5.2 KiB
📝 #214 - AVX512 build error
| Author | pt13762104 |
|---|---|
| State | ❌ Closed |
| Created | 2025-02-21 |
| Updated | 2025-02-21 |
Description
When building for AVX512, this error occurs:
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp: In member function '__m256i {anonymous}::DequantizerIQ6K::make_one(__m256i, __m256i) const':
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:2750:114: warning: overflow in conversion from 'int' to 'char' changes value from '255' to '-1' [-Woverflow]
2750 | auto mask1 = _mm256_andnot_si256(_mm256_or_si256(mask4, _mm256_or_si256(mask2, mask3)), _mm256_set1_epi8(0xff));
| ^~~~
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp: At global scope:
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:2846:48: warning: overflow in conversion from 'int' to 'short int' changes value from '65534' to '-2' [-Woverflow]
2846 | const __m256i bmask = _mm256_set1_epi16(0xfffe);
| ^~~~~~
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp: In member function 'void {anonymous}::QFT<Float, nrc_in>::load_r4(int, int, {anonymous}::QFBase::Data*) const':
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:8550:42: error: cannot convert '{anonymous}::QFBase::Data' {aka '__m512'} to '__m256'
8550 | auto t0 = _mm256_unpacklo_ps(xv[0], xv[1]);
| ~~~~^
| |
| {anonymous}::QFBase::Data {aka __m512}
In file included from /usr/lib/gcc/x86_64-redhat-linux/14/include/immintrin.h:43,
from /home/why/ik_llama.cpp/ggml/src/./ggml-impl.h:449,
from /home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:24:
/usr/lib/gcc/x86_64-redhat-linux/14/include/avxintrin.h:1100:28: note: initializing argument 1 of '__m256 _mm256_unpacklo_ps(__m256, __m256)'
1100 | _mm256_unpacklo_ps (__m256 __A, __m256 __B)
| ~~~~~~~^~~
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:8551:42: error: cannot convert '{anonymous}::QFBase::Data' {aka '__m512'} to '__m256'
8551 | auto t1 = _mm256_unpacklo_ps(xv[2], xv[3]);
| ~~~~^
| |
| {anonymous}::QFBase::Data {aka __m512}
/usr/lib/gcc/x86_64-redhat-linux/14/include/avxintrin.h:1100:28: note: initializing argument 1 of '__m256 _mm256_unpacklo_ps(__m256, __m256)'
1100 | _mm256_unpacklo_ps (__m256 __A, __m256 __B)
| ~~~~~~~^~~
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:8552:42: error: cannot convert '{anonymous}::QFBase::Data' {aka '__m512'} to '__m256'
8552 | auto t2 = _mm256_unpackhi_ps(xv[0], xv[1]);
| ~~~~^
| |
| {anonymous}::QFBase::Data {aka __m512}
/usr/lib/gcc/x86_64-redhat-linux/14/include/avxintrin.h:1094:28: note: initializing argument 1 of '__m256 _mm256_unpackhi_ps(__m256, __m256)'
1094 | _mm256_unpackhi_ps (__m256 __A, __m256 __B)
| ~~~~~~~^~~
/home/why/ik_llama.cpp/ggml/src/iqk/iqk_mul_mat.cpp:8553:42: error: cannot convert '{anonymous}::QFBase::Data' {aka '__m512'} to '__m256'
8553 | auto t3 = _mm256_unpackhi_ps(xv[2], xv[3]);
| ~~~~^
| |
| {anonymous}::QFBase::Data {aka __m512}
/usr/lib/gcc/x86_64-redhat-linux/14/include/avxintrin.h:1094:28: note: initializing argument 1 of '__m256 _mm256_unpackhi_ps(__m256, __m256)'
1094 | _mm256_unpackhi_ps (__m256 __A, __m256 __B)
| ~~~~~~~^~~
I have tried multiple copies of GCC 14, they produce the same result. The AVX2 builds fine, it's AVX512 that have trouble building.
💬 Conversation
👤 ikawrakow commented the 2025-02-21 at 07:09:29:
Does #215 fix it?
👤 pt13762104 commented the 2025-02-21 at 07:41:00:
I'll try, thanks
👤 pt13762104 commented the 2025-02-21 at 07:41:00:
I'll try
👤 pt13762104 commented the 2025-02-21 at 07:51:50:
It doesn't...
👤 ikawrakow commented the 2025-02-21 at 07:53:16:
What is the new compilation error?
👤 pt13762104 commented the 2025-02-21 at 07:59:04:
Seems like that fixed it, my bad
👤 ikawrakow commented the 2025-02-21 at 10:35:38:
@pt13762104 I think #216 really fixes it. Can you try? Thanks.
👤 pt13762104 commented the 2025-02-21 at 11:05:47:
I'll try to run a model to see if it's working
👤 pt13762104 commented the 2025-02-21 at 13:31:25:
It seemed to work fine, the models run, it compiles nicely...
👤 ikawrakow commented the 2025-02-21 at 13:33:09:
OK, thanks! I'll merge #216