//go:build !noasm && gc && amd64 package ubc // avx512Terms holds the bit tests of CalculateDvMask in the layout the // AVX-512 kernel consumes, 16 terms per group. Each term says that a set of // DVs survives only if bit ka of W[a] XOR bit kb of W[b] equals e, and names // the mask that clears those DVs when it does not. // // The six vectors of a group are, in order: a-34, b-34, ka, kb, e, and the // clear mask. Word indices are biased by 34 because the kernel keeps W[34..65] // in two registers. A group is 384 bytes, and the kernel runs one per pass. // // The terms are those of calculateDvMaskGeneric, which is a port of // ubc_check.c and has not changed since the collision detection was published // in 2017. TestAVX512MatchesGeneric holds the two implementations to the same // results over inputs that reach every DV, so a mistake here is a test // failure rather than a missed collision. The comment above each group names // its terms as a.ka^b.kb=e. var avx512Terms = [960]uint32{ // terms 0-15 // 44.29^45.29=0 49.29^50.29=0 48.29^49.29=0 47.4^50.29=0 47.29^48.29=0 46.4^49.29=0 46.29^47.29=0 45.4^48.29=0 // 45.29^46.29=0 44.4^47.29=0 43.4^46.29=0 43.29^44.29=0 42.4^45.29=0 41.4^44.29=0 40.29^41.29=0 54.29^55.29=0 0x0000000a, 0x0000000f, 0x0000000e, 0x0000000d, 0x0000000d, 0x0000000c, 0x0000000c, 0x0000000b, 0x0000000b, 0x0000000a, 0x00000009, 0x00000009, 0x00000008, 0x00000007, 0x00000006, 0x00000014, 0x0000000b, 0x00000010, 0x0000000f, 0x00000010, 0x0000000e, 0x0000000f, 0x0000000d, 0x0000000e, 0x0000000c, 0x0000000d, 0x0000000c, 0x0000000a, 0x0000000b, 0x0000000a, 0x00000007, 0x00000015, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000004, 0x0000001d, 0x00000004, 0x0000001d, 0x00000004, 0x0000001d, 0x00000004, 0x00000004, 0x0000001d, 0x00000004, 0x00000004, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0xfd7c5f7f, 0x3d7efff7, 0x9f5f7ffb, 0x7dfedddf, 0xcfcfdffd, 0xbf7f7777, 0xe7e7f7fe, 0xdfdfdddb, 0xf5f57dff, 0xefeff775, 0xf7f7fdda, 0xff5ed7df, 0xfdfd7f75, 0xff7edfda, 0x7ff5ff5d, 0x3f77dfff, // terms 16-31 // 53.29^54.29=0 52.29^53.29=0 50.4^53.29=0 50.29^51.29=0 49.4^52.29=0 48.4^51.29=0 42.29^43.29=0 41.29^42.29=0 // 40.4^43.29=0 39.4^42.29=0 38.4^41.29=0 37.4^40.29=0 55.29^56.29=0 52.4^55.29=0 51.4^54.29=0 51.29^52.29=0 0x00000013, 0x00000012, 0x00000010, 0x00000010, 0x0000000f, 0x0000000e, 0x00000008, 0x00000007, 0x00000006, 0x00000005, 0x00000004, 0x00000003, 0x00000015, 0x00000012, 0x00000011, 0x00000011, 0x00000014, 0x00000013, 0x00000013, 0x00000011, 0x00000012, 0x00000011, 0x00000009, 0x00000008, 0x00000009, 0x00000008, 0x00000007, 0x00000006, 0x00000016, 0x00000015, 0x00000014, 0x00000012, 0x0000001d, 0x0000001d, 0x00000004, 0x0000001d, 0x00000004, 0x00000004, 0x0000001d, 0x0000001d, 0x00000004, 0x00000004, 0x00000004, 0x00000004, 0x0000001d, 0x00000004, 0x00000004, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x9fddf7ff, 0xcfeefdff, 0xdfed77ff, 0x75fdffdf, 0xeff6ddff, 0xf7fd777f, 0xffcff5f7, 0xffe7fd7b, 0x7fdff7f5, 0xbfeffdfa, 0x5ff7ff7d, 0xaffdffde, 0x7def7fff, 0x7f6f7fff, 0xbfd7dfff, 0xe7f7ff7f, // terms 32-47 // 36.4^40.29=0 53.29^56.29=1 51.29^54.29=1 50.29^52.29=1 49.29^51.29=1 48.29^50.29=1 47.29^49.29=1 46.29^48.29=1 // 45.6^47.6=0 45.29^47.29=1 44.6^46.6=0 44.29^46.29=1 41.1^42.6=1 40.1^41.6=1 40.4^42.4=1 39.1^40.6=1 0x00000002, 0x00000013, 0x00000011, 0x00000010, 0x0000000f, 0x0000000e, 0x0000000d, 0x0000000c, 0x0000000b, 0x0000000b, 0x0000000a, 0x0000000a, 0x00000007, 0x00000006, 0x00000006, 0x00000005, 0x00000006, 0x00000016, 0x00000014, 0x00000012, 0x00000011, 0x00000010, 0x0000000f, 0x0000000e, 0x0000000d, 0x0000000d, 0x0000000c, 0x0000000c, 0x00000008, 0x00000007, 0x00000008, 0x00000006, 0x00000004, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000006, 0x0000001d, 0x00000006, 0x0000001d, 0x00000001, 0x00000001, 0x00000004, 0x00000001, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000006, 0x0000001d, 0x00000006, 0x0000001d, 0x00000006, 0x00000006, 0x00000004, 0x00000006, 0x00000000, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0xffeefdf7, 0xffcf7fff, 0xfff5f7ff, 0xfffeddff, 0xffff777f, 0xffffdddf, 0xfffff777, 0xfffffddb, 0xffffbbbf, 0xffffff75, 0xffffeeef, 0xffffffda, 0xfbfbfeff, 0xfeffbfbf, 0x7ffffff5, 0xffbfefef, // terms 48-63 // 39.4^41.4=1 38.4^40.4=1 37.4^39.4=1 36.1^37.6=1 35.4^39.29=0 63.0^64.5=1 63.1^64.6=1 62.0^63.5=1 // 61.0^62.5=1 61.2^62.7=1 60.0^61.5=1 58.29^59.29=0 57.29^58.29=0 56.4^59.29=0 56.29^59.29=1 56.29^57.29=0 0x00000005, 0x00000004, 0x00000003, 0x00000002, 0x00000001, 0x0000001d, 0x0000001d, 0x0000001c, 0x0000001b, 0x0000001b, 0x0000001a, 0x00000018, 0x00000017, 0x00000016, 0x00000016, 0x00000016, 0x00000007, 0x00000006, 0x00000005, 0x00000003, 0x00000005, 0x0000001e, 0x0000001e, 0x0000001d, 0x0000001c, 0x0000001c, 0x0000001b, 0x00000019, 0x00000018, 0x00000019, 0x00000019, 0x00000017, 0x00000004, 0x00000004, 0x00000004, 0x00000001, 0x00000004, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000002, 0x00000000, 0x0000001d, 0x0000001d, 0x00000004, 0x0000001d, 0x0000001d, 0x00000004, 0x00000004, 0x00000004, 0x00000006, 0x0000001d, 0x00000005, 0x00000006, 0x00000005, 0x00000005, 0x00000007, 0x00000005, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000000, 0x00000000, 0x00000000, 0x00000001, 0x00000000, 0xbffffffa, 0x5ffffffd, 0xaffffffe, 0xfffbefbf, 0xfff7ff7b, 0xffefff7f, 0xfffefffb, 0xfff7ffdf, 0xfffdfff7, 0xfffbffef, 0xfffefffb, 0xddffffff, 0xef7fffff, 0xd7ffffff, 0xf5ffffff, 0xf7dfffff, // terms 64-79 // 55.4^58.29=0 54.4^57.29=0 53.4^56.29=0 50.6^51.1=0 48.6^50.6=0 48.29^55.29=1 47.6^49.6=0 47.6^48.1=0 // 46.6^48.6=0 46.6^47.1=0 44.1^45.6=1 43.6^45.6=0 42.6^44.6=0 42.6^43.1=0 41.6^42.1=0 40.6^41.1=0 0x00000015, 0x00000014, 0x00000013, 0x00000010, 0x0000000e, 0x0000000e, 0x0000000d, 0x0000000d, 0x0000000c, 0x0000000c, 0x0000000a, 0x00000009, 0x00000008, 0x00000008, 0x00000007, 0x00000006, 0x00000018, 0x00000017, 0x00000016, 0x00000011, 0x00000010, 0x00000015, 0x0000000f, 0x0000000e, 0x0000000e, 0x0000000d, 0x0000000b, 0x0000000b, 0x0000000a, 0x00000009, 0x00000008, 0x00000007, 0x00000004, 0x00000004, 0x00000004, 0x00000006, 0x00000006, 0x0000001d, 0x00000006, 0x00000006, 0x00000006, 0x00000006, 0x00000001, 0x00000006, 0x00000006, 0x00000006, 0x00000006, 0x00000006, 0x0000001d, 0x0000001d, 0x0000001d, 0x00000001, 0x00000006, 0x0000001d, 0x00000006, 0x00000001, 0x00000006, 0x00000001, 0x00000006, 0x00000006, 0x00000006, 0x00000001, 0x00000001, 0x00000001, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0xedffffff, 0xf77fffff, 0xfddfffff, 0xfffbefff, 0xfffbefff, 0xffff5fff, 0xffffbbff, 0xfbffffbf, 0xffffeeff, 0xfeffffef, 0xffbfbfff, 0xfffffbbf, 0xfffffeef, 0xfbfbffff, 0xfeffbfff, 0xffbfefff, // terms 80-95 // 39.4^43.29=0 38.4^42.29=0 37.1^38.6=1 37.4^41.29=0 36.4^38.4=1 35.1^36.6=1 35.3^39.28=0 61.1^62.6=1 // 59.5^63.30=0 58.0^63.30=1 62.1^63.6=1 60.5^64.30=0 59.0^64.30=1 40.6^42.6=0 62.2^63.7=1 41.6^43.6=0 0x00000005, 0x00000004, 0x00000003, 0x00000003, 0x00000002, 0x00000001, 0x00000001, 0x0000001b, 0x00000019, 0x00000018, 0x0000001c, 0x0000001a, 0x00000019, 0x00000006, 0x0000001c, 0x00000007, 0x00000009, 0x00000008, 0x00000004, 0x00000007, 0x00000004, 0x00000002, 0x00000005, 0x0000001c, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001e, 0x0000001e, 0x00000008, 0x0000001d, 0x00000009, 0x00000004, 0x00000004, 0x00000001, 0x00000004, 0x00000004, 0x00000001, 0x00000003, 0x00000001, 0x00000005, 0x00000000, 0x00000001, 0x00000005, 0x00000000, 0x00000006, 0x00000002, 0x00000006, 0x0000001d, 0x0000001d, 0x00000006, 0x0000001d, 0x00000004, 0x00000006, 0x0000001c, 0x00000006, 0x0000001e, 0x0000001e, 0x00000006, 0x0000001e, 0x0000001e, 0x00000006, 0x00000007, 0x00000006, 0x00000000, 0x00000000, 0x00000001, 0x00000000, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0xfdff7fff, 0xff7fdfff, 0xffffbeff, 0xffdff7ff, 0xd7ffffff, 0xfffffbef, 0xfff7dfff, 0xfffffffe, 0xfffffffe, 0xfffffffe, 0xfffffffd, 0xfffffffd, 0xfffffffd, 0xffffffef, 0xffffffbf, 0xffffffbf, // terms 96-111 // 63.2^64.7=1 48.6^49.1=0 49.6^50.1=0 42.1^50.1=1 39.6^40.1=0 38.1^40.1=1 36.4^37.4=1 43.1^51.1=1 // 37.4^38.4=1 51.6^52.1=0 49.6^51.6=0 37.1^37.6=0 35.5^39.30=0 38.4^39.4=1 47.1^51.1=1 36.3^40.28=0 0x0000001d, 0x0000000e, 0x0000000f, 0x00000008, 0x00000005, 0x00000004, 0x00000002, 0x00000009, 0x00000003, 0x00000011, 0x0000000f, 0x00000003, 0x00000001, 0x00000004, 0x0000000d, 0x00000002, 0x0000001e, 0x0000000f, 0x00000010, 0x00000010, 0x00000006, 0x00000006, 0x00000003, 0x00000011, 0x00000004, 0x00000012, 0x00000011, 0x00000003, 0x00000005, 0x00000005, 0x00000011, 0x00000006, 0x00000002, 0x00000006, 0x00000006, 0x00000001, 0x00000006, 0x00000001, 0x00000004, 0x00000001, 0x00000004, 0x00000006, 0x00000006, 0x00000001, 0x00000005, 0x00000004, 0x00000001, 0x00000003, 0x00000007, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000004, 0x00000001, 0x00000004, 0x00000001, 0x00000006, 0x00000006, 0x0000001e, 0x00000004, 0x00000001, 0x0000001c, 0x00000001, 0x00000000, 0x00000000, 0x00000001, 0x00000000, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000001, 0x00000001, 0x00000000, 0xfffffeff, 0xfffffeff, 0xfffffbff, 0xfffffbff, 0xfffffbff, 0xfffffbff, 0xfffff7ff, 0xffffefff, 0xffffdfff, 0xffffbfff, 0xffffbfff, 0xffffbfff, 0xffffbfff, 0xffff7fff, 0xfffbffff, 0xffefffff, // terms 112-127 // 35.30^40.28=1 37.3^41.28=0 36.30^41.28=1 53.6^54.1=0 51.6^53.6=0 50.1^54.1=1 45.6^46.1=0 37.5^41.30=0 // 36.0^41.30=1 55.29^58.29=1 38.3^42.28=0 37.30^42.28=1 54.6^55.1=0 52.6^54.6=0 51.1^55.1=1 45.1^47.1=1 0x00000001, 0x00000003, 0x00000002, 0x00000013, 0x00000011, 0x00000010, 0x0000000b, 0x00000003, 0x00000002, 0x00000015, 0x00000004, 0x00000003, 0x00000014, 0x00000012, 0x00000011, 0x0000000b, 0x00000006, 0x00000007, 0x00000007, 0x00000014, 0x00000013, 0x00000014, 0x0000000c, 0x00000007, 0x00000007, 0x00000018, 0x00000008, 0x00000008, 0x00000015, 0x00000014, 0x00000015, 0x0000000d, 0x0000001e, 0x00000003, 0x0000001e, 0x00000006, 0x00000006, 0x00000001, 0x00000006, 0x00000005, 0x00000000, 0x0000001d, 0x00000003, 0x0000001e, 0x00000006, 0x00000006, 0x00000001, 0x00000001, 0x0000001c, 0x0000001c, 0x0000001c, 0x00000001, 0x00000006, 0x00000001, 0x00000001, 0x0000001e, 0x0000001e, 0x0000001d, 0x0000001c, 0x0000001c, 0x00000001, 0x00000006, 0x00000001, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000001, 0x00000001, 0xffefffff, 0xffdfffff, 0xffdfffff, 0xffbfffff, 0xffbfffff, 0xffbfffff, 0xffbfffff, 0xffbfffff, 0xffbfffff, 0xff7fffff, 0xff7fffff, 0xff7fffff, 0xfeffffff, 0xfeffffff, 0xfeffffff, 0xfeffffff, // terms 128-143 // 38.5^42.30=0 37.0^42.30=1 39.3^43.28=0 38.30^43.28=1 55.6^56.1=0 53.6^55.6=0 52.1^56.1=1 46.1^48.1=1 // 39.5^43.30=0 38.0^43.30=1 59.29^60.29=0 40.3^44.28=0 40.4^44.29=0 39.30^44.28=1 58.29^61.29=1 57.4^61.29=0 0x00000004, 0x00000003, 0x00000005, 0x00000004, 0x00000015, 0x00000013, 0x00000012, 0x0000000c, 0x00000005, 0x00000004, 0x00000019, 0x00000006, 0x00000006, 0x00000005, 0x00000018, 0x00000017, 0x00000008, 0x00000008, 0x00000009, 0x00000009, 0x00000016, 0x00000015, 0x00000016, 0x0000000e, 0x00000009, 0x00000009, 0x0000001a, 0x0000000a, 0x0000000a, 0x0000000a, 0x0000001b, 0x0000001b, 0x00000005, 0x00000000, 0x00000003, 0x0000001e, 0x00000006, 0x00000006, 0x00000001, 0x00000001, 0x00000005, 0x00000000, 0x0000001d, 0x00000003, 0x00000004, 0x0000001e, 0x0000001d, 0x00000004, 0x0000001e, 0x0000001e, 0x0000001c, 0x0000001c, 0x00000001, 0x00000006, 0x00000001, 0x00000001, 0x0000001e, 0x0000001e, 0x0000001d, 0x0000001c, 0x0000001d, 0x0000001c, 0x0000001d, 0x0000001d, 0x00000000, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000001, 0x00000001, 0x00000000, 0x00000001, 0x00000000, 0x00000000, 0x00000000, 0x00000001, 0x00000001, 0x00000000, 0xfeffffff, 0xfeffffff, 0xfdffffff, 0xfdffffff, 0xfbffffff, 0xfbffffff, 0xfbffffff, 0xfbffffff, 0xfbffffff, 0xfbffffff, 0xf7ffffff, 0xf7ffffff, 0xf7ffffff, 0xf7ffffff, 0xefffffff, 0xefffffff, // terms 144-159 // 41.3^45.28=0 41.4^45.29=0 58.4^62.29=0 42.3^46.28=0 42.4^46.29=0 59.4^63.29=0 57.4^59.29=0 43.3^47.28=0 // 43.4^47.29=0 60.4^64.29=0 44.3^48.28=0 44.4^48.29=0 35.0^35.0=0 35.0^35.0=0 35.0^35.0=0 35.0^35.0=0 0x00000007, 0x00000007, 0x00000018, 0x00000008, 0x00000008, 0x00000019, 0x00000017, 0x00000009, 0x00000009, 0x0000001a, 0x0000000a, 0x0000000a, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x0000000b, 0x0000000b, 0x0000001c, 0x0000000c, 0x0000000c, 0x0000001d, 0x00000019, 0x0000000d, 0x0000000d, 0x0000001e, 0x0000000e, 0x0000000e, 0x00000001, 0x00000001, 0x00000001, 0x00000001, 0x00000003, 0x00000004, 0x00000004, 0x00000003, 0x00000004, 0x00000004, 0x00000004, 0x00000003, 0x00000004, 0x00000004, 0x00000003, 0x00000004, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x0000001c, 0x0000001d, 0x0000001d, 0x0000001c, 0x0000001d, 0x0000001d, 0x0000001d, 0x0000001c, 0x0000001d, 0x0000001d, 0x0000001c, 0x0000001d, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0x00000000, 0xefffffff, 0xefffffff, 0xdfffffff, 0xdfffffff, 0xdfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0xffffffff, 0xffffffff, 0xffffffff, 0xffffffff, }