* Re: [dpdk-dev] [PATCH] hash: fix thash gfni implementation
2021-11-09 17:24 [dpdk-dev] [PATCH] hash: fix thash gfni implementation Vladimir Medvedkin
@ 2021-11-09 17:34 ` Lance Richardson
2021-11-09 18:08 ` Ji, Kai
` (2 subsequent siblings)
3 siblings, 0 replies; 11+ messages in thread
From: Lance Richardson @ 2021-11-09 17:34 UTC (permalink / raw)
To: Vladimir Medvedkin
Cc: dev, Thomas Monjalon, David Marchand, Ananyev, Konstantin,
Yipeng Wang, Sameh Gobriel, Bruce Richardson
On Tue, Nov 9, 2021 at 12:25 PM Vladimir Medvedkin
<vladimir.medvedkin@intel.com> wrote:
>
> This patch replaces _mm512_set_epi8 with _mm512_set_epi32
> due to the lack of support by some compilers.
> Also 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.
>
> Fixes: 4fd8c4cb0de1 ("hash: add new Toeplitz hash implementation")
> Cc: vladimir.medvedkin@intel.com
>
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
> ---
> lib/hash/rte_thash_x86_gfni.h | 31 +++++++++++++------------------
> 1 file changed, 13 insertions(+), 18 deletions(-)
>
> diff --git a/lib/hash/rte_thash_x86_gfni.h b/lib/hash/rte_thash_x86_gfni.h
> index c2889c3734..7afbbedc08 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__)
> #define RTE_THASH_GFNI_DEFINED
>
> #define RTE_THASH_FIRST_ITER_MSK 0x0f0f0f0f0f0e0c08
> @@ -56,23 +56,18 @@ 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,
> + 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
>
This fixes the gcc 8 build issue on my CentOS 8 system.
Acked-by: Lance Richardson <lance.richardson@broadcom.com>
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [dpdk-dev] [PATCH] hash: fix thash gfni implementation
2021-11-09 17:24 [dpdk-dev] [PATCH] hash: fix thash gfni implementation Vladimir Medvedkin
2021-11-09 17:34 ` Lance Richardson
@ 2021-11-09 18:08 ` Ji, Kai
2021-11-12 14:17 ` [PATCH v2] " Vladimir Medvedkin
2021-11-16 14:33 ` [PATCH v3] " Vladimir Medvedkin
3 siblings, 0 replies; 11+ messages in thread
From: Ji, Kai @ 2021-11-09 18:08 UTC (permalink / raw)
To: Medvedkin, Vladimir, dev
Cc: thomas, david.marchand, Ananyev, Konstantin, lance.richardson,
Medvedkin, Vladimir, Wang, Yipeng1, Gobriel, Sameh, Richardson,
Bruce
> -----Original Message-----
> From: dev <dev-bounces@dpdk.org> On Behalf Of Vladimir Medvedkin
> Sent: Tuesday, November 9, 2021 5:25 PM
> To: dev@dpdk.org
> Cc: thomas@monjalon.net; david.marchand@redhat.com; Ananyev,
> Konstantin <konstantin.ananyev@intel.com>;
> lance.richardson@broadcom.com; Medvedkin, Vladimir
> <vladimir.medvedkin@intel.com>; Wang, Yipeng1
> <yipeng1.wang@intel.com>; Gobriel, Sameh <sameh.gobriel@intel.com>;
> Richardson, Bruce <bruce.richardson@intel.com>
> Subject: [dpdk-dev] [PATCH] hash: fix thash gfni implementation
>
> This patch replaces _mm512_set_epi8 with _mm512_set_epi32 due to the
> lack of support by some compilers.
> Also 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.
>
> Fixes: 4fd8c4cb0de1 ("hash: add new Toeplitz hash implementation")
> Cc: vladimir.medvedkin@intel.com
>
> Signed-off-by: Vladimir Medvedkin <vladimir.medvedkin@intel.com>
Acked-by: Ji, Kai <kai.ji@intel.com>
^ permalink raw reply [flat|nested] 11+ messages in thread* [PATCH v2] hash: fix thash gfni implementation
2021-11-09 17:24 [dpdk-dev] [PATCH] hash: fix thash gfni implementation Vladimir Medvedkin
2021-11-09 17:34 ` Lance Richardson
2021-11-09 18:08 ` Ji, Kai
@ 2021-11-12 14:17 ` Vladimir Medvedkin
2021-11-16 13:53 ` David Marchand
2021-11-16 14:33 ` [PATCH v3] " Vladimir Medvedkin
3 siblings, 1 reply; 11+ messages in thread
From: Vladimir Medvedkin @ 2021-11-12 14:17 UTC (permalink / raw)
To: dev
Cc: thomas, david.marchand, konstantin.ananyev, lance.richardson,
vladimir.medvedkin, Ji, Kai, Yipeng Wang, Sameh Gobriel,
Bruce Richardson
1. This patch replaces _mm512_set_epi8 with _mm512_set_epi32
due to the lack of support by some compilers.
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.
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.
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__)
#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,
+ 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
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [PATCH v2] hash: fix thash gfni implementation
2021-11-12 14:17 ` [PATCH v2] " Vladimir Medvedkin
@ 2021-11-16 13:53 ` David Marchand
2021-11-16 14:08 ` Walsh, Conor
` (2 more replies)
0 siblings, 3 replies; 11+ messages in thread
From: David Marchand @ 2021-11-16 13:53 UTC (permalink / raw)
To: Vladimir Medvedkin
Cc: dev, Thomas Monjalon, Ananyev, Konstantin, Lance Richardson, Ji,
Kai, Yipeng Wang, Sameh Gobriel, Bruce Richardson
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?
>
> 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_.
> #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...
> + 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
>
--
David Marchand
^ permalink raw reply [flat|nested] 11+ messages in thread* RE: [PATCH v2] hash: fix thash gfni implementation
2021-11-16 13:53 ` David Marchand
@ 2021-11-16 14:08 ` Walsh, Conor
2021-11-16 14:10 ` Bruce Richardson
2021-11-16 14:18 ` Medvedkin, Vladimir
2 siblings, 0 replies; 11+ messages in thread
From: Walsh, Conor @ 2021-11-16 14:08 UTC (permalink / raw)
To: David Marchand, Medvedkin, Vladimir
Cc: dev, Thomas Monjalon, Ananyev, Konstantin, Lance Richardson, Ji,
Ji, Kai, Wang, Yipeng1, Gobriel, Sameh, Richardson, Bruce
> From: David Marchand <david.marchand@redhat.com>
> Sent: Tuesday 16 November 2021 13:54
> To: Medvedkin, Vladimir <vladimir.medvedkin@intel.com>
> Cc: dev <dev@dpdk.org>; Thomas Monjalon <thomas@monjalon.net>;
> Ananyev, Konstantin <konstantin.ananyev@intel.com>; Lance Richardson
> <lance.richardson@broadcom.com>; Ji@dpdk.org; Ji, Kai <kai.ji@intel.com>;
> Wang, Yipeng1 <yipeng1.wang@intel.com>; Gobriel, Sameh
> <sameh.gobriel@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>
> Subject: Re: [PATCH v2] hash: fix thash gfni implementation
>
> 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?
Hi David,
I ran into this issue and reported it to Vladimir.
Currently when you try to cross compile main for 32-bit (e.g. with test-meson-builds.sh) its broken because of this issue, this affects our ability to test.
So in that scenario it sees that GFNI and AVX-512 are on the system but cant build because its trying to cross compile 64-bit only functions for a 32-bit build.
Thanks,
Conor.
>
>
> >
> > 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_.
>
>
> > #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...
>
>
> > + 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
> >
>
>
> --
> David Marchand
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [PATCH v2] hash: fix thash gfni implementation
2021-11-16 13:53 ` David Marchand
2021-11-16 14:08 ` Walsh, Conor
@ 2021-11-16 14:10 ` Bruce Richardson
2021-11-16 14:17 ` David Marchand
2021-11-16 14:18 ` Medvedkin, Vladimir
2 siblings, 1 reply; 11+ messages in thread
From: Bruce Richardson @ 2021-11-16 14:10 UTC (permalink / raw)
To: David Marchand
Cc: Vladimir Medvedkin, dev, Thomas Monjalon, Ananyev, Konstantin,
Lance Richardson, Ji, Kai, Yipeng Wang, Sameh Gobriel
On Tue, Nov 16, 2021 at 02:53:49PM +0100, 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?
>
The processor would be 64-bit but the code could well be 32-bit, and even
then can still use these AVX512 instruction sets.
/Bruce
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH v2] hash: fix thash gfni implementation
2021-11-16 14:10 ` Bruce Richardson
@ 2021-11-16 14:17 ` David Marchand
0 siblings, 0 replies; 11+ messages in thread
From: David Marchand @ 2021-11-16 14:17 UTC (permalink / raw)
To: Bruce Richardson
Cc: Vladimir Medvedkin, dev, Thomas Monjalon, Ananyev, Konstantin,
Lance Richardson, Ji, Kai, Yipeng Wang, Sameh Gobriel
On Tue, Nov 16, 2021 at 3:10 PM Bruce Richardson
<bruce.richardson@intel.com> wrote:
>
> On Tue, Nov 16, 2021 at 02:53:49PM +0100, 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?
> >
> The processor would be 64-bit but the code could well be 32-bit, and even
> then can still use these AVX512 instruction sets.
Oh, thanks for the explanation.
--
David Marchand
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [PATCH v2] hash: fix thash gfni implementation
2021-11-16 13:53 ` David Marchand
2021-11-16 14:08 ` Walsh, Conor
2021-11-16 14:10 ` Bruce Richardson
@ 2021-11-16 14:18 ` Medvedkin, Vladimir
2 siblings, 0 replies; 11+ messages in thread
From: Medvedkin, Vladimir @ 2021-11-16 14:18 UTC (permalink / raw)
To: David Marchand
Cc: dev, Thomas Monjalon, Ananyev, Konstantin, Lance Richardson, Ji,
Kai, Yipeng Wang, Sameh Gobriel, Bruce Richardson
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
^ permalink raw reply [flat|nested] 11+ messages in thread
* [PATCH v3] hash: fix thash gfni implementation
2021-11-09 17:24 [dpdk-dev] [PATCH] hash: fix thash gfni implementation Vladimir Medvedkin
` (2 preceding siblings ...)
2021-11-12 14:17 ` [PATCH v2] " Vladimir Medvedkin
@ 2021-11-16 14:33 ` Vladimir Medvedkin
2021-11-17 9:20 ` David Marchand
3 siblings, 1 reply; 11+ messages in thread
From: Vladimir Medvedkin @ 2021-11-16 14:33 UTC (permalink / raw)
To: dev
Cc: thomas, david.marchand, konstantin.ananyev, lance.richardson,
vladimir.medvedkin, Ji Kai, Yipeng Wang, Sameh Gobriel,
Bruce Richardson
1. This patch replaces _mm512_set_epi8 with _mm512_set_epi32
due to the lack of support by some compilers.
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.
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.
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 | 47 ++++++++++++++++++++---------------
1 file changed, 27 insertions(+), 20 deletions(-)
diff --git a/lib/hash/rte_thash_x86_gfni.h b/lib/hash/rte_thash_x86_gfni.h
index c2889c3734..8e3da51ecd 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__)
#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 +67,19 @@ 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(0x07060504, 0x07060504,
+ 0x06050403, 0x06050403,
+ 0x05040302, 0x05040302,
+ 0x04030201, 0x04030201,
+ 0x03020100, 0x03020100,
+ 0x020100FF, 0x020100FF,
+ 0x0100FFFE, 0x0100FFFE,
+ 0x00FFFEFD, 0x00FFFEFD);
+ const __m512i rewind_idx = _mm512_set_epi32(0, 0, 0, 0, 0, 0, 0, 0,
+ 0, 0,
+ 0x0000003B, 0x0000003B,
+ 0x00003B3A, 0x00003B3A,
+ 0x003B3A39, 0x003B3A39);
const __mmask64 rewind_mask = RTE_THASH_REWIND_MSK;
const __m512i shift_8 = _mm512_set1_epi8(8);
__m512i xor_acc = _mm512_setzero_si512();
@@ -214,7 +221,7 @@ rte_thash_gfni_bulk(const uint64_t *mtrx, int len, uint8_t *tuple[],
}
}
-#endif /* _GFNI_ */
+#endif /* __GFNI__ && __AVX512F__ */
#ifdef __cplusplus
}
--
2.25.1
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [PATCH v3] hash: fix thash gfni implementation
2021-11-16 14:33 ` [PATCH v3] " Vladimir Medvedkin
@ 2021-11-17 9:20 ` David Marchand
0 siblings, 0 replies; 11+ messages in thread
From: David Marchand @ 2021-11-17 9:20 UTC (permalink / raw)
To: Vladimir Medvedkin
Cc: dev, Thomas Monjalon, Ananyev, Konstantin, Lance Richardson,
Yipeng Wang, Sameh Gobriel, Bruce Richardson, Ji, Kai
On Tue, Nov 16, 2021 at 3:33 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.
> 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.
> 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.
>
> Fixes: 4fd8c4cb0de1 ("hash: add new Toeplitz hash implementation")
>
> 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>
Afaics in git history, that's "Kai Ji".
I updated indentation and applied, thanks.
--
David Marchand
^ permalink raw reply [flat|nested] 11+ messages in thread