Compile bug: ARMv7 NEON FP16 Intrinsic Errors When Cross-Compiling with Android NDK r26b

[Namitjain07]

Git commit

b4623

Here's a refined GitHub issue description with clearer output formatting and explicit error context:

---

Description
Cross-compiling for Android ARMv7 (armeabi-v7a) fails with missing NEON FP16 intrinsics despite proper toolchain configuration. The errors occur specifically in sgemm.cpp when using NDK r26b's Clang 17 compiler.

---

Environment

Component	Version
Android NDK	r26b
Host OS	Ubuntu 22.04.4 LTS
CMake	3.28.3
llama.cpp	commit 21c84b5d
Target ABI	armeabi-v7a
Android API	24

---

Steps to Reproduce

# 1. Fresh clone
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
git checkout 21c84b5d

# 2. Configure for ARMv7
mkdir build-android && cd build-android
cmake .. \
  -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
  -DANDROID_ABI=armeabi-v7a \
  -DANDROID_ARM_NEON=ON \
  -DANDROID_NATIVE_API_LEVEL=24 \
  -DCMAKE_CXX_FLAGS="-march=armv7-a -mfpu=neon-fp16"

# 3. Build
make -j4 llama

---

Observed Errors

[42%] Building CXX object ggml/src/CMakeFiles/ggml-cpu.dir/ggml-cpu/llamafile/sgemm.cpp.o
/home/user/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp:223:12: error: use of undeclared identifier 'vld1q_f16'
    return vld1q_f16((const float16_t *)p);  // <--- Missing NEON FP16 intrinsic
           ^
/home/user/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp:226:25: error: use of undeclared identifier 'vld1_f16'
    return vcvt_f32_f16(vld1_f16((const float16_t *)p));  // <--- FP16 conversion failure
                        ^
2 errors generated.
make[2]: *** [ggml/src/CMakeFiles/ggml-cpu.dir/build.make:188] Error 1

---

Expected Behavior

Successful compilation of NEON-accelerated FP16 code paths for ARMv7 targets.

---

Diagnostic Evidence

1. Compiler Feature Check:

   -- Performing Test GGML_COMPILER_SUPPORTS_FP16_FORMAT_I3E -- Failed
   

   Indicates compiler doesn't recognize ARMv7 FP16 extensions despite -mfpu=neon-fp16.

2. Architecture Validation:

   -- CMAKE_SYSTEM_PROCESSOR: armv7-a
   -- ANDROID_ARM_NEON: TRUE
   

   Confirms ARMv7+NEON target configuration.

3. Header Dependency:
   Missing #include <arm_neon.h> in files using NEON intrinsics.

---

Workarounds Attempted

Attempt	Result
Manual #include <arm_neon.h>	Compilation progresses but fails at linking
-D__ARM_FEATURE_FP16_VECTOR_ARITHMETIC=1	No change in intrinsic availability
NDK r25c downgrade	Build succeeds but breaks OpenCL compatibility
ARM64-v8a target	Builds successfully (confirms ARMv7-specific issue)

---

Reproduction Rate

• 100% reproducible on clean builds
• Affects all ARMv7 NEON FP16 code paths
• Verified across multiple Ubuntu 22.04 machines

---

Request:
Guidance on either:

1. Correct compiler flags to enable ARMv7 FP16 vector arithmetic in NDK r26b, or
2. Source modifications to make NEON FP16 code paths compatible with NDK's Clang 17 toolchain
3. complete commands to compile for ARMv7 to create libllama.so.

---

[ag2s20150909]

-DGGML_LLAMAFILE=OFF

[Namitjain07]

Can u please tell the complete commands to compile libllama.so in armv7 and armv8