[OneDNN]Fix ocr error since pass avx512 command #64132
Conversation
|
你的PR提交成功，感谢你对开源项目的贡献! |
|
related issue: #11597 |
|
|
||
| #ifdef PADDLE_WITH_DNNL | ||
| if (!phi::backends::cpu::MayIUse( | ||
| phi::backends::cpu::cpu_isa_t::avx512_core)) { |
There was a problem hiding this comment.
For here, the limitation is a bit strict, since in self_dp_attention_kernel, the commonly used ISA was AVX512F. While this condition will block application of machines with only AVX512F.
zhanglirong1999
force-pushed
the
ocr_avx2_fix
branch
3 times, most recently
from
a2715c9
to
03877ef
Compare
|
thanks for fixing, skip avx512 implemented pass on non-avx512 platforms will solve the illegal instruction issue. I am kind of wondering the expected performance on non-avx512 platforms, since we have issues reporting unacceptable low speed on non-avx512 plaftorms, e.g.: PaddlePaddle/PaddleOCR#10346. |
Hi @jzhang533, as I mentioned in the issue, this is actually the same reason causing unexpected phenomenon. AMD seems has their mechanism to deal with incompatible higher ISA (and which also makes it hard to find the root cause since it wouldn’t report the error directly). And about performance on non-avx512 platform. If we build and install Paddle on non-avx512 machine, the performance will be as expected, which means the performance will be better when enabling oneDNN/mkldnn. |
* fix ocr avx error since pass avx512 command * replace 512 core to 512f, limit restrict * skip avx2 in new pir pass * fix
PR Category
Others
PR Types
Bug fixes
Description
Fix ocr error since pass avx512 command