RedMagic 11 Pro display fixes: IMapper5 gralloc backend and UBWC swapchain buffers

[Leb-Sun]

Makes vendor display features that consume the app's swapchain — RedMagic GameSpace upscaling and frame generation — work under Turnip. Both device-confirmed on Adreno 840 / Android 16. Supersedes #2 and #3.

Also drops the -e (eco) variant: its only behavioural change was measured worse and already removed, leaving tunables for a closed question. Variants are now b and p.

Root cause — three chained gaps

1. Turnip never requested UBWC. Android's loader asks the driver, via vkGetPhysicalDeviceImageFormatProperties2 + VkAndroidHardwareBufferUsageANDROID, what usage to allocate swapchain buffers with (frameworks/native/vulkan/libvulkan/swapchain.cpp:1450+ — it returns before the legacy GetSwapchainGrallocUsageX chain, which modern Android never calls). Measured with an app-side probe:

driver	androidHardwareBufferUsage
Turnip (stock)	0x200
system Adreno	0x10000200
exported Adreno 842.8 (adrenotools-loaded)	0x10000200

0x10000000 is the QTI vendor bit requesting UBWC. Mesa's vk_image_usage_to_ahb_usage() has no vendor-usage concept — there is an in-tree XXX acknowledging this.

2. QTI's mapper misdescribes the resulting buffer. For 1216x2688 RGBA8888 UBWC it reports fourcc=0, modifier=0 (LINEAR — wrong) and two planes with the metadata plane first: plane[0] offset=86016 stride=4864 (data, 12164) and plane[1] offset=0 stride=128 (metadata, 128672=86016).

3. Mesa rejects that. vk_gralloc_to_drm_explicit_layout() treats any plane n>0 at offset 0 as disjoint and returns VK_ERROR_INVALID_EXTERNAL_HANDLE, so vkCreateSwapchainKHR fails and the screen goes black.

Changes

• add_aimapper_gralloc.py — IMapper5 gralloc backend reached through android_load_sphal_library() + AIMapper_loadIMapper(). A driver loaded into an app process cannot link libui, so Mesa's existing IMapper backends are gated out and u_gralloc falls back to u_gralloc_fallback.c (no YCbCr, UBWC guessed from a private handle offset). Needs only dlopen/dlsym.
• add_ubwc_swapchain_usage.py — sets the QTI UBWC usage bit, scoped to the allocation query only. It must not go in vk_image_usage_to_ahb_usage(): that function also answers import validation, and demanding UBWC there rejects every linear buffer (observed: 448 nativeImportAhbToVulkan failed).
• UBWC detection via StandardMetadataType::COMPRESSION — device-measured, QTI returns name="QTI" value=10 for UBWC and name="android.hardware.graphics.common.Compression" value=0 for linear. Plane-layout signature retained as a fallback for vendors that do not populate it.

Verification

aimapper: UBWC via COMPRESSION metadata → WN-ANB-LAYOUT: result=0 modifier=0x500000000000001 planes=1, picture renders, upscaling and frame generation both working.

Symptom ladder along the way: green (no RGBA source at all) → black (source exists, swapchain fails) → working.

Not included

add_anb_gralloc_usage34.py (ANB usage3/4 + spec 10) stays on ubwc-swapchain-usage. It closes a real Mesa gap — Mesa dispatches only v1/v2 while its own header is spec 11 — but the modern loader never calls that chain, so it is unproven here and does not belong on a branch named "fixes".

Note

This exact tree has not been built or run; it is assembled from device-confirmed pieces with diagnostics stripped. This PR is the first build of it.