[v5,2/2] net: add support for AVX512/VPCLMULQDQ based CRC
Checks
Commit Message
This patch enables the optimized calculation of CRC32-Ethernet and
CRC16-CCITT using the AVX512 and VPCLMULQDQ instruction sets. This CRC
implementation is built if the compiler supports the required instruction
sets. It is selected at run-time if the host CPU, again, supports the
required instruction sets.
Signed-off-by: Mairtin o Loingsigh <mairtin.oloingsigh@intel.com>
Signed-off-by: David Coyle <david.coyle@intel.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
---
app/test/test_crc.c | 11 +-
config/x86/meson.build | 6 +-
doc/guides/rel_notes/release_20_11.rst | 2 +
lib/librte_net/meson.build | 55 +++++
lib/librte_net/net_crc.h | 11 +
lib/librte_net/net_crc_avx512.c | 423 +++++++++++++++++++++++++++++++++
lib/librte_net/rte_net_crc.c | 46 ++++
lib/librte_net/rte_net_crc.h | 4 +-
8 files changed, 554 insertions(+), 4 deletions(-)
create mode 100644 lib/librte_net/net_crc_avx512.c
Comments
> -----Original Message-----
> From: O'loingsigh, Mairtin <mairtin.oloingsigh@intel.com>
> Sent: Friday, October 9, 2020 2:51 PM
> To: Singh, Jasvinder <jasvinder.singh@intel.com>; Richardson, Bruce
> <bruce.richardson@intel.com>; De Lara Guarch, Pablo
> <pablo.de.lara.guarch@intel.com>; Ananyev, Konstantin
> <konstantin.ananyev@intel.com>
> Cc: dev@dpdk.org; Ryan, Brendan <brendan.ryan@intel.com>; O'loingsigh,
> Mairtin <mairtin.oloingsigh@intel.com>; Coyle, David
> <david.coyle@intel.com>
> Subject: [PATCH v5 2/2] net: add support for AVX512/VPCLMULQDQ based
> CRC
>
> This patch enables the optimized calculation of CRC32-Ethernet and CRC16-
> CCITT using the AVX512 and VPCLMULQDQ instruction sets. This CRC
> implementation is built if the compiler supports the required instruction sets.
> It is selected at run-time if the host CPU, again, supports the required
> instruction sets.
>
> Signed-off-by: Mairtin o Loingsigh <mairtin.oloingsigh@intel.com>
> Signed-off-by: David Coyle <david.coyle@intel.com>
> Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
> ---
> app/test/test_crc.c | 11 +-
> config/x86/meson.build | 6 +-
> doc/guides/rel_notes/release_20_11.rst | 2 +
> lib/librte_net/meson.build | 55 +++++
> lib/librte_net/net_crc.h | 11 +
> lib/librte_net/net_crc_avx512.c | 423
> +++++++++++++++++++++++++++++++++
> lib/librte_net/rte_net_crc.c | 46 ++++
> lib/librte_net/rte_net_crc.h | 4 +-
> 8 files changed, 554 insertions(+), 4 deletions(-) create mode 100644
> lib/librte_net/net_crc_avx512.c
>
> diff --git a/app/test/test_crc.c b/app/test/test_crc.c index
> f8a74e04e..bf1d34435 100644
> --- a/app/test/test_crc.c
> +++ b/app/test/test_crc.c
> @@ -1,5 +1,5 @@
> /* SPDX-License-Identifier: BSD-3-Clause
> - * Copyright(c) 2017 Intel Corporation
> + * Copyright(c) 2017-2020 Intel Corporation
> */
>
> #include "test.h"
> @@ -149,6 +149,15 @@ test_crc(void)
> return ret;
> }
>
> + /* set CRC avx512 mode */
> + rte_net_crc_set_alg(RTE_NET_CRC_AVX512);
> +
> + ret = test_crc_calc();
> + if (ret < 0) {
> + printf("test crc (x86_64 AVX512): failed (%d)\n", ret);
> + return ret;
> + }
> +
> /* set CRC neon mode */
> rte_net_crc_set_alg(RTE_NET_CRC_NEON);
>
> diff --git a/config/x86/meson.build b/config/x86/meson.build index
> fea4d5403..172b72b72 100644
> --- a/config/x86/meson.build
> +++ b/config/x86/meson.build
> @@ -1,5 +1,5 @@
> # SPDX-License-Identifier: BSD-3-Clause -# Copyright(c) 2017-2019 Intel
> Corporation
> +# Copyright(c) 2017-2020 Intel Corporation
>
> # get binutils version for the workaround of Bug 97 if not is_windows @@ -
> 23,7 +23,9 @@ endforeach
>
> optional_flags = ['AES', 'PCLMUL',
> 'AVX', 'AVX2', 'AVX512F',
> - 'RDRND', 'RDSEED']
> + 'RDRND', 'RDSEED',
> + 'AVX512BW', 'AVX512DQ',
> + 'AVX512VL', 'VPCLMULQDQ']
> foreach f:optional_flags
> if cc.get_define('__@0@__'.format(f), args: machine_args) == '1'
> if f == 'PCLMUL' # special case flags with different defines diff
> --git a/doc/guides/rel_notes/release_20_11.rst
> b/doc/guides/rel_notes/release_20_11.rst
> index b77297f7e..5eda680d5 100644
> --- a/doc/guides/rel_notes/release_20_11.rst
> +++ b/doc/guides/rel_notes/release_20_11.rst
> @@ -58,6 +58,8 @@ New Features
> * **Updated CRC modules of rte_net library.**
>
> * Added run-time selection of the optimal architecture-specific CRC path.
> + * Added optimized implementations of CRC32-Ethernet and CRC16-CCITT
> + using the AVX512 and VPCLMULQDQ instruction sets.
>
> * **Updated Broadcom bnxt driver.**
>
> diff --git a/lib/librte_net/meson.build b/lib/librte_net/meson.build index
> fa439b9e5..6c96b361a 100644
> --- a/lib/librte_net/meson.build
> +++ b/lib/librte_net/meson.build
> @@ -24,18 +24,62 @@ deps += ['mbuf']
> if dpdk_conf.has('RTE_ARCH_X86_64')
> net_crc_sse42_cpu_support = (
> cc.get_define('__PCLMUL__', args: machine_args) != '')
> + net_crc_avx512_cpu_support = (
> + cc.get_define('__AVX512F__', args: machine_args) != '' and
> + cc.get_define('__AVX512BW__', args: machine_args) != '' and
> + cc.get_define('__AVX512DQ__', args: machine_args) != '' and
> + cc.get_define('__AVX512VL__', args: machine_args) != '' and
> + cc.get_define('__VPCLMULQDQ__', args: machine_args) != '')
> +
> net_crc_sse42_cc_support = (
> cc.has_argument('-mpclmul') and cc.has_argument('-maes'))
> + net_crc_avx512_cc_support = (
> + not machine_args.contains('-mno-avx512f') and
> + cc.has_argument('-mavx512f') and
> + cc.has_argument('-mavx512bw') and
> + cc.has_argument('-mavx512dq') and
> + cc.has_argument('-mavx512vl') and
> + cc.has_argument('-mvpclmulqdq') and
> + cc.has_argument('-mavx2') and
> + cc.has_argument('-mavx'))
>
> build_static_net_crc_sse42_lib = 0
> + build_static_net_crc_avx512_lib = 0
>
> if net_crc_sse42_cpu_support == true
> sources += files('net_crc_sse.c')
> cflags += ['-DCC_X86_64_SSE42_PCLMULQDQ_SUPPORT']
> + if net_crc_avx512_cpu_support == true
> + sources += files('net_crc_avx512.c')
> + cflags += ['-
> DCC_X86_64_AVX512_VPCLMULQDQ_SUPPORT']
> + elif net_crc_avx512_cc_support == true
> + build_static_net_crc_avx512_lib = 1
> + net_crc_avx512_lib_cflags = ['-mavx512f',
> + '-mavx512bw',
> + '-mavx512dq',
> + '-mavx512vl',
> + '-mvpclmulqdq',
> + '-mavx2',
> + '-mavx']
> + cflags += ['-
> DCC_X86_64_AVX512_VPCLMULQDQ_SUPPORT']
> + endif
> elif net_crc_sse42_cc_support == true
> build_static_net_crc_sse42_lib = 1
> net_crc_sse42_lib_cflags = ['-mpclmul', '-maes']
> cflags += ['-DCC_X86_64_SSE42_PCLMULQDQ_SUPPORT']
> + if net_crc_avx512_cc_support == true
> + build_static_net_crc_avx512_lib = 1
> + net_crc_avx512_lib_cflags = ['-mpclmul',
> + '-maes',
> + '-mavx512f',
> + '-mavx512bw',
> + '-mavx512dq',
> + '-mavx512vl',
> + '-mvpclmulqdq',
> + '-mavx2',
> + '-mavx']
> + cflags += ['-
> DCC_X86_64_AVX512_VPCLMULQDQ_SUPPORT']
> + endif
> endif
>
> if build_static_net_crc_sse42_lib == 1 @@ -47,6 +91,17 @@ if
> dpdk_conf.has('RTE_ARCH_X86_64')
> net_crc_sse42_lib_cflags])
> objs += net_crc_sse42_lib.extract_objects('net_crc_sse.c')
> endif
> +
> + if build_static_net_crc_avx512_lib == 1
> + net_crc_avx512_lib = static_library(
> + 'net_crc_avx512_lib',
> + 'net_crc_avx512.c',
> + dependencies: static_rte_eal,
> + c_args: [cflags,
> + net_crc_avx512_lib_cflags])
> + objs +=
> net_crc_avx512_lib.extract_objects('net_crc_avx512.c')
> + endif
> +
> elif (dpdk_conf.has('RTE_ARCH_ARM64') and
> cc.get_define('__ARM_FEATURE_CRYPTO', args:
> machine_args) != '')
> sources += files('net_crc_neon.c')
> diff --git a/lib/librte_net/net_crc.h b/lib/librte_net/net_crc.h index
> a1578a56c..7a74d5406 100644
> --- a/lib/librte_net/net_crc.h
> +++ b/lib/librte_net/net_crc.h
> @@ -20,6 +20,17 @@ rte_crc16_ccitt_sse42_handler(const uint8_t *data,
> uint32_t data_len); uint32_t rte_crc32_eth_sse42_handler(const uint8_t
> *data, uint32_t data_len);
>
> +/* AVX512 */
> +
> +void
> +rte_net_crc_avx512_init(void);
> +
> +uint32_t
> +rte_crc16_ccitt_avx512_handler(const uint8_t *data, uint32_t data_len);
> +
> +uint32_t
> +rte_crc32_eth_avx512_handler(const uint8_t *data, uint32_t data_len);
> +
> /* NEON */
>
> void
> diff --git a/lib/librte_net/net_crc_avx512.c b/lib/librte_net/net_crc_avx512.c
> new file mode 100644 index 000000000..3740fe3c9
> --- /dev/null
> +++ b/lib/librte_net/net_crc_avx512.c
> @@ -0,0 +1,423 @@
> +/* SPDX-License-Identifier: BSD-3-Clause
> + * Copyright(c) 2020 Intel Corporation
> + */
> +
> +#include <string.h>
> +
> +#include <rte_common.h>
> +#include <rte_branch_prediction.h>
> +#include <rte_cpuflags.h>
> +
> +#include "net_crc.h"
> +
> +#include <x86intrin.h>
> +
> +/* VPCLMULQDQ CRC computation context structure */ struct
> +crc_vpclmulqdq_ctx {
> + __m512i rk1_rk2;
> + __m512i rk3_rk4;
> + __m512i fold_7x128b;
> + __m512i fold_3x128b;
> + __m128i rk5_rk6;
> + __m128i rk7_rk8;
> + __m128i fold_1x128b;
> +};
> +
> +static struct crc_vpclmulqdq_ctx crc32_eth __rte_aligned(64); static
> +struct crc_vpclmulqdq_ctx crc16_ccitt __rte_aligned(64);
> +
> +static uint16_t byte_len_to_mask_table[] = {
> + 0x0000, 0x0001, 0x0003, 0x0007,
> + 0x000f, 0x001f, 0x003f, 0x007f,
> + 0x00ff, 0x01ff, 0x03ff, 0x07ff,
> + 0x0fff, 0x1fff, 0x3fff, 0x7fff,
> + 0xffff};
> +
> +static const uint8_t shf_table[32] __rte_aligned(16) = {
> + 0x00, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
> + 0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
> + 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
> + 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f };
> +
> +static const uint32_t mask[4] __rte_aligned(16) = {
> + 0xffffffff, 0xffffffff, 0x00000000, 0x00000000 };
> +
> +static const uint32_t mask2[4] __rte_aligned(16) = {
> + 0x00000000, 0xffffffff, 0xffffffff, 0xffffffff };
> +
> +static __rte_always_inline __m512i
> +crcr32_folding_round(__m512i data_block, __m512i precomp, __m512i
> fold)
> +{
> + __m512i tmp0, tmp1;
> +
> + tmp0 = _mm512_clmulepi64_epi128(fold, precomp, 0x01);
> + tmp1 = _mm512_clmulepi64_epi128(fold, precomp, 0x10);
> +
> + return _mm512_ternarylogic_epi64(tmp0, tmp1, data_block, 0x96); }
> +
> +static __rte_always_inline __m128i
> +crc32_fold_128(__m512i fold0, __m512i fold1,
> + const struct crc_vpclmulqdq_ctx *params) {
> + __m128i res, res2;
> + __m256i a;
> + __m512i tmp0, tmp1, tmp2, tmp3;
> + __m512i tmp4;
> +
> + tmp0 = _mm512_clmulepi64_epi128(fold0, params->fold_7x128b,
> 0x01);
> + tmp1 = _mm512_clmulepi64_epi128(fold0, params->fold_7x128b,
> 0x10);
> +
> + res = _mm512_extracti64x2_epi64(fold1, 3);
> + tmp4 = _mm512_maskz_broadcast_i32x4(0xF, res);
> +
> + tmp2 = _mm512_clmulepi64_epi128(fold1, params->fold_3x128b,
> 0x01);
> + tmp3 = _mm512_clmulepi64_epi128(fold1, params->fold_3x128b,
> 0x10);
> +
> + tmp0 = _mm512_ternarylogic_epi64(tmp0, tmp1, tmp2, 0x96);
> + tmp0 = _mm512_ternarylogic_epi64(tmp0, tmp3, tmp4, 0x96);
> +
> + tmp1 = _mm512_shuffle_i64x2(tmp0, tmp0, 0x4e);
> +
> + a = _mm256_xor_si256(*(__m256i *)&tmp1, *(__m256i *)&tmp0);
> + res = _mm256_extracti64x2_epi64(a, 1);
> + res2 = _mm_xor_si128(res, *(__m128i *)&a);
> +
> + return res2;
> +}
> +
> +static __rte_always_inline __m128i
> +last_two_xmm(const uint8_t *data, uint32_t data_len, uint32_t n, __m128i
> res,
> + const struct crc_vpclmulqdq_ctx *params) {
> + uint32_t offset;
> + __m128i res2, res3, res4, pshufb_shf;
> +
> + const uint32_t mask3[4] __rte_aligned(16) = {
> + 0x80808080, 0x80808080, 0x80808080, 0x80808080
> + };
> +
> + res2 = res;
> + offset = data_len - n;
> + res3 = _mm_loadu_si128((const __m128i *)&data[n+offset-16]);
> +
> + pshufb_shf = _mm_loadu_si128((const __m128i *)
> + (shf_table + (data_len-n)));
> +
> + res = _mm_shuffle_epi8(res, pshufb_shf);
> + pshufb_shf = _mm_xor_si128(pshufb_shf,
> + _mm_load_si128((const __m128i *) mask3));
> + res2 = _mm_shuffle_epi8(res2, pshufb_shf);
> +
> + res2 = _mm_blendv_epi8(res2, res3, pshufb_shf);
> +
> + res4 = _mm_clmulepi64_si128(res, params->fold_1x128b, 0x01);
> + res = _mm_clmulepi64_si128(res, params->fold_1x128b, 0x10);
> + res = _mm_ternarylogic_epi64(res, res2, res4, 0x96);
> +
> + return res;
> +}
> +
> +static __rte_always_inline __m128i
> +done_128(__m128i res, const struct crc_vpclmulqdq_ctx *params) {
> + __m128i res1;
> +
> + res1 = res;
> +
> + res = _mm_clmulepi64_si128(res, params->rk5_rk6, 0x0);
> + res1 = _mm_srli_si128(res1, 8);
> + res = _mm_xor_si128(res, res1);
> +
> + res1 = res;
> + res = _mm_slli_si128(res, 4);
> + res = _mm_clmulepi64_si128(res, params->rk5_rk6, 0x10);
> + res = _mm_xor_si128(res, res1);
> +
> + return res;
> +}
> +
> +static __rte_always_inline uint32_t
> +barrett_reduction(__m128i data64, const struct crc_vpclmulqdq_ctx
> +*params) {
> + __m128i tmp0, tmp1;
> +
> + data64 = _mm_and_si128(data64, *(const __m128i *)mask2);
> + tmp0 = data64;
> + tmp1 = data64;
> +
> + data64 = _mm_clmulepi64_si128(tmp0, params->rk7_rk8, 0x0);
> + data64 = _mm_ternarylogic_epi64(data64, tmp1, *(const __m128i
> *)mask,
> + 0x28);
> +
> + tmp1 = data64;
> + data64 = _mm_clmulepi64_si128(data64, params->rk7_rk8, 0x10);
> + data64 = _mm_ternarylogic_epi64(data64, tmp1, tmp0, 0x96);
> +
> + return _mm_extract_epi32(data64, 2);
> +}
> +
> +static __rte_always_inline void
> +reduction_loop(__m128i *fold, int *len, const uint8_t *data, uint32_t *n,
> + const struct crc_vpclmulqdq_ctx *params) {
> + __m128i tmp, tmp1;
> +
> + tmp = _mm_clmulepi64_si128(*fold, params->fold_1x128b, 0x1);
> + *fold = _mm_clmulepi64_si128(*fold, params->fold_1x128b, 0x10);
> + *fold = _mm_xor_si128(*fold, tmp);
> + tmp1 = _mm_loadu_si128((const __m128i *)&data[*n]);
> + *fold = _mm_xor_si128(*fold, tmp1);
> + *n += 16;
> + *len -= 16;
> +}
> +
> +static __rte_always_inline uint32_t
> +crc32_eth_calc_vpclmulqdq(const uint8_t *data, uint32_t data_len, uint32_t
> crc,
> + const struct crc_vpclmulqdq_ctx *params) {
> + __m128i res, d, b;
> + __m512i temp, k;
> + __m512i qw0 = _mm512_set1_epi64(0), qw1, qw2, qw3;
> + __m512i fold0, fold1, fold2, fold3;
> + __mmask16 mask;
> + uint32_t n = 0;
> + int reduction = 0;
> +
> + /* Get CRC init value */
> + b = _mm_cvtsi32_si128(crc);
> + temp = _mm512_castsi128_si512(b);
> +
> + if (data_len > 255) {
> + fold0 = _mm512_loadu_si512((const __m512i *)data);
> + fold1 = _mm512_loadu_si512((const __m512i *)(data+64));
> + fold2 = _mm512_loadu_si512((const __m512i *)(data+128));
> + fold3 = _mm512_loadu_si512((const __m512i *)(data+192));
> + fold0 = _mm512_xor_si512(fold0, temp);
> +
> + /* Main folding loop */
> + k = params->rk1_rk2;
> + for (n = 256; (n + 256) <= data_len; n += 256) {
> + qw0 = _mm512_loadu_si512((const __m512i
> *)&data[n]);
> + qw1 = _mm512_loadu_si512((const __m512i *)
> + &(data[n+64]));
> + qw2 = _mm512_loadu_si512((const __m512i *)
> + &(data[n+128]));
> + qw3 = _mm512_loadu_si512((const __m512i *)
> + &(data[n+192]));
> + fold0 = crcr32_folding_round(qw0, k, fold0);
> + fold1 = crcr32_folding_round(qw1, k, fold1);
> + fold2 = crcr32_folding_round(qw2, k, fold2);
> + fold3 = crcr32_folding_round(qw3, k, fold3);
> + }
> +
> + /* 256 to 128 fold */
> + k = params->rk3_rk4;
> + fold0 = crcr32_folding_round(fold2, k, fold0);
> + fold1 = crcr32_folding_round(fold3, k, fold1);
> +
> + res = crc32_fold_128(fold0, fold1, params);
> +
> + reduction = 240 - ((n+256)-data_len);
> +
> + while (reduction > 0)
> + reduction_loop(&res, &reduction, data, &n,
> + params);
> +
> + reduction += 16;
> +
> + if (n != data_len)
> + res = last_two_xmm(data, data_len, n, res,
> + params);
> + } else {
> + if (data_len > 31) {
> + res = _mm_cvtsi32_si128(crc);
> + d = _mm_loadu_si128((const __m128i *)data);
> + res = _mm_xor_si128(res, d);
> + n += 16;
> +
> + reduction = 240 - ((n+256)-data_len);
> +
> + while (reduction > 0)
> + reduction_loop(&res, &reduction, data, &n,
> + params);
> +
> + if (n != data_len)
> + res = last_two_xmm(data, data_len, n, res,
> + params);
> + } else if (data_len > 16) {
> + res = _mm_cvtsi32_si128(crc);
> + d = _mm_loadu_si128((const __m128i *)data);
> + res = _mm_xor_si128(res, d);
> + n += 16;
> +
> + if (n != data_len)
> + res = last_two_xmm(data, data_len, n, res,
> + params);
> + } else if (data_len == 16) {
> + res = _mm_cvtsi32_si128(crc);
> + d = _mm_loadu_si128((const __m128i *)data);
> + res = _mm_xor_si128(res, d);
> + } else {
> + res = _mm_cvtsi32_si128(crc);
> + mask = byte_len_to_mask_table[data_len];
> + d = _mm_maskz_loadu_epi8(mask, data);
> + res = _mm_xor_si128(res, d);
> +
> + if (data_len > 3) {
> + d = _mm_loadu_si128((const __m128i *)
> + &shf_table[data_len]);
> + res = _mm_shuffle_epi8(res, d);
> + } else if (data_len > 2) {
> + res = _mm_slli_si128(res, 5);
> + goto do_barrett_reduction;
> + } else if (data_len > 1) {
> + res = _mm_slli_si128(res, 6);
> + goto do_barrett_reduction;
> + } else if (data_len > 0) {
> + res = _mm_slli_si128(res, 7);
> + goto do_barrett_reduction;
> + } else {
> + /* zero length case */
> + return crc;
> + }
> + }
> + }
> +
> + res = done_128(res, params);
> +
> +do_barrett_reduction:
> + n = barrett_reduction(res, params);
> +
> + return n;
> +}
> +
> +static void
> +crc32_load_init_constants(void)
> +{
> + __m128i a;
> + /* fold constants */
> + uint64_t c0 = 0x00000000e95c1271;
> + uint64_t c1 = 0x00000000ce3371cb;
> + uint64_t c2 = 0x00000000910eeec1;
> + uint64_t c3 = 0x0000000033fff533;
> + uint64_t c4 = 0x000000000cbec0ed;
> + uint64_t c5 = 0x0000000031f8303f;
> + uint64_t c6 = 0x0000000057c54819;
> + uint64_t c7 = 0x00000000df068dc2;
> + uint64_t c8 = 0x00000000ae0b5394;
> + uint64_t c9 = 0x000000001c279815;
> + uint64_t c10 = 0x000000001d9513d7;
> + uint64_t c11 = 0x000000008f352d95;
> + uint64_t c12 = 0x00000000af449247;
> + uint64_t c13 = 0x000000003db1ecdc;
> + uint64_t c14 = 0x0000000081256527;
> + uint64_t c15 = 0x00000000f1da05aa;
> + uint64_t c16 = 0x00000000ccaa009e;
> + uint64_t c17 = 0x00000000ae689191;
> + uint64_t c18 = 0x00000000ccaa009e;
> + uint64_t c19 = 0x00000000b8bc6765;
> + uint64_t c20 = 0x00000001f7011640;
> + uint64_t c21 = 0x00000001db710640;
> +
> + a = _mm_set_epi64x(c1, c0);
> + crc32_eth.rk1_rk2 = _mm512_broadcast_i32x4(a);
> +
> + a = _mm_set_epi64x(c3, c2);
> + crc32_eth.rk3_rk4 = _mm512_broadcast_i32x4(a);
> +
> + crc32_eth.fold_7x128b = _mm512_setr_epi64(c4, c5, c6, c7, c8,
> + c9, c10, c11);
> + crc32_eth.fold_3x128b = _mm512_setr_epi64(c12, c13, c14, c15,
> + c16, c17, 0, 0);
> + crc32_eth.fold_1x128b = _mm_setr_epi64(_mm_cvtsi64_m64(c16),
> + _mm_cvtsi64_m64(c17));
> +
> + crc32_eth.rk5_rk6 = _mm_setr_epi64(_mm_cvtsi64_m64(c18),
> + _mm_cvtsi64_m64(c19));
> + crc32_eth.rk7_rk8 = _mm_setr_epi64(_mm_cvtsi64_m64(c20),
> + _mm_cvtsi64_m64(c21));
> +}
> +
> +static void
> +crc16_load_init_constants(void)
> +{
> + __m128i a;
> + /* fold constants */
> + uint64_t c0 = 0x0000000000009a19;
> + uint64_t c1 = 0x0000000000002df8;
> + uint64_t c2 = 0x00000000000068af;
> + uint64_t c3 = 0x000000000000b6c9;
> + uint64_t c4 = 0x000000000000c64f;
> + uint64_t c5 = 0x000000000000cd95;
> + uint64_t c6 = 0x000000000000d341;
> + uint64_t c7 = 0x000000000000b8f2;
> + uint64_t c8 = 0x0000000000000842;
> + uint64_t c9 = 0x000000000000b072;
> + uint64_t c10 = 0x00000000000047e3;
> + uint64_t c11 = 0x000000000000922d;
> + uint64_t c12 = 0x0000000000000e3a;
> + uint64_t c13 = 0x0000000000004d7a;
> + uint64_t c14 = 0x0000000000005b44;
> + uint64_t c15 = 0x0000000000007762;
> + uint64_t c16 = 0x00000000000081bf;
> + uint64_t c17 = 0x0000000000008e10;
> + uint64_t c18 = 0x00000000000081bf;
> + uint64_t c19 = 0x0000000000001cbb;
> + uint64_t c20 = 0x000000011c581910;
> + uint64_t c21 = 0x0000000000010810;
> +
> + a = _mm_set_epi64x(c1, c0);
> + crc16_ccitt.rk1_rk2 = _mm512_broadcast_i32x4(a);
> +
> + a = _mm_set_epi64x(c3, c2);
> + crc16_ccitt.rk3_rk4 = _mm512_broadcast_i32x4(a);
> +
> + crc16_ccitt.fold_7x128b = _mm512_setr_epi64(c4, c5, c6, c7, c8,
> + c9, c10, c11);
> + crc16_ccitt.fold_3x128b = _mm512_setr_epi64(c12, c13, c14, c15,
> + c16, c17, 0, 0);
> + crc16_ccitt.fold_1x128b = _mm_setr_epi64(_mm_cvtsi64_m64(c16),
> + _mm_cvtsi64_m64(c17));
> +
> + crc16_ccitt.rk5_rk6 = _mm_setr_epi64(_mm_cvtsi64_m64(c18),
> + _mm_cvtsi64_m64(c19));
> + crc16_ccitt.rk7_rk8 = _mm_setr_epi64(_mm_cvtsi64_m64(c20),
> + _mm_cvtsi64_m64(c21));
> +}
> +
> +void
> +rte_net_crc_avx512_init(void)
> +{
> + crc32_load_init_constants();
> + crc16_load_init_constants();
> +
> + /*
> + * Reset the register as following calculation may
> + * use other data types such as float, double, etc.
> + */
> + _mm_empty();
> +}
> +
> +uint32_t
> +rte_crc16_ccitt_avx512_handler(const uint8_t *data, uint32_t data_len)
> +{
> + /* return 16-bit CRC value */
> + return (uint16_t)~crc32_eth_calc_vpclmulqdq(data,
> + data_len,
> + 0xffff,
> + &crc16_ccitt);
> +}
> +
> +uint32_t
> +rte_crc32_eth_avx512_handler(const uint8_t *data, uint32_t data_len) {
> + /* return 32-bit CRC value */
> + return ~crc32_eth_calc_vpclmulqdq(data,
> + data_len,
> + 0xffffffffUL,
> + &crc32_eth);
> +}
> diff --git a/lib/librte_net/rte_net_crc.c b/lib/librte_net/rte_net_crc.c index
> d271d5205..32a366590 100644
> --- a/lib/librte_net/rte_net_crc.c
> +++ b/lib/librte_net/rte_net_crc.c
> @@ -37,6 +37,12 @@ static const rte_net_crc_handler handlers_scalar[] = {
> [RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_handler,
> [RTE_NET_CRC32_ETH] = rte_crc32_eth_handler, };
> +#ifdef CC_X86_64_AVX512_VPCLMULQDQ_SUPPORT
> +static const rte_net_crc_handler handlers_avx512[] = {
> + [RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_avx512_handler,
> + [RTE_NET_CRC32_ETH] = rte_crc32_eth_avx512_handler, }; #endif
> #ifdef CC_X86_64_SSE42_PCLMULQDQ_SUPPORT static const
> rte_net_crc_handler handlers_sse42[] = {
> [RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_sse42_handler, @@ -
> 134,6 +140,39 @@ rte_crc32_eth_handler(const uint8_t *data, uint32_t
> data_len)
> crc32_eth_lut);
> }
>
> +/* AVX512/VPCLMULQDQ handling */
> +
> +#define AVX512_VPCLMULQDQ_CPU_SUPPORTED ( \
> + rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) && \
> + rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) && \
> + rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512DQ) && \
> + rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512VL) && \
> + rte_cpu_get_flag_enabled(RTE_CPUFLAG_PCLMULQDQ) && \
> + rte_cpu_get_flag_enabled(RTE_CPUFLAG_VPCLMULQDQ) \
> +)
> +
> +static const rte_net_crc_handler *
> +avx512_vpclmulqdq_get_handlers(void)
> +{
> +#ifdef CC_X86_64_AVX512_VPCLMULQDQ_SUPPORT
> + if (AVX512_VPCLMULQDQ_CPU_SUPPORTED)
> + return handlers_avx512;
> +#endif
> + return NULL;
> +}
> +
> +static uint8_t
> +avx512_vpclmulqdq_init(void)
> +{
> +#ifdef CC_X86_64_AVX512_VPCLMULQDQ_SUPPORT
> + if (AVX512_VPCLMULQDQ_CPU_SUPPORTED) {
> + rte_net_crc_avx512_init();
> + return 1;
> + }
> +#endif
> + return 0;
> +}
> +
> /* SSE4.2/PCLMULQDQ handling */
>
> #define SSE42_PCLMULQDQ_CPU_SUPPORTED \ @@ -196,6 +235,11 @@
> rte_net_crc_set_alg(enum rte_net_crc_alg alg)
> handlers = NULL;
>
> switch (alg) {
> + case RTE_NET_CRC_AVX512:
> + handlers = avx512_vpclmulqdq_get_handlers();
> + if (handlers != NULL)
> + break;
> + /* fall-through */
> case RTE_NET_CRC_SSE42:
> handlers = sse42_pclmulqdq_get_handlers();
> break; /* for x86, always break here */ @@ -235,6 +279,8
> @@ RTE_INIT(rte_net_crc_init)
>
> if (sse42_pclmulqdq_init())
> alg = RTE_NET_CRC_SSE42;
> + if (avx512_vpclmulqdq_init())
> + alg = RTE_NET_CRC_AVX512;
> if (neon_pmull_init())
> alg = RTE_NET_CRC_NEON;
>
> diff --git a/lib/librte_net/rte_net_crc.h b/lib/librte_net/rte_net_crc.h index
> 16e85ca97..72d3e10ff 100644
> --- a/lib/librte_net/rte_net_crc.h
> +++ b/lib/librte_net/rte_net_crc.h
> @@ -1,5 +1,5 @@
> /* SPDX-License-Identifier: BSD-3-Clause
> - * Copyright(c) 2017 Intel Corporation
> + * Copyright(c) 2017-2020 Intel Corporation
> */
>
> #ifndef _RTE_NET_CRC_H_
> @@ -23,6 +23,7 @@ enum rte_net_crc_alg {
> RTE_NET_CRC_SCALAR = 0,
> RTE_NET_CRC_SSE42,
> RTE_NET_CRC_NEON,
> + RTE_NET_CRC_AVX512,
> };
>
> /**
> @@ -35,6 +36,7 @@ enum rte_net_crc_alg {
> * - RTE_NET_CRC_SCALAR
> * - RTE_NET_CRC_SSE42 (Use 64-bit SSE4.2 intrinsic)
> * - RTE_NET_CRC_NEON (Use ARM Neon intrinsic)
> + * - RTE_NET_CRC_AVX512 (Use 512-bit AVX intrinsic)
> */
> void
> rte_net_crc_set_alg(enum rte_net_crc_alg alg);
> --
> 2.12.3
Reviewed-by: Jasvinder Singh <jasvinder.singh@intel.com>
@@ -1,5 +1,5 @@
/* SPDX-License-Identifier: BSD-3-Clause
- * Copyright(c) 2017 Intel Corporation
+ * Copyright(c) 2017-2020 Intel Corporation
*/
#include "test.h"
@@ -149,6 +149,15 @@ test_crc(void)
return ret;
}
+ /* set CRC avx512 mode */
+ rte_net_crc_set_alg(RTE_NET_CRC_AVX512);
+
+ ret = test_crc_calc();
+ if (ret < 0) {
+ printf("test crc (x86_64 AVX512): failed (%d)\n", ret);
+ return ret;
+ }
+
/* set CRC neon mode */
rte_net_crc_set_alg(RTE_NET_CRC_NEON);
@@ -1,5 +1,5 @@
# SPDX-License-Identifier: BSD-3-Clause
-# Copyright(c) 2017-2019 Intel Corporation
+# Copyright(c) 2017-2020 Intel Corporation
# get binutils version for the workaround of Bug 97
if not is_windows
@@ -23,7 +23,9 @@ endforeach
optional_flags = ['AES', 'PCLMUL',
'AVX', 'AVX2', 'AVX512F',
- 'RDRND', 'RDSEED']
+ 'RDRND', 'RDSEED',
+ 'AVX512BW', 'AVX512DQ',
+ 'AVX512VL', 'VPCLMULQDQ']
foreach f:optional_flags
if cc.get_define('__@0@__'.format(f), args: machine_args) == '1'
if f == 'PCLMUL' # special case flags with different defines
@@ -58,6 +58,8 @@ New Features
* **Updated CRC modules of rte_net library.**
* Added run-time selection of the optimal architecture-specific CRC path.
+ * Added optimized implementations of CRC32-Ethernet and CRC16-CCITT
+ using the AVX512 and VPCLMULQDQ instruction sets.
* **Updated Broadcom bnxt driver.**
@@ -24,18 +24,62 @@ deps += ['mbuf']
if dpdk_conf.has('RTE_ARCH_X86_64')
net_crc_sse42_cpu_support = (
cc.get_define('__PCLMUL__', args: machine_args) != '')
+ net_crc_avx512_cpu_support = (
+ cc.get_define('__AVX512F__', args: machine_args) != '' and
+ cc.get_define('__AVX512BW__', args: machine_args) != '' and
+ cc.get_define('__AVX512DQ__', args: machine_args) != '' and
+ cc.get_define('__AVX512VL__', args: machine_args) != '' and
+ cc.get_define('__VPCLMULQDQ__', args: machine_args) != '')
+
net_crc_sse42_cc_support = (
cc.has_argument('-mpclmul') and cc.has_argument('-maes'))
+ net_crc_avx512_cc_support = (
+ not machine_args.contains('-mno-avx512f') and
+ cc.has_argument('-mavx512f') and
+ cc.has_argument('-mavx512bw') and
+ cc.has_argument('-mavx512dq') and
+ cc.has_argument('-mavx512vl') and
+ cc.has_argument('-mvpclmulqdq') and
+ cc.has_argument('-mavx2') and
+ cc.has_argument('-mavx'))
build_static_net_crc_sse42_lib = 0
+ build_static_net_crc_avx512_lib = 0
if net_crc_sse42_cpu_support == true
sources += files('net_crc_sse.c')
cflags += ['-DCC_X86_64_SSE42_PCLMULQDQ_SUPPORT']
+ if net_crc_avx512_cpu_support == true
+ sources += files('net_crc_avx512.c')
+ cflags += ['-DCC_X86_64_AVX512_VPCLMULQDQ_SUPPORT']
+ elif net_crc_avx512_cc_support == true
+ build_static_net_crc_avx512_lib = 1
+ net_crc_avx512_lib_cflags = ['-mavx512f',
+ '-mavx512bw',
+ '-mavx512dq',
+ '-mavx512vl',
+ '-mvpclmulqdq',
+ '-mavx2',
+ '-mavx']
+ cflags += ['-DCC_X86_64_AVX512_VPCLMULQDQ_SUPPORT']
+ endif
elif net_crc_sse42_cc_support == true
build_static_net_crc_sse42_lib = 1
net_crc_sse42_lib_cflags = ['-mpclmul', '-maes']
cflags += ['-DCC_X86_64_SSE42_PCLMULQDQ_SUPPORT']
+ if net_crc_avx512_cc_support == true
+ build_static_net_crc_avx512_lib = 1
+ net_crc_avx512_lib_cflags = ['-mpclmul',
+ '-maes',
+ '-mavx512f',
+ '-mavx512bw',
+ '-mavx512dq',
+ '-mavx512vl',
+ '-mvpclmulqdq',
+ '-mavx2',
+ '-mavx']
+ cflags += ['-DCC_X86_64_AVX512_VPCLMULQDQ_SUPPORT']
+ endif
endif
if build_static_net_crc_sse42_lib == 1
@@ -47,6 +91,17 @@ if dpdk_conf.has('RTE_ARCH_X86_64')
net_crc_sse42_lib_cflags])
objs += net_crc_sse42_lib.extract_objects('net_crc_sse.c')
endif
+
+ if build_static_net_crc_avx512_lib == 1
+ net_crc_avx512_lib = static_library(
+ 'net_crc_avx512_lib',
+ 'net_crc_avx512.c',
+ dependencies: static_rte_eal,
+ c_args: [cflags,
+ net_crc_avx512_lib_cflags])
+ objs += net_crc_avx512_lib.extract_objects('net_crc_avx512.c')
+ endif
+
elif (dpdk_conf.has('RTE_ARCH_ARM64') and
cc.get_define('__ARM_FEATURE_CRYPTO', args: machine_args) != '')
sources += files('net_crc_neon.c')
@@ -20,6 +20,17 @@ rte_crc16_ccitt_sse42_handler(const uint8_t *data, uint32_t data_len);
uint32_t
rte_crc32_eth_sse42_handler(const uint8_t *data, uint32_t data_len);
+/* AVX512 */
+
+void
+rte_net_crc_avx512_init(void);
+
+uint32_t
+rte_crc16_ccitt_avx512_handler(const uint8_t *data, uint32_t data_len);
+
+uint32_t
+rte_crc32_eth_avx512_handler(const uint8_t *data, uint32_t data_len);
+
/* NEON */
void
new file mode 100644
@@ -0,0 +1,423 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Intel Corporation
+ */
+
+#include <string.h>
+
+#include <rte_common.h>
+#include <rte_branch_prediction.h>
+#include <rte_cpuflags.h>
+
+#include "net_crc.h"
+
+#include <x86intrin.h>
+
+/* VPCLMULQDQ CRC computation context structure */
+struct crc_vpclmulqdq_ctx {
+ __m512i rk1_rk2;
+ __m512i rk3_rk4;
+ __m512i fold_7x128b;
+ __m512i fold_3x128b;
+ __m128i rk5_rk6;
+ __m128i rk7_rk8;
+ __m128i fold_1x128b;
+};
+
+static struct crc_vpclmulqdq_ctx crc32_eth __rte_aligned(64);
+static struct crc_vpclmulqdq_ctx crc16_ccitt __rte_aligned(64);
+
+static uint16_t byte_len_to_mask_table[] = {
+ 0x0000, 0x0001, 0x0003, 0x0007,
+ 0x000f, 0x001f, 0x003f, 0x007f,
+ 0x00ff, 0x01ff, 0x03ff, 0x07ff,
+ 0x0fff, 0x1fff, 0x3fff, 0x7fff,
+ 0xffff};
+
+static const uint8_t shf_table[32] __rte_aligned(16) = {
+ 0x00, 0x81, 0x82, 0x83, 0x84, 0x85, 0x86, 0x87,
+ 0x88, 0x89, 0x8a, 0x8b, 0x8c, 0x8d, 0x8e, 0x8f,
+ 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07,
+ 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f
+};
+
+static const uint32_t mask[4] __rte_aligned(16) = {
+ 0xffffffff, 0xffffffff, 0x00000000, 0x00000000
+};
+
+static const uint32_t mask2[4] __rte_aligned(16) = {
+ 0x00000000, 0xffffffff, 0xffffffff, 0xffffffff
+};
+
+static __rte_always_inline __m512i
+crcr32_folding_round(__m512i data_block, __m512i precomp, __m512i fold)
+{
+ __m512i tmp0, tmp1;
+
+ tmp0 = _mm512_clmulepi64_epi128(fold, precomp, 0x01);
+ tmp1 = _mm512_clmulepi64_epi128(fold, precomp, 0x10);
+
+ return _mm512_ternarylogic_epi64(tmp0, tmp1, data_block, 0x96);
+}
+
+static __rte_always_inline __m128i
+crc32_fold_128(__m512i fold0, __m512i fold1,
+ const struct crc_vpclmulqdq_ctx *params)
+{
+ __m128i res, res2;
+ __m256i a;
+ __m512i tmp0, tmp1, tmp2, tmp3;
+ __m512i tmp4;
+
+ tmp0 = _mm512_clmulepi64_epi128(fold0, params->fold_7x128b, 0x01);
+ tmp1 = _mm512_clmulepi64_epi128(fold0, params->fold_7x128b, 0x10);
+
+ res = _mm512_extracti64x2_epi64(fold1, 3);
+ tmp4 = _mm512_maskz_broadcast_i32x4(0xF, res);
+
+ tmp2 = _mm512_clmulepi64_epi128(fold1, params->fold_3x128b, 0x01);
+ tmp3 = _mm512_clmulepi64_epi128(fold1, params->fold_3x128b, 0x10);
+
+ tmp0 = _mm512_ternarylogic_epi64(tmp0, tmp1, tmp2, 0x96);
+ tmp0 = _mm512_ternarylogic_epi64(tmp0, tmp3, tmp4, 0x96);
+
+ tmp1 = _mm512_shuffle_i64x2(tmp0, tmp0, 0x4e);
+
+ a = _mm256_xor_si256(*(__m256i *)&tmp1, *(__m256i *)&tmp0);
+ res = _mm256_extracti64x2_epi64(a, 1);
+ res2 = _mm_xor_si128(res, *(__m128i *)&a);
+
+ return res2;
+}
+
+static __rte_always_inline __m128i
+last_two_xmm(const uint8_t *data, uint32_t data_len, uint32_t n, __m128i res,
+ const struct crc_vpclmulqdq_ctx *params)
+{
+ uint32_t offset;
+ __m128i res2, res3, res4, pshufb_shf;
+
+ const uint32_t mask3[4] __rte_aligned(16) = {
+ 0x80808080, 0x80808080, 0x80808080, 0x80808080
+ };
+
+ res2 = res;
+ offset = data_len - n;
+ res3 = _mm_loadu_si128((const __m128i *)&data[n+offset-16]);
+
+ pshufb_shf = _mm_loadu_si128((const __m128i *)
+ (shf_table + (data_len-n)));
+
+ res = _mm_shuffle_epi8(res, pshufb_shf);
+ pshufb_shf = _mm_xor_si128(pshufb_shf,
+ _mm_load_si128((const __m128i *) mask3));
+ res2 = _mm_shuffle_epi8(res2, pshufb_shf);
+
+ res2 = _mm_blendv_epi8(res2, res3, pshufb_shf);
+
+ res4 = _mm_clmulepi64_si128(res, params->fold_1x128b, 0x01);
+ res = _mm_clmulepi64_si128(res, params->fold_1x128b, 0x10);
+ res = _mm_ternarylogic_epi64(res, res2, res4, 0x96);
+
+ return res;
+}
+
+static __rte_always_inline __m128i
+done_128(__m128i res, const struct crc_vpclmulqdq_ctx *params)
+{
+ __m128i res1;
+
+ res1 = res;
+
+ res = _mm_clmulepi64_si128(res, params->rk5_rk6, 0x0);
+ res1 = _mm_srli_si128(res1, 8);
+ res = _mm_xor_si128(res, res1);
+
+ res1 = res;
+ res = _mm_slli_si128(res, 4);
+ res = _mm_clmulepi64_si128(res, params->rk5_rk6, 0x10);
+ res = _mm_xor_si128(res, res1);
+
+ return res;
+}
+
+static __rte_always_inline uint32_t
+barrett_reduction(__m128i data64, const struct crc_vpclmulqdq_ctx *params)
+{
+ __m128i tmp0, tmp1;
+
+ data64 = _mm_and_si128(data64, *(const __m128i *)mask2);
+ tmp0 = data64;
+ tmp1 = data64;
+
+ data64 = _mm_clmulepi64_si128(tmp0, params->rk7_rk8, 0x0);
+ data64 = _mm_ternarylogic_epi64(data64, tmp1, *(const __m128i *)mask,
+ 0x28);
+
+ tmp1 = data64;
+ data64 = _mm_clmulepi64_si128(data64, params->rk7_rk8, 0x10);
+ data64 = _mm_ternarylogic_epi64(data64, tmp1, tmp0, 0x96);
+
+ return _mm_extract_epi32(data64, 2);
+}
+
+static __rte_always_inline void
+reduction_loop(__m128i *fold, int *len, const uint8_t *data, uint32_t *n,
+ const struct crc_vpclmulqdq_ctx *params)
+{
+ __m128i tmp, tmp1;
+
+ tmp = _mm_clmulepi64_si128(*fold, params->fold_1x128b, 0x1);
+ *fold = _mm_clmulepi64_si128(*fold, params->fold_1x128b, 0x10);
+ *fold = _mm_xor_si128(*fold, tmp);
+ tmp1 = _mm_loadu_si128((const __m128i *)&data[*n]);
+ *fold = _mm_xor_si128(*fold, tmp1);
+ *n += 16;
+ *len -= 16;
+}
+
+static __rte_always_inline uint32_t
+crc32_eth_calc_vpclmulqdq(const uint8_t *data, uint32_t data_len, uint32_t crc,
+ const struct crc_vpclmulqdq_ctx *params)
+{
+ __m128i res, d, b;
+ __m512i temp, k;
+ __m512i qw0 = _mm512_set1_epi64(0), qw1, qw2, qw3;
+ __m512i fold0, fold1, fold2, fold3;
+ __mmask16 mask;
+ uint32_t n = 0;
+ int reduction = 0;
+
+ /* Get CRC init value */
+ b = _mm_cvtsi32_si128(crc);
+ temp = _mm512_castsi128_si512(b);
+
+ if (data_len > 255) {
+ fold0 = _mm512_loadu_si512((const __m512i *)data);
+ fold1 = _mm512_loadu_si512((const __m512i *)(data+64));
+ fold2 = _mm512_loadu_si512((const __m512i *)(data+128));
+ fold3 = _mm512_loadu_si512((const __m512i *)(data+192));
+ fold0 = _mm512_xor_si512(fold0, temp);
+
+ /* Main folding loop */
+ k = params->rk1_rk2;
+ for (n = 256; (n + 256) <= data_len; n += 256) {
+ qw0 = _mm512_loadu_si512((const __m512i *)&data[n]);
+ qw1 = _mm512_loadu_si512((const __m512i *)
+ &(data[n+64]));
+ qw2 = _mm512_loadu_si512((const __m512i *)
+ &(data[n+128]));
+ qw3 = _mm512_loadu_si512((const __m512i *)
+ &(data[n+192]));
+ fold0 = crcr32_folding_round(qw0, k, fold0);
+ fold1 = crcr32_folding_round(qw1, k, fold1);
+ fold2 = crcr32_folding_round(qw2, k, fold2);
+ fold3 = crcr32_folding_round(qw3, k, fold3);
+ }
+
+ /* 256 to 128 fold */
+ k = params->rk3_rk4;
+ fold0 = crcr32_folding_round(fold2, k, fold0);
+ fold1 = crcr32_folding_round(fold3, k, fold1);
+
+ res = crc32_fold_128(fold0, fold1, params);
+
+ reduction = 240 - ((n+256)-data_len);
+
+ while (reduction > 0)
+ reduction_loop(&res, &reduction, data, &n,
+ params);
+
+ reduction += 16;
+
+ if (n != data_len)
+ res = last_two_xmm(data, data_len, n, res,
+ params);
+ } else {
+ if (data_len > 31) {
+ res = _mm_cvtsi32_si128(crc);
+ d = _mm_loadu_si128((const __m128i *)data);
+ res = _mm_xor_si128(res, d);
+ n += 16;
+
+ reduction = 240 - ((n+256)-data_len);
+
+ while (reduction > 0)
+ reduction_loop(&res, &reduction, data, &n,
+ params);
+
+ if (n != data_len)
+ res = last_two_xmm(data, data_len, n, res,
+ params);
+ } else if (data_len > 16) {
+ res = _mm_cvtsi32_si128(crc);
+ d = _mm_loadu_si128((const __m128i *)data);
+ res = _mm_xor_si128(res, d);
+ n += 16;
+
+ if (n != data_len)
+ res = last_two_xmm(data, data_len, n, res,
+ params);
+ } else if (data_len == 16) {
+ res = _mm_cvtsi32_si128(crc);
+ d = _mm_loadu_si128((const __m128i *)data);
+ res = _mm_xor_si128(res, d);
+ } else {
+ res = _mm_cvtsi32_si128(crc);
+ mask = byte_len_to_mask_table[data_len];
+ d = _mm_maskz_loadu_epi8(mask, data);
+ res = _mm_xor_si128(res, d);
+
+ if (data_len > 3) {
+ d = _mm_loadu_si128((const __m128i *)
+ &shf_table[data_len]);
+ res = _mm_shuffle_epi8(res, d);
+ } else if (data_len > 2) {
+ res = _mm_slli_si128(res, 5);
+ goto do_barrett_reduction;
+ } else if (data_len > 1) {
+ res = _mm_slli_si128(res, 6);
+ goto do_barrett_reduction;
+ } else if (data_len > 0) {
+ res = _mm_slli_si128(res, 7);
+ goto do_barrett_reduction;
+ } else {
+ /* zero length case */
+ return crc;
+ }
+ }
+ }
+
+ res = done_128(res, params);
+
+do_barrett_reduction:
+ n = barrett_reduction(res, params);
+
+ return n;
+}
+
+static void
+crc32_load_init_constants(void)
+{
+ __m128i a;
+ /* fold constants */
+ uint64_t c0 = 0x00000000e95c1271;
+ uint64_t c1 = 0x00000000ce3371cb;
+ uint64_t c2 = 0x00000000910eeec1;
+ uint64_t c3 = 0x0000000033fff533;
+ uint64_t c4 = 0x000000000cbec0ed;
+ uint64_t c5 = 0x0000000031f8303f;
+ uint64_t c6 = 0x0000000057c54819;
+ uint64_t c7 = 0x00000000df068dc2;
+ uint64_t c8 = 0x00000000ae0b5394;
+ uint64_t c9 = 0x000000001c279815;
+ uint64_t c10 = 0x000000001d9513d7;
+ uint64_t c11 = 0x000000008f352d95;
+ uint64_t c12 = 0x00000000af449247;
+ uint64_t c13 = 0x000000003db1ecdc;
+ uint64_t c14 = 0x0000000081256527;
+ uint64_t c15 = 0x00000000f1da05aa;
+ uint64_t c16 = 0x00000000ccaa009e;
+ uint64_t c17 = 0x00000000ae689191;
+ uint64_t c18 = 0x00000000ccaa009e;
+ uint64_t c19 = 0x00000000b8bc6765;
+ uint64_t c20 = 0x00000001f7011640;
+ uint64_t c21 = 0x00000001db710640;
+
+ a = _mm_set_epi64x(c1, c0);
+ crc32_eth.rk1_rk2 = _mm512_broadcast_i32x4(a);
+
+ a = _mm_set_epi64x(c3, c2);
+ crc32_eth.rk3_rk4 = _mm512_broadcast_i32x4(a);
+
+ crc32_eth.fold_7x128b = _mm512_setr_epi64(c4, c5, c6, c7, c8,
+ c9, c10, c11);
+ crc32_eth.fold_3x128b = _mm512_setr_epi64(c12, c13, c14, c15,
+ c16, c17, 0, 0);
+ crc32_eth.fold_1x128b = _mm_setr_epi64(_mm_cvtsi64_m64(c16),
+ _mm_cvtsi64_m64(c17));
+
+ crc32_eth.rk5_rk6 = _mm_setr_epi64(_mm_cvtsi64_m64(c18),
+ _mm_cvtsi64_m64(c19));
+ crc32_eth.rk7_rk8 = _mm_setr_epi64(_mm_cvtsi64_m64(c20),
+ _mm_cvtsi64_m64(c21));
+}
+
+static void
+crc16_load_init_constants(void)
+{
+ __m128i a;
+ /* fold constants */
+ uint64_t c0 = 0x0000000000009a19;
+ uint64_t c1 = 0x0000000000002df8;
+ uint64_t c2 = 0x00000000000068af;
+ uint64_t c3 = 0x000000000000b6c9;
+ uint64_t c4 = 0x000000000000c64f;
+ uint64_t c5 = 0x000000000000cd95;
+ uint64_t c6 = 0x000000000000d341;
+ uint64_t c7 = 0x000000000000b8f2;
+ uint64_t c8 = 0x0000000000000842;
+ uint64_t c9 = 0x000000000000b072;
+ uint64_t c10 = 0x00000000000047e3;
+ uint64_t c11 = 0x000000000000922d;
+ uint64_t c12 = 0x0000000000000e3a;
+ uint64_t c13 = 0x0000000000004d7a;
+ uint64_t c14 = 0x0000000000005b44;
+ uint64_t c15 = 0x0000000000007762;
+ uint64_t c16 = 0x00000000000081bf;
+ uint64_t c17 = 0x0000000000008e10;
+ uint64_t c18 = 0x00000000000081bf;
+ uint64_t c19 = 0x0000000000001cbb;
+ uint64_t c20 = 0x000000011c581910;
+ uint64_t c21 = 0x0000000000010810;
+
+ a = _mm_set_epi64x(c1, c0);
+ crc16_ccitt.rk1_rk2 = _mm512_broadcast_i32x4(a);
+
+ a = _mm_set_epi64x(c3, c2);
+ crc16_ccitt.rk3_rk4 = _mm512_broadcast_i32x4(a);
+
+ crc16_ccitt.fold_7x128b = _mm512_setr_epi64(c4, c5, c6, c7, c8,
+ c9, c10, c11);
+ crc16_ccitt.fold_3x128b = _mm512_setr_epi64(c12, c13, c14, c15,
+ c16, c17, 0, 0);
+ crc16_ccitt.fold_1x128b = _mm_setr_epi64(_mm_cvtsi64_m64(c16),
+ _mm_cvtsi64_m64(c17));
+
+ crc16_ccitt.rk5_rk6 = _mm_setr_epi64(_mm_cvtsi64_m64(c18),
+ _mm_cvtsi64_m64(c19));
+ crc16_ccitt.rk7_rk8 = _mm_setr_epi64(_mm_cvtsi64_m64(c20),
+ _mm_cvtsi64_m64(c21));
+}
+
+void
+rte_net_crc_avx512_init(void)
+{
+ crc32_load_init_constants();
+ crc16_load_init_constants();
+
+ /*
+ * Reset the register as following calculation may
+ * use other data types such as float, double, etc.
+ */
+ _mm_empty();
+}
+
+uint32_t
+rte_crc16_ccitt_avx512_handler(const uint8_t *data, uint32_t data_len)
+{
+ /* return 16-bit CRC value */
+ return (uint16_t)~crc32_eth_calc_vpclmulqdq(data,
+ data_len,
+ 0xffff,
+ &crc16_ccitt);
+}
+
+uint32_t
+rte_crc32_eth_avx512_handler(const uint8_t *data, uint32_t data_len)
+{
+ /* return 32-bit CRC value */
+ return ~crc32_eth_calc_vpclmulqdq(data,
+ data_len,
+ 0xffffffffUL,
+ &crc32_eth);
+}
@@ -37,6 +37,12 @@ static const rte_net_crc_handler handlers_scalar[] = {
[RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_handler,
[RTE_NET_CRC32_ETH] = rte_crc32_eth_handler,
};
+#ifdef CC_X86_64_AVX512_VPCLMULQDQ_SUPPORT
+static const rte_net_crc_handler handlers_avx512[] = {
+ [RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_avx512_handler,
+ [RTE_NET_CRC32_ETH] = rte_crc32_eth_avx512_handler,
+};
+#endif
#ifdef CC_X86_64_SSE42_PCLMULQDQ_SUPPORT
static const rte_net_crc_handler handlers_sse42[] = {
[RTE_NET_CRC16_CCITT] = rte_crc16_ccitt_sse42_handler,
@@ -134,6 +140,39 @@ rte_crc32_eth_handler(const uint8_t *data, uint32_t data_len)
crc32_eth_lut);
}
+/* AVX512/VPCLMULQDQ handling */
+
+#define AVX512_VPCLMULQDQ_CPU_SUPPORTED ( \
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512F) && \
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512BW) && \
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512DQ) && \
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_AVX512VL) && \
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_PCLMULQDQ) && \
+ rte_cpu_get_flag_enabled(RTE_CPUFLAG_VPCLMULQDQ) \
+)
+
+static const rte_net_crc_handler *
+avx512_vpclmulqdq_get_handlers(void)
+{
+#ifdef CC_X86_64_AVX512_VPCLMULQDQ_SUPPORT
+ if (AVX512_VPCLMULQDQ_CPU_SUPPORTED)
+ return handlers_avx512;
+#endif
+ return NULL;
+}
+
+static uint8_t
+avx512_vpclmulqdq_init(void)
+{
+#ifdef CC_X86_64_AVX512_VPCLMULQDQ_SUPPORT
+ if (AVX512_VPCLMULQDQ_CPU_SUPPORTED) {
+ rte_net_crc_avx512_init();
+ return 1;
+ }
+#endif
+ return 0;
+}
+
/* SSE4.2/PCLMULQDQ handling */
#define SSE42_PCLMULQDQ_CPU_SUPPORTED \
@@ -196,6 +235,11 @@ rte_net_crc_set_alg(enum rte_net_crc_alg alg)
handlers = NULL;
switch (alg) {
+ case RTE_NET_CRC_AVX512:
+ handlers = avx512_vpclmulqdq_get_handlers();
+ if (handlers != NULL)
+ break;
+ /* fall-through */
case RTE_NET_CRC_SSE42:
handlers = sse42_pclmulqdq_get_handlers();
break; /* for x86, always break here */
@@ -235,6 +279,8 @@ RTE_INIT(rte_net_crc_init)
if (sse42_pclmulqdq_init())
alg = RTE_NET_CRC_SSE42;
+ if (avx512_vpclmulqdq_init())
+ alg = RTE_NET_CRC_AVX512;
if (neon_pmull_init())
alg = RTE_NET_CRC_NEON;
@@ -1,5 +1,5 @@
/* SPDX-License-Identifier: BSD-3-Clause
- * Copyright(c) 2017 Intel Corporation
+ * Copyright(c) 2017-2020 Intel Corporation
*/
#ifndef _RTE_NET_CRC_H_
@@ -23,6 +23,7 @@ enum rte_net_crc_alg {
RTE_NET_CRC_SCALAR = 0,
RTE_NET_CRC_SSE42,
RTE_NET_CRC_NEON,
+ RTE_NET_CRC_AVX512,
};
/**
@@ -35,6 +36,7 @@ enum rte_net_crc_alg {
* - RTE_NET_CRC_SCALAR
* - RTE_NET_CRC_SSE42 (Use 64-bit SSE4.2 intrinsic)
* - RTE_NET_CRC_NEON (Use ARM Neon intrinsic)
+ * - RTE_NET_CRC_AVX512 (Use 512-bit AVX intrinsic)
*/
void
rte_net_crc_set_alg(enum rte_net_crc_alg alg);