linux/lib/checksum_kunit.c

642 lines
35 KiB
C
Raw Normal View History

x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
// SPDX-License-Identifier: GPL-2.0+
/*
* Test cases csum_partial, csum_fold, ip_fast_csum, csum_ipv6_magic
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
*/
#include <kunit/test.h>
#include <asm/checksum.h>
#include <net/ip6_checksum.h>
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
#define MAX_LEN 512
#define MAX_ALIGN 64
#define TEST_BUFLEN (MAX_LEN + MAX_ALIGN)
#define IPv4_MIN_WORDS 5
#define IPv4_MAX_WORDS 15
#define NUM_IPv6_TESTS 200
#define NUM_IP_FAST_CSUM_TESTS 181
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
/* Values for a little endian CPU. Byte swap each half on big endian CPU. */
static const u32 random_init_sum = 0x2847aab;
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
static const u8 random_buf[] = {
0xac, 0xd7, 0x76, 0x69, 0x6e, 0xf2, 0x93, 0x2c, 0x1f, 0xe0, 0xde, 0x86,
0x8f, 0x54, 0x33, 0x90, 0x95, 0xbf, 0xff, 0xb9, 0xea, 0x62, 0x6e, 0xb5,
0xd3, 0x4f, 0xf5, 0x60, 0x50, 0x5c, 0xc7, 0xfa, 0x6d, 0x1a, 0xc7, 0xf0,
0xd2, 0x2c, 0x12, 0x3d, 0x88, 0xe3, 0x14, 0x21, 0xb1, 0x5e, 0x45, 0x31,
0xa2, 0x85, 0x36, 0x76, 0xba, 0xd8, 0xad, 0xbb, 0x9e, 0x49, 0x8f, 0xf7,
0xce, 0xea, 0xef, 0xca, 0x2c, 0x29, 0xf7, 0x15, 0x5c, 0x1d, 0x4d, 0x09,
0x1f, 0xe2, 0x14, 0x31, 0x8c, 0x07, 0x57, 0x23, 0x1f, 0x6f, 0x03, 0xe1,
0x93, 0x19, 0x53, 0x03, 0x45, 0x49, 0x9a, 0x3b, 0x8e, 0x0c, 0x12, 0x5d,
0x8a, 0xb8, 0x9b, 0x8c, 0x9a, 0x03, 0xe5, 0xa2, 0x43, 0xd2, 0x3b, 0x4e,
0x7e, 0x30, 0x3c, 0x22, 0x2d, 0xc5, 0xfc, 0x9e, 0xdb, 0xc6, 0xf9, 0x69,
0x12, 0x39, 0x1f, 0xa0, 0x11, 0x0c, 0x3f, 0xf5, 0x53, 0xc9, 0x30, 0xfb,
0xb0, 0xdd, 0x21, 0x1d, 0x34, 0xe2, 0x65, 0x30, 0xf1, 0xe8, 0x1b, 0xe7,
0x55, 0x0d, 0xeb, 0xbd, 0xcc, 0x9d, 0x24, 0xa4, 0xad, 0xa7, 0x93, 0x47,
0x19, 0x2e, 0xc4, 0x5c, 0x3b, 0xc7, 0x6d, 0x95, 0x0c, 0x47, 0x60, 0xaf,
0x5b, 0x47, 0xee, 0xdc, 0x31, 0x31, 0x14, 0x12, 0x7e, 0x9e, 0x45, 0xb1,
0xc1, 0x69, 0x4b, 0x84, 0xfc, 0x88, 0xc1, 0x9e, 0x46, 0xb4, 0xc2, 0x25,
0xc5, 0x6c, 0x4c, 0x22, 0x58, 0x5c, 0xbe, 0xff, 0xea, 0x88, 0x88, 0x7a,
0xcb, 0x1c, 0x5d, 0x63, 0xa1, 0xf2, 0x33, 0x0c, 0xa2, 0x16, 0x0b, 0x6e,
0x2b, 0x79, 0x58, 0xf7, 0xac, 0xd3, 0x6a, 0x3f, 0x81, 0x57, 0x48, 0x45,
0xe3, 0x7c, 0xdc, 0xd6, 0x34, 0x7e, 0xe6, 0x73, 0xfa, 0xcb, 0x31, 0x18,
0xa9, 0x0b, 0xee, 0x6b, 0x99, 0xb9, 0x2d, 0xde, 0x22, 0x0e, 0x71, 0x57,
0x0e, 0x9b, 0x11, 0xd1, 0x15, 0x41, 0xd0, 0x6b, 0x50, 0x8a, 0x23, 0x64,
0xe3, 0x9c, 0xb3, 0x55, 0x09, 0xe9, 0x32, 0x67, 0xf9, 0xe0, 0x73, 0xf1,
0x60, 0x66, 0x0b, 0x88, 0x79, 0x8d, 0x4b, 0x52, 0x83, 0x20, 0x26, 0x78,
0x49, 0x27, 0xe7, 0x3e, 0x29, 0xa8, 0x18, 0x82, 0x41, 0xdd, 0x1e, 0xcc,
0x3b, 0xc4, 0x65, 0xd1, 0x21, 0x40, 0x72, 0xb2, 0x87, 0x5e, 0x16, 0x10,
0x80, 0x3f, 0x4b, 0x58, 0x1c, 0xc2, 0x79, 0x20, 0xf0, 0xe0, 0x80, 0xd3,
0x52, 0xa5, 0x19, 0x6e, 0x47, 0x90, 0x08, 0xf5, 0x50, 0xe2, 0xd6, 0xae,
0xe9, 0x2e, 0xdc, 0xd5, 0xb4, 0x90, 0x1f, 0x79, 0x49, 0x82, 0x21, 0x84,
0xa0, 0xb5, 0x2f, 0xff, 0x30, 0x71, 0xed, 0x80, 0x68, 0xb1, 0x6d, 0xef,
0xf6, 0xcf, 0xb8, 0x41, 0x79, 0xf5, 0x01, 0xbc, 0x0c, 0x9b, 0x0e, 0x06,
0xf3, 0xb0, 0xbb, 0x97, 0xb8, 0xb1, 0xfd, 0x51, 0x4e, 0xef, 0x0a, 0x3d,
0x7a, 0x3d, 0xbd, 0x61, 0x00, 0xa2, 0xb3, 0xf0, 0x1d, 0x77, 0x7b, 0x6c,
0x01, 0x61, 0xa5, 0xa3, 0xdb, 0xd5, 0xd5, 0xf4, 0xb5, 0x28, 0x9f, 0x0a,
0xa3, 0x82, 0x5f, 0x4b, 0x40, 0x0f, 0x05, 0x0e, 0x78, 0xed, 0xbf, 0x17,
0xf6, 0x5a, 0x8a, 0x7d, 0xf9, 0x45, 0xc1, 0xd7, 0x1b, 0x9d, 0x6c, 0x07,
0x88, 0xf3, 0xbc, 0xf1, 0xea, 0x28, 0x1f, 0xb8, 0x7a, 0x60, 0x3c, 0xce,
0x3e, 0x50, 0xb2, 0x0b, 0xcf, 0xe5, 0x08, 0x1f, 0x48, 0x04, 0xf9, 0x35,
0x29, 0x15, 0xbe, 0x82, 0x96, 0xc2, 0x55, 0x04, 0x6c, 0x19, 0x45, 0x29,
0x0b, 0xb6, 0x49, 0x12, 0xfb, 0x8d, 0x1b, 0x75, 0x8b, 0xd9, 0x6a, 0x5c,
0xbe, 0x46, 0x2b, 0x41, 0xfe, 0x21, 0xad, 0x1f, 0x75, 0xe7, 0x90, 0x3d,
0xe1, 0xdf, 0x4b, 0xe1, 0x81, 0xe2, 0x17, 0x02, 0x7b, 0x58, 0x8b, 0x92,
0x1a, 0xac, 0x46, 0xdd, 0x2e, 0xce, 0x40, 0x09
};
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
/* Values for a little endian CPU. Byte swap on big endian CPU. */
static const u16 expected_results[] = {
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
0x82d0, 0x8224, 0xab23, 0xaaad, 0x41ad, 0x413f, 0x4f3e, 0x4eab, 0x22ab,
0x228c, 0x428b, 0x41ad, 0xbbac, 0xbb1d, 0x671d, 0x66ea, 0xd6e9, 0xd654,
0x1754, 0x1655, 0x5d54, 0x5c6a, 0xfa69, 0xf9fb, 0x44fb, 0x4428, 0xf527,
0xf432, 0x9432, 0x93e2, 0x37e2, 0x371b, 0x3d1a, 0x3cad, 0x22ad, 0x21e6,
0x31e5, 0x3113, 0x0513, 0x0501, 0xc800, 0xc778, 0xe477, 0xe463, 0xc363,
0xc2b2, 0x64b2, 0x646d, 0x336d, 0x32cb, 0xadca, 0xad94, 0x3794, 0x36da,
0x5ed9, 0x5e2c, 0xa32b, 0xa28d, 0x598d, 0x58fe, 0x61fd, 0x612f, 0x772e,
0x763f, 0xac3e, 0xac12, 0x8312, 0x821b, 0x6d1b, 0x6cbf, 0x4fbf, 0x4f72,
0x4672, 0x4653, 0x6452, 0x643e, 0x333e, 0x32b2, 0x2bb2, 0x2b5b, 0x085b,
0x083c, 0x993b, 0x9938, 0xb837, 0xb7a4, 0x9ea4, 0x9e51, 0x9b51, 0x9b0c,
0x520c, 0x5172, 0x1672, 0x15e4, 0x09e4, 0x09d2, 0xacd1, 0xac47, 0xf446,
0xf3ab, 0x67ab, 0x6711, 0x6411, 0x632c, 0xc12b, 0xc0e8, 0xeee7, 0xeeac,
0xa0ac, 0xa02e, 0x702e, 0x6ff2, 0x4df2, 0x4dc5, 0x88c4, 0x87c8, 0xe9c7,
0xe8ec, 0x22ec, 0x21f3, 0xb8f2, 0xb8e0, 0x7fe0, 0x7fc1, 0xdfc0, 0xdfaf,
0xd3af, 0xd370, 0xde6f, 0xde1c, 0x151c, 0x14ec, 0x19eb, 0x193b, 0x3c3a,
0x3c19, 0x1f19, 0x1ee5, 0x3ce4, 0x3c7f, 0x0c7f, 0x0b8e, 0x238d, 0x2372,
0x3c71, 0x3c1c, 0x2f1c, 0x2e31, 0x7130, 0x7064, 0xd363, 0xd33f, 0x2f3f,
0x2e92, 0x8791, 0x86fe, 0x3ffe, 0x3fe5, 0x11e5, 0x1121, 0xb520, 0xb4e5,
0xede4, 0xed77, 0x5877, 0x586b, 0x116b, 0x110b, 0x620a, 0x61af, 0x1aaf,
0x19c1, 0x3dc0, 0x3d8f, 0x0c8f, 0x0c7b, 0xfa7a, 0xf9fc, 0x5bfc, 0x5bb7,
0xaab6, 0xa9f5, 0x40f5, 0x40aa, 0xbca9, 0xbbad, 0x33ad, 0x32ec, 0x94eb,
0x94a5, 0xe0a4, 0xdfe2, 0xbae2, 0xba1d, 0x4e1d, 0x4dd1, 0x2bd1, 0x2b79,
0xcf78, 0xceba, 0xcfb9, 0xcecf, 0x46cf, 0x4647, 0xcc46, 0xcb7b, 0xaf7b,
0xaf1e, 0x4c1e, 0x4b7d, 0x597c, 0x5949, 0x4d49, 0x4ca7, 0x36a7, 0x369c,
0xc89b, 0xc870, 0x4f70, 0x4f18, 0x5817, 0x576b, 0x846a, 0x8400, 0x4500,
0x447f, 0xed7e, 0xed36, 0xa836, 0xa753, 0x2b53, 0x2a77, 0x5476, 0x5442,
0xd641, 0xd55b, 0x625b, 0x6161, 0x9660, 0x962f, 0x7e2f, 0x7d86, 0x7286,
0x7198, 0x0698, 0x05ff, 0x4cfe, 0x4cd1, 0x6ed0, 0x6eae, 0x60ae, 0x603d,
0x093d, 0x092f, 0x6e2e, 0x6e1d, 0x9d1c, 0x9d07, 0x5c07, 0x5b37, 0xf036,
0xefe6, 0x65e6, 0x65c3, 0x01c3, 0x00e0, 0x64df, 0x642c, 0x0f2c, 0x0f23,
0x2622, 0x25f0, 0xbeef, 0xbdf6, 0xddf5, 0xdd82, 0xec81, 0xec21, 0x8621,
0x8616, 0xfe15, 0xfd9c, 0x709c, 0x7051, 0x1e51, 0x1dce, 0xfdcd, 0xfda7,
0x85a7, 0x855e, 0x5e5e, 0x5d77, 0x1f77, 0x1f4e, 0x774d, 0x7735, 0xf534,
0xf4f3, 0x17f3, 0x17d5, 0x4bd4, 0x4b99, 0x8798, 0x8733, 0xb632, 0xb611,
0x7611, 0x759f, 0xc39e, 0xc317, 0x6517, 0x6501, 0x5501, 0x5481, 0x1581,
0x1536, 0xbd35, 0xbd19, 0xfb18, 0xfa9f, 0xda9f, 0xd9af, 0xf9ae, 0xf92e,
0x262e, 0x25dc, 0x80db, 0x80c2, 0x12c2, 0x127b, 0x827a, 0x8272, 0x8d71,
0x8d21, 0xab20, 0xaa4a, 0xfc49, 0xfb60, 0xcd60, 0xcc84, 0xf783, 0xf6cf,
0x66cf, 0x66b0, 0xedaf, 0xed66, 0x6b66, 0x6b45, 0xe744, 0xe6a4, 0x31a4,
0x3175, 0x3274, 0x3244, 0xc143, 0xc056, 0x4056, 0x3fee, 0x8eed, 0x8e80,
0x9f7f, 0x9e89, 0xcf88, 0xced0, 0x8dd0, 0x8d57, 0x9856, 0x9855, 0xdc54,
0xdc48, 0x4148, 0x413a, 0x3b3a, 0x3a47, 0x8a46, 0x898b, 0xf28a, 0xf1d2,
0x40d2, 0x3fd5, 0xeed4, 0xee86, 0xff85, 0xff7b, 0xc27b, 0xc201, 0x8501,
0x8444, 0x2344, 0x2344, 0x8143, 0x8090, 0x908f, 0x9072, 0x1972, 0x18f7,
0xacf6, 0xacf5, 0x4bf5, 0x4b50, 0xa84f, 0xa774, 0xd273, 0xd19e, 0xdd9d,
0xdce8, 0xb4e8, 0xb449, 0xaa49, 0xa9a6, 0x27a6, 0x2747, 0xdc46, 0xdc06,
0xcd06, 0xcd01, 0xbf01, 0xbe89, 0xd188, 0xd0c9, 0xb9c9, 0xb8d3, 0x5ed3,
0x5e49, 0xe148, 0xe04f, 0x9b4f, 0x9a8e, 0xc38d, 0xc372, 0x2672, 0x2606,
0x1f06, 0x1e7e, 0x2b7d, 0x2ac1, 0x39c0, 0x38d6, 0x10d6, 0x10b7, 0x58b6,
0x583c, 0xf83b, 0xf7ff, 0x29ff, 0x29c1, 0xd9c0, 0xd90e, 0xce0e, 0xcd3f,
0xe83e, 0xe836, 0xc936, 0xc8ee, 0xc4ee, 0xc3f5, 0x8ef5, 0x8ecc, 0x79cc,
0x790e, 0xf70d, 0xf677, 0x3477, 0x3422, 0x3022, 0x2fb6, 0x16b6, 0x1671,
0xed70, 0xed65, 0x3765, 0x371c, 0x251c, 0x2421, 0x9720, 0x9705, 0x2205,
0x217a, 0x4879, 0x480f, 0xec0e, 0xeb50, 0xa550, 0xa525, 0x6425, 0x6327,
0x4227, 0x417a, 0x227a, 0x2205, 0x3b04, 0x3a74, 0xfd73, 0xfc92, 0x1d92,
0x1d47, 0x3c46, 0x3bc5, 0x59c4, 0x59ad, 0x57ad, 0x5732, 0xff31, 0xfea6,
0x6ca6, 0x6c8c, 0xc08b, 0xc045, 0xe344, 0xe316, 0x1516, 0x14d6,
};
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
/* Values for a little endian CPU. Byte swap each half on big endian CPU. */
static const u32 init_sums_no_overflow[] = {
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
0xffffffff, 0xfffffffb, 0xfffffbfb, 0xfffffbf7, 0xfffff7f7, 0xfffff7f3,
0xfffff3f3, 0xfffff3ef, 0xffffefef, 0xffffefeb, 0xffffebeb, 0xffffebe7,
0xffffe7e7, 0xffffe7e3, 0xffffe3e3, 0xffffe3df, 0xffffdfdf, 0xffffdfdb,
0xffffdbdb, 0xffffdbd7, 0xffffd7d7, 0xffffd7d3, 0xffffd3d3, 0xffffd3cf,
0xffffcfcf, 0xffffcfcb, 0xffffcbcb, 0xffffcbc7, 0xffffc7c7, 0xffffc7c3,
0xffffc3c3, 0xffffc3bf, 0xffffbfbf, 0xffffbfbb, 0xffffbbbb, 0xffffbbb7,
0xffffb7b7, 0xffffb7b3, 0xffffb3b3, 0xffffb3af, 0xffffafaf, 0xffffafab,
0xffffabab, 0xffffaba7, 0xffffa7a7, 0xffffa7a3, 0xffffa3a3, 0xffffa39f,
0xffff9f9f, 0xffff9f9b, 0xffff9b9b, 0xffff9b97, 0xffff9797, 0xffff9793,
0xffff9393, 0xffff938f, 0xffff8f8f, 0xffff8f8b, 0xffff8b8b, 0xffff8b87,
0xffff8787, 0xffff8783, 0xffff8383, 0xffff837f, 0xffff7f7f, 0xffff7f7b,
0xffff7b7b, 0xffff7b77, 0xffff7777, 0xffff7773, 0xffff7373, 0xffff736f,
0xffff6f6f, 0xffff6f6b, 0xffff6b6b, 0xffff6b67, 0xffff6767, 0xffff6763,
0xffff6363, 0xffff635f, 0xffff5f5f, 0xffff5f5b, 0xffff5b5b, 0xffff5b57,
0xffff5757, 0xffff5753, 0xffff5353, 0xffff534f, 0xffff4f4f, 0xffff4f4b,
0xffff4b4b, 0xffff4b47, 0xffff4747, 0xffff4743, 0xffff4343, 0xffff433f,
0xffff3f3f, 0xffff3f3b, 0xffff3b3b, 0xffff3b37, 0xffff3737, 0xffff3733,
0xffff3333, 0xffff332f, 0xffff2f2f, 0xffff2f2b, 0xffff2b2b, 0xffff2b27,
0xffff2727, 0xffff2723, 0xffff2323, 0xffff231f, 0xffff1f1f, 0xffff1f1b,
0xffff1b1b, 0xffff1b17, 0xffff1717, 0xffff1713, 0xffff1313, 0xffff130f,
0xffff0f0f, 0xffff0f0b, 0xffff0b0b, 0xffff0b07, 0xffff0707, 0xffff0703,
0xffff0303, 0xffff02ff, 0xfffffefe, 0xfffffefa, 0xfffffafa, 0xfffffaf6,
0xfffff6f6, 0xfffff6f2, 0xfffff2f2, 0xfffff2ee, 0xffffeeee, 0xffffeeea,
0xffffeaea, 0xffffeae6, 0xffffe6e6, 0xffffe6e2, 0xffffe2e2, 0xffffe2de,
0xffffdede, 0xffffdeda, 0xffffdada, 0xffffdad6, 0xffffd6d6, 0xffffd6d2,
0xffffd2d2, 0xffffd2ce, 0xffffcece, 0xffffceca, 0xffffcaca, 0xffffcac6,
0xffffc6c6, 0xffffc6c2, 0xffffc2c2, 0xffffc2be, 0xffffbebe, 0xffffbeba,
0xffffbaba, 0xffffbab6, 0xffffb6b6, 0xffffb6b2, 0xffffb2b2, 0xffffb2ae,
0xffffaeae, 0xffffaeaa, 0xffffaaaa, 0xffffaaa6, 0xffffa6a6, 0xffffa6a2,
0xffffa2a2, 0xffffa29e, 0xffff9e9e, 0xffff9e9a, 0xffff9a9a, 0xffff9a96,
0xffff9696, 0xffff9692, 0xffff9292, 0xffff928e, 0xffff8e8e, 0xffff8e8a,
0xffff8a8a, 0xffff8a86, 0xffff8686, 0xffff8682, 0xffff8282, 0xffff827e,
0xffff7e7e, 0xffff7e7a, 0xffff7a7a, 0xffff7a76, 0xffff7676, 0xffff7672,
0xffff7272, 0xffff726e, 0xffff6e6e, 0xffff6e6a, 0xffff6a6a, 0xffff6a66,
0xffff6666, 0xffff6662, 0xffff6262, 0xffff625e, 0xffff5e5e, 0xffff5e5a,
0xffff5a5a, 0xffff5a56, 0xffff5656, 0xffff5652, 0xffff5252, 0xffff524e,
0xffff4e4e, 0xffff4e4a, 0xffff4a4a, 0xffff4a46, 0xffff4646, 0xffff4642,
0xffff4242, 0xffff423e, 0xffff3e3e, 0xffff3e3a, 0xffff3a3a, 0xffff3a36,
0xffff3636, 0xffff3632, 0xffff3232, 0xffff322e, 0xffff2e2e, 0xffff2e2a,
0xffff2a2a, 0xffff2a26, 0xffff2626, 0xffff2622, 0xffff2222, 0xffff221e,
0xffff1e1e, 0xffff1e1a, 0xffff1a1a, 0xffff1a16, 0xffff1616, 0xffff1612,
0xffff1212, 0xffff120e, 0xffff0e0e, 0xffff0e0a, 0xffff0a0a, 0xffff0a06,
0xffff0606, 0xffff0602, 0xffff0202, 0xffff01fe, 0xfffffdfd, 0xfffffdf9,
0xfffff9f9, 0xfffff9f5, 0xfffff5f5, 0xfffff5f1, 0xfffff1f1, 0xfffff1ed,
0xffffeded, 0xffffede9, 0xffffe9e9, 0xffffe9e5, 0xffffe5e5, 0xffffe5e1,
0xffffe1e1, 0xffffe1dd, 0xffffdddd, 0xffffddd9, 0xffffd9d9, 0xffffd9d5,
0xffffd5d5, 0xffffd5d1, 0xffffd1d1, 0xffffd1cd, 0xffffcdcd, 0xffffcdc9,
0xffffc9c9, 0xffffc9c5, 0xffffc5c5, 0xffffc5c1, 0xffffc1c1, 0xffffc1bd,
0xffffbdbd, 0xffffbdb9, 0xffffb9b9, 0xffffb9b5, 0xffffb5b5, 0xffffb5b1,
0xffffb1b1, 0xffffb1ad, 0xffffadad, 0xffffada9, 0xffffa9a9, 0xffffa9a5,
0xffffa5a5, 0xffffa5a1, 0xffffa1a1, 0xffffa19d, 0xffff9d9d, 0xffff9d99,
0xffff9999, 0xffff9995, 0xffff9595, 0xffff9591, 0xffff9191, 0xffff918d,
0xffff8d8d, 0xffff8d89, 0xffff8989, 0xffff8985, 0xffff8585, 0xffff8581,
0xffff8181, 0xffff817d, 0xffff7d7d, 0xffff7d79, 0xffff7979, 0xffff7975,
0xffff7575, 0xffff7571, 0xffff7171, 0xffff716d, 0xffff6d6d, 0xffff6d69,
0xffff6969, 0xffff6965, 0xffff6565, 0xffff6561, 0xffff6161, 0xffff615d,
0xffff5d5d, 0xffff5d59, 0xffff5959, 0xffff5955, 0xffff5555, 0xffff5551,
0xffff5151, 0xffff514d, 0xffff4d4d, 0xffff4d49, 0xffff4949, 0xffff4945,
0xffff4545, 0xffff4541, 0xffff4141, 0xffff413d, 0xffff3d3d, 0xffff3d39,
0xffff3939, 0xffff3935, 0xffff3535, 0xffff3531, 0xffff3131, 0xffff312d,
0xffff2d2d, 0xffff2d29, 0xffff2929, 0xffff2925, 0xffff2525, 0xffff2521,
0xffff2121, 0xffff211d, 0xffff1d1d, 0xffff1d19, 0xffff1919, 0xffff1915,
0xffff1515, 0xffff1511, 0xffff1111, 0xffff110d, 0xffff0d0d, 0xffff0d09,
0xffff0909, 0xffff0905, 0xffff0505, 0xffff0501, 0xffff0101, 0xffff00fd,
0xfffffcfc, 0xfffffcf8, 0xfffff8f8, 0xfffff8f4, 0xfffff4f4, 0xfffff4f0,
0xfffff0f0, 0xfffff0ec, 0xffffecec, 0xffffece8, 0xffffe8e8, 0xffffe8e4,
0xffffe4e4, 0xffffe4e0, 0xffffe0e0, 0xffffe0dc, 0xffffdcdc, 0xffffdcd8,
0xffffd8d8, 0xffffd8d4, 0xffffd4d4, 0xffffd4d0, 0xffffd0d0, 0xffffd0cc,
0xffffcccc, 0xffffccc8, 0xffffc8c8, 0xffffc8c4, 0xffffc4c4, 0xffffc4c0,
0xffffc0c0, 0xffffc0bc, 0xffffbcbc, 0xffffbcb8, 0xffffb8b8, 0xffffb8b4,
0xffffb4b4, 0xffffb4b0, 0xffffb0b0, 0xffffb0ac, 0xffffacac, 0xffffaca8,
0xffffa8a8, 0xffffa8a4, 0xffffa4a4, 0xffffa4a0, 0xffffa0a0, 0xffffa09c,
0xffff9c9c, 0xffff9c98, 0xffff9898, 0xffff9894, 0xffff9494, 0xffff9490,
0xffff9090, 0xffff908c, 0xffff8c8c, 0xffff8c88, 0xffff8888, 0xffff8884,
0xffff8484, 0xffff8480, 0xffff8080, 0xffff807c, 0xffff7c7c, 0xffff7c78,
0xffff7878, 0xffff7874, 0xffff7474, 0xffff7470, 0xffff7070, 0xffff706c,
0xffff6c6c, 0xffff6c68, 0xffff6868, 0xffff6864, 0xffff6464, 0xffff6460,
0xffff6060, 0xffff605c, 0xffff5c5c, 0xffff5c58, 0xffff5858, 0xffff5854,
0xffff5454, 0xffff5450, 0xffff5050, 0xffff504c, 0xffff4c4c, 0xffff4c48,
0xffff4848, 0xffff4844, 0xffff4444, 0xffff4440, 0xffff4040, 0xffff403c,
0xffff3c3c, 0xffff3c38, 0xffff3838, 0xffff3834, 0xffff3434, 0xffff3430,
0xffff3030, 0xffff302c, 0xffff2c2c, 0xffff2c28, 0xffff2828, 0xffff2824,
0xffff2424, 0xffff2420, 0xffff2020, 0xffff201c, 0xffff1c1c, 0xffff1c18,
0xffff1818, 0xffff1814, 0xffff1414, 0xffff1410, 0xffff1010, 0xffff100c,
0xffff0c0c, 0xffff0c08, 0xffff0808, 0xffff0804, 0xffff0404, 0xffff0400,
0xffff0000, 0xfffffffb,
};
static const u16 expected_csum_ipv6_magic[] = {
0x18d4, 0x3085, 0x2e4b, 0xd9f4, 0xbdc8, 0x78f, 0x1034, 0x8422, 0x6fc0,
0xd2f6, 0xbeb5, 0x9d3, 0x7e2a, 0x312e, 0x778e, 0xc1bb, 0x7cf2, 0x9d1e,
0xca21, 0xf3ff, 0x7569, 0xb02e, 0xca86, 0x7e76, 0x4539, 0x45e3, 0xf28d,
0xdf81, 0x8fd5, 0x3b5d, 0x8324, 0xf471, 0x83be, 0x1daf, 0x8c46, 0xe682,
0xd1fb, 0x6b2e, 0xe687, 0x2a33, 0x4833, 0x2d67, 0x660f, 0x2e79, 0xd65e,
0x6b62, 0x6672, 0x5dbd, 0x8680, 0xbaa5, 0x2229, 0x2125, 0x2d01, 0x1cc0,
0x6d36, 0x33c0, 0xee36, 0xd832, 0x9820, 0x8a31, 0x53c5, 0x2e2, 0xdb0e,
0x49ed, 0x17a7, 0x77a0, 0xd72e, 0x3d72, 0x7dc8, 0x5b17, 0xf55d, 0xa4d9,
0x1446, 0x5d56, 0x6b2e, 0x69a5, 0xadb6, 0xff2a, 0x92e, 0xe044, 0x3402,
0xbb60, 0xec7f, 0xe7e6, 0x1986, 0x32f4, 0x8f8, 0x5e00, 0x47c6, 0x3059,
0x3969, 0xe957, 0x4388, 0x2854, 0x3334, 0xea71, 0xa6de, 0x33f9, 0x83fc,
0x37b4, 0x5531, 0x3404, 0x1010, 0xed30, 0x610a, 0xc95, 0x9aed, 0x6ff,
0x5136, 0x2741, 0x660e, 0x8b80, 0xf71, 0xa263, 0x88af, 0x7a73, 0x3c37,
0x1908, 0x6db5, 0x2e92, 0x1cd2, 0x70c8, 0xee16, 0xe80, 0xcd55, 0x6e6,
0x6434, 0x127, 0x655d, 0x2ea0, 0xb4f4, 0xdc20, 0x5671, 0xe462, 0xe52b,
0xdb44, 0x3589, 0xc48f, 0xe60b, 0xd2d2, 0x66ad, 0x498, 0x436, 0xb917,
0xf0ca, 0x1a6e, 0x1cb7, 0xbf61, 0x2870, 0xc7e8, 0x5b30, 0xe4a5, 0x168,
0xadfc, 0xd035, 0xe690, 0xe283, 0xfb27, 0xe4ad, 0xb1a5, 0xf2d5, 0xc4b6,
0x8a30, 0xd7d5, 0x7df9, 0x91d5, 0x63ed, 0x2d21, 0x312b, 0xab19, 0xa632,
0x8d2e, 0xef06, 0x57b9, 0xc373, 0xbd1f, 0xa41f, 0x8444, 0x9975, 0x90cb,
0xc49c, 0xe965, 0x4eff, 0x5a, 0xef6d, 0xe81a, 0xe260, 0x853a, 0xff7a,
0x99aa, 0xb06b, 0xee19, 0xcc2c, 0xf34c, 0x7c49, 0xdac3, 0xa71e, 0xc988,
0x3845, 0x1014
};
static const u16 expected_fast_csum[] = {
0xda83, 0x45da, 0x4f46, 0x4e4f, 0x34e, 0xe902, 0xa5e9, 0x87a5, 0x7187,
0x5671, 0xf556, 0x6df5, 0x816d, 0x8f81, 0xbb8f, 0xfbba, 0x5afb, 0xbe5a,
0xedbe, 0xabee, 0x6aac, 0xe6b, 0xea0d, 0x67ea, 0x7e68, 0x8a7e, 0x6f8a,
0x3a70, 0x9f3a, 0xe89e, 0x75e8, 0x7976, 0xfa79, 0x2cfa, 0x3c2c, 0x463c,
0x7146, 0x7a71, 0x547a, 0xfd53, 0x99fc, 0xb699, 0x92b6, 0xdb91, 0xe8da,
0x5fe9, 0x1e60, 0xae1d, 0x39ae, 0xf439, 0xa1f4, 0xdda1, 0xede, 0x790f,
0x579, 0x1206, 0x9012, 0x2490, 0xd224, 0x5cd2, 0xa65d, 0xca7, 0x220d,
0xf922, 0xbf9, 0x920b, 0x1b92, 0x361c, 0x2e36, 0x4d2e, 0x24d, 0x2,
0xcfff, 0x90cf, 0xa591, 0x93a5, 0x7993, 0x9579, 0xc894, 0x50c8, 0x5f50,
0xd55e, 0xcad5, 0xf3c9, 0x8f4, 0x4409, 0x5043, 0x5b50, 0x55b, 0x2205,
0x1e22, 0x801e, 0x3780, 0xe137, 0x7ee0, 0xf67d, 0x3cf6, 0xa53c, 0x2ea5,
0x472e, 0x5147, 0xcf51, 0x1bcf, 0x951c, 0x1e95, 0xc71e, 0xe4c7, 0xc3e4,
0x3dc3, 0xee3d, 0xa4ed, 0xf9a4, 0xcbf8, 0x75cb, 0xb375, 0x50b4, 0x3551,
0xf835, 0x19f8, 0x8c1a, 0x538c, 0xad52, 0xa3ac, 0xb0a3, 0x5cb0, 0x6c5c,
0x5b6c, 0xc05a, 0x92c0, 0x4792, 0xbe47, 0x53be, 0x1554, 0x5715, 0x4b57,
0xe54a, 0x20e5, 0x21, 0xd500, 0xa1d4, 0xa8a1, 0x57a9, 0xca57, 0x5ca,
0x1c06, 0x4f1c, 0xe24e, 0xd9e2, 0xf0d9, 0x4af1, 0x474b, 0x8146, 0xe81,
0xfd0e, 0x84fd, 0x7c85, 0xba7c, 0x17ba, 0x4a17, 0x964a, 0xf595, 0xff5,
0x5310, 0x3253, 0x6432, 0x4263, 0x2242, 0xe121, 0x32e1, 0xf632, 0xc5f5,
0x21c6, 0x7d22, 0x8e7c, 0x418e, 0x5641, 0x3156, 0x7c31, 0x737c, 0x373,
0x2503, 0xc22a, 0x3c2, 0x4a04, 0x8549, 0x5285, 0xa352, 0xe8a3, 0x6fe8,
0x1a6f, 0x211a, 0xe021, 0x38e0, 0x7638, 0xf575, 0x9df5, 0x169e, 0xf116,
0x23f1, 0xcd23, 0xece, 0x660f, 0x4866, 0x6a48, 0x716a, 0xee71, 0xa2ee,
0xb8a2, 0x61b9, 0xa361, 0xf7a2, 0x26f7, 0x1127, 0x6611, 0xe065, 0x36e0,
0x1837, 0x3018, 0x1c30, 0x721b, 0x3e71, 0xe43d, 0x99e4, 0x9e9a, 0xb79d,
0xa9b7, 0xcaa, 0xeb0c, 0x4eb, 0x1305, 0x8813, 0xb687, 0xa9b6, 0xfba9,
0xd7fb, 0xccd8, 0x2ecd, 0x652f, 0xae65, 0x3fae, 0x3a40, 0x563a, 0x7556,
0x2776, 0x1228, 0xef12, 0xf9ee, 0xcef9, 0x56cf, 0xa956, 0x24a9, 0xba24,
0x5fba, 0x665f, 0xf465, 0x8ff4, 0x6d8f, 0x346d, 0x5f34, 0x385f, 0xd137,
0xb8d0, 0xacb8, 0x55ac, 0x7455, 0xe874, 0x89e8, 0xd189, 0xa0d1, 0xb2a0,
0xb8b2, 0x36b8, 0x5636, 0xd355, 0x8d3, 0x1908, 0x2118, 0xc21, 0x990c,
0x8b99, 0x158c, 0x7815, 0x9e78, 0x6f9e, 0x4470, 0x1d44, 0x341d, 0x2634,
0x3f26, 0x793e, 0xc79, 0xcc0b, 0x26cc, 0xd126, 0x1fd1, 0xb41f, 0xb6b4,
0x22b7, 0xa122, 0xa1, 0x7f01, 0x837e, 0x3b83, 0xaf3b, 0x6fae, 0x916f,
0xb490, 0xffb3, 0xceff, 0x50cf, 0x7550, 0x7275, 0x1272, 0x2613, 0xaa26,
0xd5aa, 0x7d5, 0x9607, 0x96, 0xb100, 0xf8b0, 0x4bf8, 0xdd4c, 0xeddd,
0x98ed, 0x2599, 0x9325, 0xeb92, 0x8feb, 0xcc8f, 0x2acd, 0x392b, 0x3b39,
0xcb3b, 0x6acb, 0xd46a, 0xb8d4, 0x6ab8, 0x106a, 0x2f10, 0x892f, 0x789,
0xc806, 0x45c8, 0x7445, 0x3c74, 0x3a3c, 0xcf39, 0xd7ce, 0x58d8, 0x6e58,
0x336e, 0x1034, 0xee10, 0xe9ed, 0xc2e9, 0x3fc2, 0xd53e, 0xd2d4, 0xead2,
0x8fea, 0x2190, 0x1162, 0xbe11, 0x8cbe, 0x6d8c, 0xfb6c, 0x6dfb, 0xd36e,
0x3ad3, 0xf3a, 0x870e, 0xc287, 0x53c3, 0xc54, 0x5b0c, 0x7d5a, 0x797d,
0xec79, 0x5dec, 0x4d5e, 0x184e, 0xd618, 0x60d6, 0xb360, 0x98b3, 0xf298,
0xb1f2, 0x69b1, 0xf969, 0xef9, 0xab0e, 0x21ab, 0xe321, 0x24e3, 0x8224,
0x5481, 0x5954, 0x7a59, 0xff7a, 0x7dff, 0x1a7d, 0xa51a, 0x46a5, 0x6b47,
0xe6b, 0x830e, 0xa083, 0xff9f, 0xd0ff, 0xffd0, 0xe6ff, 0x7de7, 0xc67d,
0xd0c6, 0x61d1, 0x3a62, 0xc3b, 0x150c, 0x1715, 0x4517, 0x5345, 0x3954,
0xdd39, 0xdadd, 0x32db, 0x6a33, 0xd169, 0x86d1, 0xb687, 0x3fb6, 0x883f,
0xa487, 0x39a4, 0x2139, 0xbe20, 0xffbe, 0xedfe, 0x8ded, 0x368e, 0xc335,
0x51c3, 0x9851, 0xf297, 0xd6f2, 0xb9d6, 0x95ba, 0x2096, 0xea1f, 0x76e9,
0x4e76, 0xe04d, 0xd0df, 0x80d0, 0xa280, 0xfca2, 0x75fc, 0xef75, 0x32ef,
0x6833, 0xdf68, 0xc4df, 0x76c4, 0xb77, 0xb10a, 0xbfb1, 0x58bf, 0x5258,
0x4d52, 0x6c4d, 0x7e6c, 0xb67e, 0xccb5, 0x8ccc, 0xbe8c, 0xc8bd, 0x9ac8,
0xa99b, 0x52a9, 0x2f53, 0xc30, 0x3e0c, 0xb83d, 0x83b7, 0x5383, 0x7e53,
0x4f7e, 0xe24e, 0xb3e1, 0x8db3, 0x618e, 0xc861, 0xfcc8, 0x34fc, 0x9b35,
0xaa9b, 0xb1aa, 0x5eb1, 0x395e, 0x8639, 0xd486, 0x8bd4, 0x558b, 0x2156,
0xf721, 0x4ef6, 0x14f, 0x7301, 0xdd72, 0x49de, 0x894a, 0x9889, 0x8898,
0x7788, 0x7b77, 0x637b, 0xb963, 0xabb9, 0x7cab, 0xc87b, 0x21c8, 0xcb21,
0xdfca, 0xbfdf, 0xf2bf, 0x6af2, 0x626b, 0xb261, 0x3cb2, 0xc63c, 0xc9c6,
0xc9c9, 0xb4c9, 0xf9b4, 0x91f9, 0x4091, 0x3a40, 0xcc39, 0xd1cb, 0x7ed1,
0x537f, 0x6753, 0xa167, 0xba49, 0x88ba, 0x7789, 0x3877, 0xf037, 0xd3ef,
0xb5d4, 0x55b6, 0xa555, 0xeca4, 0xa1ec, 0xb6a2, 0x7b7, 0x9507, 0xfd94,
0x82fd, 0x5c83, 0x765c, 0x9676, 0x3f97, 0xda3f, 0x6fda, 0x646f, 0x3064,
0x5e30, 0x655e, 0x6465, 0xcb64, 0xcdca, 0x4ccd, 0x3f4c, 0x243f, 0x6f24,
0x656f, 0x6065, 0x3560, 0x3b36, 0xac3b, 0x4aac, 0x714a, 0x7e71, 0xda7e,
0x7fda, 0xda7f, 0x6fda, 0xff6f, 0xc6ff, 0xedc6, 0xd4ed, 0x70d5, 0xeb70,
0xa3eb, 0x80a3, 0xca80, 0x3fcb, 0x2540, 0xf825, 0x7ef8, 0xf87e, 0x73f8,
0xb474, 0xb4b4, 0x92b5, 0x9293, 0x93, 0x3500, 0x7134, 0x9071, 0xfa8f,
0x51fa, 0x1452, 0xba13, 0x7ab9, 0x957a, 0x8a95, 0x6e8a, 0x6d6e, 0x7c6d,
0x447c, 0x9744, 0x4597, 0x8945, 0xef88, 0x8fee, 0x3190, 0x4831, 0x8447,
0xa183, 0x1da1, 0xd41d, 0x2dd4, 0x4f2e, 0xc94e, 0xcbc9, 0xc9cb, 0x9ec9,
0x319e, 0xd531, 0x20d5, 0x4021, 0xb23f, 0x29b2, 0xd828, 0xecd8, 0x5ded,
0xfc5d, 0x4dfc, 0xd24d, 0x6bd2, 0x5f6b, 0xb35e, 0x7fb3, 0xee7e, 0x56ee,
0xa657, 0x68a6, 0x8768, 0x7787, 0xb077, 0x4cb1, 0x764c, 0xb175, 0x7b1,
0x3d07, 0x603d, 0x3560, 0x3e35, 0xb03d, 0xd6b0, 0xc8d6, 0xd8c8, 0x8bd8,
0x3e8c, 0x303f, 0xd530, 0xf1d4, 0x42f1, 0xca42, 0xddca, 0x41dd, 0x3141,
0x132, 0xe901, 0x8e9, 0xbe09, 0xe0bd, 0x2ce0, 0x862d, 0x3986, 0x9139,
0x6d91, 0x6a6d, 0x8d6a, 0x1b8d, 0xac1b, 0xedab, 0x54ed, 0xc054, 0xcebf,
0xc1ce, 0x5c2, 0x3805, 0x6038, 0x5960, 0xd359, 0xdd3, 0xbe0d, 0xafbd,
0x6daf, 0x206d, 0x2c20, 0x862c, 0x8e86, 0xec8d, 0xa2ec, 0xa3a2, 0x51a3,
0x8051, 0xfd7f, 0x91fd, 0xa292, 0xaf14, 0xeeae, 0x59ef, 0x535a, 0x8653,
0x3986, 0x9539, 0xb895, 0xa0b8, 0x26a0, 0x2227, 0xc022, 0x77c0, 0xad77,
0x46ad, 0xaa46, 0x60aa, 0x8560, 0x4785, 0xd747, 0x45d7, 0x2346, 0x5f23,
0x25f, 0x1d02, 0x71d, 0x8206, 0xc82, 0x180c, 0x3018, 0x4b30, 0x4b,
0x3001, 0x1230, 0x2d12, 0x8c2d, 0x148d, 0x4015, 0x5f3f, 0x3d5f, 0x6b3d,
0x396b, 0x473a, 0xf746, 0x44f7, 0x8945, 0x3489, 0xcb34, 0x84ca, 0xd984,
0xf0d9, 0xbcf0, 0x63bd, 0x3264, 0xf332, 0x45f3, 0x7346, 0x5673, 0xb056,
0xd3b0, 0x4ad4, 0x184b, 0x7d18, 0x6c7d, 0xbb6c, 0xfeba, 0xe0fe, 0x10e1,
0x5410, 0x2954, 0x9f28, 0x3a9f, 0x5a3a, 0xdb59, 0xbdc, 0xb40b, 0x1ab4,
0x131b, 0x5d12, 0x6d5c, 0xe16c, 0xb0e0, 0x89b0, 0xba88, 0xbb, 0x3c01,
0xe13b, 0x6fe1, 0x446f, 0xa344, 0x81a3, 0xfe81, 0xc7fd, 0x38c8, 0xb38,
0x1a0b, 0x6d19, 0xf36c, 0x47f3, 0x6d48, 0xb76d, 0xd3b7, 0xd8d2, 0x52d9,
0x4b53, 0xa54a, 0x34a5, 0xc534, 0x9bc4, 0xed9b, 0xbeed, 0x3ebe, 0x233e,
0x9f22, 0x4a9f, 0x774b, 0x4577, 0xa545, 0x64a5, 0xb65, 0x870b, 0x487,
0x9204, 0x5f91, 0xd55f, 0x35d5, 0x1a35, 0x71a, 0x7a07, 0x4e7a, 0xfc4e,
0x1efc, 0x481f, 0x7448, 0xde74, 0xa7dd, 0x1ea7, 0xaa1e, 0xcfaa, 0xfbcf,
0xedfb, 0x6eee, 0x386f, 0x4538, 0x6e45, 0xd96d, 0x11d9, 0x7912, 0x4b79,
0x494b, 0x6049, 0xac5f, 0x65ac, 0x1366, 0x5913, 0xe458, 0x7ae4, 0x387a,
0x3c38, 0xb03c, 0x76b0, 0x9376, 0xe193, 0x42e1, 0x7742, 0x6476, 0x3564,
0x3c35, 0x6a3c, 0xcc69, 0x94cc, 0x5d95, 0xe5e, 0xee0d, 0x4ced, 0xce4c,
0x52ce, 0xaa52, 0xdaaa, 0xe4da, 0x1de5, 0x4530, 0x5445, 0x3954, 0xb639,
0x81b6, 0x7381, 0x1574, 0xc215, 0x10c2, 0x3f10, 0x6b3f, 0xe76b, 0x7be7,
0xbc7b, 0xf7bb, 0x41f7, 0xcc41, 0x38cc, 0x4239, 0xa942, 0x4a9, 0xc504,
0x7cc4, 0x437c, 0x6743, 0xea67, 0x8dea, 0xe88d, 0xd8e8, 0xdcd8, 0x17dd,
0x5718, 0x958, 0xa609, 0x41a5, 0x5842, 0x159, 0x9f01, 0x269f, 0x5a26,
0x405a, 0xc340, 0xb4c3, 0xd4b4, 0xf4d3, 0xf1f4, 0x39f2, 0xe439, 0x67e4,
0x4168, 0xa441, 0xdda3, 0xdedd, 0x9df, 0xab0a, 0xa5ab, 0x9a6, 0xba09,
0x9ab9, 0xad9a, 0x5ae, 0xe205, 0xece2, 0xecec, 0x14ed, 0xd614, 0x6bd5,
0x916c, 0x3391, 0x6f33, 0x206f, 0x8020, 0x780, 0x7207, 0x2472, 0x8a23,
0xb689, 0x3ab6, 0xf739, 0x97f6, 0xb097, 0xa4b0, 0xe6a4, 0x88e6, 0x2789,
0xb28, 0x350b, 0x1f35, 0x431e, 0x1043, 0xc30f, 0x79c3, 0x379, 0x5703,
0x3256, 0x4732, 0x7247, 0x9d72, 0x489d, 0xd348, 0xa4d3, 0x7ca4, 0xbf7b,
0x45c0, 0x7b45, 0x337b, 0x4034, 0x843f, 0xd083, 0x35d0, 0x6335, 0x4d63,
0xe14c, 0xcce0, 0xfecc, 0x35ff, 0x5636, 0xf856, 0xeef8, 0x2def, 0xfc2d,
0x4fc, 0x6e04, 0xb66d, 0x78b6, 0xbb78, 0x3dbb, 0x9a3d, 0x839a, 0x9283,
0x593, 0xd504, 0x23d5, 0x5424, 0xd054, 0x61d0, 0xdb61, 0x17db, 0x1f18,
0x381f, 0x9e37, 0x679e, 0x1d68, 0x381d, 0x8038, 0x917f, 0x491, 0xbb04,
0x23bb, 0x4124, 0xd41, 0xa30c, 0x8ba3, 0x8b8b, 0xc68b, 0xd2c6, 0xebd2,
0x93eb, 0xbd93, 0x99bd, 0x1a99, 0xea19, 0x58ea, 0xcf58, 0x73cf, 0x1073,
0x9e10, 0x139e, 0xea13, 0xcde9, 0x3ecd, 0x883f, 0xf89, 0x180f, 0x2a18,
0x212a, 0xce20, 0x73ce, 0xf373, 0x60f3, 0xad60, 0x4093, 0x8e40, 0xb98e,
0xbfb9, 0xf1bf, 0x8bf1, 0x5e8c, 0xe95e, 0x14e9, 0x4e14, 0x1c4e, 0x7f1c,
0xe77e, 0x6fe7, 0xf26f, 0x13f2, 0x8b13, 0xda8a, 0x5fda, 0xea5f, 0x4eea,
0xa84f, 0x88a8, 0x1f88, 0x2820, 0x9728, 0x5a97, 0x3f5b, 0xb23f, 0x70b2,
0x2c70, 0x232d, 0xf623, 0x4f6, 0x905, 0x7509, 0xd675, 0x28d7, 0x9428,
0x3794, 0xf036, 0x2bf0, 0xba2c, 0xedb9, 0xd7ed, 0x59d8, 0xed59, 0x4ed,
0xe304, 0x18e3, 0x5c19, 0x3d5c, 0x753d, 0x6d75, 0x956d, 0x7f95, 0xc47f,
0x83c4, 0xa84, 0x2e0a, 0x5f2e, 0xb95f, 0x77b9, 0x6d78, 0xf46d, 0x1bf4,
0xed1b, 0xd6ed, 0xe0d6, 0x5e1, 0x3905, 0x5638, 0xa355, 0x99a2, 0xbe99,
0xb4bd, 0x85b4, 0x2e86, 0x542e, 0x6654, 0xd765, 0x73d7, 0x3a74, 0x383a,
0x2638, 0x7826, 0x7677, 0x9a76, 0x7e99, 0x2e7e, 0xea2d, 0xa6ea, 0x8a7,
0x109, 0x3300, 0xad32, 0x5fad, 0x465f, 0x2f46, 0xc62f, 0xd4c5, 0xad5,
0xcb0a, 0x4cb, 0xb004, 0x7baf, 0xe47b, 0x92e4, 0x8e92, 0x638e, 0x1763,
0xc17, 0xf20b, 0x1ff2, 0x8920, 0x5889, 0xcb58, 0xf8cb, 0xcaf8, 0x84cb,
0x9f84, 0x8a9f, 0x918a, 0x4991, 0x8249, 0xff81, 0x46ff, 0x5046, 0x5f50,
0x725f, 0xf772, 0x8ef7, 0xe08f, 0xc1e0, 0x1fc2, 0x9e1f, 0x8b9d, 0x108b,
0x411, 0x2b04, 0xb02a, 0x1fb0, 0x1020, 0x7a0f, 0x587a, 0x8958, 0xb188,
0xb1b1, 0x49b2, 0xb949, 0x7ab9, 0x917a, 0xfc91, 0xe6fc, 0x47e7, 0xbc47,
0x8fbb, 0xea8e, 0x34ea, 0x2635, 0x1726, 0x9616, 0xc196, 0xa6c1, 0xf3a6,
0x11f3, 0x4811, 0x3e48, 0xeb3e, 0xf7ea, 0x1bf8, 0xdb1c, 0x8adb, 0xe18a,
0x42e1, 0x9d42, 0x5d9c, 0x6e5d, 0x286e, 0x4928, 0x9a49, 0xb09c, 0xa6b0,
0x2a7, 0xe702, 0xf5e6, 0x9af5, 0xf9b, 0x810f, 0x8080, 0x180, 0x1702,
0x5117, 0xa650, 0x11a6, 0x1011, 0x550f, 0xd554, 0xbdd5, 0x6bbe, 0xc66b,
0xfc7, 0x5510, 0x5555, 0x7655, 0x177, 0x2b02, 0x6f2a, 0xb70, 0x9f0b,
0xcf9e, 0xf3cf, 0x3ff4, 0xcb40, 0x8ecb, 0x768e, 0x5277, 0x8652, 0x9186,
0x9991, 0x5099, 0xd350, 0x93d3, 0x6d94, 0xe6d, 0x530e, 0x3153, 0xa531,
0x64a5, 0x7964, 0x7c79, 0x467c, 0x1746, 0x3017, 0x3730, 0x538, 0x5,
0x1e00, 0x5b1e, 0x955a, 0xae95, 0x3eaf, 0xff3e, 0xf8ff, 0xb2f9, 0xa1b3,
0xb2a1, 0x5b2, 0xad05, 0x7cac, 0x2d7c, 0xd32c, 0x80d2, 0x7280, 0x8d72,
0x1b8e, 0x831b, 0xac82, 0xfdac, 0xa7fd, 0x15a8, 0xd614, 0xe0d5, 0x7be0,
0xb37b, 0x61b3, 0x9661, 0x9d95, 0xc79d, 0x83c7, 0xd883, 0xead7, 0xceb,
0xf60c, 0xa9f5, 0x19a9, 0xa019, 0x8f9f, 0xd48f, 0x3ad5, 0x853a, 0x985,
0x5309, 0x6f52, 0x1370, 0x6e13, 0xa96d, 0x98a9, 0x5198, 0x9f51, 0xb69f,
0xa1b6, 0x2ea1, 0x672e, 0x2067, 0x6520, 0xaf65, 0x6eaf, 0x7e6f, 0xee7e,
0x17ef, 0xa917, 0xcea8, 0x9ace, 0xff99, 0x5dff, 0xdf5d, 0x38df, 0xa39,
0x1c0b, 0xe01b, 0x46e0, 0xcb46, 0x90cb, 0xba90, 0x4bb, 0x9104, 0x9d90,
0xc89c, 0xf6c8, 0x6cf6, 0x886c, 0x1789, 0xbd17, 0x70bc, 0x7e71, 0x17e,
0x1f01, 0xa01f, 0xbaa0, 0x14bb, 0xfc14, 0x7afb, 0xa07a, 0x3da0, 0xbf3d,
0x48bf, 0x8c48, 0x968b, 0x9d96, 0xfd9d, 0x96fd, 0x9796, 0x6b97, 0xd16b,
0xf4d1, 0x3bf4, 0x253c, 0x9125, 0x6691, 0xc166, 0x34c1, 0x5735, 0x1a57,
0xdc19, 0x77db, 0x8577, 0x4a85, 0x824a, 0x9182, 0x7f91, 0xfd7f, 0xb4c3,
0xb5b4, 0xb3b5, 0x7eb3, 0x617e, 0x4e61, 0xa4f, 0x530a, 0x3f52, 0xa33e,
0x34a3, 0x9234, 0xf091, 0xf4f0, 0x1bf5, 0x311b, 0x9631, 0x6a96, 0x386b,
0x1d39, 0xe91d, 0xe8e9, 0x69e8, 0x426a, 0xee42, 0x89ee, 0x368a, 0x2837,
0x7428, 0x5974, 0x6159, 0x1d62, 0x7b1d, 0xf77a, 0x7bf7, 0x6b7c, 0x696c,
0xf969, 0x4cf9, 0x714c, 0x4e71, 0x6b4e, 0x256c, 0x6e25, 0xe96d, 0x94e9,
0x8f94, 0x3e8f, 0x343e, 0x4634, 0xb646, 0x97b5, 0x8997, 0xe8a, 0x900e,
0x8090, 0xfd80, 0xa0fd, 0x16a1, 0xf416, 0xebf4, 0x95ec, 0x1196, 0x8911,
0x3d89, 0xda3c, 0x9fd9, 0xd79f, 0x4bd7, 0x214c, 0x3021, 0x4f30, 0x994e,
0x5c99, 0x6f5d, 0x326f, 0xab31, 0x6aab, 0xe969, 0x90e9, 0x1190, 0xff10,
0xa2fe, 0xe0a2, 0x66e1, 0x4067, 0x9e3f, 0x2d9e, 0x712d, 0x8170, 0xd180,
0xffd1, 0x25ff, 0x3826, 0x2538, 0x5f24, 0xc45e, 0x1cc4, 0xdf1c, 0x93df,
0xc793, 0x80c7, 0x2380, 0xd223, 0x7ed2, 0xfc7e, 0x22fd, 0x7422, 0x1474,
0xb714, 0x7db6, 0x857d, 0xa85, 0xa60a, 0x88a6, 0x4289, 0x7842, 0xc278,
0xf7c2, 0xcdf7, 0x84cd, 0xae84, 0x8cae, 0xb98c, 0x1aba, 0x4d1a, 0x884c,
0x4688, 0xcc46, 0xd8cb, 0x2bd9, 0xbe2b, 0xa2be, 0x72a2, 0xf772, 0xd2f6,
0x75d2, 0xc075, 0xa3c0, 0x63a3, 0xae63, 0x8fae, 0x2a90, 0x5f2a, 0xef5f,
0x5cef, 0xa05c, 0x89a0, 0x5e89, 0x6b5e, 0x736b, 0x773, 0x9d07, 0xe99c,
0x27ea, 0x2028, 0xc20, 0x980b, 0x4797, 0x2848, 0x9828, 0xc197, 0x48c2,
0x2449, 0x7024, 0x570, 0x3e05, 0xd3e, 0xf60c, 0xbbf5, 0x69bb, 0x3f6a,
0x740, 0xf006, 0xe0ef, 0xbbe0, 0xadbb, 0x56ad, 0xcf56, 0xbfce, 0xa9bf,
0x205b, 0x6920, 0xae69, 0x50ae, 0x2050, 0xf01f, 0x27f0, 0x9427, 0x8993,
0x8689, 0x4087, 0x6e40, 0xb16e, 0xa1b1, 0xe8a1, 0x87e8, 0x6f88, 0xfe6f,
0x4cfe, 0xe94d, 0xd5e9, 0x47d6, 0x3148, 0x5f31, 0xc35f, 0x13c4, 0xa413,
0x5a5, 0x2405, 0xc223, 0x66c2, 0x3667, 0x5e37, 0x5f5e, 0x2f5f, 0x8c2f,
0xe48c, 0xd0e4, 0x4d1, 0xd104, 0xe4d0, 0xcee4, 0xfcf, 0x480f, 0xa447,
0x5ea4, 0xff5e, 0xbefe, 0x8dbe, 0x1d8e, 0x411d, 0x1841, 0x6918, 0x5469,
0x1155, 0xc611, 0xaac6, 0x37ab, 0x2f37, 0xca2e, 0x87ca, 0xbd87, 0xabbd,
0xb3ab, 0xcb4, 0xce0c, 0xfccd, 0xa5fd, 0x72a5, 0xf072, 0x83f0, 0xfe83,
0x97fd, 0xc997, 0xb0c9, 0xadb0, 0xe6ac, 0x88e6, 0x1088, 0xbe10, 0x16be,
0xa916, 0xa3a8, 0x46a3, 0x5447, 0xe953, 0x84e8, 0x2085, 0xa11f, 0xfa1,
0xdd0f, 0xbedc, 0x5abe, 0x805a, 0xc97f, 0x6dc9, 0x826d, 0x4a82, 0x934a,
0x5293, 0xd852, 0xd3d8, 0xadd3, 0xf4ad, 0xf3f4, 0xfcf3, 0xfefc, 0xcafe,
0xb7ca, 0x3cb8, 0xa13c, 0x18a1, 0x1418, 0xea13, 0x91ea, 0xf891, 0x53f8,
0xa254, 0xe9a2, 0x87ea, 0x4188, 0x1c41, 0xdc1b, 0xf5db, 0xcaf5, 0x45ca,
0x6d45, 0x396d, 0xde39, 0x90dd, 0x1e91, 0x1e, 0x7b00, 0x6a7b, 0xa46a,
0xc9a3, 0x9bc9, 0x389b, 0x1139, 0x5211, 0x1f52, 0xeb1f, 0xabeb, 0x48ab,
0x9348, 0xb392, 0x17b3, 0x1618, 0x5b16, 0x175b, 0xdc17, 0xdedb, 0x1cdf,
0xeb1c, 0xd1ea, 0x4ad2, 0xd4b, 0xc20c, 0x24c2, 0x7b25, 0x137b, 0x8b13,
0x618b, 0xa061, 0xff9f, 0xfffe, 0x72ff, 0xf572, 0xe2f5, 0xcfe2, 0xd2cf,
0x75d3, 0x6a76, 0xc469, 0x1ec4, 0xfc1d, 0x59fb, 0x455a, 0x7a45, 0xa479,
0xb7a4
};
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
static u8 tmp_buf[TEST_BUFLEN];
#define full_csum(buff, len, sum) csum_fold(csum_partial(buff, len, sum))
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
#define CHECK_EQ(lhs, rhs) KUNIT_ASSERT_EQ(test, (__force u64)lhs, (__force u64)rhs)
static __sum16 to_sum16(u16 x)
{
return (__force __sum16)le16_to_cpu((__force __le16)x);
}
/* This function swaps the bytes inside each half of a __wsum */
static __wsum to_wsum(u32 x)
{
u16 hi = le16_to_cpu((__force __le16)(x >> 16));
u16 lo = le16_to_cpu((__force __le16)x);
return (__force __wsum)((hi << 16) | lo);
}
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
static void assert_setup_correct(struct kunit *test)
{
CHECK_EQ(sizeof(random_buf) / sizeof(random_buf[0]), MAX_LEN);
CHECK_EQ(sizeof(expected_results) / sizeof(expected_results[0]),
MAX_LEN);
CHECK_EQ(sizeof(init_sums_no_overflow) /
sizeof(init_sums_no_overflow[0]),
MAX_LEN);
}
/*
* Test with randomized input (pre determined random with known results).
*/
static void test_csum_fixed_random_inputs(struct kunit *test)
{
int len, align;
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
__wsum sum;
__sum16 result, expec;
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
assert_setup_correct(test);
for (align = 0; align < TEST_BUFLEN; ++align) {
memcpy(&tmp_buf[align], random_buf,
min(MAX_LEN, TEST_BUFLEN - align));
for (len = 0; len < MAX_LEN && (align + len) < TEST_BUFLEN;
++len) {
/*
* Test the precomputed random input.
*/
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
sum = to_wsum(random_init_sum);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
result = full_csum(&tmp_buf[align], len, sum);
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
expec = to_sum16(expected_results[len]);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
CHECK_EQ(result, expec);
}
}
}
/*
* All ones input test. If there are any missing carry operations, it fails.
*/
static void test_csum_all_carry_inputs(struct kunit *test)
{
int len, align;
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
__wsum sum;
__sum16 result, expec;
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
assert_setup_correct(test);
memset(tmp_buf, 0xff, TEST_BUFLEN);
for (align = 0; align < TEST_BUFLEN; ++align) {
for (len = 0; len < MAX_LEN && (align + len) < TEST_BUFLEN;
++len) {
/*
* All carries from input and initial sum.
*/
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
sum = to_wsum(0xffffffff);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
result = full_csum(&tmp_buf[align], len, sum);
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
expec = to_sum16((len & 1) ? 0xff00 : 0);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
CHECK_EQ(result, expec);
/*
* All carries from input.
*/
sum = 0;
result = full_csum(&tmp_buf[align], len, sum);
if (len & 1)
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
expec = to_sum16(0xff00);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
else if (len)
expec = 0;
else
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
expec = to_sum16(0xffff);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
CHECK_EQ(result, expec);
}
}
}
/*
* Test with input that alone doesn't cause any carries. By selecting the
* maximum initial sum, this allows us to test that there are no carries
* where there shouldn't be.
*/
static void test_csum_no_carry_inputs(struct kunit *test)
{
int len, align;
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
__wsum sum;
__sum16 result, expec;
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
assert_setup_correct(test);
memset(tmp_buf, 0x4, TEST_BUFLEN);
for (align = 0; align < TEST_BUFLEN; ++align) {
for (len = 0; len < MAX_LEN && (align + len) < TEST_BUFLEN;
++len) {
/*
* Expect no carries.
*/
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
sum = to_wsum(init_sums_no_overflow[len]);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
result = full_csum(&tmp_buf[align], len, sum);
expec = 0;
CHECK_EQ(result, expec);
/*
* Expect one carry.
*/
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
sum = to_wsum(init_sums_no_overflow[len] + 1);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
result = full_csum(&tmp_buf[align], len, sum);
kunit: Fix checksum tests on big endian CPUs On powerpc64le checksum kunit tests work: [ 2.011457][ T1] KTAP version 1 [ 2.011662][ T1] # Subtest: checksum [ 2.011848][ T1] 1..3 [ 2.034710][ T1] ok 1 test_csum_fixed_random_inputs [ 2.079325][ T1] ok 2 test_csum_all_carry_inputs [ 2.127102][ T1] ok 3 test_csum_no_carry_inputs [ 2.127202][ T1] # checksum: pass:3 fail:0 skip:0 total:3 [ 2.127533][ T1] # Totals: pass:3 fail:0 skip:0 total:3 [ 2.127956][ T1] ok 1 checksum But on powerpc64 and powerpc32 they fail: [ 1.859890][ T1] KTAP version 1 [ 1.860041][ T1] # Subtest: checksum [ 1.860201][ T1] 1..3 [ 1.861927][ T58] # test_csum_fixed_random_inputs: ASSERTION FAILED at lib/checksum_kunit.c:243 [ 1.861927][ T58] Expected result == expec, but [ 1.861927][ T58] result == 54991 (0xd6cf) [ 1.861927][ T58] expec == 33316 (0x8224) [ 1.863742][ T1] not ok 1 test_csum_fixed_random_inputs [ 1.864520][ T60] # test_csum_all_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:267 [ 1.864520][ T60] Expected result == expec, but [ 1.864520][ T60] result == 255 (0xff) [ 1.864520][ T60] expec == 65280 (0xff00) [ 1.868820][ T1] not ok 2 test_csum_all_carry_inputs [ 1.869977][ T62] # test_csum_no_carry_inputs: ASSERTION FAILED at lib/checksum_kunit.c:306 [ 1.869977][ T62] Expected result == expec, but [ 1.869977][ T62] result == 64515 (0xfc03) [ 1.869977][ T62] expec == 0 (0x0) [ 1.872060][ T1] not ok 3 test_csum_no_carry_inputs [ 1.872102][ T1] # checksum: pass:0 fail:3 skip:0 total:3 [ 1.872458][ T1] # Totals: pass:0 fail:3 skip:0 total:3 [ 1.872791][ T1] not ok 3 checksum This is because all expected values were calculated for X86 which is little endian. On big endian systems all precalculated 16 bits halves must be byte swapped. And this is confirmed by a huge amount of sparse errors when building with C=2 So fix all sparse errors and it will naturally work on all endianness. Fixes: 688eb8191b47 ("x86/csum: Improve performance of `csum_partial`") Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu> Signed-off-by: David S. Miller <davem@davemloft.net>
2023-08-23 13:21:43 +00:00
expec = to_sum16(len ? 0xfffe : 0xffff);
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
CHECK_EQ(result, expec);
}
}
}
static void test_ip_fast_csum(struct kunit *test)
{
__sum16 csum_result;
u16 expected;
for (int len = IPv4_MIN_WORDS; len < IPv4_MAX_WORDS; len++) {
for (int index = 0; index < NUM_IP_FAST_CSUM_TESTS; index++) {
csum_result = ip_fast_csum(random_buf + index, len);
expected =
expected_fast_csum[(len - IPv4_MIN_WORDS) *
NUM_IP_FAST_CSUM_TESTS +
index];
CHECK_EQ(to_sum16(expected), csum_result);
}
}
}
static void test_csum_ipv6_magic(struct kunit *test)
{
#if defined(CONFIG_NET)
const struct in6_addr *saddr;
const struct in6_addr *daddr;
unsigned int len;
unsigned char proto;
__wsum csum;
const int daddr_offset = sizeof(struct in6_addr);
const int len_offset = sizeof(struct in6_addr) + sizeof(struct in6_addr);
const int proto_offset = sizeof(struct in6_addr) + sizeof(struct in6_addr) +
sizeof(int);
const int csum_offset = sizeof(struct in6_addr) + sizeof(struct in6_addr) +
sizeof(int) + sizeof(char);
for (int i = 0; i < NUM_IPv6_TESTS; i++) {
saddr = (const struct in6_addr *)(random_buf + i);
daddr = (const struct in6_addr *)(random_buf + i +
daddr_offset);
len = le32_to_cpu(*(__le32 *)(random_buf + i + len_offset));
proto = *(random_buf + i + proto_offset);
csum = *(__wsum *)(random_buf + i + csum_offset);
CHECK_EQ(to_sum16(expected_csum_ipv6_magic[i]),
csum_ipv6_magic(saddr, daddr, len, proto, csum));
}
#endif /* !CONFIG_NET */
}
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
static struct kunit_case __refdata checksum_test_cases[] = {
KUNIT_CASE(test_csum_fixed_random_inputs),
KUNIT_CASE(test_csum_all_carry_inputs),
KUNIT_CASE(test_csum_no_carry_inputs),
KUNIT_CASE(test_ip_fast_csum),
KUNIT_CASE(test_csum_ipv6_magic),
x86/csum: Improve performance of `csum_partial` 1) Add special case for len == 40 as that is the hottest value. The nets a ~8-9% latency improvement and a ~30% throughput improvement in the len == 40 case. 2) Use multiple accumulators in the 64-byte loop. This dramatically improves ILP and results in up to a 40% latency/throughput improvement (better for more iterations). Results from benchmarking on Icelake. Times measured with rdtsc() len lat_new lat_old r tput_new tput_old r 8 3.58 3.47 1.032 3.58 3.51 1.021 16 4.14 4.02 1.028 3.96 3.78 1.046 24 4.99 5.03 0.992 4.23 4.03 1.050 32 5.09 5.08 1.001 4.68 4.47 1.048 40 5.57 6.08 0.916 3.05 4.43 0.690 48 6.65 6.63 1.003 4.97 4.69 1.059 56 7.74 7.72 1.003 5.22 4.95 1.055 64 6.65 7.22 0.921 6.38 6.42 0.994 96 9.43 9.96 0.946 7.46 7.54 0.990 128 9.39 12.15 0.773 8.90 8.79 1.012 200 12.65 18.08 0.699 11.63 11.60 1.002 272 15.82 23.37 0.677 14.43 14.35 1.005 440 24.12 36.43 0.662 21.57 22.69 0.951 952 46.20 74.01 0.624 42.98 53.12 0.809 1024 47.12 78.24 0.602 46.36 58.83 0.788 1552 72.01 117.30 0.614 71.92 96.78 0.743 2048 93.07 153.25 0.607 93.28 137.20 0.680 2600 114.73 194.30 0.590 114.28 179.32 0.637 3608 156.34 268.41 0.582 154.97 254.02 0.610 4096 175.01 304.03 0.576 175.89 292.08 0.602 There is no such thing as a free lunch, however, and the special case for len == 40 does add overhead to the len != 40 cases. This seems to amount to be ~5% throughput and slightly less in terms of latency. Testing: Part of this change is a new kunit test. The tests check all alignment X length pairs in [0, 64) X [0, 512). There are three cases. 1) Precomputed random inputs/seed. The expected results where generated use the generic implementation (which is assumed to be non-buggy). 2) An input of all 1s. The goal of this test is to catch any case a carry is missing. 3) An input that never carries. The goal of this test si to catch any case of incorrectly carrying. More exhaustive tests that test all alignment X length pairs in [0, 8192) X [0, 8192] on random data are also available here: https://github.com/goldsteinn/csum-reproduction The reposity also has the code for reproducing the above benchmark numbers. Signed-off-by: Noah Goldstein <goldstein.w.n@gmail.com> Signed-off-by: Dave Hansen <dave.hansen@linux.intel.com> Link: https://lore.kernel.org/all/20230511011002.935690-1-goldstein.w.n%40gmail.com
2023-05-11 01:10:02 +00:00
{}
};
static struct kunit_suite checksum_test_suite = {
.name = "checksum",
.test_cases = checksum_test_cases,
};
kunit_test_suites(&checksum_test_suite);
MODULE_AUTHOR("Noah Goldstein <goldstein.w.n@gmail.com>");
MODULE_LICENSE("GPL");