| Message ID | 20240507152450.3781323-1-emma.finn@intel.com |
|---|---|
| State | Superseded |
| Headers | show |
| Series | [ovs-dev] odp-execute: Fix AVX checksum calculation. | expand |
| Context | Check | Description |
|---|---|---|
| ovsrobot/apply-robot | success | apply and check: success |
| ovsrobot/github-robot-_Build_and_Test | success | github build: passed |
| ovsrobot/intel-ovs-compilation | success | test: success |
On 7 May 2024, at 17:24, Emma Finn wrote: > The AVX implementation for calcualting checksum was not > handling carry-over addition correctly in some cases. > This patch adds an additional shuffle to add 16-bit padding > to the final part of the calculation to handle such cases. > > Fixes: 92eb03f7b03a ("odp-execute: Add ISA implementation of set_masked IPv4 action") > Signed-off-by: Emma Finn <emma.finn@intel.com> > Reported-by: Eelco Chaudron <echaudro@redhat.com> Hi Emma, Thanks for fixing this, however, it looks like the same issue also exists in IPv6. In addition, can you also add a test case for both IPv4 and IPv6? Thanks, Eelco > --- > lib/odp-execute-avx512.c | 2 ++ > 1 file changed, 2 insertions(+) > > diff --git a/lib/odp-execute-avx512.c b/lib/odp-execute-avx512.c > index 50c48bfd4..911fdd3ea 100644 > --- a/lib/odp-execute-avx512.c > +++ b/lib/odp-execute-avx512.c > @@ -366,6 +366,8 @@ avx512_get_delta(__m256i old_header, __m256i new_header) > 0xF, 0xF, 0xF, 0xF); > v_delta = _mm256_permutexvar_epi32(v_swap32a, v_delta); > > + v_delta = _mm256_hadd_epi32(v_delta, v_zeros); > + v_delta = _mm256_shuffle_epi8(v_delta, v_swap16a); > v_delta = _mm256_hadd_epi32(v_delta, v_zeros); > v_delta = _mm256_hadd_epi16(v_delta, v_zeros); > > -- > 2.25.1
diff --git a/lib/odp-execute-avx512.c b/lib/odp-execute-avx512.c index 50c48bfd4..911fdd3ea 100644 --- a/lib/odp-execute-avx512.c +++ b/lib/odp-execute-avx512.c @@ -366,6 +366,8 @@ avx512_get_delta(__m256i old_header, __m256i new_header) 0xF, 0xF, 0xF, 0xF); v_delta = _mm256_permutexvar_epi32(v_swap32a, v_delta); + v_delta = _mm256_hadd_epi32(v_delta, v_zeros); + v_delta = _mm256_shuffle_epi8(v_delta, v_swap16a); v_delta = _mm256_hadd_epi32(v_delta, v_zeros); v_delta = _mm256_hadd_epi16(v_delta, v_zeros);
The AVX implementation for calcualting checksum was not handling carry-over addition correctly in some cases. This patch adds an additional shuffle to add 16-bit padding to the final part of the calculation to handle such cases. Fixes: 92eb03f7b03a ("odp-execute: Add ISA implementation of set_masked IPv4 action") Signed-off-by: Emma Finn <emma.finn@intel.com> Reported-by: Eelco Chaudron <echaudro@redhat.com> --- lib/odp-execute-avx512.c | 2 ++ 1 file changed, 2 insertions(+)