Re: [PATCH v2] hash: fix thash gfni implementation

["Medvedkin, Vladimir" <vladimir.medvedkin@intel.com>]

Hi David,

On 16/11/2021 14:53, David Marchand wrote:
> On Fri, Nov 12, 2021 at 3:17 PM Vladimir Medvedkin
> <vladimir.medvedkin@intel.com> wrote:
>>
>> 1. This patch replaces _mm512_set_epi8 with _mm512_set_epi32
>> due to the lack of support by some compilers.
> 
> Ok, it was the initial report from Lance.
> 
>> 2. This patch checks if AVX512F is supported along with GFNI.
>> This is done if the code is built on a platform that supports GFNI,
>> but does not support AVX512.
> 
> Ok.
> 
>> 3. Also this patch fixes compilation problems on 32bit arch due to
>> lack of support for _mm_extract_epi64() by implementing XOR folding
>> with _mm_extract_epi32() on 32-bit arch.
> 
> This code is under a #if defined(__GFNI__) && defined(__AVX512F__).
> 
> Does such a 32 bits processor exist, that supports AVX512 and GFNI?
> 
> 

This breaks the 32 bit build.

>>
>> Fixes: 4fd8c4cb0de1 ("hash: add new Toeplitz hash implementation")
>> Cc: vladimir.medvedkin@intel.com
>>
>> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
>> Acked-by: Lance Richardson <lance.richardson@broadcom.com>
>> Acked-by: Ji, Kai <kai.ji@intel.com>
>> ---
>>   lib/hash/rte_thash_x86_gfni.h | 44 ++++++++++++++++++++---------------
>>   1 file changed, 25 insertions(+), 19 deletions(-)
>>
>> diff --git a/lib/hash/rte_thash_x86_gfni.h b/lib/hash/rte_thash_x86_gfni.h
>> index c2889c3734..987dec4988 100644
>> --- a/lib/hash/rte_thash_x86_gfni.h
>> +++ b/lib/hash/rte_thash_x86_gfni.h
>> @@ -18,7 +18,7 @@
>>   extern "C" {
>>   #endif
>>
>> -#ifdef __GFNI__
>> +#if defined(__GFNI__) && defined(__AVX512F__)
> 
> Please update #endif comments accordingly, or remove invalid/obsolete
> comment about _GFNI_.
> 

Sure, will do.

> 
>>   #define RTE_THASH_GFNI_DEFINED
>>
>>   #define RTE_THASH_FIRST_ITER_MSK       0x0f0f0f0f0f0e0c08
>> @@ -33,7 +33,6 @@ __rte_thash_xor_reduce(__m512i xor_acc, uint32_t *val_1, uint32_t *val_2)
>>   {
>>          __m256i tmp_256_1, tmp_256_2;
>>          __m128i tmp128_1, tmp128_2;
>> -       uint64_t tmp_1, tmp_2;
>>
>>          tmp_256_1 = _mm512_castsi512_si256(xor_acc);
>>          tmp_256_2 = _mm512_extracti32x8_epi32(xor_acc, 1);
>> @@ -43,12 +42,24 @@ __rte_thash_xor_reduce(__m512i xor_acc, uint32_t *val_1, uint32_t *val_2)
>>          tmp128_2 = _mm256_extracti32x4_epi32(tmp_256_1, 1);
>>          tmp128_1 = _mm_xor_si128(tmp128_1, tmp128_2);
>>
>> +#ifdef RTE_ARCH_X86_64
>> +       uint64_t tmp_1, tmp_2;
>>          tmp_1 = _mm_extract_epi64(tmp128_1, 0);
>>          tmp_2 = _mm_extract_epi64(tmp128_1, 1);
>>          tmp_1 ^= tmp_2;
>>
>>          *val_1 = (uint32_t)tmp_1;
>>          *val_2 = (uint32_t)(tmp_1 >> 32);
>> +#else
>> +       uint32_t tmp_1, tmp_2;
>> +       tmp_1 = _mm_extract_epi32(tmp128_1, 0);
>> +       tmp_2 = _mm_extract_epi32(tmp128_1, 1);
>> +       tmp_1 ^= _mm_extract_epi32(tmp128_1, 2);
>> +       tmp_2 ^= _mm_extract_epi32(tmp128_1, 3);
>> +
>> +       *val_1 = tmp_1;
>> +       *val_2 = tmp_2;
>> +#endif
>>   }
>>
>>   __rte_internal
>> @@ -56,23 +67,18 @@ static inline __m512i
>>   __rte_thash_gfni(const uint64_t *mtrx, const uint8_t *tuple,
>>          const uint8_t *secondary_tuple, int len)
>>   {
>> -       __m512i permute_idx = _mm512_set_epi8(7, 6, 5, 4, 7, 6, 5, 4,
>> -                                               6, 5, 4, 3, 6, 5, 4, 3,
>> -                                               5, 4, 3, 2, 5, 4, 3, 2,
>> -                                               4, 3, 2, 1, 4, 3, 2, 1,
>> -                                               3, 2, 1, 0, 3, 2, 1, 0,
>> -                                               2, 1, 0, -1, 2, 1, 0, -1,
>> -                                               1, 0, -1, -2, 1, 0, -1, -2,
>> -                                               0, -1, -2, -3, 0, -1, -2, -3);
>> -
>> -       const __m512i rewind_idx = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
>> -                                               0, 0, 0, 0, 0, 0, 0, 0,
>> -                                               0, 0, 0, 0, 0, 0, 0, 0,
>> -                                               0, 0, 0, 0, 0, 0, 0, 0,
>> -                                               0, 0, 0, 0, 0, 0, 0, 0,
>> -                                               0, 0, 0, 59, 0, 0, 0, 59,
>> -                                               0, 0, 59, 58, 0, 0, 59, 58,
>> -                                               0, 59, 58, 57, 0, 59, 58, 57);
>> +       __m512i permute_idx = _mm512_set_epi32(0x7060504, 0x7060504,
> 
> Nit: it is easier to read fully expanded 32 bits values, like
> 0x07060504 instead of 0x7060504
> Etc...
> 

Will fix in v3.

> 
>> +                                               0x6050403, 0x6050403,
>> +                                               0x5040302, 0x5040302,
>> +                                               0x4030201, 0x4030201,
>> +                                               0x3020100, 0x3020100,
>> +                                               0x20100FF, 0x20100FF,
>> +                                               0x100FFFE, 0x100FFFE,
>> +                                               0xFFFEFD, 0xFFFEFD);
>> +       const __m512i rewind_idx = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0,
>> +                                                       0, 0, 0x3B, 0x3B,
>> +                                                       0x3B3A, 0x3B3A,
>> +                                                       0x3B3A39, 0x3B3A39);
>>          const __mmask64 rewind_mask = RTE_THASH_REWIND_MSK;
>>          const __m512i shift_8 = _mm512_set1_epi8(8);
>>          __m512i xor_acc = _mm512_setzero_si512();
>> --
>> 2.25.1
>>
> 
> 

-- 
Regards,
Vladimir