Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
Hi Martin, On Tue, Nov 20, 2018 at 5:29 PM Martin Willi wrote: > Thanks for the offer, no need at this time. But I certainly would > welcome if you could do some (Wireguard) benching with that code to see > if it works for you. I certainly will test it in a few different network circumstances, especially since real testing like this is sometimes more telling than busy-loop benchmarks. > > Actually, similarly here, a 10nm Cannon Lake machine should be > > arriving at my house this week, which should make for some > > interesting testing ground for non-throttled zmm, if you'd like to > > play with it. > > Maybe in a future iteration, thanks. In fact would it be interesting to > know if Cannon Lake can handle that throttling better. Everything I've read on the Internet seems to indicate that's the case, so one of the first things I'll be doing is seeing if that's true. There are also the AVX512 IFMA instructions to play with! Jason
Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
Hi Jason, > [...] I have a massive Xeon Gold 5120 machine that I can give you > access to if you'd like to do some testing and benching. Thanks for the offer, no need at this time. But I certainly would welcome if you could do some (Wireguard) benching with that code to see if it works for you. > Actually, similarly here, a 10nm Cannon Lake machine should be > arriving at my house this week, which should make for some > interesting testing ground for non-throttled zmm, if you'd like to > play with it. Maybe in a future iteration, thanks. In fact would it be interesting to know if Cannon Lake can handle that throttling better. Regards Martin
Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
Hi Martin, On Mon, Nov 19, 2018 at 8:52 AM Martin Willi wrote: > > Adding AVX-512VL support is relatively simple. I have a patchset mostly > ready that is more than competitive with the code from Zinc. I'll clean > that up and do more testing before posting it later this week. Terrific. Depending on how it turns out, it'll be nice to try integrating this into Zinc. I have a massive Xeon Gold 5120 machine that I can give you access to if you'd like to do some testing and benching. Poke me on IRC -- I'm zx2c4. > I don't think that having AVX-512F is that important until it is really > usable on CPUs in the market. Actually, similarly here, a 10nm Cannon Lake machine should be arriving at my house this week, which should make for some interesting testing ground for non-throttled zmm, if you'd like to play with it. Jason
Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
Hi Jason, > I'd be inclined to roll with your implementation if it can eventually > become competitive with Andy Polyakov's, [...] I think for the SSSE3/AVX2 code paths it is competitive; especially for small sizes it is faster, which is not that unimportant when implementing layer 3 VPNs. > there are still no AVX-512 paths, which means it's considerably > slower on all newer generation Intel chips. Andy's has the AVX-512VL > implementation for Skylake (using ymm, so as not to hit throttling) > and AVX-512F for Cannon Lake and beyond (using zmm). I don't think that having AVX-512F is that important until it is really usable on CPUs in the market. Adding AVX-512VL support is relatively simple. I have a patchset mostly ready that is more than competitive with the code from Zinc. I'll clean that up and do more testing before posting it later this week. Best regards Martin
Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
On Sun, Nov 11, 2018 at 10:36:24AM +0100, Martin Willi wrote: > This patchset improves performance of the ChaCha20 SIMD implementations > for x86_64. For some specific encryption lengths, performance is more > than doubled. Two mechanisms are used to achieve this: > > * Instead of calculating the minimal number of required blocks for a > given encryption length, functions producing more blocks are used > more aggressively. Calculating a 4-block function can be faster than > calculating a 2-block and a 1-block function, even if only three > blocks are actually required. > > * In addition to the 8-block AVX2 function, a 4-block and a 2-block > function are introduced. > > Patches 1-3 add support for partial lengths to the existing 1-, 4- and > 8-block functions. Patch 4 makes use of that by engaging the next higher > level block functions more aggressively. Patch 5 and 6 add the new AVX2 > functions for 2 and 4 blocks. Patches are based on cryptodev and would > need adjustments to apply on top of the Adiantum patchset. > > Note that the more aggressive use of larger block functions calculate > blocks that may get discarded. This may have a negative impact on energy > usage or the processors thermal budget. However, with the new block > functions we can avoid this over-calculation for many lengths, so the > performance win can be considered more important. > > Below are performance numbers measured with tcrypt using additional > encryption lengths; numbers in kOps/s, on my i7-5557U. old is the > existing, new the implementation with this patchset. As comparison > the numbers for zinc in v6: > > len old new zinc >8 5908 5818 5818 > 16 5917 5828 5726 > 24 5916 5869 5757 > 32 5920 5789 5813 > 40 5868 5799 5710 > 48 5877 5761 5761 > 56 5869 5797 5742 > 64 5897 5862 5685 > 72 3381 4979 3520 > 80 3364 5541 3475 > 88 3350 4977 3424 > 96 3342 5530 3371 > 104 3328 4923 3313 > 112 3317 5528 3207 > 120 3313 4970 3150 > 128 3492 5535 3568 > 136 2487 4570 3690 > 144 2481 5047 3599 > 152 2473 4565 3566 > 160 2459 5022 3515 > 168 2461 4550 3437 > 176 2454 5020 3325 > 184 2449 4535 3279 > 192 2538 5011 3762 > 200 1962 4537 3702 > 208 1962 4971 3622 > 216 1954 4487 3518 > 224 1949 4936 3445 > 232 1948 4497 3422 > 240 1941 4947 3317 > 248 1940 4481 3279 > 256 3798 4964 3723 > 264 2638 3577 3639 > 272 2637 3567 3597 > 280 2628 3563 3565 > 288 2630 3795 3484 > 296 2621 3580 3422 > 304 2612 3569 3352 > 312 2602 3599 3308 > 320 2694 3821 3694 > 328 2060 3538 3681 > 336 2054 3565 3599 > 344 2054 3553 3523 > 352 2049 3809 3419 > 360 2045 3575 3403 > 368 2035 3560 3334 > 376 2036 3555 3257 > 384 2092 3785 3715 > 392 1691 3505 3612 > 400 1684 3527 3553 > 408 1686 3527 3496 > 416 1684 3804 3430 > 424 1681 3555 3402 > 432 1675 3559 3311 > 440 1672 3558 3275 > 448 1710 3780 3689 > 456 1431 3541 3618 > 464 1428 3538 3576 > 472 1430 3527 3509 > 480 1426 3788 3405 > 488 1423 3502 3397 > 496 1423 3519 3298 > 504 1418 3519 3277 > 512 3694 3736 3735 > 520 2601 2571 2209 > 528 2601 2677 2148 > 536 2587 2534 2164 > 544 2578 2659 2138 > 552 2570 2552 2126 > 560 2566 2661 2035 > 568 2567 2542 2041 > 576 2639 2674 2199 > 584 2031 2531 2183 > 592 2027 2660 2145 > 600 2016 2513 2155 > 608 2009 2638 2133 > 616 2006 2522 2115 > 624 2000 2649 2064 > 632 1996 2518 2045 > 640 2053 2651 2188 > 648 1666 2402 2182 > 656 1663 2517 2158 > 664 1659 2397 2147 > 672 1657 2510 2139 > 680 1656 2394 2114 > 688 1653 2497 2077 > 696 1646 2393 2043 > 704 1678 2510 2208 > 712 1414 2391 2189 > 720 1412 2506 2169 > 728 1411 2384 2145 > 736 1408 2494 2142 > 744 1408 2379 2081 > 752 1405 2485 2064 > 760 1403 2376 2043 > 768 2189 2498 2211 > 776 1756 2137 2192 > 784 1746 2145 2146 > 792 1744 2141 2141 > 800 1743 2094 > 808 1742 2140 2100 > 816 1735 2134 2061 > 824 1731 2135 2045 > 832 1778 2223 > 840 1480 2132 2184 > 848 1480 2134 2173 > 856 1476 2124 2145 > 864 1474 2210 2126 > 872 1472 2127 2105 > 880 1463 2123 2056 > 888 1468 2123 2043 > 896 1494 2208 2219 > 904 1278 2120 2192 > 912 1277 2121 2170 > 920 1273 2118 2149 > 928 1272 2207 2125 > 936 1267 2125 2098 > 944 1265 2127 2060 > 952 1267 2126 2049 > 960 1289 2213 2204 > 968 1125 2123 2187 > 976 1122 2127 2166 > 984 1120 2123 2136 > 992 1118 2207 2119 > 1000 1118 2120 2101 > 1008 1117 2122 2042 > 1016 1115 2121 2048 > 1024 2174 2191 2195 > 1032 1748 1724 1565 > 1040 1745 1782 1544 > 1048 1736 1737 1554 > 1056 1738 1802 1541 > 1064 1735 1728 1523 > 1072 1730 1780 1507 > 1080 1729 1724 1497 > 1088 1757 1783 1592 > 1096 1475 1723 1575 > 1104 1474 1778 1563 > 1112 1472 1708 1544 > 1120 1468 1774 1521 > 1128 1466 1718 1521 > 1136 1462 1780 1501 > 1144 1460 1719 1491 > 1152 1481 1782 1575 > 1160 1271 1647 1558 > 1168 1271 1706 1554 > 1176 1268 1645 1545 > 1184 1265 1711 1538 > 1192 1265 1648 1530 > 1200 1264 1705 1493 > 1208 1262 1647 1498 > 1216 1277 1695 1581
Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
Hi Martin, This is nice work, and given that it's quite clean -- and that it's usually hard to screw up chacha in subtle ways when test vectors pass (unlike, say, poly1305 or curve25519), I'd be inclined to roll with your implementation if it can eventually become competitive with Andy Polyakov's, which I'm currently working on for Zinc (which no longer has pre-generated code, addressing the biggest hurdle; v9 will be sent shortly). Specifically, I'm not quite sure the improvements here tip the balance apply to all avx2 microarchitectures, and most importantly, there are still no AVX-512 paths, which means it's considerably slower on all newer generation Intel chips. Andy's has the AVX-512VL implementation for Skylake (using ymm, so as not to hit throttling) and AVX-512F for Cannon Lake and beyond (using zmm). I've attached some measurements below showing how stark the difference is. The take away is that while Andy's implementation is still ahead in terms of performance today, I'd certainly encourage your efforts to gain parity with that, and I'd be happy have that when the performance and fuzzing time is right for it. So please do keep chipping away at it; I think it's a potentially useful effort. Regards, Jason size old zinc 0 64 54 16 386 372 32 388 396 48 388 420 64 366 350 80 708 666 96 708 692 112 706 736 128 692 648 144 1036 682 160 1036 708 176 1036 730 192 1016 658 208 1360 684 224 1362 708 240 1360 732 256 644 500 272 990 526 288 988 556 304 988 576 320 972 500 336 1314 532 352 1316 558 368 1318 578 384 1308 506 400 1644 532 416 1644 556 432 1644 594 448 1624 508 464 1970 534 480 1970 556 496 1968 582 512 660 624 528 1016 682 544 1016 702 560 1018 728 576 998 654 592 1344 680 608 1344 708 624 1344 730 640 1326 654 656 1670 686 672 1670 708 688 1670 732 704 1652 658 720 1998 682 736 1998 710 752 1996 734 768 1256 662 784 1606 688 800 1606 714 816 1606 736 832 1584 660 848 1948 688 864 1950 714 880 1948 736 896 1912 688 912 2258 718 928 2258 744 944 2256 768 960 2238 692 976 2584 718 992 2584 744 1008 2584 770 On Thu, Nov 15, 2018 at 6:21 PM Herbert Xu wrote: > > On Sun, Nov 11, 2018 at 10:36:24AM +0100, Martin Willi wrote: > > This patchset improves performance of the ChaCha20 SIMD implementations > > for x86_64. For some specific encryption lengths, performance is more > > than doubled. Two mechanisms are used to achieve this: > > > > * Instead of calculating the minimal number of required blocks for a > > given encryption length, functions producing more blocks are used > > more aggressively. Calculating a 4-block function can be faster than > > calculating a 2-block and a 1-block function, even if only three > > blocks are actually required. > > > > * In addition to the 8-block AVX2 function, a 4-block and a 2-block > > function are introduced. > > > > Patches 1-3 add support for partial lengths to the existing 1-, 4- and > > 8-block functions. Patch 4 makes use of that by engaging the next higher > > level block functions more aggressively. Patch 5 and 6 add the new AVX2 > > functions for 2 and 4 blocks. Patches are based on cryptodev and would > > need adjustments to apply on top of the Adiantum patchset. > > > > Note that the more aggressive use of larger block functions calculate > > blocks that may get discarded. This may have a negative impact on energy > > usage or the processors thermal budget. However, with the new block > > functions we can avoid this over-calculation for many lengths, so the > > performance win can be considered more important. > > > > Below are performance numbers measured with tcrypt using additional > > encryption lengths; numbers in kOps/s, on my i7-5557U. old is the > > existing, new the implementation with this patchset. As comparison > > the numbers for zinc in v6: > > > > len old new zinc > >8 5908 5818 5818 > > 16 5917 5828 5726 > > 24 5916 5869 5757 > > 32 5920 5789 5813 > > 40 5868 5799 5710 > > 48 5877 5761 5761 > > 56 5869 5797 5742 > > 64 5897 5862 5685 > > 72 3381 4979 3520 > > 80 3364 5541 3475 > > 88 3350 4977 3424 > > 96 3342 5530 3371 > > 104 3328 4923 3313 > > 112 3317 5528 3207 > > 120 3313 4970 3150 > > 128 3492 5535 3568 > > 136 2487 4570 3690 > > 144 2481 5047 3599 > > 152 2473 4565 3566 > > 160 2459 5022 3515 > > 168 2461 4550 3437 > > 176 2454 5020 3325 > > 184 2449 4535 3279 > > 192 2538 5011 3762 > > 200 1962 4537 3702 > > 208 1962 4971 3622 > > 216 1954 4487 3518 > > 224 1949 4936 3445 > > 232 1948 4497 3422 > > 240 1941 4947 3317 > > 248 1940 4481 3279 > > 256 3798 4964 3723 > > 264 2638 3577 3639 > > 272 2637 3567 3597 > > 280 2628 3563 3565 > > 288 2630 3795 3484 > > 296 2621 3580 3422 > > 304 2612 3569 3352 > > 312 2602 3599 3308 > > 320 2694 3821 3694 > > 328 2060 3538 3681 > > 336 2054 3565 3599 > > 344 2054 3553 3523 > > 352 2049 3809 3419 > > 360 2045 3575 3403 > > 368 2035 3560 3334 > > 376 2036 3555 3257 > > 384 2092 3785 3715 > >
Re: [PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
On Sun, Nov 11, 2018 at 10:36:24AM +0100, Martin Willi wrote: > This patchset improves performance of the ChaCha20 SIMD implementations > for x86_64. For some specific encryption lengths, performance is more > than doubled. Two mechanisms are used to achieve this: > > * Instead of calculating the minimal number of required blocks for a > given encryption length, functions producing more blocks are used > more aggressively. Calculating a 4-block function can be faster than > calculating a 2-block and a 1-block function, even if only three > blocks are actually required. > > * In addition to the 8-block AVX2 function, a 4-block and a 2-block > function are introduced. > > Patches 1-3 add support for partial lengths to the existing 1-, 4- and > 8-block functions. Patch 4 makes use of that by engaging the next higher > level block functions more aggressively. Patch 5 and 6 add the new AVX2 > functions for 2 and 4 blocks. Patches are based on cryptodev and would > need adjustments to apply on top of the Adiantum patchset. > > Note that the more aggressive use of larger block functions calculate > blocks that may get discarded. This may have a negative impact on energy > usage or the processors thermal budget. However, with the new block > functions we can avoid this over-calculation for many lengths, so the > performance win can be considered more important. > > Below are performance numbers measured with tcrypt using additional > encryption lengths; numbers in kOps/s, on my i7-5557U. old is the > existing, new the implementation with this patchset. As comparison > the numbers for zinc in v6: > > len old new zinc >8 5908 5818 5818 > 16 5917 5828 5726 > 24 5916 5869 5757 > 32 5920 5789 5813 > 40 5868 5799 5710 > 48 5877 5761 5761 > 56 5869 5797 5742 > 64 5897 5862 5685 > 72 3381 4979 3520 > 80 3364 5541 3475 > 88 3350 4977 3424 > 96 3342 5530 3371 > 104 3328 4923 3313 > 112 3317 5528 3207 > 120 3313 4970 3150 > 128 3492 5535 3568 > 136 2487 4570 3690 > 144 2481 5047 3599 > 152 2473 4565 3566 > 160 2459 5022 3515 > 168 2461 4550 3437 > 176 2454 5020 3325 > 184 2449 4535 3279 > 192 2538 5011 3762 > 200 1962 4537 3702 > 208 1962 4971 3622 > 216 1954 4487 3518 > 224 1949 4936 3445 > 232 1948 4497 3422 > 240 1941 4947 3317 > 248 1940 4481 3279 > 256 3798 4964 3723 > 264 2638 3577 3639 > 272 2637 3567 3597 > 280 2628 3563 3565 > 288 2630 3795 3484 > 296 2621 3580 3422 > 304 2612 3569 3352 > 312 2602 3599 3308 > 320 2694 3821 3694 > 328 2060 3538 3681 > 336 2054 3565 3599 > 344 2054 3553 3523 > 352 2049 3809 3419 > 360 2045 3575 3403 > 368 2035 3560 3334 > 376 2036 3555 3257 > 384 2092 3785 3715 > 392 1691 3505 3612 > 400 1684 3527 3553 > 408 1686 3527 3496 > 416 1684 3804 3430 > 424 1681 3555 3402 > 432 1675 3559 3311 > 440 1672 3558 3275 > 448 1710 3780 3689 > 456 1431 3541 3618 > 464 1428 3538 3576 > 472 1430 3527 3509 > 480 1426 3788 3405 > 488 1423 3502 3397 > 496 1423 3519 3298 > 504 1418 3519 3277 > 512 3694 3736 3735 > 520 2601 2571 2209 > 528 2601 2677 2148 > 536 2587 2534 2164 > 544 2578 2659 2138 > 552 2570 2552 2126 > 560 2566 2661 2035 > 568 2567 2542 2041 > 576 2639 2674 2199 > 584 2031 2531 2183 > 592 2027 2660 2145 > 600 2016 2513 2155 > 608 2009 2638 2133 > 616 2006 2522 2115 > 624 2000 2649 2064 > 632 1996 2518 2045 > 640 2053 2651 2188 > 648 1666 2402 2182 > 656 1663 2517 2158 > 664 1659 2397 2147 > 672 1657 2510 2139 > 680 1656 2394 2114 > 688 1653 2497 2077 > 696 1646 2393 2043 > 704 1678 2510 2208 > 712 1414 2391 2189 > 720 1412 2506 2169 > 728 1411 2384 2145 > 736 1408 2494 2142 > 744 1408 2379 2081 > 752 1405 2485 2064 > 760 1403 2376 2043 > 768 2189 2498 2211 > 776 1756 2137 2192 > 784 1746 2145 2146 > 792 1744 2141 2141 > 800 1743 2094 > 808 1742 2140 2100 > 816 1735 2134 2061 > 824 1731 2135 2045 > 832 1778 2223 > 840 1480 2132 2184 > 848 1480 2134 2173 > 856 1476 2124 2145 > 864 1474 2210 2126 > 872 1472 2127 2105 > 880 1463 2123 2056 > 888 1468 2123 2043 > 896 1494 2208 2219 > 904 1278 2120 2192 > 912 1277 2121 2170 > 920 1273 2118 2149 > 928 1272 2207 2125 > 936 1267 2125 2098 > 944 1265 2127 2060 > 952 1267 2126 2049 > 960 1289 2213 2204 > 968 1125 2123 2187 > 976 1122 2127 2166 > 984 1120 2123 2136 > 992 1118 2207 2119 > 1000 1118 2120 2101 > 1008 1117 2122 2042 > 1016 1115 2121 2048 > 1024 2174 2191 2195 > 1032 1748 1724 1565 > 1040 1745 1782 1544 > 1048 1736 1737 1554 > 1056 1738 1802 1541 > 1064 1735 1728 1523 > 1072 1730 1780 1507 > 1080 1729 1724 1497 > 1088 1757 1783 1592 > 1096 1475 1723 1575 > 1104 1474 1778 1563 > 1112 1472 1708 1544 > 1120 1468 1774 1521 > 1128 1466 1718 1521 > 1136 1462 1780 1501 > 1144 1460 1719 1491 > 1152 1481 1782 1575 > 1160 1271 1647 1558 > 1168 1271 1706 1554 > 1176 1268 1645 1545 > 1184 1265 1711 1538 > 1192 1265 1648 1530 > 1200 1264 1705 1493 > 1208 1262 1647 1498 > 1216 1277 1695 1581
[PATCH 0/6] crypto: x86/chacha20 - SIMD performance improvements
This patchset improves performance of the ChaCha20 SIMD implementations for x86_64. For some specific encryption lengths, performance is more than doubled. Two mechanisms are used to achieve this: * Instead of calculating the minimal number of required blocks for a given encryption length, functions producing more blocks are used more aggressively. Calculating a 4-block function can be faster than calculating a 2-block and a 1-block function, even if only three blocks are actually required. * In addition to the 8-block AVX2 function, a 4-block and a 2-block function are introduced. Patches 1-3 add support for partial lengths to the existing 1-, 4- and 8-block functions. Patch 4 makes use of that by engaging the next higher level block functions more aggressively. Patch 5 and 6 add the new AVX2 functions for 2 and 4 blocks. Patches are based on cryptodev and would need adjustments to apply on top of the Adiantum patchset. Note that the more aggressive use of larger block functions calculate blocks that may get discarded. This may have a negative impact on energy usage or the processors thermal budget. However, with the new block functions we can avoid this over-calculation for many lengths, so the performance win can be considered more important. Below are performance numbers measured with tcrypt using additional encryption lengths; numbers in kOps/s, on my i7-5557U. old is the existing, new the implementation with this patchset. As comparison the numbers for zinc in v6: len old new zinc 8 5908 5818 5818 16 5917 5828 5726 24 5916 5869 5757 32 5920 5789 5813 40 5868 5799 5710 48 5877 5761 5761 56 5869 5797 5742 64 5897 5862 5685 72 3381 4979 3520 80 3364 5541 3475 88 3350 4977 3424 96 3342 5530 3371 104 3328 4923 3313 112 3317 5528 3207 120 3313 4970 3150 128 3492 5535 3568 136 2487 4570 3690 144 2481 5047 3599 152 2473 4565 3566 160 2459 5022 3515 168 2461 4550 3437 176 2454 5020 3325 184 2449 4535 3279 192 2538 5011 3762 200 1962 4537 3702 208 1962 4971 3622 216 1954 4487 3518 224 1949 4936 3445 232 1948 4497 3422 240 1941 4947 3317 248 1940 4481 3279 256 3798 4964 3723 264 2638 3577 3639 272 2637 3567 3597 280 2628 3563 3565 288 2630 3795 3484 296 2621 3580 3422 304 2612 3569 3352 312 2602 3599 3308 320 2694 3821 3694 328 2060 3538 3681 336 2054 3565 3599 344 2054 3553 3523 352 2049 3809 3419 360 2045 3575 3403 368 2035 3560 3334 376 2036 3555 3257 384 2092 3785 3715 392 1691 3505 3612 400 1684 3527 3553 408 1686 3527 3496 416 1684 3804 3430 424 1681 3555 3402 432 1675 3559 3311 440 1672 3558 3275 448 1710 3780 3689 456 1431 3541 3618 464 1428 3538 3576 472 1430 3527 3509 480 1426 3788 3405 488 1423 3502 3397 496 1423 3519 3298 504 1418 3519 3277 512 3694 3736 3735 520 2601 2571 2209 528 2601 2677 2148 536 2587 2534 2164 544 2578 2659 2138 552 2570 2552 2126 560 2566 2661 2035 568 2567 2542 2041 576 2639 2674 2199 584 2031 2531 2183 592 2027 2660 2145 600 2016 2513 2155 608 2009 2638 2133 616 2006 2522 2115 624 2000 2649 2064 632 1996 2518 2045 640 2053 2651 2188 648 1666 2402 2182 656 1663 2517 2158 664 1659 2397 2147 672 1657 2510 2139 680 1656 2394 2114 688 1653 2497 2077 696 1646 2393 2043 704 1678 2510 2208 712 1414 2391 2189 720 1412 2506 2169 728 1411 2384 2145 736 1408 2494 2142 744 1408 2379 2081 752 1405 2485 2064 760 1403 2376 2043 768 2189 2498 2211 776 1756 2137 2192 784 1746 2145 2146 792 1744 2141 2141 800 1743 2094 808 1742 2140 2100 816 1735 2134 2061 824 1731 2135 2045 832 1778 2223 840 1480 2132 2184 848 1480 2134 2173 856 1476 2124 2145 864 1474 2210 2126 872 1472 2127 2105 880 1463 2123 2056 888 1468 2123 2043 896 1494 2208 2219 904 1278 2120 2192 912 1277 2121 2170 920 1273 2118 2149 928 1272 2207 2125 936 1267 2125 2098 944 1265 2127 2060 952 1267 2126 2049 960 1289 2213 2204 968 1125 2123 2187 976 1122 2127 2166 984 1120 2123 2136 992 1118 2207 2119 1000 1118 2120 2101 1008 1117 2122 2042 1016 1115 2121 2048 1024 2174 2191 2195 1032 1748 1724 1565 1040 1745 1782 1544 1048 1736 1737 1554 1056 1738 1802 1541 1064 1735 1728 1523 1072 1730 1780 1507 1080 1729 1724 1497 1088 1757 1783 1592 1096 1475 1723 1575 1104 1474 1778 1563 1112 1472 1708 1544 1120 1468 1774 1521 1128 1466 1718 1521 1136 1462 1780 1501 1144 1460 1719 1491 1152 1481 1782 1575 1160 1271 1647 1558 1168 1271 1706 1554 1176 1268 1645 1545 1184 1265 1711 1538 1192 1265 1648 1530 1200 1264 1705 1493 1208 1262 1647 1498 1216 1277 1695 1581 1224 1120 1642 1563 1232 1115 1702 1549 1240 1121 1646 1538 1248 1119 1703 1527 1256 1115 1640 1520 1264 1114 1693 1505 1272 1112 1642 1492 1280 1552 1699 1574 1288 1314 1525 1573 1296 1315 1522 1551 1304 1312 1521 1548 1312 1311 1564 1535 1320 1309 1518 1524 1328 1302 1527 1508 1336 1303 1521 1500 1344 1333 1561 1579 1352 1157 1524 1573 1360 1152 1520 1546 1368 1154 1522 1545 1376 1153 1562 1536 1384 1151 1525 1526 1392
