ACL cross-compilation leads to KleidiAI aarch64-specific files compilation on non-aarch64 platforms

[alvoron]

The issue is reproduced on v25.02 where KleidiAI integration has been introduced.
2 files kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c and kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c have a platform check:

#if !defined(__aarch64__)
#error This file must be compiled for AArch64.
#else  // Architectural features check.

The issue is reproduced in OpenVINO cross-compile Android ARM 64 CI:
https://github.com/openvinotoolkit/openvino/actions/runs/13652458750/job/38164073920?pr=28669#step:12:3114

third_party/kleidiai/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:8:2: error: This file must be compiled for AArch64.
#error This file must be compiled for AArch64.
 ^
1 error generated.
scons: *** [build/third_party/kleidiai/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.o] Error 1

ComputeLibrary compilation command used in CI:
https://github.com/openvinotoolkit/openvino/actions/runs/13652458750/job/38164073920?pr=28669#step:12:3109

/usr/local/lib/python3.10/dist-packages/cmake/data/bin/cmake -E env /bin/scons neon=1 opencl=0 examples=0 Werror=0 gemm_tuner=0 reference_openmp=0 validation_tests=0 benchmark_tests=0 data_layout_support=all arch=arm64-v8.2-a openmp=0 cppthreads=1 estate=64 install_dir=install --jobs=8 os=android build=cross_compile compiler_cache=sccache toolchain_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-" compiler_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/" extra_link_flags=\ \ --target=aarch64-none-linux-android29\ --gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ --sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot extra_cxx_flags=-Wsuggest-override\ -g\ -DANDROID\ -fdata-sections\ -ffunction-sections\ -funwind-tables\ -fstack-protector-strong\ -no-canonical-prefixes\ -D_FORTIFY_SOURCE=2\ -Wformat\ -Werror=format-security\ \ \ -fsigned-char\ -ffunction-sections\ -fdata-sections\ -fdiagnostics-show-option\ -Wall\ -Wundef\ -Wmissing-declarations\ -Wno-error=deprecated-declarations\ -Wno-cpp\ -Wno-delete-non-abstract-non-virtual-dtor\ -Wno-error\ -Wno-undef\ -Wno-missing-declarations\ -Wno-macro-redefined\ -Wno-suggest-override\ -fvisibility-inlines-hidden\ \ -Wall\ -Wno-unknown-pragmas\ -fvisibility=internal\ -O3\ -Wno-undef\ -fPIC\ \ --target=aarch64-none-linux-android29\ --gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ --sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot\ -Wno-suggest-override --silent fixed_format_kernels=True build/libarm_compute-static.a

Should I update cross-compilation scons command or some changes is required on ACL side? Probably, these 2 files should be moved from common section of cc_library_static to arm64 section in Android.pb file?

[morgolock]

Hi @alvoron

Looking at the job shared above, it seems you're building for aarch64 so there should be no problem building KAI.

For some reason the build system gets confused and thinks it's not building for aarch64.

I'm trying to reproduce the problem using the same options present in the job but I can't get the error.

It would be helpful to see what command the build system is using when building kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c. I can't find the output in the logs, can you try to include the output of scons when building ACL so that we can understand what's going on here?

FAILED: /__w/openvino/openvino/openvino/src/plugins/intel_cpu/thirdparty/ComputeLibrary/build/libarm_compute-static.a 
cd /__w/openvino/openvino/openvino/src/plugins/intel_cpu/thirdparty/ComputeLibrary && /usr/local/lib/python3.10/dist-packages/
cmake/data/bin/cmake -E env /bin/scons neon=1 opencl=0 examples=0 Werror=0 gemm_tuner=0 reference_openmp=0 
validation_tests=0 benchmark_tests=0 data_layout_support=all arch=arm64-v8.2-a openmp=0 cppthreads=1 estate=64 
install_dir=install --jobs=8 os=android build=cross_compile compiler_cache=sccache toolchain_prefix="/deps/android_tools/ndk-
bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-" compiler_prefix="/deps/android_tools/ndk-bundle/toolchains/
llvm/prebuilt/linux-x86_64/bin/" extra_link_flags=\ \ --target=aarch64-none-linux-android29\ --gcc-toolchain=/deps/android_tools/ndk-
bundle/toolchains/llvm/prebuilt/linux-x86_64\ --sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot 
extra_cxx_flags=-Wsuggest-override\ -g\ -DANDROID\ -fdata-sections\ -ffunction-sections\ -funwind-tables\ -fstack-protector-strong\ -
no-canonical-prefixes\ -D_FORTIFY_SOURCE=2\ -Wformat\ -Werror=format-security\ \ \ -fsigned-char\ -ffunction-sections\ -fdata-
sections\ -fdiagnostics-show-option\ -Wall\ -Wundef\ -Wmissing-declarations\ -Wno-error=deprecated-declarations\ -Wno-cpp\ -Wno-
delete-non-abstract-non-virtual-dtor\ -Wno-error\ -Wno-undef\ -Wno-missing-declarations\ -Wno-macro-redefined\ -Wno-suggest-
override\ -fvisibility-inlines-hidden\ \ -Wall\ -Wno-unknown-pragmas\ -fvisibility=internal\ -O3\ -Wno-undef\ -fPIC\ \ --target=aarch64-
none-linux-android29\ --gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ --sysroot=/deps/
android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot\ -Wno-suggest-override --silent fixed_format_kernels=True 
build/libarm_compute-static.a
Using compilers:
CC sccache /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang
CXX sccache /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang++
third_party/kleidiai/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c:8:2: error: This file must be compiled for AArch64.
#error This file must be compiled for AArch64.

[alvoron]

@morgolock thank you for the prompt response
This is the complete command:

sccache /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang -o 
build/third_party/kleidiai/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.o -c 
-DARM_COMPUTE_CPP_SCHEDULER=1 
-DENABLE_NEON 
-DARM_COMPUTE_ENABLE_NEON 
-DARM_COMPUTE_ENABLE_FP16 
-DANDROID 
-DARM_COMPUTE_CPU_ENABLED 
-DARM_COMPUTE_VERSION_MAJOR=47 
-DARM_COMPUTE_VERSION_MINOR=0 
-DARM_COMPUTE_VERSION_PATCH=0 
-DARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS 
-Iinclude 
-I. -Ibuild/build/src/core 
-Ibuild/src/core 
-Ibuild/src/core/NEON/kernels/arm_gemm 
-Isrc/core/NEON/kernels/arm_gemm 
-Ibuild/src/core/NEON/kernels/convolution/common 
-Isrc/core/NEON/kernels/convolution/common 
-Ibuild/src/core/NEON/kernels/convolution/winograd 
-Isrc/core/NEON/kernels/convolution/winograd 
-Ibuild/src/core/NEON/kernels/arm_conv/depthwise 
-Isrc/core/NEON/kernels/arm_conv/depthwise 
-Ibuild/src/core/NEON/kernels/arm_conv/pooling 
-Isrc/core/NEON/kernels/arm_conv/pooling 
-Ibuild/src/core/NEON/kernels/arm_conv 
-Isrc/core/NEON/kernels/arm_conv 
-Ibuild/src/core/NEON/kernels/assembly 
-Isrc/core/NEON/kernels/assembly 
-Ibuild/arm_compute/core/NEON/kernels/assembly 
-Iarm_compute/core/NEON/kernels/assembly 
-Ibuild/src/cpu/kernels/assembly 
-Isrc/cpu/kernels/assembly 
-Ibuild/third_party/kleidiai 
-Ithird_party/kleidiai 
third_party/kleidiai/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c

https://github.com/openvinotoolkit/openvino/actions/runs/13671766795/job/38223436758?pr=28669#step:12:4013

scons command:

/usr/local/lib/python3.10/dist-packages/cmake/data/bin/cmake 
-E env SHCXXCOMSTR="Building\ " SHCCCOMSTR="Building\ " CCCOMSTR="Compiling\ " CXXCOMSTR="Compiling\ " LINKCOMSTR="Linking\ " 
/bin/scons 
neon=1 
opencl=0 
examples=0 
Werror=0 
gemm_tuner=0 
reference_openmp=0 
validation_tests=0 benchmark_tests=0 
data_layout_support=all 
arch=arm64-v8.2-a 
openmp=0 
cppthreads=1 
estate=64 
install_dir=install 
--jobs=8 
os=android 
build=cross_compile 
compiler_cache=sccache 
toolchain_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-" 
compiler_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/" 
extra_link_flags=\ \ 
--target=aarch64-none-linux-android29\ 
--gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ 
--sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot 
extra_cxx_flags=-Wsuggest-override\ -g\ -DANDROID\ -fdata-sections\ 
-ffunction-sections\ -funwind-tables\ -fstack-protector-strong\ 
-no-canonical-prefixes\ -D_FORTIFY_SOURCE=2\ 
-Wformat\ -Werror=format-security\ \ \ -fsigned-char\ -ffunction-sections\ 
-fdata-sections\ -fdiagnostics-show-option\ 
-Wall\ -Wundef\ -Wmissing-declarations\ -Wno-error=deprecated-declarations\ 
-Wno-cpp\ -Wno-delete-non-abstract-non-virtual-dtor\ -Wno-error\ 
-Wno-undef\ -Wno-missing-declarations\ -Wno-macro-redefined\ -Wno-suggest-override\ 
-fvisibility-inlines-hidden\ \ 
-Wall\ -Wno-unknown-pragmas\ -fvisibility=internal\ -O3\ -Wno-undef\ -fPIC\ \ 
--target=aarch64-none-linux-android29\ 
--gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ 
--sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot\ 
-Wno-suggest-override 
fixed_format_kernels=True build/libarm_compute-static.a

https://github.com/openvinotoolkit/openvino/actions/runs/13671766795/job/38223436758?pr=28669#step:12:3491

[morgolock]

Hi @alvoron

Thanks for the additional details.

I see the scons command is using arch=arm64-v8.2-a and estate=64 so I would expect __aarch64__ to be defined when building .c and .cpp files.

This seems a problem with the toolchain or the extra args used in the build command: --target , --gcc-toolchain or --sysroot

We could try using clang -print-targets to see what targets the toolchain supports. If it supports aarch64, for some reason it's not selecting the correct target and it's building for aarch32

[alvoron]

@morgolock I run clang -print-targets in CI

  Registered Targets:
    aarch64    - AArch64 (little endian)
    aarch64_32 - AArch64 (little endian ILP32)
    aarch64_be - AArch64 (big endian)
    arm        - ARM
    arm64      - ARM64 (little endian)
    arm64_32   - ARM64 (little endian ILP32)
    armeb      - ARM (big endian)
    bpf        - BPF (host endian)
    bpfeb      - BPF (big endian)
    bpfel      - BPF (little endian)
    thumb      - Thumb
    thumbeb    - Thumb (big endian)
    x86        - 32-bit X86: Pentium-Pro and above
    x86-64     - 64-bit X86: EM64T and AMD64

https://github.com/openvinotoolkit/openvino/actions/runs/13699585832/job/38310081339?pr=28669#step:12:2278

[morgolock]

Hi @alvoron

Thanks for sharing the details above. Your toolchain supports aarch64, it's not clear why then the build system is compiling kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c for aarch32.

I built ACL for android using the same options as above without problems:

AR=llvm-ar CXX=clang++ CC=clang PATH=../../toolchains/android-ndk-r26d/toolchains/llvm/prebuilt/linux-x86_64/bin/:$PATH scons
 neon=1 opencl=0   benchmark_tests=0 validation_tests=1 standalone=0 os=android Werror=0 neon=1 opencl=0  
compiler_prefix=aarch64-linux-android31- toolchain_prefix=llvm- neon=1 opencl=0 examples=0 Werror=0 gemm_tuner=0 
reference_openmp=0 validation_tests=0 benchmark_tests=0 data_layout_support=all arch=arm64-v8.2-a openmp=0 cppthreads=1 
estate=64 install_dir=install --jobs=8 os=android build=cross_compile

I think the problem may be caused by the additional arguments shown below :

toolchain_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android-" 
compiler_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/" 
extra_link_flags=\ \ 
--target=aarch64-none-linux-android29\ 
--gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ 
--sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot 
extra_cxx_flags=-Wsuggest-override\ -g\ -DANDROID\ -fdata-sections\ 
-ffunction-sections\ -funwind-tables\ -fstack-protector-strong\ 
-no-canonical-prefixes\ -D_FORTIFY_SOURCE=2\ 
-Wformat\ -Werror=format-security\ \ \ -fsigned-char\ -ffunction-sections\ 
-fdata-sections\ -fdiagnostics-show-option\ 
-Wall\ -Wundef\ -Wmissing-declarations\ -Wno-error=deprecated-declarations\ 
-Wno-cpp\ -Wno-delete-non-abstract-non-virtual-dtor\ -Wno-error\ 
-Wno-undef\ -Wno-missing-declarations\ -Wno-macro-redefined\ -Wno-suggest-override\ 
-fvisibility-inlines-hidden\ \ 
-Wall\ -Wno-unknown-pragmas\ -fvisibility=internal\ -O3\ -Wno-undef\ -fPIC\ \ 
--target=aarch64-none-linux-android29\ 
--gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64\ 
--sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot\ 
-Wno-suggest-override 

For troubleshooting, could you please try removing all, or as many as possible, of these arguments? Especially --target, --gcc-toolchain, --sysroot ? maybe even extra_cxx_flags
I'd also remove the full path in toolchain_prefix and compiler_prefix? Basically I'd try to remove options to identify where the problem is.

Instead of having the path to the toolchain in the toolchain_prefix I suggest appending it to the PATH variable for example:
PATH=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/:$PATH scons ...

For compiler_prefix I use compiler_prefix=aarch64-linux-android31- whereas you use compiler_prefix="/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/" and your target aarch64-none-linux-android29.

What version of the NDK are you using?

[alvoron]

@morgolock
I added the same architecture check into OpenVINO source code that compiled before KleiaiAI compilation issue:

#if !defined(__aarch64__)
#error This file must be compiled for AArch64.
#else  // Architectural features check.
#pragma message("Architectural features check passed")
#endif

I've got Architectural features check passed, so __aarch64__ is defined for OpenVINO compilation:
https://github.com/openvinotoolkit/openvino/actions/runs/13721320718/job/38379085423?pr=28669#step:12:131

The line above contains compilation command:

2025-03-07T13:50:17.2406427Z [102/7311] sccache /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang++ 
--target=aarch64-none-linux-android29 
--gcc-toolchain=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64 
--sysroot=/deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/sysroot 
-DIN_OV_COMPONENT -DOV_BUILD_POSTFIX=\"\" 
-DOV_NATIVE_PARENT_PROJECT_ROOT_DIR=\"openvino\" 
-I/__w/openvino/openvino/openvino/src/common/util/include 
-isystem /__w/openvino/openvino/build/vcpkg_installed/arm64-android/include -Wsuggest-override 
-g 
-DANDROID 
-fdata-sections -ffunction-sections -funwind-tables -fstack-protector-strong -no-canonical-prefixes -D_FORTIFY_SOURCE=2 -Wformat -Werror=format-security   -fsigned-char -ffunction-sections -fdata-sections -fdiagnostics-show-option -Wall -Wundef -Wmissing-declarations -O2 -DNDEBUG   -Wformat -Wformat-security -D_FORTIFY_SOURCE=2 
-fstack-protector-all 
-ffile-prefix-map=/__w/openvino/openvino/openvino/= 
-ffile-prefix-map=../openvino/= 
-std=c++17 
-fPIC 
-fvisibility=hidden -fvisibility-inlines-hidden -Werror -MD -MT src/common/util/CMakeFiles/openvino_util.dir/src/common_util.cpp.o -MF src/common/util/CMakeFiles/openvino_util.dir/src/common_util.cpp.o.d -o src/common/util/CMakeFiles/openvino_util.dir/src/common_util.cpp.o -c /__w/openvino/openvino/openvino/src/common/util/src/common_util.cpp
2025-03-07T13:50:17.2419962Z /__w/openvino/openvino/openvino/src/common/util/src/common_util.cpp:16:9: warning: Architectural features check passed [-W#pragma-messages]

[morgolock]

Hi @alvoron

Thanks for trying that and sharing the details above.

I see that works because in the previous version of ACL there were no .c files. In the new version of ACL we have .c files from KAI and the build system is not properly compiling these files: we get the error This file must be compiled for AArch64.

We found a problem in ACL. Could you please cherry-pick this patch ? That will solve the problem. I was discussing the issue with a colleague and we realized that the march setting is missing from the build command for the .c files.

sccache /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang -o build/third_party/kleidiai/kai/ukernels/
matmul/matmul_clamp_f32_f32_f32p/kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.o -c -
DARM_COMPUTE_CPP_SCHEDULER=1 -DENABLE_NEON -DARM_COMPUTE_ENABLE_NEON -
DARM_COMPUTE_ENABLE_FP16 -DANDROID -DARM_COMPUTE_CPU_ENABLED -DARM_COMPUTE_VERSION_MAJOR=47 
-DARM_COMPUTE_VERSION_MINOR=0 -DARM_COMPUTE_VERSION_PATCH=0 -
DARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS -Iinclude -I. -Ibuild/build/src/core -Ibuild/src/core -Ibuild/src/core/NEON/
kernels/arm_gemm -Isrc/core/NEON/kernels/arm_gemm -Ibuild/src/core/NEON/kernels/convolution/common -Isrc/core/NEON/
kernels/convolution/common -Ibuild/src/core/NEON/kernels/convolution/winograd -Isrc/core/NEON/kernels/convolution/winograd -
Ibuild/src/core/NEON/kernels/arm_conv/depthwise -Isrc/core/NEON/kernels/arm_conv/depthwise -Ibuild/src/core/NEON/kernels/
arm_conv/pooling -Isrc/core/NEON/kernels/arm_conv/pooling -Ibuild/src/core/NEON/kernels/arm_conv -Isrc/core/NEON/kernels/
arm_conv -Ibuild/src/core/NEON/kernels/assembly -Isrc/core/NEON/kernels/assembly -Ibuild/arm_compute/core/NEON/kernels/
assembly -Iarm_compute/core/NEON/kernels/assembly -Ibuild/src/cpu/kernels/assembly -Isrc/cpu/kernels/assembly -Ibuild/
third_party/kleidiai -Ithird_party/kleidiai third_party/kleidiai/kai/ukernels/matmul/matmul_clamp_f32_f32_f32p/
kai_matmul_clamp_f32_f32_f32p8x1biasf32_6x8x4_neon_mla.c

Here we should see the march option as in the .cpp files: -march=armv8.2-a+fp16

Hope this helps

[alvoron]

@morgolock thank you for the prompt patch!
The initial issue with kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c compilation has been solved: the error This file must be compiled for AArch64. is not raised anymore.

However, the new issue is unknown target CPU 'armv8.2-a+fp16' for kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c compilation:

sccache /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang 
-o build/third_party/kleidiai/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.o -c 
-march=armv8.2-a+fp16 -no-integrated-as -O3 
-DARM_COMPUTE_CPP_SCHEDULER=1 -DENABLE_NEON -DARM_COMPUTE_ENABLE_NEON 
-DARM_COMPUTE_ENABLE_FP16 -DANDROID -DARM_COMPUTE_CPU_ENABLED 
-DARM_COMPUTE_VERSION_MAJOR=47 -DARM_COMPUTE_VERSION_MINOR=0 
-DARM_COMPUTE_VERSION_PATCH=0 -DARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS 
-Iinclude -I. -Ibuild/build/src/core -Ibuild/src/core -Ibuild/src/core/NEON/kernels/arm_gemm 
-Isrc/core/NEON/kernels/arm_gemm -Ibuild/src/core/NEON/kernels/convolution/common 
-Isrc/core/NEON/kernels/convolution/common -Ibuild/src/core/NEON/kernels/convolution/winograd 
-Isrc/core/NEON/kernels/convolution/winograd -Ibuild/src/core/NEON/kernels/arm_conv/depthwise 
-Isrc/core/NEON/kernels/arm_conv/depthwise -Ibuild/src/core/NEON/kernels/arm_conv/pooling 
-Isrc/core/NEON/kernels/arm_conv/pooling -Ibuild/src/core/NEON/kernels/arm_conv 
-Isrc/core/NEON/kernels/arm_conv 
-Ibuild/src/core/NEON/kernels/assembly -Isrc/core/NEON/kernels/assembly 
-Ibuild/arm_compute/core/NEON/kernels/assembly 
-Iarm_compute/core/NEON/kernels/assembly -Ibuild/src/cpu/kernels/assembly 
-Isrc/cpu/kernels/assembly -Ibuild/third_party/kleidiai 
-Ithird_party/kleidiai third_party/kleidiai/kai/ukernels/matmul/pack/kai_rhs_pack_kxn_f32p8x1biasf32_f32_f32_neon.c
2025-03-07T16:32:29.1271938Z error: unknown target CPU 'armv8.2-a+fp16'

It seems that --target=aarch64-none-linux-android29 is also missing for C files. Adding env.Append(CFLAGS = ['--target=aarch64-none-linux-android29']) to Android section in SConstruct helps me to finish Android build successfully.

[alvoron]

@morgolock did you have a chance to take a look at target flag?

[morgolock]

Hi @alvoron

The error you see: unknown target CPU 'armv8.2-a+fp16' suggests that the toolchain does not support the build target v8.2.

My guess is that if you call directly /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang the default toolchain can't build for the aarch64 target. We can verify this by calling to /deps/android_tools/ndk-bundle/toolchains/llvm/prebuilt/linux-x86_64/bin/clang --print-targets

If you see below, this is how I build ACL, I specify the option compiler_prefix=aarch64-linux-android31-. This is then appended as a prefix to the toolchain.

 AR=llvm-ar CXX=clang++ CC=clang PATH=../../toolchains/android-ndk-r26d/toolchains/llvm/prebuilt/linux-x86_64/bin/:$PATH scons neon=1 opencl=0   benchmark_tests=0 validation_tests=1 standalone=0 os=android Werror=0 neon=1 opencl=0  compiler_prefix=aarch64-linux-android31- toolchain_prefix=llvm- neon=1 opencl=0 examples=0 Werror=0 gemm_tuner=0 reference_openmp=0 validation_tests=0 benchmark_tests=0 data_layout_support=all arch=arm64-v8.2-a openmp=0 cppthreads=1 estate=64 install_dir=install --jobs=8 os=android build=cross_compile compiler_cache=ccache

As you can see below the output build command from scons is then aarch64-linux-android31-clang++

ccache aarch64-linux-android31-clang++ -o build/src/runtime/NEON/functions/NEPReluLayer.o -c -DARCH_ARM -Wextra -Wdisabled-optimization -Wformat=2 -Winit-self -Wstrict-overflow=2 -Wswitch-default -Woverloaded-virtual -Wformat-security -Wctor-dtor-privacy -Wsign-promo -Weffc++ -Wno-overlength-strings -Wall -std=c++14 -pedantic -Wno-vla-extension -DENABLE_FP16_KERNELS -DARM_COMPUTE_ENABLE_FP16 -DENABLE_FP32_KERNELS -DENABLE_QASYMM8_KERNELS -DENABLE_QASYMM8_SIGNED_KERNELS -DENABLE_QSYMM16_KERNELS -DENABLE_INTEGER_KERNELS -DENABLE_NHWC_KERNELS -DENABLE_NCHW_KERNELS -march=armv8.2-a+fp16 -O3 -DARM_COMPUTE_CPP_SCHEDULER=1 -DENABLE_NEON -DARM_COMPUTE_ENABLE_NEON -DARM_COMPUTE_ENABLE_FP16 -DANDROID -DARM_COMPUTE_CPU_ENABLED -DARM_COMPUTE_VERSION_MAJOR=49 -DARM_COMPUTE_VERSION_MINOR=0 -DARM_COMPUTE_VERSION_PATCH=0 -Iinclude -I. -Ibuild/build/src/core -Ibuild/src/core -Ibuild/src/core/NEON/kernels/arm_gemm -Isrc/core/NEON/kernels/arm_gemm -Ibuild/src/core/NEON/kernels/convolution/common -Isrc/core/NEON/kernels/convolution/common -Ibuild/src/core/NEON/kernels/convolution/winograd -Isrc/core/NEON/kernels/convolution/winograd -Ibuild/src/core/NEON/kernels/arm_conv/depthwise -Isrc/core/NEON/kernels/arm_conv/depthwise -Ibuild/src/core/NEON/kernels/arm_conv/pooling -Isrc/core/NEON/kernels/arm_conv/pooling -Ibuild/src/core/NEON/kernels/arm_conv -Isrc/core/NEON/kernels/arm_conv -Ibuild/src/core/NEON/kernels/assembly -Isrc/core/NEON/kernels/assembly -Ibuild/arm_compute/core/NEON/kernels/assembly -Iarm_compute/core/NEON/kernels/assembly -Ibuild/src/cpu/kernels/assembly -Isrc/cpu/kernels/assembly -Ibuild/third_party/kleidiai -Ithird_party/kleidiai src/runtime/NEON/functions/NEPReluLayer.cpp

Can you please try using the same approach and specify compiler_prefix=aarch64-linux-android31- ? That should solve the problem. This has the same effect as using the option -target=aarch64-none-linux-android31

Hope this helps