sleeping_kangaroo12 0.0.3 → 0.0.4
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/README.md +34 -67
- data/ext/Rakefile +12 -37
- data/ext/binding/sleeping_kangaroo12.c +1 -16
- data/ext/{xkcp → k12}/Makefile +0 -0
- data/ext/k12/Makefile.build +118 -0
- data/ext/k12/README.markdown +86 -0
- data/ext/k12/lib/ARMv8Asha3/KeccakP-1600-ARMv8Asha3.S +623 -0
- data/ext/k12/lib/ARMv8Asha3/KeccakP-1600-SnP.h +65 -0
- data/ext/k12/lib/ARMv8Asha3/KeccakP-1600-opt64.c +227 -0
- data/ext/{xkcp/lib/low/KeccakP-1600/compact → k12/lib/Inplace32BI}/KeccakP-1600-SnP.h +4 -9
- data/ext/{xkcp/lib/low/KeccakP-1600/plain-32bits-inplace → k12/lib/Inplace32BI}/KeccakP-1600-inplace32BI.c +65 -160
- data/ext/k12/lib/KangarooTwelve.c +332 -0
- data/ext/{xkcp/lib/high/KangarooTwelve → k12/lib}/KangarooTwelve.h +53 -16
- data/ext/{xkcp/lib/low/KeccakP-1600/AVX2 → k12/lib/Optimized64}/KeccakP-1600-AVX2.s +122 -558
- data/ext/k12/lib/Optimized64/KeccakP-1600-AVX512-plainC.c +241 -0
- data/ext/k12/lib/Optimized64/KeccakP-1600-AVX512.s +551 -0
- data/ext/k12/lib/Optimized64/KeccakP-1600-SnP.h +74 -0
- data/ext/{xkcp/lib/low/KeccakP-1600/common/KeccakP-1600-64.macros → k12/lib/Optimized64/KeccakP-1600-opt64.c} +447 -169
- data/ext/k12/lib/Optimized64/KeccakP-1600-runtimeDispatch.c +406 -0
- data/ext/k12/lib/Optimized64/KeccakP-1600-timesN-AVX2.c +419 -0
- data/ext/k12/lib/Optimized64/KeccakP-1600-timesN-AVX512.c +458 -0
- data/ext/k12/lib/Optimized64/KeccakP-1600-timesN-SSSE3.c +438 -0
- data/ext/{xkcp/lib/low/KeccakP-1600/plain-64bits → k12/lib/Plain64}/KeccakP-1600-SnP.h +14 -20
- data/ext/{xkcp/lib/low/KeccakP-1600/ref-64bits/KeccakP-1600-reference.h → k12/lib/Plain64/KeccakP-1600-plain64.c} +9 -8
- data/ext/{xkcp/lib/common → k12/lib}/align.h +3 -2
- data/ext/{xkcp/lib/common → k12/lib}/brg_endian.h +0 -0
- data/ext/{xkcp → k12}/support/Build/ExpandProducts.xsl +0 -0
- data/ext/{xkcp → k12}/support/Build/ToGlobalMakefile.xsl +0 -0
- data/ext/{xkcp → k12}/support/Build/ToOneTarget.xsl +0 -0
- data/ext/{xkcp → k12}/support/Build/ToTargetConfigFile.xsl +0 -0
- data/ext/{xkcp → k12}/support/Build/ToTargetMakefile.xsl +10 -16
- data/ext/{xkcp → k12}/support/Build/ToVCXProj.xsl +0 -0
- data/lib/sleeping_kangaroo12/version.rb +1 -1
- metadata +33 -276
- data/ext/config/xkcp.build +0 -17
- data/ext/xkcp/LICENSE +0 -1
- data/ext/xkcp/Makefile.build +0 -200
- data/ext/xkcp/README.markdown +0 -296
- data/ext/xkcp/lib/HighLevel.build +0 -143
- data/ext/xkcp/lib/LowLevel.build +0 -757
- data/ext/xkcp/lib/high/KangarooTwelve/KangarooTwelve.c +0 -301
- data/ext/xkcp/lib/high/Keccak/FIPS202/KeccakHash.c +0 -81
- data/ext/xkcp/lib/high/Keccak/FIPS202/KeccakHash.h +0 -125
- data/ext/xkcp/lib/high/Keccak/FIPS202/SimpleFIPS202.c +0 -48
- data/ext/xkcp/lib/high/Keccak/FIPS202/SimpleFIPS202.h +0 -79
- data/ext/xkcp/lib/high/Keccak/KeccakDuplex.c +0 -81
- data/ext/xkcp/lib/high/Keccak/KeccakDuplex.h +0 -73
- data/ext/xkcp/lib/high/Keccak/KeccakDuplex.inc +0 -195
- data/ext/xkcp/lib/high/Keccak/KeccakSponge.c +0 -111
- data/ext/xkcp/lib/high/Keccak/KeccakSponge.h +0 -76
- data/ext/xkcp/lib/high/Keccak/KeccakSponge.inc +0 -314
- data/ext/xkcp/lib/high/Keccak/PRG/KeccakPRG.c +0 -61
- data/ext/xkcp/lib/high/Keccak/PRG/KeccakPRG.h +0 -67
- data/ext/xkcp/lib/high/Keccak/PRG/KeccakPRG.inc +0 -128
- data/ext/xkcp/lib/high/Keccak/SP800-185/SP800-185.c +0 -93
- data/ext/xkcp/lib/high/Keccak/SP800-185/SP800-185.h +0 -599
- data/ext/xkcp/lib/high/Keccak/SP800-185/SP800-185.inc +0 -573
- data/ext/xkcp/lib/high/Ketje/Ketjev2.c +0 -87
- data/ext/xkcp/lib/high/Ketje/Ketjev2.h +0 -88
- data/ext/xkcp/lib/high/Ketje/Ketjev2.inc +0 -274
- data/ext/xkcp/lib/high/Keyak/Keyakv2.c +0 -132
- data/ext/xkcp/lib/high/Keyak/Keyakv2.h +0 -217
- data/ext/xkcp/lib/high/Keyak/Keyakv2.inc +0 -81
- data/ext/xkcp/lib/high/Keyak/Motorist.inc +0 -953
- data/ext/xkcp/lib/high/Kravatte/Kravatte.c +0 -533
- data/ext/xkcp/lib/high/Kravatte/Kravatte.h +0 -115
- data/ext/xkcp/lib/high/Kravatte/KravatteModes.c +0 -557
- data/ext/xkcp/lib/high/Kravatte/KravatteModes.h +0 -247
- data/ext/xkcp/lib/high/Xoodyak/Cyclist.h +0 -66
- data/ext/xkcp/lib/high/Xoodyak/Cyclist.inc +0 -336
- data/ext/xkcp/lib/high/Xoodyak/Xoodyak-parameters.h +0 -26
- data/ext/xkcp/lib/high/Xoodyak/Xoodyak.c +0 -55
- data/ext/xkcp/lib/high/Xoodyak/Xoodyak.h +0 -35
- data/ext/xkcp/lib/high/Xoofff/Xoofff.c +0 -634
- data/ext/xkcp/lib/high/Xoofff/Xoofff.h +0 -147
- data/ext/xkcp/lib/high/Xoofff/XoofffModes.c +0 -483
- data/ext/xkcp/lib/high/Xoofff/XoofffModes.h +0 -241
- data/ext/xkcp/lib/high/common/Phases.h +0 -25
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-SnP.h +0 -41
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-inplace-32bi-armv6m-le-armcc.s +0 -1666
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-inplace-32bi-armv6m-le-gcc.s +0 -1655
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-inplace-32bi-armv7a-le-armcc.s +0 -1268
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-inplace-32bi-armv7a-le-gcc.s +0 -1264
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-inplace-32bi-armv7m-le-armcc.s +0 -1178
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-inplace-32bi-armv7m-le-gcc.s +0 -1175
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-u1-32bi-armv6m-le-armcc.s +0 -1338
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-u1-32bi-armv6m-le-gcc.s +0 -1336
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-u2-32bi-armv6m-le-armcc.s +0 -1343
- data/ext/xkcp/lib/low/KeccakP-1600/ARM/KeccakP-1600-u2-32bi-armv6m-le-gcc.s +0 -1339
- data/ext/xkcp/lib/low/KeccakP-1600/ARMv7A-NEON/KeccakP-1600-SnP.h +0 -42
- data/ext/xkcp/lib/low/KeccakP-1600/ARMv7A-NEON/KeccakP-1600-armv7a-le-neon-armcc.s +0 -823
- data/ext/xkcp/lib/low/KeccakP-1600/ARMv7A-NEON/KeccakP-1600-armv7a-le-neon-gcc.s +0 -831
- data/ext/xkcp/lib/low/KeccakP-1600/ARMv8A/KeccakP-1600-SnP.h +0 -31
- data/ext/xkcp/lib/low/KeccakP-1600/ARMv8A/KeccakP-1600-armv8a-neon.s +0 -540
- data/ext/xkcp/lib/low/KeccakP-1600/AVR8/KeccakP-1600-SnP.h +0 -42
- data/ext/xkcp/lib/low/KeccakP-1600/AVR8/KeccakP-1600-avr8-compact.s +0 -733
- data/ext/xkcp/lib/low/KeccakP-1600/AVR8/KeccakP-1600-avr8-fast.s +0 -1121
- data/ext/xkcp/lib/low/KeccakP-1600/AVX2/KeccakP-1600-SnP.h +0 -52
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/C/KeccakP-1600-AVX512.c +0 -623
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/C/KeccakP-1600-SnP.h +0 -47
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/C/u12/KeccakP-1600-AVX512-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/C/u6/KeccakP-1600-AVX512-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/C/ua/KeccakP-1600-AVX512-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/KeccakP-1600-AVX512.s +0 -1031
- data/ext/xkcp/lib/low/KeccakP-1600/AVX512/KeccakP-1600-SnP.h +0 -53
- data/ext/xkcp/lib/low/KeccakP-1600/XOP/KeccakP-1600-SnP.h +0 -44
- data/ext/xkcp/lib/low/KeccakP-1600/XOP/KeccakP-1600-XOP.c +0 -476
- data/ext/xkcp/lib/low/KeccakP-1600/XOP/u6/KeccakP-1600-XOP-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/XOP/ua/KeccakP-1600-XOP-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/common/KeccakP-1600-unrolling.macros +0 -305
- data/ext/xkcp/lib/low/KeccakP-1600/compact/KeccakP-1600-compact64.c +0 -420
- data/ext/xkcp/lib/low/KeccakP-1600/plain-32bits-inplace/KeccakP-1600-SnP.h +0 -43
- data/ext/xkcp/lib/low/KeccakP-1600/plain-64bits/KeccakP-1600-opt64.c +0 -565
- data/ext/xkcp/lib/low/KeccakP-1600/plain-64bits/lcu6/KeccakP-1600-opt64-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600/plain-64bits/lcua/KeccakP-1600-opt64-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600/plain-64bits/lcua-shld/KeccakP-1600-opt64-config.h +0 -8
- data/ext/xkcp/lib/low/KeccakP-1600/plain-64bits/u6/KeccakP-1600-opt64-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/plain-64bits/ua/KeccakP-1600-opt64-config.h +0 -6
- data/ext/xkcp/lib/low/KeccakP-1600/ref-32bits/KeccakP-1600-SnP.h +0 -44
- data/ext/xkcp/lib/low/KeccakP-1600/ref-32bits/KeccakP-1600-reference.h +0 -23
- data/ext/xkcp/lib/low/KeccakP-1600/ref-32bits/KeccakP-1600-reference32BI.c +0 -625
- data/ext/xkcp/lib/low/KeccakP-1600/ref-64bits/KeccakP-1600-SnP.h +0 -44
- data/ext/xkcp/lib/low/KeccakP-1600/ref-64bits/KeccakP-1600-reference.c +0 -440
- data/ext/xkcp/lib/low/KeccakP-1600/x86-64/KeccakP-1600-SnP.h +0 -42
- data/ext/xkcp/lib/low/KeccakP-1600/x86-64/KeccakP-1600-x86-64-gas.s +0 -1196
- data/ext/xkcp/lib/low/KeccakP-1600/x86-64/KeccakP-1600-x86-64-gas_Apple.s +0 -1124
- data/ext/xkcp/lib/low/KeccakP-1600/x86-64/KeccakP-1600-x86-64-shld-gas.s +0 -1196
- data/ext/xkcp/lib/low/KeccakP-1600-times2/ARMv7A-NEON/KeccakP-1600-inplace-pl2-armv7a-neon-le-armcc.s +0 -1392
- data/ext/xkcp/lib/low/KeccakP-1600-times2/ARMv7A-NEON/KeccakP-1600-inplace-pl2-armv7a-neon-le-gcc.s +0 -1394
- data/ext/xkcp/lib/low/KeccakP-1600-times2/ARMv7A-NEON/KeccakP-1600-times2-SnP.h +0 -42
- data/ext/xkcp/lib/low/KeccakP-1600-times2/AVX512/AVX512u12/SIMD512-2-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times2/AVX512/AVX512u4/SIMD512-2-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times2/AVX512/AVX512ufull/SIMD512-2-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times2/AVX512/KeccakP-1600-times2-SIMD512.c +0 -850
- data/ext/xkcp/lib/low/KeccakP-1600-times2/AVX512/KeccakP-1600-times2-SnP.h +0 -51
- data/ext/xkcp/lib/low/KeccakP-1600-times2/SIMD128/KeccakP-1600-times2-SIMD128.c +0 -957
- data/ext/xkcp/lib/low/KeccakP-1600-times2/SIMD128/KeccakP-1600-times2-SnP.h +0 -49
- data/ext/xkcp/lib/low/KeccakP-1600-times2/SIMD128/SSSE3-u2/SIMD128-config.h +0 -8
- data/ext/xkcp/lib/low/KeccakP-1600-times2/SIMD128/SSSE3-ua/SIMD128-config.h +0 -8
- data/ext/xkcp/lib/low/KeccakP-1600-times2/SIMD128/XOP-u2/SIMD128-config.h +0 -9
- data/ext/xkcp/lib/low/KeccakP-1600-times2/SIMD128/XOP-ua/SIMD128-config.h +0 -9
- data/ext/xkcp/lib/low/KeccakP-1600-times2/fallback-on1/KeccakP-1600-times2-SnP.h +0 -45
- data/ext/xkcp/lib/low/KeccakP-1600-times2/fallback-on1/KeccakP-1600-times2-on1.c +0 -37
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX2/KeccakP-1600-times4-SIMD256.c +0 -1321
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX2/KeccakP-1600-times4-SnP.h +0 -55
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX2/u12/SIMD256-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX2/u6/SIMD256-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX2/ua/SIMD256-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX512/AVX512u12/SIMD512-4-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX512/AVX512u4/SIMD512-4-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX512/AVX512ufull/SIMD512-4-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX512/KeccakP-1600-times4-SIMD512.c +0 -881
- data/ext/xkcp/lib/low/KeccakP-1600-times4/AVX512/KeccakP-1600-times4-SnP.h +0 -51
- data/ext/xkcp/lib/low/KeccakP-1600-times4/fallback-on1/KeccakP-1600-times4-SnP.h +0 -45
- data/ext/xkcp/lib/low/KeccakP-1600-times4/fallback-on1/KeccakP-1600-times4-on1.c +0 -37
- data/ext/xkcp/lib/low/KeccakP-1600-times4/fallback-on2/KeccakP-1600-times4-SnP.h +0 -45
- data/ext/xkcp/lib/low/KeccakP-1600-times4/fallback-on2/KeccakP-1600-times4-on2.c +0 -38
- data/ext/xkcp/lib/low/KeccakP-1600-times8/AVX512/KeccakP-1600-times8-SIMD512.c +0 -1615
- data/ext/xkcp/lib/low/KeccakP-1600-times8/AVX512/KeccakP-1600-times8-SnP.h +0 -57
- data/ext/xkcp/lib/low/KeccakP-1600-times8/AVX512/u12/SIMD512-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times8/AVX512/u4/SIMD512-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times8/AVX512/ua/SIMD512-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-1600-times8/fallback-on1/KeccakP-1600-times8-SnP.h +0 -45
- data/ext/xkcp/lib/low/KeccakP-1600-times8/fallback-on1/KeccakP-1600-times8-on1.c +0 -37
- data/ext/xkcp/lib/low/KeccakP-1600-times8/fallback-on2/KeccakP-1600-times8-SnP.h +0 -45
- data/ext/xkcp/lib/low/KeccakP-1600-times8/fallback-on2/KeccakP-1600-times8-on2.c +0 -38
- data/ext/xkcp/lib/low/KeccakP-1600-times8/fallback-on4/KeccakP-1600-times8-SnP.h +0 -45
- data/ext/xkcp/lib/low/KeccakP-1600-times8/fallback-on4/KeccakP-1600-times8-on4.c +0 -38
- data/ext/xkcp/lib/low/KeccakP-200/ARM/KeccakP-200-SnP.h +0 -41
- data/ext/xkcp/lib/low/KeccakP-200/ARM/KeccakP-200-armv6m-le-armcc.s +0 -442
- data/ext/xkcp/lib/low/KeccakP-200/ARM/KeccakP-200-armv6m-le-gcc.s +0 -446
- data/ext/xkcp/lib/low/KeccakP-200/ARM/KeccakP-200-armv7m-le-armcc.s +0 -419
- data/ext/xkcp/lib/low/KeccakP-200/ARM/KeccakP-200-armv7m-le-gcc.s +0 -427
- data/ext/xkcp/lib/low/KeccakP-200/AVR8/KeccakP-200-SnP.h +0 -41
- data/ext/xkcp/lib/low/KeccakP-200/AVR8/KeccakP-200-avr8-fast.s +0 -647
- data/ext/xkcp/lib/low/KeccakP-200/compact/KeccakP-200-SnP.h +0 -39
- data/ext/xkcp/lib/low/KeccakP-200/compact/KeccakP-200-compact.c +0 -190
- data/ext/xkcp/lib/low/KeccakP-200/ref/KeccakP-200-SnP.h +0 -43
- data/ext/xkcp/lib/low/KeccakP-200/ref/KeccakP-200-reference.c +0 -412
- data/ext/xkcp/lib/low/KeccakP-200/ref/KeccakP-200-reference.h +0 -23
- data/ext/xkcp/lib/low/KeccakP-400/ARM/KeccakP-400-SnP.h +0 -41
- data/ext/xkcp/lib/low/KeccakP-400/ARM/KeccakP-400-armv6m-le-armcc.s +0 -454
- data/ext/xkcp/lib/low/KeccakP-400/ARM/KeccakP-400-armv6m-le-gcc.s +0 -458
- data/ext/xkcp/lib/low/KeccakP-400/ARM/KeccakP-400-armv7m-le-armcc.s +0 -455
- data/ext/xkcp/lib/low/KeccakP-400/ARM/KeccakP-400-armv7m-le-gcc.s +0 -458
- data/ext/xkcp/lib/low/KeccakP-400/AVR8/KeccakP-400-SnP.h +0 -41
- data/ext/xkcp/lib/low/KeccakP-400/AVR8/KeccakP-400-avr8-fast.s +0 -728
- data/ext/xkcp/lib/low/KeccakP-400/ref/KeccakP-400-SnP.h +0 -43
- data/ext/xkcp/lib/low/KeccakP-400/ref/KeccakP-400-reference.c +0 -414
- data/ext/xkcp/lib/low/KeccakP-400/ref/KeccakP-400-reference.h +0 -23
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-SnP.h +0 -42
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u1-armv6m-le-armcc.s +0 -527
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u1-armv6m-le-gcc.s +0 -533
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u2-armv6m-le-armcc.s +0 -528
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u2-armv6m-le-gcc.s +0 -534
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u2-armv7a-le-armcc.s +0 -521
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u2-armv7a-le-gcc.s +0 -527
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u2-armv7m-le-armcc.s +0 -517
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-u2-armv7m-le-gcc.s +0 -523
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-uf-armv7m-le-armcc.s +0 -550
- data/ext/xkcp/lib/low/KeccakP-800/ARM/KeccakP-800-uf-armv7m-le-gcc.s +0 -556
- data/ext/xkcp/lib/low/KeccakP-800/ARMv8A/KeccakP-800-SnP.h +0 -32
- data/ext/xkcp/lib/low/KeccakP-800/ARMv8A/KeccakP-800-armv8a-neon.s +0 -432
- data/ext/xkcp/lib/low/KeccakP-800/AVR8/KeccakP-800-SnP.h +0 -42
- data/ext/xkcp/lib/low/KeccakP-800/AVR8/KeccakP-800-avr8-fast.s +0 -929
- data/ext/xkcp/lib/low/KeccakP-800/compact/KeccakP-800-SnP.h +0 -40
- data/ext/xkcp/lib/low/KeccakP-800/compact/KeccakP-800-compact.c +0 -244
- data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-SnP.h +0 -46
- data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-opt32-bis.macros +0 -184
- data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-opt32.c +0 -454
- data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-opt32.macros +0 -459
- data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-unrolling-bis.macros +0 -83
- data/ext/xkcp/lib/low/KeccakP-800/plain/KeccakP-800-unrolling.macros +0 -88
- data/ext/xkcp/lib/low/KeccakP-800/plain/lcu2/KeccakP-800-opt32-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-800/plain/lcua/KeccakP-800-opt32-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-800/plain/u2/KeccakP-800-opt32-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-800/plain/ua/KeccakP-800-opt32-config.h +0 -7
- data/ext/xkcp/lib/low/KeccakP-800/ref/KeccakP-800-SnP.h +0 -44
- data/ext/xkcp/lib/low/KeccakP-800/ref/KeccakP-800-reference.c +0 -437
- data/ext/xkcp/lib/low/KeccakP-800/ref/KeccakP-800-reference.h +0 -23
- data/ext/xkcp/lib/low/Ketje/OptimizedAsmARM/Ket.h +0 -57
- data/ext/xkcp/lib/low/Ketje/OptimizedAsmARM/KetjeJr-armv7m-le-armcc.s +0 -475
- data/ext/xkcp/lib/low/Ketje/OptimizedAsmARM/KetjeJr-armv7m-le-gcc.s +0 -480
- data/ext/xkcp/lib/low/Ketje/OptimizedAsmARM/KetjeSr-armv7m-le-armcc.s +0 -590
- data/ext/xkcp/lib/low/Ketje/OptimizedAsmARM/KetjeSr-armv7m-le-gcc.s +0 -590
- data/ext/xkcp/lib/low/Ketje/OptimizedLE/Ket.c +0 -126
- data/ext/xkcp/lib/low/Ketje/OptimizedLE/Ket.h +0 -68
- data/ext/xkcp/lib/low/Ketje/OptimizedLE/Ket.inc +0 -174
- data/ext/xkcp/lib/low/Ketje/SnP-compliant/Ket.c +0 -80
- data/ext/xkcp/lib/low/Ketje/SnP-compliant/Ket.h +0 -68
- data/ext/xkcp/lib/low/Ketje/SnP-compliant/Ket.inc +0 -142
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-SnP.h +0 -55
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-u1-armv6m-le-armcc.s +0 -1086
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-u1-armv6m-le-gcc.s +0 -1092
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-uf-armv6-le-armcc.s +0 -721
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-uf-armv6-le-gcc.s +0 -726
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-uf-armv7m-le-armcc.s +0 -723
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodoo-uf-armv7m-le-gcc.s +0 -729
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodyak-u1-armv6m-le-armcc.s +0 -1164
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodyak-u1-armv6m-le-gcc.s +0 -1165
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodyak-uf-armv6-le-armcc.s +0 -562
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodyak-uf-armv6-le-gcc.s +0 -563
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodyak-uf-armv7m-le-armcc.s +0 -563
- data/ext/xkcp/lib/low/Xoodoo/ARM/Xoodyak-uf-armv7m-le-gcc.s +0 -565
- data/ext/xkcp/lib/low/Xoodoo/ARMv7A-NEON/Xoodoo-SnP.h +0 -55
- data/ext/xkcp/lib/low/Xoodoo/ARMv7A-NEON/Xoodoo-uf-armv7a-neon-le-armcc.s +0 -476
- data/ext/xkcp/lib/low/Xoodoo/ARMv7A-NEON/Xoodoo-uf-armv7a-neon-le-gcc.s +0 -485
- data/ext/xkcp/lib/low/Xoodoo/ARMv7A-NEON/Xoodyak-uf-armv7a-neon-le-armcc.s +0 -362
- data/ext/xkcp/lib/low/Xoodoo/ARMv7A-NEON/Xoodyak-uf-armv7a-neon-le-gcc.s +0 -367
- data/ext/xkcp/lib/low/Xoodoo/AVR8/Xoodoo-SnP.h +0 -43
- data/ext/xkcp/lib/low/Xoodoo/AVR8/Xoodoo-avr8-u1.s +0 -1341
- data/ext/xkcp/lib/low/Xoodoo/AVX512/Xoodoo-SIMD512.c +0 -581
- data/ext/xkcp/lib/low/Xoodoo/AVX512/Xoodoo-SnP.h +0 -58
- data/ext/xkcp/lib/low/Xoodoo/AVX512/Xoodyak-full-block-SIMD512.c +0 -332
- data/ext/xkcp/lib/low/Xoodoo/SSE2/Xoodoo-SIMD128.c +0 -329
- data/ext/xkcp/lib/low/Xoodoo/SSE2/Xoodoo-SnP.h +0 -53
- data/ext/xkcp/lib/low/Xoodoo/SSE2/Xoodyak-full-block-SIMD128.c +0 -355
- data/ext/xkcp/lib/low/Xoodoo/Xoodoo.h +0 -79
- data/ext/xkcp/lib/low/Xoodoo/plain/Xoodoo-SnP.h +0 -56
- data/ext/xkcp/lib/low/Xoodoo/plain/Xoodoo-optimized.c +0 -399
- data/ext/xkcp/lib/low/Xoodoo/plain/Xoodyak-full-blocks.c +0 -127
- data/ext/xkcp/lib/low/Xoodoo/ref/Xoodoo-SnP.h +0 -43
- data/ext/xkcp/lib/low/Xoodoo/ref/Xoodoo-reference.c +0 -253
- data/ext/xkcp/lib/low/Xoodoo-times16/AVX512/Xoodoo-times16-SIMD512.c +0 -1044
- data/ext/xkcp/lib/low/Xoodoo-times16/AVX512/Xoodoo-times16-SnP.h +0 -49
- data/ext/xkcp/lib/low/Xoodoo-times16/fallback-on1/Xoodoo-times16-SnP.h +0 -45
- data/ext/xkcp/lib/low/Xoodoo-times16/fallback-on1/Xoodoo-times16-on1.c +0 -37
- data/ext/xkcp/lib/low/Xoodoo-times4/ARMv7A-NEON/Xoodoo-times4-ARMv7A.s +0 -1587
- data/ext/xkcp/lib/low/Xoodoo-times4/ARMv7A-NEON/Xoodoo-times4-SnP.h +0 -48
- data/ext/xkcp/lib/low/Xoodoo-times4/AVX512/Xoodoo-times4-SIMD512.c +0 -1202
- data/ext/xkcp/lib/low/Xoodoo-times4/AVX512/Xoodoo-times4-SnP.h +0 -48
- data/ext/xkcp/lib/low/Xoodoo-times4/SSSE3/Xoodoo-times4-SIMD128.c +0 -484
- data/ext/xkcp/lib/low/Xoodoo-times4/SSSE3/Xoodoo-times4-SnP.h +0 -44
- data/ext/xkcp/lib/low/Xoodoo-times4/fallback-on1/Xoodoo-times4-SnP.h +0 -45
- data/ext/xkcp/lib/low/Xoodoo-times4/fallback-on1/Xoodoo-times4-on1.c +0 -37
- data/ext/xkcp/lib/low/Xoodoo-times8/AVX2/Xoodoo-times8-SIMD256.c +0 -939
- data/ext/xkcp/lib/low/Xoodoo-times8/AVX2/Xoodoo-times8-SnP.h +0 -49
- data/ext/xkcp/lib/low/Xoodoo-times8/AVX512/Xoodoo-times8-SIMD512.c +0 -1216
- data/ext/xkcp/lib/low/Xoodoo-times8/AVX512/Xoodoo-times8-SnP.h +0 -48
- data/ext/xkcp/lib/low/Xoodoo-times8/fallback-on1/Xoodoo-times8-SnP.h +0 -45
- data/ext/xkcp/lib/low/Xoodoo-times8/fallback-on1/Xoodoo-times8-on1.c +0 -37
- data/ext/xkcp/lib/low/common/PlSnP-Fallback.inc +0 -290
- data/ext/xkcp/lib/low/common/SnP-Relaned.h +0 -141
- data/ext/xkcp/support/Kernel-PMU/Kernel-pmu.md +0 -133
- data/ext/xkcp/support/Kernel-PMU/Makefile +0 -8
- data/ext/xkcp/support/Kernel-PMU/enable_arm_pmu.c +0 -129
- data/ext/xkcp/support/Kernel-PMU/load-module +0 -1
- data/ext/xkcp/util/KeccakSum/KeccakSum.c +0 -394
- data/ext/xkcp/util/KeccakSum/base64.c +0 -86
- data/ext/xkcp/util/KeccakSum/base64.h +0 -12
@@ -1,5 +1,5 @@
|
|
1
1
|
/*
|
2
|
-
|
2
|
+
K12 based on the eXtended Keccak Code Package (XKCP)
|
3
3
|
https://github.com/XKCP/XKCP
|
4
4
|
|
5
5
|
The Keccak-p permutations, designed by Guido Bertoni, Joan Daemen, Michaël Peeters and Gilles Van Assche.
|
@@ -12,8 +12,241 @@ https://keccak.team/
|
|
12
12
|
To the extent possible under law, the implementer has waived all copyright
|
13
13
|
and related or neighboring rights to the source code in this file.
|
14
14
|
http://creativecommons.org/publicdomain/zero/1.0/
|
15
|
+
|
16
|
+
---
|
17
|
+
|
18
|
+
Please refer to the XKCP for more details.
|
15
19
|
*/
|
16
20
|
|
21
|
+
#include <stdint.h>
|
22
|
+
#include <stdlib.h>
|
23
|
+
#include <string.h>
|
24
|
+
#include "brg_endian.h"
|
25
|
+
#include <KeccakP-1600-SnP.h>
|
26
|
+
|
27
|
+
#define KeccakP1600_opt64_implementation_config "all rounds unrolled"
|
28
|
+
#define KeccakP1600_opt64_fullUnrolling
|
29
|
+
/* Or */
|
30
|
+
/*
|
31
|
+
#define KeccakP1600_opt64_implementation_config "6 rounds unrolled"
|
32
|
+
#define KeccakP1600_opt64_unrolling 6
|
33
|
+
*/
|
34
|
+
/* Or */
|
35
|
+
/*
|
36
|
+
#define KeccakP1600_opt64_implementation_config "lane complementing, 6 rounds unrolled"
|
37
|
+
#define KeccakP1600_opt64_unrolling 6
|
38
|
+
#define KeccakP1600_opt64_useLaneComplementing
|
39
|
+
*/
|
40
|
+
/* Or */
|
41
|
+
/*
|
42
|
+
#define KeccakP1600_opt64_implementation_config "lane complementing, all rounds unrolled"
|
43
|
+
#define KeccakP1600_opt64_fullUnrolling
|
44
|
+
#define KeccakP1600_opt64_useLaneComplementing
|
45
|
+
*/
|
46
|
+
/* Or */
|
47
|
+
/*
|
48
|
+
#define KeccakP1600_opt64_implementation_config "lane complementing, all rounds unrolled, using SHLD for rotations"
|
49
|
+
#define KeccakP1600_opt64_fullUnrolling
|
50
|
+
#define KeccakP1600_opt64_useLaneComplementing
|
51
|
+
#define KeccakP1600_opt64_useSHLD
|
52
|
+
*/
|
53
|
+
|
54
|
+
#if defined(KeccakP1600_opt64_useLaneComplementing)
|
55
|
+
#define UseBebigokimisa
|
56
|
+
#endif
|
57
|
+
|
58
|
+
#if defined(_MSC_VER)
|
59
|
+
#define ROL64(a, offset) _rotl64(a, offset)
|
60
|
+
#elif defined(KeccakP1600_opt64_useSHLD)
|
61
|
+
#define ROL64(x,N) ({ \
|
62
|
+
register uint64_t __out; \
|
63
|
+
register uint64_t __in = x; \
|
64
|
+
__asm__ ("shld %2,%0,%0" : "=r"(__out) : "0"(__in), "i"(N)); \
|
65
|
+
__out; \
|
66
|
+
})
|
67
|
+
#else
|
68
|
+
#define ROL64(a, offset) ((((uint64_t)a) << offset) ^ (((uint64_t)a) >> (64-offset)))
|
69
|
+
#endif
|
70
|
+
|
71
|
+
#ifdef KeccakP1600_opt64_fullUnrolling
|
72
|
+
#define FullUnrolling
|
73
|
+
#else
|
74
|
+
#define Unrolling KeccakP1600_opt64_unrolling
|
75
|
+
#endif
|
76
|
+
|
77
|
+
static const uint64_t KeccakF1600RoundConstants[24] = {
|
78
|
+
0x0000000000000001ULL,
|
79
|
+
0x0000000000008082ULL,
|
80
|
+
0x800000000000808aULL,
|
81
|
+
0x8000000080008000ULL,
|
82
|
+
0x000000000000808bULL,
|
83
|
+
0x0000000080000001ULL,
|
84
|
+
0x8000000080008081ULL,
|
85
|
+
0x8000000000008009ULL,
|
86
|
+
0x000000000000008aULL,
|
87
|
+
0x0000000000000088ULL,
|
88
|
+
0x0000000080008009ULL,
|
89
|
+
0x000000008000000aULL,
|
90
|
+
0x000000008000808bULL,
|
91
|
+
0x800000000000008bULL,
|
92
|
+
0x8000000000008089ULL,
|
93
|
+
0x8000000000008003ULL,
|
94
|
+
0x8000000000008002ULL,
|
95
|
+
0x8000000000000080ULL,
|
96
|
+
0x000000000000800aULL,
|
97
|
+
0x800000008000000aULL,
|
98
|
+
0x8000000080008081ULL,
|
99
|
+
0x8000000000008080ULL,
|
100
|
+
0x0000000080000001ULL,
|
101
|
+
0x8000000080008008ULL };
|
102
|
+
|
103
|
+
/* ---------------------------------------------------------------- */
|
104
|
+
|
105
|
+
void KeccakP1600_opt64_Initialize(void *state)
|
106
|
+
{
|
107
|
+
memset(state, 0, 200);
|
108
|
+
#ifdef KeccakP1600_opt64_useLaneComplementing
|
109
|
+
((uint64_t*)state)[ 1] = ~(uint64_t)0;
|
110
|
+
((uint64_t*)state)[ 2] = ~(uint64_t)0;
|
111
|
+
((uint64_t*)state)[ 8] = ~(uint64_t)0;
|
112
|
+
((uint64_t*)state)[12] = ~(uint64_t)0;
|
113
|
+
((uint64_t*)state)[17] = ~(uint64_t)0;
|
114
|
+
((uint64_t*)state)[20] = ~(uint64_t)0;
|
115
|
+
#endif
|
116
|
+
}
|
117
|
+
|
118
|
+
/* ---------------------------------------------------------------- */
|
119
|
+
|
120
|
+
void KeccakP1600_opt64_AddBytesInLane(void *state, unsigned int lanePosition, const unsigned char *data, unsigned int offset, unsigned int length)
|
121
|
+
{
|
122
|
+
#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
|
123
|
+
uint64_t lane;
|
124
|
+
if (length == 0)
|
125
|
+
return;
|
126
|
+
if (length == 1)
|
127
|
+
lane = data[0];
|
128
|
+
else {
|
129
|
+
lane = 0;
|
130
|
+
memcpy(&lane, data, length);
|
131
|
+
}
|
132
|
+
lane <<= offset*8;
|
133
|
+
#else
|
134
|
+
uint64_t lane = 0;
|
135
|
+
unsigned int i;
|
136
|
+
for(i=0; i<length; i++)
|
137
|
+
lane |= ((uint64_t)data[i]) << ((i+offset)*8);
|
138
|
+
#endif
|
139
|
+
((uint64_t*)state)[lanePosition] ^= lane;
|
140
|
+
}
|
141
|
+
|
142
|
+
/* ---------------------------------------------------------------- */
|
143
|
+
|
144
|
+
static void KeccakP1600_opt64_AddLanes(void *state, const unsigned char *data, unsigned int laneCount)
|
145
|
+
{
|
146
|
+
#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
|
147
|
+
unsigned int i = 0;
|
148
|
+
#ifdef NO_MISALIGNED_ACCESSES
|
149
|
+
/* If either pointer is misaligned, fall back to byte-wise xor. */
|
150
|
+
if (((((uintptr_t)state) & 7) != 0) || ((((uintptr_t)data) & 7) != 0)) {
|
151
|
+
for (i = 0; i < laneCount * 8; i++) {
|
152
|
+
((unsigned char*)state)[i] ^= data[i];
|
153
|
+
}
|
154
|
+
}
|
155
|
+
else
|
156
|
+
#endif
|
157
|
+
{
|
158
|
+
/* Otherwise... */
|
159
|
+
for( ; (i+8)<=laneCount; i+=8) {
|
160
|
+
((uint64_t*)state)[i+0] ^= ((uint64_t*)data)[i+0];
|
161
|
+
((uint64_t*)state)[i+1] ^= ((uint64_t*)data)[i+1];
|
162
|
+
((uint64_t*)state)[i+2] ^= ((uint64_t*)data)[i+2];
|
163
|
+
((uint64_t*)state)[i+3] ^= ((uint64_t*)data)[i+3];
|
164
|
+
((uint64_t*)state)[i+4] ^= ((uint64_t*)data)[i+4];
|
165
|
+
((uint64_t*)state)[i+5] ^= ((uint64_t*)data)[i+5];
|
166
|
+
((uint64_t*)state)[i+6] ^= ((uint64_t*)data)[i+6];
|
167
|
+
((uint64_t*)state)[i+7] ^= ((uint64_t*)data)[i+7];
|
168
|
+
}
|
169
|
+
for( ; (i+4)<=laneCount; i+=4) {
|
170
|
+
((uint64_t*)state)[i+0] ^= ((uint64_t*)data)[i+0];
|
171
|
+
((uint64_t*)state)[i+1] ^= ((uint64_t*)data)[i+1];
|
172
|
+
((uint64_t*)state)[i+2] ^= ((uint64_t*)data)[i+2];
|
173
|
+
((uint64_t*)state)[i+3] ^= ((uint64_t*)data)[i+3];
|
174
|
+
}
|
175
|
+
for( ; (i+2)<=laneCount; i+=2) {
|
176
|
+
((uint64_t*)state)[i+0] ^= ((uint64_t*)data)[i+0];
|
177
|
+
((uint64_t*)state)[i+1] ^= ((uint64_t*)data)[i+1];
|
178
|
+
}
|
179
|
+
if (i<laneCount) {
|
180
|
+
((uint64_t*)state)[i+0] ^= ((uint64_t*)data)[i+0];
|
181
|
+
}
|
182
|
+
}
|
183
|
+
#else
|
184
|
+
unsigned int i;
|
185
|
+
const uint8_t *curData = data;
|
186
|
+
for(i=0; i<laneCount; i++, curData+=8) {
|
187
|
+
uint64_t lane = (uint64_t)curData[0]
|
188
|
+
| ((uint64_t)curData[1] << 8)
|
189
|
+
| ((uint64_t)curData[2] << 16)
|
190
|
+
| ((uint64_t)curData[3] << 24)
|
191
|
+
| ((uint64_t)curData[4] << 32)
|
192
|
+
| ((uint64_t)curData[5] << 40)
|
193
|
+
| ((uint64_t)curData[6] << 48)
|
194
|
+
| ((uint64_t)curData[7] << 56);
|
195
|
+
((uint64_t*)state)[i] ^= lane;
|
196
|
+
}
|
197
|
+
#endif
|
198
|
+
}
|
199
|
+
|
200
|
+
/* ---------------------------------------------------------------- */
|
201
|
+
|
202
|
+
void KeccakP1600_opt64_AddByte(void *state, unsigned char byte, unsigned int offset)
|
203
|
+
{
|
204
|
+
#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
|
205
|
+
((unsigned char*)(state))[offset] ^= byte;
|
206
|
+
#else
|
207
|
+
uint64_t lane = byte;
|
208
|
+
lane <<= (offset%8)*8;
|
209
|
+
((uint64_t*)state)[offset/8] ^= lane;
|
210
|
+
#endif
|
211
|
+
}
|
212
|
+
|
213
|
+
/* ---------------------------------------------------------------- */
|
214
|
+
|
215
|
+
#define SnP_AddBytes(state, data, offset, length, SnP_AddLanes, SnP_AddBytesInLane, SnP_laneLengthInBytes) \
|
216
|
+
{ \
|
217
|
+
if ((offset) == 0) { \
|
218
|
+
SnP_AddLanes(state, data, (length)/SnP_laneLengthInBytes); \
|
219
|
+
SnP_AddBytesInLane(state, \
|
220
|
+
(length)/SnP_laneLengthInBytes, \
|
221
|
+
(data)+((length)/SnP_laneLengthInBytes)*SnP_laneLengthInBytes, \
|
222
|
+
0, \
|
223
|
+
(length)%SnP_laneLengthInBytes); \
|
224
|
+
} \
|
225
|
+
else { \
|
226
|
+
unsigned int _sizeLeft = (length); \
|
227
|
+
unsigned int _lanePosition = (offset)/SnP_laneLengthInBytes; \
|
228
|
+
unsigned int _offsetInLane = (offset)%SnP_laneLengthInBytes; \
|
229
|
+
const unsigned char *_curData = (data); \
|
230
|
+
while(_sizeLeft > 0) { \
|
231
|
+
unsigned int _bytesInLane = SnP_laneLengthInBytes - _offsetInLane; \
|
232
|
+
if (_bytesInLane > _sizeLeft) \
|
233
|
+
_bytesInLane = _sizeLeft; \
|
234
|
+
SnP_AddBytesInLane(state, _lanePosition, _curData, _offsetInLane, _bytesInLane); \
|
235
|
+
_sizeLeft -= _bytesInLane; \
|
236
|
+
_lanePosition++; \
|
237
|
+
_offsetInLane = 0; \
|
238
|
+
_curData += _bytesInLane; \
|
239
|
+
} \
|
240
|
+
} \
|
241
|
+
}
|
242
|
+
|
243
|
+
void KeccakP1600_opt64_AddBytes(void *state, const unsigned char *data, unsigned int offset, unsigned int length)
|
244
|
+
{
|
245
|
+
SnP_AddBytes(state, data, offset, length, KeccakP1600_opt64_AddLanes, KeccakP1600_opt64_AddBytesInLane, 8);
|
246
|
+
}
|
247
|
+
|
248
|
+
/* ---------------------------------------------------------------- */
|
249
|
+
|
17
250
|
#define declareABCDE \
|
18
251
|
uint64_t Aba, Abe, Abi, Abo, Abu; \
|
19
252
|
uint64_t Aga, Age, Agi, Ago, Agu; \
|
@@ -539,6 +772,193 @@ http://creativecommons.org/publicdomain/zero/1.0/
|
|
539
772
|
X##so = Y##so; \
|
540
773
|
X##su = Y##su; \
|
541
774
|
|
775
|
+
#if ((defined(FullUnrolling)) || (Unrolling == 12))
|
776
|
+
#define rounds12 \
|
777
|
+
prepareTheta \
|
778
|
+
thetaRhoPiChiIotaPrepareTheta(12, A, E) \
|
779
|
+
thetaRhoPiChiIotaPrepareTheta(13, E, A) \
|
780
|
+
thetaRhoPiChiIotaPrepareTheta(14, A, E) \
|
781
|
+
thetaRhoPiChiIotaPrepareTheta(15, E, A) \
|
782
|
+
thetaRhoPiChiIotaPrepareTheta(16, A, E) \
|
783
|
+
thetaRhoPiChiIotaPrepareTheta(17, E, A) \
|
784
|
+
thetaRhoPiChiIotaPrepareTheta(18, A, E) \
|
785
|
+
thetaRhoPiChiIotaPrepareTheta(19, E, A) \
|
786
|
+
thetaRhoPiChiIotaPrepareTheta(20, A, E) \
|
787
|
+
thetaRhoPiChiIotaPrepareTheta(21, E, A) \
|
788
|
+
thetaRhoPiChiIotaPrepareTheta(22, A, E) \
|
789
|
+
thetaRhoPiChiIota(23, E, A) \
|
790
|
+
|
791
|
+
#elif (Unrolling == 6)
|
792
|
+
#define rounds12 \
|
793
|
+
prepareTheta \
|
794
|
+
for(i=12; i<24; i+=6) { \
|
795
|
+
thetaRhoPiChiIotaPrepareTheta(i , A, E) \
|
796
|
+
thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
|
797
|
+
thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
|
798
|
+
thetaRhoPiChiIotaPrepareTheta(i+3, E, A) \
|
799
|
+
thetaRhoPiChiIotaPrepareTheta(i+4, A, E) \
|
800
|
+
thetaRhoPiChiIotaPrepareTheta(i+5, E, A) \
|
801
|
+
} \
|
802
|
+
|
803
|
+
#elif (Unrolling == 4)
|
804
|
+
#define rounds12 \
|
805
|
+
prepareTheta \
|
806
|
+
for(i=12; i<24; i+=4) { \
|
807
|
+
thetaRhoPiChiIotaPrepareTheta(i , A, E) \
|
808
|
+
thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
|
809
|
+
thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
|
810
|
+
thetaRhoPiChiIotaPrepareTheta(i+3, E, A) \
|
811
|
+
} \
|
812
|
+
|
813
|
+
#elif (Unrolling == 3)
|
814
|
+
#define rounds12 \
|
815
|
+
prepareTheta \
|
816
|
+
for(i=12; i<24; i+=3) { \
|
817
|
+
thetaRhoPiChiIotaPrepareTheta(i , A, E) \
|
818
|
+
thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
|
819
|
+
thetaRhoPiChiIotaPrepareTheta(i+2, A, E) \
|
820
|
+
copyStateVariables(A, E) \
|
821
|
+
} \
|
822
|
+
|
823
|
+
#elif (Unrolling == 2)
|
824
|
+
#define rounds12 \
|
825
|
+
prepareTheta \
|
826
|
+
for(i=12; i<24; i+=2) { \
|
827
|
+
thetaRhoPiChiIotaPrepareTheta(i , A, E) \
|
828
|
+
thetaRhoPiChiIotaPrepareTheta(i+1, E, A) \
|
829
|
+
} \
|
830
|
+
|
831
|
+
#elif (Unrolling == 1)
|
832
|
+
#define rounds12 \
|
833
|
+
prepareTheta \
|
834
|
+
for(i=12; i<24; i++) { \
|
835
|
+
thetaRhoPiChiIotaPrepareTheta(i , A, E) \
|
836
|
+
copyStateVariables(A, E) \
|
837
|
+
} \
|
838
|
+
|
839
|
+
#else
|
840
|
+
#error "Unrolling is not correctly specified!"
|
841
|
+
#endif
|
842
|
+
|
843
|
+
void KeccakP1600_opt64_Permute_12rounds(void *state)
|
844
|
+
{
|
845
|
+
declareABCDE
|
846
|
+
#ifndef KeccakP1600_opt64_fullUnrolling
|
847
|
+
unsigned int i;
|
848
|
+
#endif
|
849
|
+
uint64_t *stateAsLanes = (uint64_t*)state;
|
850
|
+
|
851
|
+
copyFromState(A, stateAsLanes)
|
852
|
+
rounds12
|
853
|
+
copyToState(stateAsLanes, A)
|
854
|
+
}
|
855
|
+
|
856
|
+
/* ---------------------------------------------------------------- */
|
857
|
+
|
858
|
+
void KeccakP1600_opt64_ExtractBytesInLane(const void *state, unsigned int lanePosition, unsigned char *data, unsigned int offset, unsigned int length)
|
859
|
+
{
|
860
|
+
uint64_t lane = ((uint64_t*)state)[lanePosition];
|
861
|
+
#ifdef KeccakP1600_opt64_useLaneComplementing
|
862
|
+
if ((lanePosition == 1) || (lanePosition == 2) || (lanePosition == 8) || (lanePosition == 12) || (lanePosition == 17) || (lanePosition == 20))
|
863
|
+
lane = ~lane;
|
864
|
+
#endif
|
865
|
+
#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
|
866
|
+
{
|
867
|
+
uint64_t lane1[1];
|
868
|
+
lane1[0] = lane;
|
869
|
+
memcpy(data, (uint8_t*)lane1+offset, length);
|
870
|
+
}
|
871
|
+
#else
|
872
|
+
unsigned int i;
|
873
|
+
lane >>= offset*8;
|
874
|
+
for(i=0; i<length; i++) {
|
875
|
+
data[i] = lane & 0xFF;
|
876
|
+
lane >>= 8;
|
877
|
+
}
|
878
|
+
#endif
|
879
|
+
}
|
880
|
+
|
881
|
+
/* ---------------------------------------------------------------- */
|
882
|
+
|
883
|
+
#if (PLATFORM_BYTE_ORDER != IS_LITTLE_ENDIAN)
|
884
|
+
static void fromWordToBytes(uint8_t *bytes, const uint64_t word)
|
885
|
+
{
|
886
|
+
unsigned int i;
|
887
|
+
|
888
|
+
for(i=0; i<(64/8); i++)
|
889
|
+
bytes[i] = (word >> (8*i)) & 0xFF;
|
890
|
+
}
|
891
|
+
#endif
|
892
|
+
|
893
|
+
void KeccakP1600_opt64_ExtractLanes(const void *state, unsigned char *data, unsigned int laneCount)
|
894
|
+
{
|
895
|
+
#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
|
896
|
+
memcpy(data, state, laneCount*8);
|
897
|
+
#else
|
898
|
+
unsigned int i;
|
899
|
+
|
900
|
+
for(i=0; i<laneCount; i++)
|
901
|
+
fromWordToBytes(data+(i*8), ((const uint64_t*)state)[i]);
|
902
|
+
#endif
|
903
|
+
#ifdef KeccakP1600_opt64_useLaneComplementing
|
904
|
+
if (laneCount > 1) {
|
905
|
+
((uint64_t*)data)[ 1] = ~((uint64_t*)data)[ 1];
|
906
|
+
if (laneCount > 2) {
|
907
|
+
((uint64_t*)data)[ 2] = ~((uint64_t*)data)[ 2];
|
908
|
+
if (laneCount > 8) {
|
909
|
+
((uint64_t*)data)[ 8] = ~((uint64_t*)data)[ 8];
|
910
|
+
if (laneCount > 12) {
|
911
|
+
((uint64_t*)data)[12] = ~((uint64_t*)data)[12];
|
912
|
+
if (laneCount > 17) {
|
913
|
+
((uint64_t*)data)[17] = ~((uint64_t*)data)[17];
|
914
|
+
if (laneCount > 20) {
|
915
|
+
((uint64_t*)data)[20] = ~((uint64_t*)data)[20];
|
916
|
+
}
|
917
|
+
}
|
918
|
+
}
|
919
|
+
}
|
920
|
+
}
|
921
|
+
}
|
922
|
+
#endif
|
923
|
+
}
|
924
|
+
|
925
|
+
/* ---------------------------------------------------------------- */
|
926
|
+
|
927
|
+
#define SnP_ExtractBytes(state, data, offset, length, SnP_ExtractLanes, SnP_ExtractBytesInLane, SnP_laneLengthInBytes) \
|
928
|
+
{ \
|
929
|
+
if ((offset) == 0) { \
|
930
|
+
SnP_ExtractLanes(state, data, (length)/SnP_laneLengthInBytes); \
|
931
|
+
SnP_ExtractBytesInLane(state, \
|
932
|
+
(length)/SnP_laneLengthInBytes, \
|
933
|
+
(data)+((length)/SnP_laneLengthInBytes)*SnP_laneLengthInBytes, \
|
934
|
+
0, \
|
935
|
+
(length)%SnP_laneLengthInBytes); \
|
936
|
+
} \
|
937
|
+
else { \
|
938
|
+
unsigned int _sizeLeft = (length); \
|
939
|
+
unsigned int _lanePosition = (offset)/SnP_laneLengthInBytes; \
|
940
|
+
unsigned int _offsetInLane = (offset)%SnP_laneLengthInBytes; \
|
941
|
+
unsigned char *_curData = (data); \
|
942
|
+
while(_sizeLeft > 0) { \
|
943
|
+
unsigned int _bytesInLane = SnP_laneLengthInBytes - _offsetInLane; \
|
944
|
+
if (_bytesInLane > _sizeLeft) \
|
945
|
+
_bytesInLane = _sizeLeft; \
|
946
|
+
SnP_ExtractBytesInLane(state, _lanePosition, _curData, _offsetInLane, _bytesInLane); \
|
947
|
+
_sizeLeft -= _bytesInLane; \
|
948
|
+
_lanePosition++; \
|
949
|
+
_offsetInLane = 0; \
|
950
|
+
_curData += _bytesInLane; \
|
951
|
+
} \
|
952
|
+
} \
|
953
|
+
}
|
954
|
+
|
955
|
+
void KeccakP1600_opt64_ExtractBytes(const void *state, unsigned char *data, unsigned int offset, unsigned int length)
|
956
|
+
{
|
957
|
+
SnP_ExtractBytes(state, data, offset, length, KeccakP1600_opt64_ExtractLanes, KeccakP1600_opt64_ExtractBytesInLane, 8);
|
958
|
+
}
|
959
|
+
|
960
|
+
/* ---------------------------------------------------------------- */
|
961
|
+
|
542
962
|
#if (PLATFORM_BYTE_ORDER == IS_LITTLE_ENDIAN)
|
543
963
|
#define HTOLE64(x) (x)
|
544
964
|
#else
|
@@ -577,172 +997,30 @@ http://creativecommons.org/publicdomain/zero/1.0/
|
|
577
997
|
X##mu ^= HTOLE64(input[19]); \
|
578
998
|
X##sa ^= HTOLE64(input[20]); \
|
579
999
|
} \
|
580
|
-
|
581
|
-
|
582
|
-
|
583
|
-
|
584
|
-
|
585
|
-
|
586
|
-
|
587
|
-
|
588
|
-
|
589
|
-
|
590
|
-
|
591
|
-
|
592
|
-
|
593
|
-
|
594
|
-
|
595
|
-
|
596
|
-
|
597
|
-
|
598
|
-
|
599
|
-
|
600
|
-
|
601
|
-
|
602
|
-
X##be ^= HTOLE64(input[ 1]); \
|
603
|
-
X##bi ^= HTOLE64(input[ 2]); \
|
604
|
-
X##bo ^= HTOLE64(input[ 3]); \
|
605
|
-
if (laneCount < 6) { \
|
606
|
-
if (laneCount < 5) { \
|
607
|
-
} \
|
608
|
-
else { \
|
609
|
-
X##bu ^= HTOLE64(input[ 4]); \
|
610
|
-
} \
|
611
|
-
} \
|
612
|
-
else { \
|
613
|
-
X##bu ^= HTOLE64(input[ 4]); \
|
614
|
-
X##ga ^= HTOLE64(input[ 5]); \
|
615
|
-
if (laneCount < 7) { \
|
616
|
-
} \
|
617
|
-
else { \
|
618
|
-
X##ge ^= HTOLE64(input[ 6]); \
|
619
|
-
} \
|
620
|
-
} \
|
621
|
-
} \
|
622
|
-
} \
|
623
|
-
else { \
|
624
|
-
X##ba ^= HTOLE64(input[ 0]); \
|
625
|
-
X##be ^= HTOLE64(input[ 1]); \
|
626
|
-
X##bi ^= HTOLE64(input[ 2]); \
|
627
|
-
X##bo ^= HTOLE64(input[ 3]); \
|
628
|
-
X##bu ^= HTOLE64(input[ 4]); \
|
629
|
-
X##ga ^= HTOLE64(input[ 5]); \
|
630
|
-
X##ge ^= HTOLE64(input[ 6]); \
|
631
|
-
X##gi ^= HTOLE64(input[ 7]); \
|
632
|
-
if (laneCount < 12) { \
|
633
|
-
if (laneCount < 10) { \
|
634
|
-
if (laneCount < 9) { \
|
635
|
-
} \
|
636
|
-
else { \
|
637
|
-
X##go ^= HTOLE64(input[ 8]); \
|
638
|
-
} \
|
639
|
-
} \
|
640
|
-
else { \
|
641
|
-
X##go ^= HTOLE64(input[ 8]); \
|
642
|
-
X##gu ^= HTOLE64(input[ 9]); \
|
643
|
-
if (laneCount < 11) { \
|
644
|
-
} \
|
645
|
-
else { \
|
646
|
-
X##ka ^= HTOLE64(input[10]); \
|
647
|
-
} \
|
648
|
-
} \
|
649
|
-
} \
|
650
|
-
else { \
|
651
|
-
X##go ^= HTOLE64(input[ 8]); \
|
652
|
-
X##gu ^= HTOLE64(input[ 9]); \
|
653
|
-
X##ka ^= HTOLE64(input[10]); \
|
654
|
-
X##ke ^= HTOLE64(input[11]); \
|
655
|
-
if (laneCount < 14) { \
|
656
|
-
if (laneCount < 13) { \
|
657
|
-
} \
|
658
|
-
else { \
|
659
|
-
X##ki ^= HTOLE64(input[12]); \
|
660
|
-
} \
|
661
|
-
} \
|
662
|
-
else { \
|
663
|
-
X##ki ^= HTOLE64(input[12]); \
|
664
|
-
X##ko ^= HTOLE64(input[13]); \
|
665
|
-
if (laneCount < 15) { \
|
666
|
-
} \
|
667
|
-
else { \
|
668
|
-
X##ku ^= HTOLE64(input[14]); \
|
669
|
-
} \
|
670
|
-
} \
|
671
|
-
} \
|
672
|
-
} \
|
673
|
-
} \
|
674
|
-
else { \
|
675
|
-
X##ba ^= HTOLE64(input[ 0]); \
|
676
|
-
X##be ^= HTOLE64(input[ 1]); \
|
677
|
-
X##bi ^= HTOLE64(input[ 2]); \
|
678
|
-
X##bo ^= HTOLE64(input[ 3]); \
|
679
|
-
X##bu ^= HTOLE64(input[ 4]); \
|
680
|
-
X##ga ^= HTOLE64(input[ 5]); \
|
681
|
-
X##ge ^= HTOLE64(input[ 6]); \
|
682
|
-
X##gi ^= HTOLE64(input[ 7]); \
|
683
|
-
X##go ^= HTOLE64(input[ 8]); \
|
684
|
-
X##gu ^= HTOLE64(input[ 9]); \
|
685
|
-
X##ka ^= HTOLE64(input[10]); \
|
686
|
-
X##ke ^= HTOLE64(input[11]); \
|
687
|
-
X##ki ^= HTOLE64(input[12]); \
|
688
|
-
X##ko ^= HTOLE64(input[13]); \
|
689
|
-
X##ku ^= HTOLE64(input[14]); \
|
690
|
-
X##ma ^= HTOLE64(input[15]); \
|
691
|
-
if (laneCount < 24) { \
|
692
|
-
if (laneCount < 20) { \
|
693
|
-
if (laneCount < 18) { \
|
694
|
-
if (laneCount < 17) { \
|
695
|
-
} \
|
696
|
-
else { \
|
697
|
-
X##me ^= HTOLE64(input[16]); \
|
698
|
-
} \
|
699
|
-
} \
|
700
|
-
else { \
|
701
|
-
X##me ^= HTOLE64(input[16]); \
|
702
|
-
X##mi ^= HTOLE64(input[17]); \
|
703
|
-
if (laneCount < 19) { \
|
704
|
-
} \
|
705
|
-
else { \
|
706
|
-
X##mo ^= HTOLE64(input[18]); \
|
707
|
-
} \
|
708
|
-
} \
|
709
|
-
} \
|
710
|
-
else { \
|
711
|
-
X##me ^= HTOLE64(input[16]); \
|
712
|
-
X##mi ^= HTOLE64(input[17]); \
|
713
|
-
X##mo ^= HTOLE64(input[18]); \
|
714
|
-
X##mu ^= HTOLE64(input[19]); \
|
715
|
-
if (laneCount < 22) { \
|
716
|
-
if (laneCount < 21) { \
|
717
|
-
} \
|
718
|
-
else { \
|
719
|
-
X##sa ^= HTOLE64(input[20]); \
|
720
|
-
} \
|
721
|
-
} \
|
722
|
-
else { \
|
723
|
-
X##sa ^= HTOLE64(input[20]); \
|
724
|
-
X##se ^= HTOLE64(input[21]); \
|
725
|
-
if (laneCount < 23) { \
|
726
|
-
} \
|
727
|
-
else { \
|
728
|
-
X##si ^= HTOLE64(input[22]); \
|
729
|
-
} \
|
730
|
-
} \
|
731
|
-
} \
|
732
|
-
} \
|
733
|
-
else { \
|
734
|
-
X##me ^= HTOLE64(input[16]); \
|
735
|
-
X##mi ^= HTOLE64(input[17]); \
|
736
|
-
X##mo ^= HTOLE64(input[18]); \
|
737
|
-
X##mu ^= HTOLE64(input[19]); \
|
738
|
-
X##sa ^= HTOLE64(input[20]); \
|
739
|
-
X##se ^= HTOLE64(input[21]); \
|
740
|
-
X##si ^= HTOLE64(input[22]); \
|
741
|
-
X##so ^= HTOLE64(input[23]); \
|
742
|
-
if (laneCount < 25) { \
|
743
|
-
} \
|
744
|
-
else { \
|
745
|
-
X##su ^= HTOLE64(input[24]); \
|
746
|
-
} \
|
747
|
-
} \
|
1000
|
+
|
1001
|
+
#include <assert.h>
|
1002
|
+
|
1003
|
+
size_t KeccakP1600_opt64_12rounds_FastLoop_Absorb(void *state, unsigned int laneCount, const unsigned char *data, size_t dataByteLen)
|
1004
|
+
{
|
1005
|
+
size_t originalDataByteLen = dataByteLen;
|
1006
|
+
declareABCDE
|
1007
|
+
#ifndef KeccakP1600_opt64_fullUnrolling
|
1008
|
+
unsigned int i;
|
1009
|
+
#endif
|
1010
|
+
uint64_t *stateAsLanes = (uint64_t*)state;
|
1011
|
+
uint64_t *inDataAsLanes = (uint64_t*)data;
|
1012
|
+
|
1013
|
+
assert(laneCount == 21);
|
1014
|
+
|
1015
|
+
#define laneCount 21
|
1016
|
+
copyFromState(A, stateAsLanes)
|
1017
|
+
while(dataByteLen >= laneCount*8) {
|
1018
|
+
addInput(A, inDataAsLanes, laneCount)
|
1019
|
+
rounds12
|
1020
|
+
inDataAsLanes += laneCount;
|
1021
|
+
dataByteLen -= laneCount*8;
|
748
1022
|
}
|
1023
|
+
#undef laneCount
|
1024
|
+
copyToState(stateAsLanes, A)
|
1025
|
+
return originalDataByteLen - dataByteLen;
|
1026
|
+
}
|