Fix SIMD calls #59
Fix SIMD calls #59
Conversation
Fast SIMD calls and secure align for result.
fabiangreffrath
left a comment
fabiangreffrath
left a comment
There was a problem hiding this comment.
Looks good, thank you! But what was the exact issue with the former code that needed to get fixed?
fabiangreffrath
left a comment
fabiangreffrath
left a comment
There was a problem hiding this comment.
Please fix the failing builds.
libfaac/quantize.c
Outdated
| x = _mm_max_ps(x, _mm_sub_ps((__m128){0, 0, 0, 0}, x)); | ||
| x = _mm_mul_ps(x, (__m128){sfacfix, sfacfix, sfacfix, sfacfix}); | ||
| const __m256d x_d = _mm256_load_pd(&xr[cnt]); | ||
| __m128 x = _mm256_cvtpd_ps(x_d); |
There was a problem hiding this comment.
These are AVX intrinsics, but this is SSE2 code. This will crash on systems without AVX support.
You need to use two _mm_load_pd and _mm_cvtpd_ps calls and one _mm_movelh_ps.
The main problem appears when unalligned data is used. The implicit Something like this. For make an fix-only PR, I will change this part. |
Potential unaligned memory location.
Only fix unaligned memory access.
3 participants
Fast SIMD calls and secure align for result.