Hi Jason,
> Now that ChaCha20 is in Zinc, we can have the crypto API code simply
> call into it.
> delete mode 100644 arch/x86/crypto/chacha20-avx2-x86_64.S
> delete mode 100644 arch/x86/crypto/chacha20-ssse3-x86_64.S
I did some more testing with that new Zinc ChaCha20 code on x64, and
I'm still not convinced that it is an improvement compared to the
existing implementation.
>From a performance perspective, Zinc is in faster when working on sizes
that are not a multiple of chacha block sizes. This is due to the more
aggressive use of SSE/AVX code paths compared to the conservative use
in the existing implementation; instead of calculating two separate
blocks that are actually required, one can calculate four of them and
just discards two. This certainly improves benchmark results, but also
has some side effects regarding energy usage, thermal budget or even
shared hyper-threading resources.
One can certainly argue that the more aggressive approach is
preferable. However, I did some fairly trivial (non-optimized) changes
to the existing implementations to use a similar aggressive approach.
Numbers for SSE are slightly in favor of the existing implementation,
while the AVX path is almost on par, see below (produces some
interesting graphs, btw.).
When looking at your code, the assembly generated from Perl is
certainly harder to work with. The plain C version does make some heavy
use of macros and other tricks, but with a very questionable effect at
least on my system.
That being said, I think that whole mystic Zinc thing does not really
help in having a common base to work with or handling questions like
these above. In the end, these are just some crypto function that it
provides, and this IMHO can very well live under where it belongs to.
Best regards
Martin
---
ChaCha20 benchmark using tcrypt, numbers in kOps/s, current
implementation with a more aggressive SSE/AVX use vs. zinc:
size crnt zinc
8 5750 5818
16 5843 5726
24 5746 5757
32 5820 5813
40 5761 5710
48 5735 5761
56 5723 5742
64 5871 5685
72 3714 3520
80 3587 3475
88 3686 3424
96 3580 3371
104 3712 3313
112 3582 3207
120 3679 3150
128 3567 3568
136 3674 3690
144 3525 3599
152 3684 3566
160 3593 3515
168 3682 3437
176 3564 3325
184 3671 3279
192 3573 3762
200 3667 3702
208 3576 3622
216 3662 3518
224 3566 3445
232 3654 3422
240 3565 3317
248 3640 3279
256 3720 3723
264 3615 3639
272 3594 3597
280 3587 3565
288 3502 3484
296 3605 3422
304 3620 3352
312 3592 3308
320 3488 3694
328 3580 3681
336 3585 3599
344 3587 3523
352 3486 3419
360 3579 3403
368 3601 3334
376 3581 3257
384 3498 3715
392 3601 3612
400 3600 3553
408 3596 3496
416 3495 3430
424 3591 3402
432 3568 3311
440 3576 3275
448 3501 3689
456 3563 3618
464 3592 3576
472 3581 3509
480 3480 3405
488 3556 3397
496 3563 3298
504 3567 3277
512 3656 3735
520 2575 2209
528 2524 2148
536 2571 2164
544 2519 2138
552 2570 2126
560 2510 2035
568 2526 2041
576 2633 2199
584 2151 2183
592 2113 2145
600 2159 2155
608 2108 2133
616 2157 2115
624 2104 2064
632 2159 2045
640 2104 2188
648 2142 2182
656 2115 2158
664 2151 2147
672 2113 2139
680 2146 2114
688 2097 2077
696 2137 2043
704 2101 2208
712 2137 2189
720 2117 2169
728 2132 2145
736 2107 2142
744 2136 2081
752 2105 2064
760 2136 2043
768 2166 2211
776 2122 2192
784 2129 2146
792 2126 2141
800 2094 2094
808 2126 2100
816 2133 2061
824 2134 2045
832 2103 2223
840 2143 2184
848 2130 2173
856 2135 2145
864 2084 2126
872 2134 2105
880 2128 2056
888 2131 2043
896 2093 2219
904 2127 2192
912 2130 2170
920 2127 2149
928 2082 2125
936 2113 2098
944 2126 2060
952 2120 2049
960 2085 2204
968 2088 2187
976 1927 2166
984 1943 2136
992 1911 2119
1000 1959 2101
1008 2116 2042
1016 2124 2048
1024 2152 2195
1032 1729 1565
1040 1708 1544
1048 1726 1554
1056 1702 1541
1064 1724 1523
1072 1699 1507
1080 1719 1497
1088 1767 1592
1096 1536 1575
1104 1506 1563
1112 1529 1544
1120 1518 1521
1128 1526 1521
1136 1518 1501
1144 1535 1491
1152 1507 1575
1160 1525 1558
1168 1500 1554
1176 1524 1545
1184 1516 1538
1192 1532 1530
1200 1511 1493
1208 1512 1498
1216 1505 1581
1224 1518 1563
1232 1513 1549
1240 1533 1538
1248 1504 1527
1256 1532 1520
1264 1510 1505
1272 1525 1492
1280 1539 1574
1288 1518 1573
1296 1522 1551
1304 1520 1548
1312 1508 1535
1320 1524 1524
1328 1522 1508
1336 1515 1500
1344 1496 1579
1352 1517 1573
1360 1522 1546
1368 1515 1545
1376 1494 1536
1384 1516 1526
1392 1522 1504
1400 1520 1480
1408 1501 1589
1416 1511 1558
1424 1516 1546
1432 1516 1537
1440 1502 1523
1448 1516 1512
1456 1510 1491
1464 1509 1481
1472 1496 1577
1480 1514 1559
1488 1512 1548
1496 1513 1534
