ggml-backend-reg: Adreno-aware OpenCL backend selection on Android (QVAC-18993)

[Zbig9000]

Summary

Ports the Android GPU backend-selection policy from qvac-fabric-llm.cpp's ggml fork (the LLM stack) into the speech ggml fork, so whisper / parakeet / tts select GPU backends the same way the LLM stack does.

Problem

ggml_backend_load_all_from_path() loaded Vulkan and OpenCL unconditionally. On Adreno, ggml registers both for the same GPU and Vulkan loads first, so whisper.cpp's default gpu_device=0 lands on the Adreno Vulkan driver, which SIGSEGVs in vkCmdBindPipeline during ggml compute (observed on Samsung S25 Ultra, device-farm run 26646220900). On non-Adreno GPUs (e.g. Mali, proven working on Pixel 9), Vulkan is the correct/stable path and OpenCL must not be used.

Fix

On Android, after loading Vulkan, use it to detect the GPU and decide whether to load OpenCL:

GPU	{load_opencl, unload_vulkan}	Result
not Adreno (e.g. Mali)	{false, false}	skip OpenCL → Vulkan/CPU
Adreno > 700	{true, false}	load OpenCL (Vulkan kept; consumer picks OpenCL)
Adreno 1–700 (incl. ≤ 600)	{false, true}	unload Vulkan, CPU only

Only Adreno 700+ has a stable ggml GPU path. This is stricter than fabric-llm, which loads OpenCL on Adreno ≤ 600 — we treat ≤ 600 the same as 601–700 (CPU only), since older Adrenos are no more capable than the 601–700 tier we already exclude.

Implemented with ggml_backend_reg_by_name("vulkan") + ggml_backend_min_adreno_version() (public ggml_backend_reg_dev_* / ggml_backend_dev_description APIs) + ggml_backend_unload(). The threshold decision is factored into a pure ggml_adreno_resolve_backend_policy() so it's unit-testable.

Regression safety

The entire Adreno block is #ifdef __ANDROID__. Off Android the behaviour is byte-identical: load_opencl stays true and OpenCL is loaded in exactly the same position as before (after virtgpu, before hexagon).

Tests

• Pure GPU-string parser + policy decision extracted to src/ggml-adreno.h (inline, dependency-free); the parser also lowercases the description (small hardening over fabric-llm's raw substring check).
• New tests/test-adreno-version.cpp — 29 cases: 18 parser (Adreno variants; Mali/NVIDIA/AMD/Apple/Intel/llvmpipe non-Adreno; empty; unparseable) + 11 policy, incl. the > 700 boundary (Adreno exactly 700 → CPU only) and the merged ≤ 600 tier (600/500/1 → CPU only). Wired into ctest as test-adreno-version (always built, no ggml link).

Test plan

• ggml lib builds clean (x64-linux)
• ggml-backend-reg.cpp -D__ANDROID__ -fsyntax-only: clean (Android block compiles)
• test-adreno-version: 29/29 pass via ctest
• Device farm (after the ggml-speech port bump → whisper rebuild): Samsung S25 (Adreno) → OpenCL, Pixel 9 (Mali) → Vulkan, via transcription-whispercpp's GPU test

Relationship to transcription-whispercpp #2343 (complementary — both required)

This change controls which backends are available; the addon picks among them (same split as the LLM stack: engine availability + BackendSelection). On Adreno 700+ this keeps both Vulkan and OpenCL loaded (like fabric-llm keeps Vulkan), so whisper's default gpu_device=0 would still pick Vulkan → crash unless #2343's adrenoOpenclGpuDeviceIndex → gpu_device guard redirects to OpenCL. So #2343's guard is not redundant; #18 additionally handles the Adreno 1–700 (CPU-only) tier and the non-Adreno OpenCL-skip that the addon guard can't.

Downstream (separate, follow-up)

Once merged: bump the ggml-speech vcpkg port → whisper-cpp (+ tts-cpp / parakeet-cpp) deps → validate via the transcription-whispercpp addon overlay → CI → registry → addon, per the standard merge protocol.