The K1’s GPU Finally Has a Device Tree Node: SpacemiT PowerVR Bring-Up Reaches Mainline Review
Every SpacemiT K1 board on the market ships a working GPU — and almost none of them runs it through mainline Linux. The K1 carries an Imagination PowerVR Rogue BXE-2-32 (BVNC 36.29.52.182), and until now the only way to touch it was the vendor BSP with Imagination’s out-of-tree pvrsrvkm module. On 18 November 2026 Iker Pedrosa posted version 2 of a patch that changes the shape of the problem: an 11-line addition to arch/riscv/boot/dts/spacemit/k1.dtsi that describes the GPU to mainline Linux for the first time — address, interrupt, clock, reset and power domain.
It is tempting to read an 11-line device tree patch as paperwork. It is not. For this particular block it is the last piece of a chain that three different patch series had to line up before a K1 could print /dev/dri/card0 at all.
What the patch actually adds
The node lands inside the multimedia-bus node, which matters for reasons covered below. Everything the GPU needs at boot is declared there:
gpu: gpu@cac00000 {
compatible = "spacemit,k1-gpu", "img,img-bxe-2-32", "img,img-rogue";
reg = <0x0 0xcac00000 0x0 0x80000>;
interrupts = <75>;
clocks = <&syscon_apmu CLK_GPU>;
clock-names = "core";
resets = <&syscon_apmu RESET_GPU>;
power-domains = <&syscon_apmu K1_PMDOMAIN_GPU>;
};
| Property | Value | What it unlocks |
|---|---|---|
| compatible | spacemit,k1-gpu, img,img-bxe-2-32, img,img-rogue | Binds the mainline Imagination DRM driver to a K1-specific entry |
| reg | 0xcac00000, 512 KiB (0x80000) | MMIO register window for the Rogue core |
| interrupts | 75 | GPU IRQ line; also carries firmware watchdog events |
| clocks / clock-names | CLK_GPU as "core" via syscon_apmu | Single core clock, gated through the CCU |
| resets | RESET_GPU | Deasserts the GPU reset line during probe |
| power-domains | K1_PMDOMAIN_GPU | Powers the GPU island on through the APMU genpd driver |
| status | none (enabled) | Enabled for every K1 board, not one vendor’s |
Version 1 of the series (15 October) shipped the node disabled and let each board enable it, with a two-patch split. Yixun Lan reviewed it on 11 November and pushed back on exactly that point: “As GPU is integrated within SoC, I’d rather see it enabled by default, so will not only one specific board benefit from this patch, besides, I see no reason why people should disable it by default?” He also asked for the RFC prefix to be dropped and the commit message trimmed. Version 2 is that feedback applied: one patch, no board enablement, no status = "disabled".
A DT node is one layer of four
The single most misleading sentence you can write about this patch is “the K1 now has mainline GPU support.” What exists, and what does not, splits cleanly into four layers:
| Layer | Component | Status after v2 |
|---|---|---|
| Hardware description | spacemit,k1-gpu node in k1.dtsi | This patch — under review on linux-riscv / devicetree |
| DT binding | Documentation/devicetree/bindings/gpu/img,powervr-rogue.yaml | Separate patch by Baihui Liang (SpacemiT), 15 Oct 2026 — adds spacemit,k1-gpu and spacemit,k3-gpu |
| Power sequencing | APMU genpd driver, Yixun Lan’s K3 PM series | Required for probe; still pending mainline at the time of posting |
| Kernel driver | drivers/gpu/drm/imagination | Already mainline (merged in Linux 6.8), RISC-V SoCs now describable |
| Firmware | rogue_36.29.52.182_v1.fw | Not upstream — only in Imagination’s powervr branch of imagination/linux-firmware |
| Userspace | Mesa PowerVR (Vulkan-only) | BXE-2-32 is not in Mesa’s supported set |
That table is the whole story. The kernel side — the part that has historically been the blocker on RISC-V GPUs — is converging. The userspace side is not, and no device tree patch will change it.
The bring-up path, verified end to end
The cover letter is unusually concrete about what “works” means here, and it is a checklist rather than a benchmark:
- Power domain — the GPU power island switches on and off through the APMU
- Clock — the GPU clock is enabled and gated through the CCU
- Reset — the GPU reset line is deasserted
- Firmware — the PowerVR RISC-V firmware loads and signals ready
- DMA —
dma-rangescorrectly maps CPU-to-GPU physical addresses - DRM — the kernel exposes
/dev/dri/card0and/dev/dri/renderD128 - IRQ — the GPU interrupt is registered and the firmware watchdog is running
Note the fourth item: the PowerVR RISC-V firmware. The Rogue core on the K1 is not driven by host software alone; it runs its own firmware image, and that image is exactly what is missing from linux-firmware. The patch’s own reference points at a GitLab blob, which means anyone reproducing this has to fetch firmware out of tree.
multimedia-bus because that bus node defines the address translation the GPU needs: dma-ranges = <0x0 0x00000000 0x0 0x00000000 0x0 0x80000000>, <0x0 0x80000000 0x1 0x00000000 0x3 0x80000000>. Moving the GPU node to the root of the tree — a tempting “cleanup” — would leave the device describing a physical address space it cannot actually reach. This is why the patch is one file with eleven lines and not a drive-by addition.
The part that will still disappoint you
With all four series applied, the kernel is happy. Userspace is not. Mesa’s PowerVR driver is Vulkan-only — there is no OpenGL driver for these parts — and the K1’s BXE-2-32 (36.29.52.182) sits in Mesa’s “device info and firmware exist, but no support is guaranteed” bucket, not in the actively developed set. The actively developed parts are the AXE-1-16M, the BXM-4-64 and the BXS-4-64. On a not-yet-conformant GPU, Mesa refuses to run unless PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1 is exported — a variable name that is itself the most honest documentation in the stack.
The patch says the same thing in one sentence: “Userspace rendering requires Mesa’s PowerVR Vulkan driver (imagination-experimental).” So the realistic outcome on a K1 today is a working DRM device, a loaded firmware, a running watchdog — and a Vulkan driver you have to force-start, with simple applications expected to run and instability or corruption a real possibility.
| Claim | Reality check |
|---|---|
| “K1 GPU is mainline now” | Device tree node + binding are in review; APMU power domains still pending; firmware still out of tree |
| “Vulkan works” | Mesa lists BXE-2-32 as unsupported-but-present; forcing the driver is required |
| “OpenGL works” | Mesa has no PowerVR OpenGL driver at all; the vendor DDK is the only GL path |
| “Any performance numbers?” | None. The cover letter contains zero benchmark results — by design, it describes bring-up, not throughput |
| “K3 is the same situation” | Not quite: the K3’s BXM-4-64 is in Mesa’s actively developed set, so its userspace odds are better |
Why the K3 half of the binding matters more
The same binding patch that adds spacemit,k1-gpu also adds spacemit,k3-gpu for the K3’s IMG BXM-4-64, and reuses the single-power-domain rule already written for the T-Head TH1520 — the K3 entry is literally grouped with thead,th1520-gpu because they need the same configuration. Because BXM-4-64 is one of the GPUs Mesa actually maintains (Vulkan 1.2 conformant per Imagination’s published data), the K3 stands a much better chance of going from “device node exists” to “you can render on it” without a vendor DDK in the middle.
For anyone shipping K1-based products — gateways, edge boxes, handhelds — the practical read is unchanged: the vendor BSP remains the production path for GPU work, and mainline remains the path for everything else. What changed this week is that the mainline path now has a place to put the GPU, which is the precondition for every fix that comes after it.
Reproducing it
# three dependencies, applied in order
# 1) Yixun Lan: spacemit K3 PM / APMU power domains (10 Oct 2026)
# 2) Baihui Liang: dt-bindings img,powervr-rogue + spacemit,k1-gpu|k3-gpu (15 Oct 2026)
# 3) Iker Pedrosa: k1.dtsi GPU node, v2 (18 Nov 2026)
make ARCH=riscv defconfig # plus your board dts
make ARCH=riscv -j$(nproc)
# firmware is NOT in linux-firmware yet - fetch from Imagination's powervr branch
# rogue_36.29.52.182_v1.fw -> /lib/firmware/
# verify the bring-up checklist
dmesg | grep -iE 'powervr|rogue|imagination'
ls -l /dev/dri/ # expect card0 and renderD128
# userspace: Mesa PowerVR is Vulkan-only, and BXE-2-32 is unsupported
export PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1
vulkaninfo --summary
rogue_36.29.52.182_v1.fw reaches linux-firmware, and whether Mesa promotes BXE-2-32 out of the unsupported bucket. Those three events, not this patch, decide when a stock distro image boots a K1 with a usable GPU.
Sources
- [PATCH v2] riscv: dts: spacemit: k1: add PowerVR GPU device node — Iker Pedrosa, 18 Nov 2026 (full diff and cover letter) — lore.kernel.org/all/20261118-orangepi-gpu-v2-1-f68676f844eb@gmail.com
- [PATCH v1] same series, 15 Oct 2026 (board-specific enablement, two patches) — lore.kernel.org/r/20261015-orangepi-gpu-v1-0-03566a615660@gmail.com
- Review by Yixun Lan, 11 Nov 2026 (drop RFC, enable by default) — marc.info/?m=179435784658986
- [PATCH] dt-bindings: gpu: img,powervr-rogue: add spacemit,k1-gpu and spacemit,k3-gpu — Baihui Liang, 15 Oct 2026 — lore.kernel.org/linux-riscv/20261015-spacemit-gpu-binding-v1-1-a7a23c7e7788@linux.spacemit.com
- Patchwork series 1215361 (K1 GPU node) and 1214827 (JHB100 CRG) on linux-riscv — patchwork.kernel.org/project/linux-riscv/list/
- PowerVR GPU support matrix (mainline
drm/imaginationsince Linux 6.8; Mesa PowerVR per-BVNC status) — sbcwiki.com/docs/ip-vendors/imagination/powervr-gpu - Firmware blob referenced by the patch — gitlab.freedesktop.org/imagination/linux-firmware (powervr branch)
2026 年 11 月 18 日,Iker Pedrosa 提交了 v2 版补丁:在 arch/riscv/boot/dts/spacemit/k1.dtsi 里加入 11 行,第一次向主线 Linux 描述 K1 的 GPU — Imagination PowerVR Rogue BXE-2-32(BVNC 36.29.52.182)。节点 gpu@cac00000 位于 multimedia-bus 之下,寄存器窗 512 KiB,中断 75,时钟 CLK_GPU(core)、复位 RESET_GPU、电源域 K1_PMDOMAIN_GPU,且默认为启用状态。
v1(10 月 15 日)原本是「默认关闭、由各板子自行启用」的两补丁形式,维护者 Yixun Lan 在 11 月 11 日评审时明确反对:GPU 集成在 SoC 内,没有理由默认关闭,否则只有一块板子受益。v2 就是这条意见的落实 — 单补丁、无板级开关、去掉 status = "disabled"。
必须分清层次:补丁描述的是硬件,不是「K1 现在有主线 GPU 支持」。完整链条为 ①本补丁(DT 节点,评审中)②Baihui Liang 的 img,powervr-rogue.yaml 绑定(10-15,新增 spacemit,k1-gpu 与 spacemit,k3-gpu)③Yixun Lan 的 APMU 电源域系列(未进入主线则节点无法 probe)④内核驱动 drivers/gpu/drm/imagination(Linux 6.8 起已在主线)⑤固件 rogue_36.29.52.182_v1.fw(不在 linux-firmware,只在 Imagination 的 powervr 分支)⑥用户态 Mesa PowerVR(仅 Vulkan,且 BXE-2-32 不在支持列表)。
已验证的 bring-up 路径:电源岛经 APMU 上下电、GPU 时钟经 CCU 开关、复位解除、RISC-V 固件加载并报 ready、dma-ranges 完成 CPU→GPU 物理地址映射、内核暴露 /dev/dri/card0 与 renderD128、GPU 中断注册且固件看门狗运行。封面信里没有任何性能数字 — 它描述的是点亮,不是吞吐。
诚实点:Mesa 的 PowerVR 驱动只有 Vulkan,没有 OpenGL;BXE-2-32 属于「有 device info 与固件、但不保证支持」一档,需要 PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1 才能强行启动;K3 的 BXM-4-64 才是 Mesa 活跃维护的型号,因此 K3 的用户态前景明显更好(同一绑定补丁把 K3 与 T-Head TH1520 归为同一组,共用单电源域规则)。对量产项目而言,GPU 相关仍应走厂商 BSP;主线这条路的进展是「终于有了挂载点」,而不是「可以交付了」。
18 ноября 2026 года Iker Pedrosa отправил v2 патча, добавляющего в arch/riscv/boot/dts/spacemit/k1.dtsi 11 строк, которые впервые описывают GPU SoC SpacemiT K1 для mainline Linux: Imagination PowerVR Rogue BXE-2-32 (BVNC 36.29.52.182). Узел gpu@cac00000 находится внутри multimedia-bus: окно регистров 512 КиБ, прерывание 75, такт CLK_GPU («core»), сброс RESET_GPU, домен питания K1_PMDOMAIN_GPU — и включён по умолчанию.
В v1 (15 октября) узел был выключен и включался на каждой плате отдельно; ревьюер Yixun Lan 11 ноября возразил: GPU интегрирован в SoC, нет причин держать его выключенным, иначе выигрывает лишь одна плата. v2 — это учтённая правка: один патч, без платформенного включения, без status = "disabled".
Важно разделять слои. Патч описывает железо, а не «теперь у K1 есть GPU в mainline». Цепочка: (1) сам DT-узел — на ревью; (2) биндинг img,powervr-rogue.yaml от Baihui Liang (15.10, добавляет spacemit,k1-gpu и spacemit,k3-gpu); (3) серия APMU power domain от Yixun Lan — без неё узел не пройдёт probe; (4) драйвер drivers/gpu/drm/imagination — уже в mainline с Linux 6.8; (5) прошивка rogue_36.29.52.182_v1.fw — не в linux-firmware, только в ветке powervr у Imagination; (6) пользовательский стек Mesa PowerVR — только Vulkan, и BXE-2-32 не входит в поддерживаемый набор.
Проверенный путь bring-up: включение острова питания через APMU, тактирование через CCU, снятие сброса, загрузка RISC-V прошивки и сигнал ready, корректная трансляция адресов CPU→GPU через dma-ranges, появление /dev/dri/card0 и renderD128, регистрация прерывания и работа watchdog прошивки. В сопроводительном письме нет ни одного числа производительности — речь о запуске, не о пропускной способности.
Честно: в Mesa для PowerVR есть только Vulkan (OpenGL-драйвера нет вовсе); BXE-2-32 попадает в категорию «device info и прошивка есть, поддержка не гарантируется», поэтому нужен PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1; активно поддерживаются AXE-1-16M, BXM-4-64 и BXS-4-64 — то есть у K3 (BXM-4-64) шансы на пользовательский стек заметно выше, и тот же биндинг объединяет K3 с T-Head TH1520 по правилу единственного домена питания. Для серийных изделий GPU по-прежнему остаётся на стороне вендорского BSP: в mainline появилась точка крепления, а не готовый к поставке стек.
El 18 de noviembre de 2026 Iker Pedrosa envió la v2 de un parche que añade 11 líneas a arch/riscv/boot/dts/spacemit/k1.dtsi para describir por primera vez la GPU del SpacemiT K1 en el Linux mainline: una Imagination PowerVR Rogue BXE-2-32 (BVNC 36.29.52.182). El nodo gpu@cac00000 vive dentro de multimedia-bus: ventana de registros de 512 KiB, interrupción 75, reloj CLK_GPU («core»), reset RESET_GPU y dominio de energía K1_PMDOMAIN_GPU, habilitado por defecto.
La v1 (15 de octubre) lo dejaba deshabilitado y obligaba a cada placa a activarlo; el revisor Yixun Lan lo rechazó el 11 de noviembre: la GPU está integrada en el SoC, no hay motivo para desactivarla por defecto si eso solo beneficia a una placa. La v2 aplica ese criterio: un solo parche, sin activación por placa y sin status = "disabled".
Hay que separar capas. El parche describe hardware, no significa «el K1 ya tiene GPU en mainline». La cadena es: (1) este nodo DT, en revisión; (2) el binding img,powervr-rogue.yaml de Baihui Liang (15-10), que añade spacemit,k1-gpu y spacemit,k3-gpu; (3) la serie de dominios de energía APMU de Yixun Lan, sin la cual el nodo no hace probe; (4) el driver drivers/gpu/drm/imagination, ya en mainline desde Linux 6.8; (5) el firmware rogue_36.29.52.182_v1.fw, que no está en linux-firmware; (6) el stack de usuario Mesa PowerVR, solo Vulkan y con BXE-2-32 fuera del conjunto soportado.
Ruta de bring-up verificada: encendido de la isla de energía vía APMU, reloj por CCU, liberación de reset, carga del firmware RISC-V con señal de listo, traducción de direcciones CPU→GPU mediante dma-ranges, aparición de /dev/dri/card0 y renderD128, interrupción registrada y watchdog del firmware activo. La carta de presentación no incluye ni una cifra de rendimiento: describe arranque, no throughput.
Honestidad: Mesa no tiene driver OpenGL para PowerVR; BXE-2-32 figura como «hay device info y firmware, pero sin garantía de soporte», así que hace falta PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1. Los modelos en desarrollo activo son AXE-1-16M, BXM-4-64 y BXS-4-64, de modo que el K3 (BXM-4-64) tiene mejores perspectivas en espacio de usuario — el mismo binding agrupa al K3 con el T-Head TH1520 bajo la regla de un único dominio de energía. Para producto en volumen, la GPU sigue dependiendo del BSP del fabricante: en mainline ya existe el punto de anclaje, no un stack entregable.
Le 18 novembre 2026, Iker Pedrosa a envoyé la v2 d'un correctif ajoutant 11 lignes à arch/riscv/boot/dts/spacemit/k1.dtsi pour décrire pour la première fois le GPU du SpacemiT K1 dans le noyau mainline : un Imagination PowerVR Rogue BXE-2-32 (BVNC 36.29.52.182). Le nœud gpu@cac00000 se trouve dans multimedia-bus : fenêtre de registres de 512 Kio, interruption 75, horloge CLK_GPU (« core »), reset RESET_GPU, domaine d'alimentation K1_PMDOMAIN_GPU — et activé par défaut.
La v1 (15 octobre) le laissait désactivé, chaque carte devant l'activer ; le relecteur Yixun Lan s'y est opposé le 11 novembre : le GPU est intégré au SoC, rien ne justifie de le désactiver par défaut si cela ne profite qu'à une carte. La v2 applique ce retour : un seul correctif, sans activation par carte ni status = "disabled".
Il faut séparer les couches. Le correctif décrit du matériel, il ne signifie pas « le K1 a désormais un GPU mainline ». La chaîne : (1) ce nœud DT, en relecture ; (2) le binding img,powervr-rogue.yaml de Baihui Liang (15-10), qui ajoute spacemit,k1-gpu et spacemit,k3-gpu ; (3) la série de domaines d'alimentation APMU de Yixun Lan, sans laquelle le nœud ne sonde pas ; (4) le pilote drivers/gpu/drm/imagination, déjà mainline depuis Linux 6.8 ; (5) le firmware rogue_36.29.52.182_v1.fw, absent de linux-firmware ; (6) la pile utilisateur Mesa PowerVR, uniquement Vulkan, avec BXE-2-32 hors du périmètre supporté.
Chemin de bring-up vérifié : mise sous tension de l'îlot via APMU, horloge via CCU, libération du reset, chargement du firmware RISC-V avec signal prêt, traduction d'adresses CPU→GPU par dma-ranges, apparition de /dev/dri/card0 et renderD128, interruption enregistrée et watchdog du firmware actif. La lettre d'accompagnement ne contient aucun chiffre de performance : elle décrit l'allumage, pas le débit.
Honnêteté : Mesa n'a pas de pilote OpenGL pour PowerVR ; BXE-2-32 est classé « device info et firmware présents, support non garanti », d'où la nécessité de PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1. Les modèles activement développés sont AXE-1-16M, BXM-4-64 et BXS-4-64 : le K3 (BXM-4-64) a donc de meilleures chances côté espace utilisateur, le même binding le regroupant avec le T-Head TH1520 sous la règle du domaine d'alimentation unique. Pour un produit en volume, le GPU reste du ressort du BSP constructeur : le mainline fournit désormais un point d'ancrage, pas une pile livrable.
Am 18. November 2026 hat Iker Pedrosa v2 eines Patches geschickt, der 11 Zeilen in arch/riscv/boot/dts/spacemit/k1.dtsi ergänzt und damit erstmals die GPU des SpacemiT K1 im Mainline-Linux beschreibt: eine Imagination PowerVR Rogue BXE-2-32 (BVNC 36.29.52.182). Der Knoten gpu@cac00000 liegt im multimedia-bus: 512 KiB Registerfenster, Interrupt 75, Takt CLK_GPU („core“), Reset RESET_GPU, Power-Domain K1_PMDOMAIN_GPU — und standardmäßig aktiviert.
In v1 (15. Oktober) war der Knoten deaktiviert und musste pro Board eingeschaltet werden; Reviewer Yixun Lan widersprach am 11. November: Die GPU ist im SoC integriert, es gibt keinen Grund, sie standardmäßig abzuschalten, wenn dadurch nur ein Board profitiert. v2 setzt das um: ein Patch, keine Board-Aktivierung, kein status = "disabled".
Schichten sauber trennen. Der Patch beschreibt Hardware — nicht „der K1 hat jetzt Mainline-GPU-Support“. Die Kette: (1) dieser DT-Knoten, im Review; (2) das Binding img,powervr-rogue.yaml von Baihui Liang (15.10), das spacemit,k1-gpu und spacemit,k3-gpu ergänzt; (3) Yixun Lans APMU-Power-Domain-Serie, ohne die der Knoten nicht probt; (4) der Treiber drivers/gpu/drm/imagination, seit Linux 6.8 in Mainline; (5) die Firmware rogue_36.29.52.182_v1.fw, nicht in linux-firmware; (6) der Mesa-PowerVR-Userspace-Stack, nur Vulkan, mit BXE-2-32 außerhalb des unterstützten Sets.
Verifizierter Bring-up-Pfad: Power-Insel über APMU, Takt über CCU, Reset freigegeben, RISC-V-Firmware lädt und meldet bereit, CPU→GPU-Adresstranslation über dma-ranges, /dev/dri/card0 und renderD128 erscheinen, Interrupt registriert und Firmware-Watchdog läuft. Das Begleitschreiben enthält keine einzige Performance-Zahl — es beschreibt das Anlaufen, nicht den Durchsatz.
Ehrlichkeit: Mesa hat für PowerVR nur Vulkan, keinen OpenGL-Treiber; BXE-2-32 fällt in die Kategorie „Device-Info und Firmware vorhanden, Support nicht garantiert“ und braucht PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1. Aktiv entwickelt werden AXE-1-16M, BXM-4-64 und BXS-4-64 — der K3 (BXM-4-64) hat also bessere Aussichten im Userspace; dasselbe Binding gruppiert ihn mit dem T-Head TH1520 unter der Regel der einzelnen Power-Domain. Für Serienprodukte bleibt die GPU Sache des Vendor-BSP: Mainline hat jetzt einen Anknüpfungspunkt, keinen lieferbaren Stack.
در ۱۸ نوامبر ۲۰۲۶ Iker Pedrosa نسخهٔ دوم وصلهای را فرستاد که ۱۱ خط به arch/riscv/boot/dts/spacemit/k1.dtsi میافزاید و برای نخستین بار GPU تراشهٔ SpacemiT K1 را در هستهٔ اصلی لینوکس توصیف میکند: یک Imagination PowerVR Rogue BXE-2-32 با BVNC برابر 36.29.52.182. گرهٔ gpu@cac00000 داخل multimedia-bus قرار دارد: پنجرهٔ ثبات ۵۱۲ کیلوبایت، وقفهٔ ۷۵، ساعت CLK_GPU با نام «core»، ریست RESET_GPU و دامنهٔ توان K1_PMDOMAIN_GPU — و بهصورت پیشفرض فعال.
در نسخهٔ اول (۱۵ اکتبر) گره غیرفعال بود و هر برد باید آن را روشن میکرد؛ بازبین Yixun Lan در ۱۱ نوامبر به این نکته ایراد گرفت: GPU درون SoC یکپارچه است و دلیلی ندارد پیشفرض خاموش بماند تا فقط یک برد سود ببرد. نسخهٔ دوم همان بازخورد است: یک وصله، بدون فعالسازی در سطح برد و بدون status = "disabled".
لایهها را باید جدا دید. این وصله سختافزار را توصیف میکند، نه اینکه «K1 اکنون پشتیبانی GPU در هستهٔ اصلی دارد». زنجیره: ۱) همین گرهٔ درخت دستگاه، در حال بازبینی؛ ۲) بایندینگ img,powervr-rogue.yaml از Baihui Liang (۱۵ اکتبر) که spacemit,k1-gpu و spacemit,k3-gpu را میافزاید؛ ۳) مجموعهٔ دامنهٔ توان APMU از Yixun Lan که بدون آن گره probe نمیشود؛ ۴) درایور drivers/gpu/drm/imagination که از لینوکس ۶٫۸ در هستهٔ اصلی است؛ ۵) فریمور rogue_36.29.52.182_v1.fw که در linux-firmware نیست؛ ۶) پشتهٔ کاربری Mesa PowerVR که فقط Vulkan است و BXE-2-32 در فهرست پشتیبانیشده قرار ندارد.
مسیر bring-up آزمودهشده: روشنشدن جزیرهٔ توان از طریق APMU، ساعت از طریق CCU، آزادسازی ریست، بارگذاری فریمور RISC-V و اعلام آمادگی، ترجمهٔ نشانی CPU→GPU با dma-ranges، پیدایش /dev/dri/card0 و renderD128، ثبت وقفه و اجرای watchdog فریمور. در نامهٔ همراه هیچ عدد کارایی وجود ندارد — سخن از روشنشدن است، نه گذردهی.
صداقت: در Mesa برای PowerVR فقط Vulkan وجود دارد و درایور OpenGL اصلاً نیست؛ BXE-2-32 در دستهٔ «اطلاعات دستگاه و فریمور هست، اما پشتیبانی تضمین نمیشود» است و به PVR_I_WANT_A_BROKEN_VULKAN_DRIVER=1 نیاز دارد. مدلهای فعال AXE-1-16M، BXM-4-64 و BXS-4-64 هستند، پس K3 (BXM-4-64) شانس بهتری در فضاي کاربری دارد؛ همان بایندینگ K3 را با T-Head TH1520 در یک گروه با قاعدهٔ دامنهٔ توان واحد قرار میدهد. برای محصول انبوه، GPU همچنان به BSP سازنده وابسته است: در هستهٔ اصلی اکنون نقطهٔ اتصال هست، نه پشتهٔ آمادهٔ تحویل.