[Bug] OpenCL Kernel compilation fails on Qualcomm Adreno 642L (no detailed error output)**

[Mr5b]

Git commit

5e4579c

Operating System & Version

HarmonyOS 4.2.0

GGML backends

OpenCL

Command-line arguments used

/data/local/tmp/stable-diffusion/bin/sd-cli -m "/data/local/tmp/model.safetensors" -p ""

Steps to reproduce

cmake .. -G Ninja -DCMAKE_TOOLCHAIN_FILE='android-ndk-r29-linux./android-ndk-r29/build/cmake/android.toolchain.cmake' -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-28 -DSD_OPENCL=ON -DGGML_OPENMP=OFF -DOpenCL_INCLUDE_DIR=... -DCMAKE_INSTALL_PREFIX=stable-diffusion -DCMAKE_BUILD_TYPE=Debug -DCMAKE_EXE_LINKER_FLAGS="-static-libstdc++ -static-libgcc"

ninja

ninja install

zip ...

adb push "stable-diffusion.zip" /data/local/tmp

adb push "model.safetensors" /data/local/tmp

adb shell

cd .... unzip ....

../../gdbserver :🔢 ./sd-cli -m "/../../model.safetensors" -p ""

What you expected to happen

nullptr

What actually happened

kernel compile error / Segmentation fault

Logs / error messages / stack trace

(ggml_opencl: device: 'QUALCOMM Adreno(TM) (OpenCL 2.0 Adreno(TM) 642L)'
[INFO ] ggml_extend.hpp:78 - ggml_opencl: OpenCL driver: OpenCL 2.0 QUALCOMM build: commit #272cf717f5 changeid #Iee40f504b5 Date: 10/05/21 Tue Local Branch: Remote Branch: Compiler E031.38.01.08
[INFO ] ggml_extend.hpp:78 - ggml_opencl: vector subgroup broadcast support: false
[INFO ] ggml_extend.hpp:78 - ggml_opencl: device FP16 support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: mem base addr align: 128
[INFO ] ggml_extend.hpp:78 - ggml_opencl: max mem alloc size: 920 MB
[INFO ] ggml_extend.hpp:78 - ggml_opencl: device max workgroup size: 1024
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM coarse grain buffer support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM fine grain buffer support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM fine grain system support: false
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM atomics support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: flattening quantized weights representation as struct of arrays (GGML_OPENCL_SOA_Q)
[INFO ] ggml_extend.hpp:78 - ggml_opencl: using kernels optimized for Adreno (GGML_OPENCL_USE_ADRENO_KERNELS)
[INFO ] ggml_extend.hpp:78 - ggml_opencl: loading OpenCL kernels
Segmentation fault)

Or

([INFO ] ggml_extend.hpp:78 - ggml_opencl: selected platform: 'QUALCOMM Snapdragon(TM)'
[INFO ] ggml_extend.hpp:78 -
ggml_opencl: device: 'QUALCOMM Adreno(TM) (OpenCL 2.0 Adreno(TM) 642L)'
[INFO ] ggml_extend.hpp:78 - ggml_opencl: OpenCL driver: OpenCL 2.0 QUALCOMM build: commit #272cf717f5 changeid #Iee40f504b5 Date: 10/05/21 Tue Local Branch: Remote Branch: Compiler E031.38.01.08
[INFO ] ggml_extend.hpp:78 - ggml_opencl: vector subgroup broadcast support: false
[INFO ] ggml_extend.hpp:78 - ggml_opencl: device FP16 support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: mem base addr align: 128
[INFO ] ggml_extend.hpp:78 - ggml_opencl: max mem alloc size: 920 MB
[INFO ] ggml_extend.hpp:78 - ggml_opencl: device max workgroup size: 1024
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM coarse grain buffer support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM fine grain buffer support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM fine grain system support: false
[INFO ] ggml_extend.hpp:78 - ggml_opencl: SVM atomics support: true
[INFO ] ggml_extend.hpp:78 - ggml_opencl: flattening quantized weights representation as struct of arrays (GGML_OPENCL_SOA_Q)
[INFO ] ggml_extend.hpp:78 - ggml_opencl: using kernels optimized for Adreno (GGML_OPENCL_USE_ADRENO_KERNELS)
[INFO ] ggml_extend.hpp:78 - ggml_opencl: loading OpenCL kernels
[ERROR] ggml_extend.hpp:84 - ggml_opencl: kernel compile error:

Pass)

Additional context / environment details

Adreno(TM) 642L