Skip to content

asahi, hk: G15C (M3 Max) on the native AGX3 path - #25

Merged
DjDeveloperr merged 4 commits into
aurora-silicon:j514s/g15s-20261017from
malik-na:g15c-native
Oct 18, 2026
Merged

DjDeveloperr merged 4 commits into
aurora-silicon:j514s/g15s-20261017from
malik-na:g15c-native

Conversation

@malik-na

Copy link
Copy Markdown
Member

Admits the M3 Max GPU (G15C, chip 0x6031, 1 die, 4 clusters x 10 cores) on the native path from #23, with the G15S
library programs. G15S is unchanged: under the drm-shim, all 215 compute and 3 render submissions are byte-identical to
#23.

Stacked on #23.

Commits

  • asahi/drm-shim: publish the platform bus link: without it, this tree's drm-shim found no Vulkan device for any chip
    and GL fell back to softpipe.
  • asahi/drm-shim: add a G15C (T6031) identity.
  • asahi,hk: admit G15C on the native path: the native-path predicates, the driver and parameter admission,
    libagx_g15s, wide-visibility barrier grounding, and tests.
  • asahi: return every out-of-table USC helper call on G15C: the profile prefix's computed helper call can index past
    the base page's two ten-entry helper tables for some G15C cores. Those calls ran on into the unmapped page after the
    base page and skipped the uniform loads, so half of every dispatch and draw went wrong. Every slot after the tables
    now holds a return. G15C only.

Tested

drm-shim: 14/14 smoke checks for G15C, plus the G15S byte-identity check above. Unit tests: 33 pass, and 1
(hk_g17p_compression) fails the same way on #23.

On the J516C (M3 Max, 40 cores), with aurora-silicon/linux #358 plus the T6031 row (PR aurora-silicon/linux#363):

check runs in 5 s wrong
compute-check 10750 0
vk-render-check 8309 0 pixels
gl-render-check (native GL) 10083 0 pixels
gl-render-check (zink) 6418 0 pixels

The helper-call fix was first measured on the legacy kit path. There, compute-check went from 4096 of 65536 words
wrong to 0, and the render checks from 2048 of 4096 pixels wrong to 0.

Notes

drm_shim now learns the bus of the fake device from a subsystem link
override, which drivers set up through drm_shim_override_link() or
drm_shim_platform_device_setup(). The asahi shim sets the bus type and
the uevent file but no link, so libdrm's drmGetDevices2() cannot classify
the node and returns no device: vkEnumeratePhysicalDevices() fails with
VK_ERROR_INITIALIZATION_FAILED for every ASAHI_SHIM_CHIP, and the EGL
loader falls back to softpipe. Tests that open the node directly were not
affected.

Add the platform subsystem link.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
ASAHI_SHIM_CHIP=g15c presents the M3 Max (T6031, G15C) as an M3 kernel
would: chip 0x6031, revision B2, one die, four clusters of ten cores with
every core present (masks 0x3ff), AGX3 shaders on the legacy HAL, the
G15S user VA range and a 1380 MHz top clock. The feature word is empty;
the ASAHI_SHIM_* switches can add bits on top as for any identity.

The G15 kernel's bind refusal for [0x7000000000, 0x7100000000) and its
submission gate now apply to both G15 identities. The gate checks prior
barriers against the feature word the shim reports instead of the G15S
constant, which is the same word for G15S.

The driver does not admit G15C yet, so a g15c run is still refused at
device open. G15S behaves as before.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
Put the M3 Max (T6031, G15C) on the native AGX3 path beside G15S. It is
the same G15 generation with four clusters instead of two: Gen9
descriptors in a 32-byte sampler heap, the G15 command stream, the Gen4
sampler, the classic queue model and the AGX3B encoding.

- agx_chip_native_agx3(), agx_chip_native_imageblock() and
  agx_chip_native_indirect_compute() name G15C, so it takes every G15S
  choice in Honeykrisp and Gallium.
- agx_chip_driver_supported() checks G15C's axes as it does G15S's, and
  agx_params_driver_supported() requires the exact T6031 parameters:
  chip 0x6031, one die, four clusters, AGX3 shaders on the legacy HAL.
  The core masks are not checked, so binned parts are admitted. Both
  land together, since an unlisted chip would pass the parameter check.
- The device loads the G15S build of libagx: G15C executes the same
  encoding and needs no scheduling stalls.
- The CDM barrier packer accepts the wide-visibility form for G15C, which
  agx_chip_cdm_wide_visibility() asks for on every native G15 part.
- A refused G15C names the parameters it expected.

Ray-query encodings stay refused on G15C. A new admission test covers
both G15 identities and the parameters that must be refused, and the
CDM barrier test covers the wide form.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.com>
On the M3 Max (G15C), half of each compute dispatch and of each draw ran
without its uniform state: compute-check got 4096 of 65536 words wrong,
and the vk and gl render checks got 2048 of 4096 pixels wrong. Without
kernel-side backing for it, the shader cores also faulted reading the
page after the USC window base page.

The profile prefix ends in a computed helper call. On G15C that call can
index past the base page's two ten-entry helper tables for some shader
cores, and the slots after the tables were zero, so the call did not
return. It ran on through the rest of the page and into the following
unbound guard page, and the uniform state it was meant to load was
skipped. On the native path every G15 program carries that prefix.

Build the base page with agx3_build_usc_base_page(), which on G15C fills
every 16-byte slot from +0x2c0 to +0x340 and from +0x350 to the end of
the page with the helper table's terminal record, so any such call
returns inside the page. The zero block-size dword at +0x340 that ends
the page's container chain is left alone. Only G15C is changed; G15S,
G17P and every per-shader container get the same bytes as before. A host
test checks the page for G15C, G15S and G17P.

The same fill was tested on J516C (M3 Max, 40 cores) with the earlier
G15C driver, where compute-check, vk-render-check and gl-render-check
passed (0 wrong over ~2700-5700 runs each) with no kernel-side page
backing.

Signed-off-by: Naeem Malik <malik2015naeem@gmail.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.

2 participants