opencl: fix Q4_0 SOA rows and Adreno paths

[happyyzy]

Summary

Fix OpenCL Q4_0 correctness issues seen when GGML_OPENCL_SOA_Q and Adreno OpenCL kernels are enabled:

• allow GET_ROWS(Q4_0) to read SOA quant buffers instead of assuming an AoS block_q4_0 buffer
• handle Adreno-transposed and Adreno-MoE Q4_0 layouts in GET_ROWS(Q4_0)
• avoid a shared static temporary buffer in transpose_2d()
• fix the fused RMSNorm/MUL offset to use src1->view_offs
• preserve F32 activation/accumulation for selected Adreno Q4_0 GEMM paths that otherwise produce invalid diffusion output

The same fix was also prepared for the older leejet/ggml fork used by stable-diffusion.cpp, but this branch was re-applied and build-checked against upstream master.

Image evidence

The image-level reproduction was run through stable-diffusion.cpp using ggml OpenCL, because the observed failure requires full diffusion-model execution.

Test device/config:

QUALCOMM Adreno 830, OpenCL 3.0 QUALCOMM build: 0800.71 Compiler E031.47.18.49
GGML_OPENCL_SOA_Q=ON
GGML_OPENCL_USE_ADRENO_KERNELS=ON
VAE conv direct enabled
Diffusion flash attention disabled

Individual images:

Case	Before	After
Flux.2 Klein Q4_0, 512x512
Z-Image Turbo Q4_0, 512x512

Reproduction commands

The same model files and generation parameters were used before and after the patch.

Flux.2 Klein Q4_0, before

git checkout 3af5f5760e19a96427f5f7a93b79cbdf3d4b265b
./bin/sd-cli \
  --diffusion-model "$KLEIN_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$FLUX2_VAE_SAFETENSORS" \
  -p "a lovely cat" \
  --cfg-scale 1.0 \
  --steps 4 \
  --seed 42 \
  -W 512 -H 512 \
  --vae-conv-direct \
  -o klein_before.png \
  -v

Result: command completes and saves an image, but the output is blank.

Flux.2 Klein Q4_0, after

git checkout e30fbfe5
./bin/sd-cli \
  --diffusion-model "$KLEIN_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$FLUX2_VAE_SAFETENSORS" \
  -p "a lovely cat" \
  --cfg-scale 1.0 \
  --steps 4 \
  --seed 42 \
  -W 512 -H 512 \
  --vae-conv-direct \
  -o klein_after.png \
  -v

Result: valid image output.

Z-Image Turbo Q4_0, before

git checkout 3af5f5760e19a96427f5f7a93b79cbdf3d4b265b
./bin/sd-cli \
  --diffusion-model "$Z_IMAGE_TURBO_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$Z_IMAGE_AE_SAFETENSORS" \
  -p "a lovely cat" \
  --cfg-scale 1.0 \
  --steps 8 \
  --seed 42 \
  -W 512 -H 512 \
  --vae-conv-direct \
  -o zimage_before.png \
  -v

Result: command completes and saves an image, but the output is blank.

Z-Image Turbo Q4_0, after

git checkout e30fbfe5
./bin/sd-cli \
  --diffusion-model "$Z_IMAGE_TURBO_Q4_0_GGUF" \
  --llm "$QWEN_3_4B_Q4_0_GGUF" \
  --vae "$Z_IMAGE_AE_SAFETENSORS" \
  -p "a lovely cat wearing black sunglasses, studio photo" \
  --cfg-scale 1.0 \
  --steps 4 \
  --seed 42 \
  -W 512 -H 512 \
  --vae-conv-direct \
  -o zimage_after.png \
  -v

Result: valid image output.

Checks

git diff --check ggmlorg/master..HEAD

cmake -S . -B /tmp/ggml_opencl_upstream_pr_build -G Ninja \
  -DCMAKE_BUILD_TYPE=Release \
  -DGGML_OPENCL=ON \
  -DGGML_OPENCL_USE_ADRENO_KERNELS=ON \
  -DGGML_OPENCL_EMBED_KERNELS=ON \
  -DGGML_NATIVE=OFF

cmake --build /tmp/ggml_opencl_upstream_pr_build --target ggml -j

The OpenCL-enabled ggml target builds successfully on this branch.

[happyyzy]

Could @lhez @max-krasnyansky @shaofeiqi help review or route this Adreno OpenCL Q4_0 correctness fix when you have a chance?

[akleine]

The author has done an excellent job here !

I have succesful tested this PR using a different (older) sd.cpp commit and different models, e.g NitroFusion and Diff-InstructStar. For both I earlier noticed similar mono-color image problems as mentioned for Flux and Z-Image.

It should be emphasized that this PR now allows the use of multiple threads to use during computation in sd.cpp when using OpenCl / Adreno and Q4_0. (I had to use the -t 1 option beforehand.)

Please merge as soon as possible @lhez @max-krasnyansky @shaofeiqi
Thank You

[lhez]

@happyyzy Thank you for bringing this up.

I wonder if we can move this PR (and your xmem FA PR #1562) to llama.cpp repo since currently almost all development happens there and this way we can have all updates in the same place.

By the way, I think you can have a separate PR for fix the fused RMSNorm/MUL offset to use src1->view_offs and we can quickly merge it.