DrewAk49

T95 H616 Mali-G31 Panfrost GPU Fix Problem: The Mali-G31 GPU was detected by the Panfrost driver, but GPU workloads would hang and fail with errors such as: gpu sched timeout gpu soft reset timed out gpu hard reset timed out error powering up gpu L2/shader/tiler The issue was not missing Panfrost support. The GPU was not being powered and initialized correctly due to device-tree power configuration. Fixes: 1. Increased GPU voltage The board was running the GPU regulator at 900mV. Changing vdd-gpu-sys to a fixed 950mV stabilized GPU operation. Device tree: regulator-min-microvolt = <950000>; regulator-max-microvolt = <950000>; regulator-name = "vdd-gpu-sys"; 2. Restored GPU power domain A previous workaround removed the GPU power domain: /delete-property/ power-domains; This allowed Panfrost to load but caused unstable GPU power management. Restored: power-domains = <&prcm_ppu 2>; 3. Linked the GPU regulator Added: mali-supply = <&reg_dcdcc>; This allows the Mali GPU driver to correctly control the GPU power supply. 4. Increased CMA memory Added kernel boot parameter: cma=128M The default 32MB CMA memory was increased to 128MB for improved DRM/display buffer reliability. Final working setup: - Kernel: Linux 7.3 - GPU: Mali-G31 Bifrost - Driver: Panfrost - GPU voltage: 950mV - PRCM GPU power domain enabled - CMA: 128MB Validation: GL_VENDOR: Panfrost GL_RENDERER: Mali-G31 (Panfrost) GL_VERSION: OpenGL ES 3.1 Repeated glmark2 tests completed successfully with no GPU scheduler timeouts. Conclusion: The T95 H616 Mali-G31 Panfrost issue was caused by incorrect GPU power sequencing in the device tree, not a lack of driver support. Correcting the regulator voltage, restoring the GPU power domain, linking the regulator, and increasing CMA memory resulted in stable hardware acceleration.

T95 H616 Mali-G31 Panfrost GPU Fix Problem: The Mali-G31 GPU was detected by the Panfrost driver, but GPU workloads would hang and fail with errors such as: gpu sched timeout gpu soft reset timed out gpu hard reset timed out error powering up gpu L2/shader/tiler The issue was not missing Panfrost support. The GPU was not being powered and initialized correctly due to device-tree power configuration. Fixes: 1. Increased GPU voltage The board was running the GPU regulator at 900mV. Changing vdd-gpu-sys to a fixed 950mV stabilized GPU operation. Device tree: regulator-min-microvolt = <950000>; regulator-max-microvolt = <950000>; regulator-name = "vdd-gpu-sys"; 2. Restored GPU power domain A previous workaround removed the GPU power domain: /delete-property/ power-domains; This allowed Panfrost to load but caused unstable GPU power management. Restored: power-domains = <&prcm_ppu 2>; 3. Linked the GPU regulator Added: mali-supply = <&reg_dcdcc>; This allows the Mali GPU driver to correctly control the GPU power supply. 4. Increased CMA memory Added kernel boot parameter: cma=128M The default 32MB CMA memory was increased to 128MB for improved DRM/display buffer reliability. Final working setup: - Kernel: Linux 7.3 - GPU: Mali-G31 Bifrost - Driver: Panfrost - GPU voltage: 950mV - PRCM GPU power domain enabled - CMA: 128MB Validation: GL_VENDOR: Panfrost GL_RENDERER: Mali-G31 (Panfrost) GL_VERSION: OpenGL ES 3.1 Repeated glmark2 tests completed successfully with no GPU scheduler timeouts. Conclusion: The T95 H616 Mali-G31 Panfrost issue was caused by incorrect GPU power sequencing in the device tree, not a lack of driver support. Correcting the regulator voltage, restoring the GPU power domain, linking the regulator, and increasing CMA memory resulted in stable hardware acceleration.

My T95 hardware: * **Model:** T95 Android TV Box * **SoC:** Allwinner H616 * **Board:** H616-T95MAX-V4.0 * **RAM:** 4GB DDR3 * **Internal storage:** 64GB eMMC * **Wi-Fi chip:** AW859A * **PMIC:** AXP305 * **Ethernet:** 100Mbps, using the AC200/AC300 internal PHY * **USB:** 2x USB 2.0 host ports * **Video:** HDMI output * **UART:** 3.3V logic * **Boot method:** Debian from microSD using mainline U-Boot * **Current OS:** Debian 12 Bookworm * **Current test kernel:** Linux 7.3-rc3 * **Stock Android:** still intact on eMMC Important detail: this is the **V4.0 board revision**, and some upstream T95 device-tree assumptions were based on other revisions. That mattered for things like the second USB port. T95 Box T95 H616 Linux issues and fixes so far: * **Only 2GB RAM showing** * Fixed with a U-Boot DRAM-size patch. Box now sees the full **4GB**. * **SD boot problems / wrong card detect** * Fixed by using the correct T95 device tree with the **PI16 SD card-detect GPIO**. * **Debian 7.3 kernel would not mount root** * `root=UUID=` failed without initramfs. * Fixed by using: `root=PARTUUID=fe8d38d0-eebf-4bb1-aa05-baafaf8c3f50` * **No HDMI output** * Added the H616 display-engine/HDMI patches and enabled HDMI in the T95 DTS. * HDMI now works at **1080p**. * **Ethernet not working** * Fixed H616 EMAC/AC200 PHY support. * Ethernet now links at **100Mbps full duplex**. * **One USB port had power but no data** * Stock Android DTS showed the V4.0 board also uses USB host 1. * Added: ```dts &ehci1 { status = "okay"; }; &ohci1 { status = "okay"; }; ``` * Both USB ports now work. * **Mali/Panfrost GPU acceleration** * Still being worked on. * Panfrost times out because of the H616 PRCM GPU power-domain path. This is the last major issue. Board: **H616-T95MAX-V4.0** OS: **Debian 12** Kernel testing: **Linux 7.3-rc3**