Build executorch using android ndk with optimized kernels shows unsupported architecture 'armv8.2-a+bf16' and unknown type name 'bfloat16x8_t'

[elegracer]

🐛 Describe the bug

For the core problem, please ignore the background and trials sections, and go to Main issue part directly.

A little background

I was originally intended to test inference performance with executorch on my android device. For my custom resnet based model, in torchscript c++ cpu, average inference time on-device is 10ms. I heard torchscript is not on active development and executorch is a replacement. So I built executorch with android ndk r21b and test my model.

First trial: build executorch and test inference time

I followed this page to successfully build and run executorch in python mode (build from source). https://pytorch.org/executorch/stable/getting-started-setup.html

Then I followed this page to convert my nn model to a .pte file. https://pytorch.org/executorch/stable/tutorials/export-to-executorch-tutorial.html

The export steps are as follows, only with model variable and example_args fit to my model.

import torch
from torch.export import export, export_for_inference, ExportedProgram
import exir
from executorch.backends.xnnpack.partition.xnnpack_partitioner import XnnpackPartitioner

class M(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.param = torch.nn.Parameter(torch.rand(3, 4))
        self.linear = torch.nn.Linear(4, 5)

    def forward(self, x):
        return self.linear(x + self.param).clamp(min=0.0, max=1.0)


example_args = (torch.randn(3, 4),)

pre_autograd_aten_dialect = export_for_inference(M(), example_args).module()

aten_dialect: ExportedProgram = export(pre_autograd_aten_dialect, example_args)

edge_program: exir.EdgeProgramManager = exir.to_edge(aten_dialect)

# Optionally do delegation:
# edge_delegated = edge_program.to_backend(XnnpackPartitioner())

executorch_program: exir.ExecutorchProgramManager = edge_program.to_executorch()

with open("model.pte", "wb") as file:
    file.write(executorch_program.buffer)

Then I followed this page to build examples/portable/executor_runner/executor_runner.cpp with android ndk r21b, with the following cmake commands. https://pytorch.org/executorch/stable/demo-apps-android.html

I modified executor_runner.cpp a bit to measure time used by the line Error status = method->execute();, simply with std::chrono utility.

cmake -S . -B build_android -DCMAKE_INSTALL_PREFIX=build_android -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON
cmake --build build_android -j`nproc` --target install

Then push the executable and the .pte model into my android device, then run it to test inference time, it shows average inference time 80ms.

Second trial: convert model with xnnpack and build executorch c++ executable with xnnpack backend

For model conversion part, I just uncomment the delegation line above.

For c++ executor part, I modified the configuration to the following. (i.e. added -DEXECUTORCH_BUILD_XNNPACK=ON). Then run the xnn_executor_runner on my android device.

cmake -S . -B build_android -DCMAKE_INSTALL_PREFIX=build_android -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON -DEXECUTORCH_BUILD_XNNPACK=ON

But inference time is still 80ms.

Main issue

So I tried building optimized kernels by adding -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON

cmake -S . -B build_android -DCMAKE_INSTALL_PREFIX=build_android -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake -DANDROID_ABI="arm64-v8a" -DEXECUTORCH_BUILD_EXTENSION_DATA_LOADER=ON -DEXECUTORCH_BUILD_EXTENSION_MODULE=ON -DEXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL=ON -DEXECUTORCH_BUILD_EXTENSION_TENSOR=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_OPTIMIZED=ON

By the way, building xnnpack itself is ok. So I didn't run into this issue. #6924
And adding -DXNNPACK_ENABLE_ARM_BF16=OFF doesn't change the result.

It's the optimized kernel part that gives the compiling error.

Cmake configuration summary (full txt file is uploaded below):

-- ******** Summary ********
--   CMAKE_BUILD_TYPE              : Release
--   CMAKE_CXX_STANDARD            : 17
--   CMAKE_CXX_COMPILER_ID         : Clang
--   CMAKE_TOOLCHAIN_FILE          : [ANDROID_NDK_ROOT]/android-ndk-r21b/build/cmake/android.toolchain.cmake
--   BUCK2                         : [HOME]/codes/executorch/buck2-bin/buck2-3bbde7daa94987db468d021ad625bc93dc62ba7fcb16945cb09b64aab077f284
--   PYTHON_EXECUTABLE             : python3
--   FLATC_EXECUTABLE              : flatc
--   EXECUTORCH_ENABLE_LOGGING              : OFF
--   EXECUTORCH_ENABLE_PROGRAM_VERIFICATION : OFF
--   EXECUTORCH_LOG_LEVEL                   : Info
--   EXECUTORCH_BUILD_ANDROID_JNI           : OFF
--   EXECUTORCH_BUILD_ARM_BAREMETAL         : OFF
--   EXECUTORCH_BUILD_COREML                : OFF
--   EXECUTORCH_BUILD_KERNELS_CUSTOM        : OFF
--   EXECUTORCH_BUILD_EXECUTOR_RUNNER       : ON
--   EXECUTORCH_BUILD_EXTENSION_DATA_LOADER : ON
--   EXECUTORCH_BUILD_EXTENSION_MODULE      : ON
--   EXECUTORCH_BUILD_EXTENSION_RUNNER_UTIL : ON
--   EXECUTORCH_BUILD_EXTENSION_TENSOR      : ON
--   EXECUTORCH_BUILD_EXTENSION_TRAINING      : OFF
--   EXECUTORCH_BUILD_FLATC                 : ON
--   EXECUTORCH_BUILD_GFLAGS                : ON
--   EXECUTORCH_BUILD_GTESTS                :
--   EXECUTORCH_BUILD_HOST_TARGETS          : ON
--   EXECUTORCH_BUILD_MPS                   : OFF
--   EXECUTORCH_BUILD_PYBIND                : OFF
--   EXECUTORCH_BUILD_QNN                   : OFF
--   EXECUTORCH_BUILD_KERNELS_OPTIMIZED     : ON
--   EXECUTORCH_BUILD_KERNELS_QUANTIZED     : OFF
--   EXECUTORCH_BUILD_DEVTOOLS              : OFF
--   EXECUTORCH_BUILD_SIZE_TEST             : OFF
--   EXECUTORCH_BUILD_XNNPACK               : ON
--   EXECUTORCH_BUILD_VULKAN                : OFF
--   EXECUTORCH_BUILD_PTHREADPOOL           : ON
--   EXECUTORCH_BUILD_CPUINFO               : ON
-- Configuring done (7.0s)
-- Generating done (0.7s)
-- Build files have been written to: [HOME]/codes/executorch/build_android

Main compiling error (full building output is uploaded below):

[ 79%] Building CXX object kernels/optimized/CMakeFiles/cpublas.dir/blas/BlasKernel.cpp.o
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:80:29: error: unknown type name 'bfloat16x8_t'; did you mean 'float16x8_t'?
f32_dot_bf16(float32x4_t a, bfloat16x8_t b, bfloat16x8_t c) {
                            ^~~~~~~~~~~~
                            float16x8_t
[ANDROID_NDK_ROOT]/android-ndk-r21b/toolchains/llvm/prebuilt/linux-x86_64/lib64/clang/9.0.8/include/arm_neon.h:65:56: note: 'float16x8_t' declared here
typedef __attribute__((neon_vector_type(8))) float16_t float16x8_t;
                                                       ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:80:45: error: unknown type name 'bfloat16x8_t'; did you mean 'float16x8_t'?
f32_dot_bf16(float32x4_t a, bfloat16x8_t b, bfloat16x8_t c) {
                                            ^~~~~~~~~~~~
                                            float16x8_t
[ANDROID_NDK_ROOT]/android-ndk-r21b/toolchains/llvm/prebuilt/linux-x86_64/lib64/clang/9.0.8/include/arm_neon.h:65:56: note: 'float16x8_t' declared here
typedef __attribute__((neon_vector_type(8))) float16_t float16x8_t;
                                                       ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:79:1: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
ET_TARGET_ARM_BF16_ATTRIBUTE static ET_INLINE float32x4_t
^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:81:10: error: use of undeclared identifier 'vbfdotq_f32'
  return vbfdotq_f32(a, b, c);
         ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:84:1: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
ET_TARGET_ARM_BF16_ATTRIBUTE static ET_INLINE void
^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:90:9: error: unknown type name 'bfloat16x8_t'; did you mean 'float16x8_t'?
  const bfloat16x8_t temp_vec1 = vld1q_bf16(reinterpret_cast<const __bf16*>(
        ^~~~~~~~~~~~
        float16x8_t
[ANDROID_NDK_ROOT]/android-ndk-r21b/toolchains/llvm/prebuilt/linux-x86_64/lib64/clang/9.0.8/include/arm_neon.h:65:56: note: 'float16x8_t' declared here
typedef __attribute__((neon_vector_type(8))) float16_t float16x8_t;
                                                       ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:90:68: error: unknown type name '__bf16'; did you mean '__be16'?
  const bfloat16x8_t temp_vec1 = vld1q_bf16(reinterpret_cast<const __bf16*>(
                                                                   ^~~~~~
                                                                   __be16
[ANDROID_NDK_ROOT]/android-ndk-r21b/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/linux/types.h:27:25: note: '__be16' declared here
typedef __u16 __bitwise __be16;
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:90:34: error: use of undeclared identifier 'vld1q_bf16'
  const bfloat16x8_t temp_vec1 = vld1q_bf16(reinterpret_cast<const __bf16*>(
                                 ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:92:9: error: unknown type name 'bfloat16x8_t'; did you mean 'float16x8_t'?
  const bfloat16x8_t temp_vec2 = vld1q_bf16(reinterpret_cast<const __bf16*>(
        ^~~~~~~~~~~~
        float16x8_t
[ANDROID_NDK_ROOT]/android-ndk-r21b/toolchains/llvm/prebuilt/linux-x86_64/lib64/clang/9.0.8/include/arm_neon.h:65:56: note: 'float16x8_t' declared here
typedef __attribute__((neon_vector_type(8))) float16_t float16x8_t;
                                                       ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:92:68: error: unknown type name '__bf16'; did you mean '__be16'?
  const bfloat16x8_t temp_vec2 = vld1q_bf16(reinterpret_cast<const __bf16*>(
                                                                   ^~~~~~
                                                                   __be16
[ANDROID_NDK_ROOT]/android-ndk-r21b/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include/linux/types.h:27:25: note: '__be16' declared here
typedef __u16 __bitwise __be16;
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:92:34: error: use of undeclared identifier 'vld1q_bf16'
  const bfloat16x8_t temp_vec2 = vld1q_bf16(reinterpret_cast<const __bf16*>(
                                 ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:119:1: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
ET_TARGET_ARM_BF16_ATTRIBUTE static ET_INLINE void
^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:150:3: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
  ET_TARGET_ARM_BF16_ATTRIBUTE ET_INLINE void operator()(const Func& f) const {
  ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:159:3: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
  ET_TARGET_ARM_BF16_ATTRIBUTE ET_INLINE void operator()(const Func& f) const {
  ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:167:1: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
ET_TARGET_ARM_BF16_ATTRIBUTE float
^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:176:33: warning: unsupported architecture 'armv8.2-a+bf16' in the 'target' attribute string; 'target' attribute ignored [-Wignored-attributes]
            ET_INLINE_ATTRIBUTE ET_TARGET_ARM_BF16_ATTRIBUTE {
                                ^
[HOME]/codes/executorch/kernels/optimized/blas/BlasKernel.cpp:78:25: note: expanded from macro 'ET_TARGET_ARM_BF16_ATTRIBUTE'
  __attribute__((target("arch=armv8.2-a+bf16")))
                        ^
7 warnings and 9 errors generated.
make[2]: *** [kernels/optimized/CMakeFiles/cpublas.dir/build.make:121: kernels/optimized/CMakeFiles/cpublas.dir/blas/BlasKernel.cpp.o] Error 1
make[2]: *** Waiting for unfinished jobs....

PS

Thank you for developing executorch which I think is just brilliant !! I'm new to this, and tried many ways to fix this error. Now it's just nothing else I can do with this. Please help me T_T !

Versions

Collecting environment information...
PyTorch version: 2.6.0+cu126
Is debug build: False
CUDA used to build PyTorch: 12.6
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu120.04.2) 9.4.0
Clang version: 21.0.0 (++20250203053216+6dfe20dbbd65-1exp1~20250203173345.1303)
CMake version: version 3.31.4
Libc version: glibc-2.31

Python version: 3.10.0 (default, Feb 14 2025, 22:44:04) [GCC 9.4.0] (64-bit runtime)
Python platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 3090
Nvidia driver version: 572.42
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True

CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 48 bits physical, 48 bits virtual
CPU(s): 12
On-line CPU(s) list: 0-11
Thread(s) per core: 2
Core(s) per socket: 6
Socket(s): 1
Vendor ID: AuthenticAMD
CPU family: 23
Model: 96
Model name: AMD Ryzen 5 PRO 4650G with Radeon Graphics
Stepping: 1
CPU MHz: 3692.984
BogoMIPS: 7385.96
Virtualization: AMD-V
Hypervisor vendor: Microsoft
Virtualization type: full
L1d cache: 192 KiB
L1i cache: 192 KiB
L2 cache: 3 MiB
L3 cache: 4 MiB
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Mitigation; untrained return thunk; SMT enabled with STIBP protection
Vulnerability Spec rstack overflow: Mitigation; safe RET
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl tsc_reliable nonstop_tsc cpuid extd_apicid pni pclmulqdq ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand hypervisor lahf_lm cmp_legacy svm cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw topoext perfctr_core ssbd ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 clzero xsaveerptr arat npt nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold v_vmsave_vmload umip rdpid

Versions of relevant libraries:
[pip3] executorch==0.5.0a0+1bc0699
[pip3] numpy==2.0.0
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] torch==2.6.0+cu126
[pip3] torchao==0.8.0+gitebc43034
[pip3] torchaudio==2.6.0+cu126
[pip3] torchsr==1.0.4
[pip3] torchvision==0.21.0+cu126
[pip3] triton==3.2.0
[conda] Could not collect

cc @larryliu0820 @lucylq

[elegracer]

cmake_configuration_output.txt

cmake_building_output.txt

[kirklandsign]

Hi @elegracer

So I built executorch with android ndk r21b and test my model.

You probably need the latest NDK version. Early version has early clang and it doesn't have bf16.

[kirklandsign]

Please feel free to re-open if you still have issue with NDK 27 or 28

[spalatinate]

@kirklandsign Actually I am running in a similar issue. I was working on executorch for a while. When I tried to reinstall executorch via install_requirements.sh --pybind xnnpack, I am running in the same error as above. This has not happened before. Thank you!

[kirklandsign]

Hi @spalatinate

Please

rm -rf pip-out cmake-out*
./install_requirements.sh
./install_executorch.sh --pybind xnnpack

Also please use a new Android NDK version.

Please re-open if it's still not working.

[elegracer]

With NDK 27c, the compilation succeeded!

[spalatinate]

@kirklandsign Sorry, I haven't specified my problem precisely... So I was trying to build executorch locally on my RPi5. It worked fine using the Clang compiler (version 14.0.6) and the release/0.4 branch. Now, with the release/0.5 and main branch, I am running into the error above. I guess it is related to the Clang compiler because when I switch to g++/gcc building executorch works just fine.

Just to be sure: I do not build anything on or for Android... :)

PS I do not have the permission to reopen this issue. Shall I open a new one? Could be relevant to others as well.

[kirklandsign]

@spalatinate Please feel free to open a new one. Thank you!

And include your environment. Honestly I haven't tried building it with RPi. For arm I only tried using macos.