, fault or timeout in the kernel
  log.

Signed-off-by: Jeff Queisser <funkzy@users.noreply.github.com>
…d), so the Mac Studio's GPU can be described

The M5 Ultra has the T6050's GPU on two dies (G17D). The driver matches
the family name "apple,agx-g17s" and tells the M5 parts apart by the
GPU's ID registers, so the Ultra needs no entry of its own there, but
its device tree should say which part it is.

Allow "apple,agx-g17d" with "apple,agx-g17s" as its fallback, beside the
M5 Max's "apple,agx-g17c" that ewb's commit 4c7fdf1 ("dt-bindings:
gpu: apple: Describe the J716c GPU and panel handoff") added: his entry
becomes a list of two names, the form the G13 and G14 parts already use
(apple,agx-g13c and apple,agx-g13d, apple,agx-g14c and apple,agx-g14d).

The M5 Ultra's GPU node has the M5 Pro's hand-off shape: one "sgx"
range, the two firmware mailboxes and the six reserved regions, and no
operating-point tables. Since sithglan's commit 8f418a4
("dt-bindings: gpu: apple,agx: Tell the two G17S layouts apart") that
entry applies to a node whose only compatible is "apple,agx-g17s", so
apply it to a node with "apple,agx-g17d" as well. A node with only
"apple,agx-g17s" and a "apple,agx-g17c" node take the entries they took
before, and no existing device tree uses the new name.

Signed-off-by: Jeff Queisser <funkzy@users.noreply.github.com>
…ing GPU: describe it for the boot loader's handoff

The GPU driver has the M5 Ultra's configuration, but the J775D device
tree leaves the GPU and both GPU coprocessor nodes of the inventory
disabled, with no mailboxes and no reservations, so the Mac Studio boots
without a render node and its desktop session cannot start.

Add t6050-j775d-gpu.dtsi in the shape of dj's
t6050-j714s-gpu-panel.dtsi, commit 32a31cf ("arm64: dts: apple:
t6050-j714s: Describe the GPU and panel handoff"): the "gpu" alias, the
six reservations the boot loader fills, the two firmware mailboxes in
place of the inventory's coprocessor nodes, GPX and AFR held on, and the
GPU node's handoff properties. The J775D ADT gives the same addresses
and interrupts as J714s's: both firmware roles and their mailboxes are
on die 0. Two things differ from the J714s file:
- no panel or display part: the Studio's display chain is described
  already;
- the part's name, "apple,agx-g17d", with the "apple,agx-g17s" fallback
  the driver and the boot loader match.

The GPU part now reads much the same in four board files: the J714s,
J716c and K114c GPU files and this one. It could be one shared file
that they include; this leaves the others as they are.

Everything stays disabled in the source: the boot loader enables the
nodes after checking the ADT and the GPU firmware build. A loader that
does not know the Ultra's build leaves the GPU off, as today. Only
t6050-j775d.dts includes the new file.

Testing, on J775D, integration/aurora's head of 2026-11-14 (the merge
of aurora-silicon#1901) with this series, in six boots: one through aura main with the
loader's own pull request ("kboot: T6050: hand the M5 Ultra's GPU to
Linux instead of refusing its firmware build"), the kernel uploaded over
USB, and five through a boot loader that also hands over the display. In
each the loader reports the GPU enabled, the driver probes
2300000000.gpu with both firmware roles' initdata acknowledged, and a
render node appears, where integration's head has none; the compute and
chained-submit checks pass, 46 concurrent clients pass, and a 628-case
render CTS subset reads 571 pass and 57 not supported, with no GPU
recovery, fault or timeout in the kernel log. The ten other T6050 board
device trees (J714c, J714s, J716c, J716s, J775c, J873s, K114c, K114s,
K116c, K116s) and T8142's J704, J813 and J815 are byte-identical with
and without the series.

Signed-off-by: Jeff Queisser <funkzy@users.noreply.github.com>
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