Build fails on Centos7 with make: *** [all] Error 2

[makermotion]

When I try to follow the instructions building ggml to use mpt here: https://github.com/ggerganov/ggml/tree/79ea7a41a388e9e7c85e426e912ab38399c44bfa/examples/mpt

distro: Centos7
gcc: 13.2.0
g++: 13.2.0
command for "cmake": cmake -D CMAKE_C_COMPILER=/usr/local/bin/gcc -D CMAKE_CXX_COMPILER=/usr/local/bin/g++ ..
command for "make": make -j 4 CC=/usr/local/bin/gcc CXX=/usr/local/bin/g++

I get the following error:

[ 48%] Linking CXX executable ../bin/test-quantize-fns
repos/ggml/tests/test-vec1.c: In function ‘reduce_vector8_0’:
repos/ggml/tests/test-vec1.c:38:7: note: the ABI for passing parameters with 32-byte alignment has changed in GCC 4.6
   38 | float reduce_vector8_0(__m256 v) {
      |       ^~~~~~~~~~~~~~~~
repos/ggml/tests/test-vec1.c: In function ‘mul_mat_vec_f32_1’:
repos/ggml/tests/test-vec1.c:60:16: warning: AVX vector return without AVX enabled changes the ABI [-Wpsabi]
   60 |         __m256 sum = _mm256_setzero_ps();
      |                ^~~
In file included from /usr/local/lib/gcc/x86_64-pc-linux-gnu/13.2.0/include/immintrin.h:43,
                 from repos/ggml/tests/test-vec1.c:10:
/usr/local/lib/gcc/x86_64-pc-linux-gnu/13.2.0/include/avxintrin.h: In function ‘reduce_vector8_0’:
/usr/local/lib/gcc/x86_64-pc-linux-gnu/13.2.0/include/avxintrin.h:524:1: error: inlining failed in call to ‘always_inline’ ‘_mm256_extractf128_ps’: target specific option mismatch
  524 | _mm256_extractf128_ps (__m256 __X, const int __N)
      | ^~~~~~~~~~~~~~~~~~~~~
repos/ggml/tests/test-vec1.c:40:17: note: called from here
   40 |     __m128 v2 = _mm256_extractf128_ps(v, 1);
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~
/usr/local/lib/gcc/x86_64-pc-linux-gnu/13.2.0/include/avxintrin.h:524:1: error: inlining failed in call to ‘always_inline’ ‘_mm256_extractf128_ps’: target specific option mismatch
  524 | _mm256_extractf128_ps (__m256 __X, const int __N)
      | ^~~~~~~~~~~~~~~~~~~~~
repos/ggml/tests/test-vec1.c:39:17: note: called from here
   39 |     __m128 v1 = _mm256_extractf128_ps(v, 0);
      |                 ^~~~~~~~~~~~~~~~~~~~~~~~~~~
make[2]: *** [tests/CMakeFiles/test-vec1.dir/test-vec1.c.o] Error 1
make[1]: *** [tests/CMakeFiles/test-vec1.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....
[ 48%] Built target test-quantize-fns
[ 49%] Linking C executable ../bin/test-mul-mat2
[ 50%] Linking CXX executable ../../bin/mnist-cnn
[ 51%] Linking CXX executable ../../bin/mnist-cpu
[ 51%] Built target test-mul-mat2
[ 51%] Built target mnist-cnn
[ 51%] Built target mnist-cpu
[ 52%] Linking CXX executable ../../bin/mnist
[ 52%] Built target mnist
[ 54%] Linking CXX executable ../bin/test-grad0
[ 54%] Built target test-grad0
[ 55%] Linking CXX executable ../bin/test-quantize-perf
[ 55%] Built target test-quantize-perf
[ 56%] Linking CXX static library libcommon-ggml.a
[ 56%] Built target common-ggml
[ 57%] Linking CXX executable ../../bin/sam
[ 57%] Built target sam
[ 58%] Linking CXX static library libwhisper-cpp.a
[ 58%] Built target whisper-cpp
make: *** [all] Error 2`

[slaren]

Does your CPU support AVX?

[makermotion]

Does your CPU support AVX?

this returns nothing grep avx /proc/cpuinfo so no.

[slaren]

That's probably the issue then. test-vec1.c requires AVX, but it is built unconditionally on x86. It shouldn't be built on systems without AVX support. Can you test if this change to tests/CMakeLists.txt fixes the issue?

diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt
index a1cedf0..35e8e67 100644
--- a/tests/CMakeLists.txt
+++ b/tests/CMakeLists.txt
@@ -159,7 +159,7 @@ target_link_libraries(${TEST_TARGET} PRIVATE ggml)

 #
 # test-vec1 (x86)
-if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86")
+if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "x86" AND ${CMAKE_C_FLAGS} MATCHES "avx")
     set(TEST_TARGET test-vec1)
     add_executable(${TEST_TARGET} ${TEST_TARGET}.c)
     target_link_libraries(${TEST_TARGET} PRIVATE ggml)

Edit: updated the diff.

[makermotion]

that works thanks!

[slaren]

Let's keep this open until it is fixed here.