fzorb/REDACTED-rig
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Integrats Hosp and adds ASM code for RTO/HOSP

Browse source
This commit is contained in:
Ben Gräf 2019-02-08 20:11:11 +01:00
parent 8c544aa08c
commit a6c88e2d7e
13 changed files with 625 additions and 44 deletions
Download patch file Download diff file Expand all files Collapse all files

14
src/Options.cpp
View file

@ -330,7 +330,8 @@ constexpr static const char *pow_variant_names[] = {
"xfh",
"fast2",
"upx",
"turtle"
"turtle",
"hosp"
};
constexpr static const char *asm_optimization_names[] = {
@ -1115,6 +1116,12 @@ bool Options::setAlgo(const char *algo)
break;
}
if (i == ARRAY_SIZE(algo_names) - 1 && (!strcmp(algo, "cryptonight-hospital") || !strcmp(algo, "cryptonight-hosp"))) {
m_algo = ALGO_CRYPTONIGHT;
m_powVariant = POW_HOSP;
break;
}
if (i == ARRAY_SIZE(algo_names) - 1) {
showUsage(1);
return false;
@ -1194,6 +1201,11 @@ bool Options::parsePowVariant(const char *powVariant)
break;
}
if (i == ARRAY_SIZE(pow_variant_names) - 1 && (!strcmp(powVariant, "hosp") || !strcmp(powVariant, "hospital"))) {
m_powVariant = POW_HOSP;
break;
}
if (i == ARRAY_SIZE(pow_variant_names) - 1) {
showUsage(1);
return false;

5
src/PowVariant.h
View file

@ -38,6 +38,7 @@ enum PowVariant
POW_FAST_2,
POW_UPX,
POW_TURTLE,
POW_HOSP,
LAST_ITEM
};
@ -71,6 +72,8 @@ inline std::string getPowVariantName(PowVariant powVariant)
return "upx";
case POW_TURTLE:
return "turtle";
case POW_HOSP:
return "hosp";
case POW_AUTODETECT:
default:
return "-1";
@ -144,6 +147,8 @@ inline PowVariant parseVariant(const std::string variant)
powVariant = PowVariant::POW_UPX;
} else if (variant == "turtle" || variant == "trtl" || variant == "pico" || variant == "turtlev2") {
powVariant = PowVariant::POW_TURTLE;
} else if (variant == "hosp" || variant == "hospital") {
powVariant = PowVariant::POW_HOSP;
}
return powVariant;

54
src/crypto/CryptoNight.cpp
View file

@ -41,7 +41,7 @@ static void cryptonight_aesni(AsmOptimization asmOptimization, PowVariant powVer
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -53,7 +53,7 @@ static void cryptonight_aesni(AsmOptimization asmOptimization, PowVariant powVer
if ((asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS <= 2) ||
(asmOptimization == AsmOptimization::ASM_RYZEN && NUM_HASH_BLOCKS == 1) ||
(asmOptimization == AsmOptimization::ASM_BULLDOZER && NUM_HASH_BLOCKS == 1)) {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
}
@ -65,7 +65,7 @@ static void cryptonight_aesni(AsmOptimization asmOptimization, PowVariant powVer
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -77,7 +77,7 @@ static void cryptonight_aesni(AsmOptimization asmOptimization, PowVariant powVer
if ((asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS <= 2) ||
(asmOptimization == AsmOptimization::ASM_RYZEN && NUM_HASH_BLOCKS == 1) ||
(asmOptimization == AsmOptimization::ASM_BULLDOZER && NUM_HASH_BLOCKS == 1)) {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
}
@ -87,13 +87,17 @@ static void cryptonight_aesni(AsmOptimization asmOptimization, PowVariant powVer
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
#endif
} else if (powVersion == PowVariant::POW_RTO) {
} else if (powVersion == PowVariant::POW_RTO || powVersion == PowVariant::POW_HOSP) {
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashLiteTube(input, size, output, scratchPad);
}
} else if (powVersion == PowVariant::POW_XFH) {
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashHeavyHaven(input, size, output, scratchPad);
} else {
@ -109,7 +113,7 @@ static void cryptonight_softaes(AsmOptimization asmOptimization, PowVariant powV
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -119,7 +123,7 @@ static void cryptonight_softaes(AsmOptimization asmOptimization, PowVariant powV
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
}
@ -129,7 +133,7 @@ static void cryptonight_softaes(AsmOptimization asmOptimization, PowVariant powV
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, false, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
}
@ -141,7 +145,7 @@ static void cryptonight_softaes(AsmOptimization asmOptimization, PowVariant powV
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_XLT_V4_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -151,13 +155,17 @@ static void cryptonight_softaes(AsmOptimization asmOptimization, PowVariant powV
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
#endif
} else if (powVersion == PowVariant::POW_RTO) {
} else if (powVersion == PowVariant::POW_RTO || powVersion == PowVariant::POW_HOSP) {
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x80000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashLiteTube(input, size, output, scratchPad);
}
} else if (powVersion == PowVariant::POW_XFH) {
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY, 0x1FFFF0, true, NUM_HASH_BLOCKS>::hashHeavyHaven(input, size, output, scratchPad);
} else {
@ -173,7 +181,7 @@ static void cryptonight_lite_aesni(AsmOptimization asmOptimization, PowVariant p
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -185,7 +193,7 @@ static void cryptonight_lite_aesni(AsmOptimization asmOptimization, PowVariant p
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, false, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -203,7 +211,7 @@ static void cryptonight_lite_softaes(AsmOptimization asmOptimization, PowVariant
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x40000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -215,7 +223,7 @@ static void cryptonight_lite_softaes(AsmOptimization asmOptimization, PowVariant
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x20000, POW_DEFAULT_INDEX_SHIFT, MEMORY_LITE, 0xFFFF0, true, NUM_HASH_BLOCKS>::hashPowV2(input, size, output, scratchPad);
}
@ -244,7 +252,7 @@ static void cryptonight_ultra_lite_aesni(AsmOptimization asmOptimization, PowVar
if ((asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS <= 2) ||
(asmOptimization == AsmOptimization::ASM_RYZEN && NUM_HASH_BLOCKS == 1) ||
(asmOptimization == AsmOptimization::ASM_BULLDOZER && NUM_HASH_BLOCKS == 1)) {
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, false, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, false, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, false, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
}
@ -258,7 +266,7 @@ static void cryptonight_ultra_lite_softaes(AsmOptimization asmOptimization, PowV
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, true, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
#else
if (asmOptimization == AsmOptimization::ASM_INTEL && NUM_HASH_BLOCKS == 1) {
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, true, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization);
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, true, NUM_HASH_BLOCKS>::hashPowV3_asm(input, size, output, scratchPad, asmOptimization, powVersion);
} else {
CryptoNightMultiHash<0x10000, POW_DEFAULT_INDEX_SHIFT, MEMORY_ULTRA_LITE, 0x1FFF0, true, NUM_HASH_BLOCKS>::hashPowV3(input, size, output, scratchPad);
}
@ -617,6 +625,11 @@ bool CryptoNight::selfTest(int algo)
cryptonight_hash_ctx[0](asmOptimization, PowVariant::POW_ALLOY,test_input, 76, output, scratchPads);
result = result && memcmp(output, test_output_alloy, 32) == 0;
// cnv7 + hosp/rto
cryptonight_hash_ctx[0](asmOptimization, PowVariant::POW_HOSP,test_input, 76, output, scratchPads);
result = result && memcmp(output, test_output_hosp, 32) == 0;
// cnv8 aka cnv2
cryptonight_hash_ctx[0](asmOptimization, PowVariant::POW_V2, test_input, 76, output, scratchPads);
@ -647,11 +660,6 @@ bool CryptoNight::selfTest(int algo)
cryptonight_hash_ctx[0](asmOptimization, PowVariant::POW_XFH, test_input, 76, output, scratchPads);
result = result && memcmp(output, test_output_xfh, 32) == 0;
#if MAX_NUM_HASH_BLOCKS > 1
cryptonight_hash_ctx[1](asmOptimization, PowVariant::POW_XFH, test_input, 76, output, scratchPads);
result = result && memcmp(output, test_output_xfh, 64) == 0;
#endif
// cnv8 + xtl aka cn-fast2
cryptonight_hash_ctx[0](asmOptimization, PowVariant::POW_FAST_2, test_input, 76, output, scratchPads);

6
src/crypto/CryptoNight_test.h
View file

@ -116,6 +116,12 @@ const static uint8_t test_output_alloy[32] = {
0x5D, 0x77, 0x16, 0x21, 0x42, 0x97, 0x5C, 0xB8, 0x50, 0xC0, 0xA5, 0x1F, 0x64, 0x07, 0xBD, 0x33
};
// CN RTO/HOSP
const static uint8_t test_output_hosp[32] = {
0x82, 0x66, 0x1E, 0x1C, 0x6E, 0x64, 0x36, 0x66, 0x84, 0x06, 0x32, 0x7A, 0x9B, 0xB1, 0x13, 0x19,
0xA5, 0x56, 0x16, 0x15, 0xDF, 0xEC, 0x1C, 0x9E, 0xE3, 0x88, 0x4A, 0x6C, 0x1C, 0xEB, 0x76, 0xA5
};
// CN XFH
const static uint8_t test_output_xfh[64] = {
0x40, 0x86, 0x5A, 0xA8, 0x87, 0x41, 0xEC, 0x1D, 0xCC, 0xBD, 0x2B, 0xC6, 0xFF, 0x36, 0xB9, 0x4D,

50
src/crypto/CryptoNight_x86.h
View file

@ -54,6 +54,7 @@ extern "C"
void cnv1_main_loop_lite_sandybridge_asm(ScratchPad* ctx0);
void cnv1_main_loop_fast_sandybridge_asm(ScratchPad* ctx0);
void cnv1_main_loop_upx_sandybridge_asm(ScratchPad* ctx0);
void cnv1_main_loop_rto_sandybridge_asm(ScratchPad* ctx0);
void cnv2_main_loop_ivybridge_asm(ScratchPad* ctx0);
void cnv2_main_loop_ryzen_asm(ScratchPad* ctx0);
@ -74,6 +75,7 @@ extern "C"
void cnv1_main_loop_lite_soft_aes_sandybridge_asm(ScratchPad* ctx0);
void cnv1_main_loop_fast_soft_aes_sandybridge_asm(ScratchPad* ctx0);
void cnv1_main_loop_upx_soft_aes_sandybridge_asm(ScratchPad* ctx0);
void cnv1_main_loop_rto_soft_aes_sandybridge_asm(ScratchPad* ctx0);
void cnv2_main_loop_soft_aes_sandybridge_asm(ScratchPad* ctx0);
void cnv2_main_loop_fastv2_soft_aes_sandybridge_asm(ScratchPad* ctx0);
@ -769,7 +771,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -887,7 +890,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -1421,7 +1425,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
keccak(static_cast<const uint8_t*>(input), (int) size, scratchPad[0]->state, 200);
@ -1443,29 +1448,37 @@ public:
scratchPad[0]->t_fn = (const uint32_t*)saes_table;
if (ITERATIONS == 0x40000) {
if (MASK == 0x1FFFF0) {
if (powVariant == PowVariant::POW_MSR) {
cnv1_main_loop_fast_soft_aes_sandybridge_asm(scratchPad[0]);
} else {
cnv1_main_loop_lite_soft_aes_sandybridge_asm(scratchPad[0]);
}
} else if (ITERATIONS == 0x20000) {
cnv1_main_loop_upx_soft_aes_sandybridge_asm(scratchPad[0]);
} else {
if (powVariant == PowVariant::POW_HOSP || powVariant == PowVariant::POW_RTO) {
cnv1_main_loop_rto_soft_aes_sandybridge_asm(scratchPad[0]);
} else {
cnv1_main_loop_soft_aes_sandybridge_asm(scratchPad[0]);
}
}
} else {
if (ITERATIONS == 0x40000) {
if (MASK == 0x1FFFF0) {
if (powVariant == PowVariant::POW_MSR) {
cnv1_main_loop_fast_sandybridge_asm(scratchPad[0]);
} else {
cnv1_main_loop_lite_sandybridge_asm(scratchPad[0]);
}
} else if (ITERATIONS == 0x20000) {
cnv1_main_loop_upx_sandybridge_asm(scratchPad[0]);
} else {
if (powVariant == PowVariant::POW_HOSP || powVariant == PowVariant::POW_RTO) {
cnv1_main_loop_rto_sandybridge_asm(scratchPad[0]);
} else {
cnv1_main_loop_sandybridge_asm(scratchPad[0]);
}
}
}
#endif
cn_implode_scratchpad<MEM, SOFT_AES>((__m128i*) l, (__m128i*) h);
@ -1551,7 +1564,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
const uint8_t* l = scratchPad[0]->memory;
uint64_t* h = reinterpret_cast<uint64_t*>(scratchPad[0]->state);
@ -2109,7 +2123,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -2297,7 +2312,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
keccak((const uint8_t*) input, (int) size, scratchPad[0]->state, 200);
keccak((const uint8_t*) input + size, (int) size, scratchPad[1]->state, 200);
@ -3104,7 +3120,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -3281,7 +3298,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -4315,7 +4333,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -4539,7 +4558,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -5196,7 +5216,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}
@ -5464,7 +5485,8 @@ public:
size_t size,
uint8_t* __restrict__ output,
ScratchPad** __restrict__ scratchPad,
AsmOptimization asmOptimization)
AsmOptimization asmOptimization,
PowVariant powVariant)
{
// not supported
}

26
src/crypto/asm/cn_main_loop.S
View file

@ -11,6 +11,7 @@
.global FN_PREFIX(cnv1_main_loop_lite_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_fast_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_upx_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_rto_sandybridge_asm)
.global FN_PREFIX(cnv2_main_loop_ivybridge_asm)
.global FN_PREFIX(cnv2_main_loop_ryzen_asm)
@ -31,6 +32,7 @@
.global FN_PREFIX(cnv1_main_loop_lite_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_fast_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_upx_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_rto_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv2_main_loop_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv2_main_loop_fastv2_soft_aes_sandybridge_asm)
@ -84,6 +86,18 @@ FN_PREFIX(cnv1_main_loop_upx_sandybridge_asm):
add rsp, 48
ret 0
#ifdef __APPLE__
ALIGN 16
#else
ALIGN 64
#endif
FN_PREFIX(cnv1_main_loop_rto_sandybridge_asm):
sub rsp, 48
mov rcx, rdi
#include "cnv1_main_loop_rto_sandybridge.inc"
add rsp, 48
ret 0
#ifdef __APPLE__
ALIGN 16
#else
@ -279,6 +293,18 @@ FN_PREFIX(cnv1_main_loop_upx_soft_aes_sandybridge_asm):
add rsp, 48
ret 0
#ifdef __APPLE__
ALIGN 16
#else
ALIGN 64
#endif
FN_PREFIX(cnv1_main_loop_rto_soft_aes_sandybridge_asm):
sub rsp, 48
mov rcx, rdi
#include "cnv1_main_loop_rto_soft_aes_sandybridge.inc"
add rsp, 48
ret 0
#ifdef __APPLE__
ALIGN 16
#else

75
src/crypto/asm/cnv1_main_loop_rto_sandybridge.inc Normal file
View file

@ -0,0 +1,75 @@
mov QWORD PTR [rsp+8], rbx
mov QWORD PTR [rsp+16], rbp
mov QWORD PTR [rsp+24], rsi
mov QWORD PTR [rsp+32], rdi
push r14
push r15
mov rax, QWORD PTR [rcx+48]
mov ebp, 524288
xor rax, QWORD PTR [rcx+16]
mov rdx, QWORD PTR [rcx+56]
xor rdx, QWORD PTR [rcx+24]
mov r8, QWORD PTR [rcx+32]
xor r8, QWORD PTR [rcx]
movq xmm3, rax
mov rax, QWORD PTR [rcx+256]
mov rdi, QWORD PTR [rcx+40]
movq xmm0, rdx
xor rdi, QWORD PTR [rcx+8]
mov rdx, r8
mov r15, QWORD PTR [rcx+264]
and edx, 2097136
mov r14, QWORD PTR [rax+35]
xor r14, QWORD PTR [rcx+192]
mov rsi, QWORD PTR [rcx+224]
punpcklqdq xmm3, xmm0
movdqu xmm2, XMMWORD PTR [rdx+rsi]
#ifdef __APPLE__
ALIGN 16
#else
ALIGN 64
#endif
cnv1_main_loop_rto_sandybridge:
movq xmm0, rdi
movq xmm1, r8
punpcklqdq xmm1, xmm0
aesenc xmm2, xmm1
movq r10, xmm2
mov r9d, r10d
and r9d, 2097136
add r9, rsi
movdqa xmm0, xmm2
pxor xmm0, xmm3
movdqa xmm3, xmm2
movdqu XMMWORD PTR [rdx+rsi], xmm0
psrldq xmm0, 11
movq rax, xmm0
movzx eax, al
movzx eax, BYTE PTR [rax+r15]
mov BYTE PTR [rsi+rdx+11], al
mov rbx, QWORD PTR [r9]
mov r11, QWORD PTR [r9+8]
mov rax, rbx
mul r10
add r8, rdx
mov QWORD PTR [r9], r8
add rdi, rax
mov rax, r14
xor rax, rdi
xor rax, r8
mov QWORD PTR [r9+8], rax
xor r8, rbx
mov rdx, r8
and edx, 2097136
movdqu xmm2, XMMWORD PTR [rdx+rsi]
xor rdi, r11
dec ebp
jne cnv1_main_loop_rto_sandybridge
mov rbx, QWORD PTR [rsp+24]
mov rbp, QWORD PTR [rsp+32]
mov rsi, QWORD PTR [rsp+40]
mov rdi, QWORD PTR [rsp+48]
pop r15
pop r14

167
src/crypto/asm/cnv1_main_loop_rto_soft_aes_sandybridge.inc Normal file
View file

@ -0,0 +1,167 @@
push rbx
push rbp
push rsi
push rdi
push r12
push r13
push r14
push r15
sub rsp, 72
movaps XMMWORD PTR [rsp], xmm6
movaps XMMWORD PTR [rsp+16], xmm7
movaps XMMWORD PTR [rsp+32], xmm8
movaps XMMWORD PTR [rsp+48], xmm9
mov rax, QWORD PTR [rcx+48]
xor rax, QWORD PTR [rcx+16]
mov rdx, QWORD PTR [rcx+56]
xor rdx, QWORD PTR [rcx+24]
mov r8, QWORD PTR [rcx+32]
xor r8, QWORD PTR [rcx]
movq xmm4, rax
mov rax, QWORD PTR [rcx+256]
mov r13, QWORD PTR [rcx+40]
movq xmm0, rdx
xor r13, QWORD PTR [rcx+8]
mov rdx, r8
mov rdi, QWORD PTR [rcx+224]
and edx, 2097136
mov rax, QWORD PTR [rax+35]
xor rax, QWORD PTR [rcx+192]
movq xmm5, rax
movq xmm8, rdi
punpcklqdq xmm4, xmm0
mov QWORD PTR [rsp+64], rdx
movq xmm6, rcx
mov rax, QWORD PTR [rcx+264]
movq xmm7, rax
mov eax, 524288
#ifdef __APPLE__
ALIGN 16
#else
ALIGN 64
#endif
cnv1_main_loop_rto_soft_aes_sandybridge:
movq xmm9, rax
mov r12, QWORD PTR [rcx+272]
mov esi, DWORD PTR [rdx+rdi]
mov r10d, DWORD PTR [rdx+rdi+4]
mov ebp, DWORD PTR [rdx+rdi+12]
mov r14d, DWORD PTR [rdx+rdi+8]
mov rdx, QWORD PTR [rsp+64]
movzx ecx, sil
shr esi, 8
mov r15d, DWORD PTR [r12+rcx*4]
movzx ecx, r10b
shr r10d, 8
mov edi, DWORD PTR [r12+rcx*4]
movzx ecx, r14b
shr r14d, 8
mov ebx, DWORD PTR [r12+rcx*4]
movzx ecx, bpl
shr ebp, 8
mov r9d, DWORD PTR [r12+rcx*4]
movzx ecx, r10b
shr r10d, 8
xor r15d, DWORD PTR [r12+rcx*4+1024]
movzx ecx, r14b
shr r14d, 8
mov eax, r14d
shr eax, 8
xor edi, DWORD PTR [r12+rcx*4+1024]
add eax, 256
movzx ecx, bpl
shr ebp, 8
xor ebx, DWORD PTR [r12+rcx*4+1024]
movzx ecx, sil
shr esi, 8
xor r9d, DWORD PTR [r12+rcx*4+1024]
add r12, 2048
movzx ecx, r10b
shr r10d, 8
add r10d, 256
mov r11d, DWORD PTR [r12+rax*4]
xor r11d, DWORD PTR [r12+rcx*4]
xor r11d, r9d
movzx ecx, sil
mov r10d, DWORD PTR [r12+r10*4]
shr esi, 8
add esi, 256
xor r10d, DWORD PTR [r12+rcx*4]
movzx ecx, bpl
xor r10d, ebx
shr ebp, 8
add ebp, 256
movd xmm1, r11d
mov r9d, DWORD PTR [r12+rcx*4]
xor r9d, DWORD PTR [r12+rsi*4]
mov eax, DWORD PTR [r12+rbp*4]
xor r9d, edi
movq rdi, xmm8
movzx ecx, r14b
movd xmm0, r10d
movd xmm2, r9d
punpckldq xmm2, xmm1
movq xmm1, r8
xor eax, DWORD PTR [r12+rcx*4]
xor eax, r15d
movd xmm3, eax
movq rax, xmm7
punpckldq xmm3, xmm0
movq xmm0, r13
punpcklqdq xmm1, xmm0
punpckldq xmm3, xmm2
pxor xmm3, xmm1
movq r9, xmm3
mov r10d, r9d
and r10d, 2097136
movdqa xmm0, xmm3
pxor xmm0, xmm4
movdqu XMMWORD PTR [rdx+rdi], xmm0
psrldq xmm0, 11
movq rcx, xmm0
movzx ecx, cl
mov cl, BYTE PTR [rcx+rax]
mov BYTE PTR [rdi+rdx+11], cl
mov rbx, QWORD PTR [r10+rdi]
mov rcx, r9
lea r9, QWORD PTR [r10+rdi]
mov r11, QWORD PTR [r9+8]
mov rax, rbx
movdqa xmm4, xmm3
mul rcx
movq rcx, xmm6
add r8, rdx
add r13, rax
movq rax, xmm5
xor rax, r13
mov QWORD PTR [r9], r8
xor rax, r8
xor r8, rbx
mov QWORD PTR [r9+8], rax
movq rax, xmm9
mov rdx, r8
xor r13, r11
and edx, 2097136
mov QWORD PTR [rsp+64], rdx
sub eax, 1
jne cnv1_main_loop_rto_soft_aes_sandybridge
movaps xmm6, XMMWORD PTR [rsp]
movaps xmm7, XMMWORD PTR [rsp+16]
movaps xmm8, XMMWORD PTR [rsp+32]
movaps xmm9, XMMWORD PTR [rsp+48]
add rsp, 72
pop r15
pop r14
pop r13
pop r12
pop rdi
pop rsi
pop rbp
pop rbx

14
src/crypto/asm/win/cn_main_loop.asm
View file

@ -4,6 +4,7 @@ PUBLIC cnv1_main_loop_sandybridge_asm
PUBLIC cnv1_main_loop_lite_sandybridge_asm
PUBLIC cnv1_main_loop_fast_sandybridge_asm
PUBLIC cnv1_main_loop_upx_sandybridge_asm
PUBLIC cnv1_main_loop_rto_sandybridge_asm
PUBLIC cnv2_main_loop_ivybridge_asm
PUBLIC cnv2_main_loop_ryzen_asm
@ -24,6 +25,7 @@ PUBLIC cnv1_main_loop_soft_aes_sandybridge_asm
PUBLIC cnv1_main_loop_lite_soft_aes_sandybridge_asm
PUBLIC cnv1_main_loop_fast_soft_aes_sandybridge_asm
PUBLIC cnv1_main_loop_upx_soft_aes_sandybridge_asm
PUBLIC cnv1_main_loop_rto_soft_aes_sandybridge_asm
PUBLIC cnv2_main_loop_soft_aes_sandybridge_asm
PUBLIC cnv2_main_loop_fastv2_soft_aes_sandybridge_asm
@ -53,6 +55,12 @@ cnv1_main_loop_upx_sandybridge_asm PROC
ret 0
cnv1_main_loop_upx_sandybridge_asm ENDP
ALIGN 64
cnv1_main_loop_rto_sandybridge_asm PROC
INCLUDE cnv1_main_loop_rto_sandybridge.inc
ret 0
cnv1_main_loop_rto_sandybridge_asm ENDP
ALIGN 64
cnv2_main_loop_ivybridge_asm PROC
INCLUDE cnv2_main_loop_ivybridge.inc
@ -149,6 +157,12 @@ cnv1_main_loop_upx_soft_aes_sandybridge_asm PROC
ret 0
cnv1_main_loop_upx_soft_aes_sandybridge_asm ENDP
ALIGN 64
cnv1_main_loop_rto_soft_aes_sandybridge_asm PROC
INCLUDE cnv1_main_loop_rto_soft_aes_sandybridge.inc
ret 0
cnv1_main_loop_rto_soft_aes_sandybridge_asm ENDP
ALIGN 64
cnv2_main_loop_soft_aes_sandybridge_asm PROC
INCLUDE cnv2_main_loop_soft_aes_sandybridge.inc

12
src/crypto/asm/win/cn_main_loop_win_gcc.S
View file

@ -7,6 +7,7 @@
.global FN_PREFIX(cnv1_main_loop_lite_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_fast_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_upx_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_rto_sandybridge_asm)
.global FN_PREFIX(cnv2_main_loop_ivybridge_asm)
.global FN_PREFIX(cnv2_main_loop_ryzen_asm)
@ -27,6 +28,7 @@
.global FN_PREFIX(cnv1_main_loop_lite_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_fast_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_upx_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv1_main_loop_rto_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv2_main_loop_soft_aes_sandybridge_asm)
.global FN_PREFIX(cnv2_main_loop_fastv2_soft_aes_sandybridge_asm)
@ -52,6 +54,11 @@ FN_PREFIX(cnv1_main_loop_upx_sandybridge_asm):
#include "../cnv1_main_loop_upx_sandybridge.inc"
ret 0
ALIGN 64
FN_PREFIX(cnv1_main_loop_rto_sandybridge_asm):
#include "../cnv1_main_loop_rto_sandybridge.inc"
ret 0
ALIGN 64
FN_PREFIX(cnv2_main_loop_ivybridge_asm):
#include "../cnv2_main_loop_ivybridge.inc"
@ -132,6 +139,11 @@ FN_PREFIX(cnv1_main_loop_upx_soft_aes_sandybridge_asm):
#include "../cnv1_main_loop_upx_soft_aes_sandybridge.inc"
ret 0
ALIGN 64
FN_PREFIX(cnv1_main_loop_rto_soft_aes_sandybridge_asm):
#include "../cnv1_main_loop_rto_soft_aes_sandybridge.inc"
ret 0
ALIGN 64
FN_PREFIX(cnv2_main_loop_soft_aes_sandybridge_asm):
#include "../cnv2_main_loop_soft_aes_sandybridge.inc"

71
src/crypto/asm/win/cnv1_main_loop_rto_sandybridge.inc Normal file
View file

@ -0,0 +1,71 @@
mov QWORD PTR [rsp+8], rbx
mov QWORD PTR [rsp+16], rbp
mov QWORD PTR [rsp+24], rsi
mov QWORD PTR [rsp+32], rdi
push r14
push r15
mov rax, QWORD PTR [rcx+48]
mov ebp, 524288
xor rax, QWORD PTR [rcx+16]
mov rdx, QWORD PTR [rcx+56]
xor rdx, QWORD PTR [rcx+24]
mov r8, QWORD PTR [rcx+32]
xor r8, QWORD PTR [rcx]
movq xmm3, rax
mov rax, QWORD PTR [rcx+256]
mov rdi, QWORD PTR [rcx+40]
movq xmm0, rdx
xor rdi, QWORD PTR [rcx+8]
mov rdx, r8
mov r15, QWORD PTR [rcx+264]
and edx, 2097136
mov r14, QWORD PTR [rax+35]
xor r14, QWORD PTR [rcx+192]
mov rsi, QWORD PTR [rcx+224]
punpcklqdq xmm3, xmm0
movdqu xmm2, XMMWORD PTR [rdx+rsi]
ALIGN 64
cnv1_main_loop_rto_sandybridge:
movq xmm0, rdi
movq xmm1, r8
punpcklqdq xmm1, xmm0
aesenc xmm2, xmm1
movq r10, xmm2
mov r9d, r10d
and r9d, 2097136
add r9, rsi
movdqa xmm0, xmm2
pxor xmm0, xmm3
movdqa xmm3, xmm2
movdqu XMMWORD PTR [rdx+rsi], xmm0
psrldq xmm0, 11
movq rax, xmm0
movzx eax, al
movzx eax, BYTE PTR [rax+r15]
mov BYTE PTR [rsi+rdx+11], al
mov rbx, QWORD PTR [r9]
mov r11, QWORD PTR [r9+8]
mov rax, rbx
mul r10
add r8, rdx
mov QWORD PTR [r9], r8
add rdi, rax
mov rax, r14
xor rax, rdi
xor rax, r8
mov QWORD PTR [r9+8], rax
xor r8, rbx
mov rdx, r8
and edx, 2097136
movdqu xmm2, XMMWORD PTR [rdx+rsi]
xor rdi, r11
dec ebp
jne cnv1_main_loop_rto_sandybridge
mov rbx, QWORD PTR [rsp+24]
mov rbp, QWORD PTR [rsp+32]
mov rsi, QWORD PTR [rsp+40]
mov rdi, QWORD PTR [rsp+48]
pop r15
pop r14

163
src/crypto/asm/win/cnv1_main_loop_rto_soft_aes_sandybridge.inc Normal file
View file

@ -0,0 +1,163 @@
push rbx
push rbp
push rsi
push rdi
push r12
push r13
push r14
push r15
sub rsp, 72
movaps XMMWORD PTR [rsp], xmm6
movaps XMMWORD PTR [rsp+16], xmm7
movaps XMMWORD PTR [rsp+32], xmm8
movaps XMMWORD PTR [rsp+48], xmm9
mov rax, QWORD PTR [rcx+48]
xor rax, QWORD PTR [rcx+16]
mov rdx, QWORD PTR [rcx+56]
xor rdx, QWORD PTR [rcx+24]
mov r8, QWORD PTR [rcx+32]
xor r8, QWORD PTR [rcx]
movq xmm4, rax
mov rax, QWORD PTR [rcx+256]
mov r13, QWORD PTR [rcx+40]
movq xmm0, rdx
xor r13, QWORD PTR [rcx+8]
mov rdx, r8
mov rdi, QWORD PTR [rcx+224]
and edx, 2097136
mov rax, QWORD PTR [rax+35]
xor rax, QWORD PTR [rcx+192]
movq xmm5, rax
movq xmm8, rdi
punpcklqdq xmm4, xmm0
mov QWORD PTR [rsp+64], rdx
movq xmm6, rcx
mov rax, QWORD PTR [rcx+264]
movq xmm7, rax
mov eax, 524288
ALIGN 64
cnv1_main_loop_rto_soft_aes_sandybridge:
movq xmm9, rax
mov r12, QWORD PTR [rcx+272]
mov esi, DWORD PTR [rdx+rdi]
mov r10d, DWORD PTR [rdx+rdi+4]
mov ebp, DWORD PTR [rdx+rdi+12]
mov r14d, DWORD PTR [rdx+rdi+8]
mov rdx, QWORD PTR [rsp+64]
movzx ecx, sil
shr esi, 8
mov r15d, DWORD PTR [r12+rcx*4]
movzx ecx, r10b
shr r10d, 8
mov edi, DWORD PTR [r12+rcx*4]
movzx ecx, r14b
shr r14d, 8
mov ebx, DWORD PTR [r12+rcx*4]
movzx ecx, bpl
shr ebp, 8
mov r9d, DWORD PTR [r12+rcx*4]
movzx ecx, r10b
shr r10d, 8
xor r15d, DWORD PTR [r12+rcx*4+1024]
movzx ecx, r14b
shr r14d, 8
mov eax, r14d
shr eax, 8
xor edi, DWORD PTR [r12+rcx*4+1024]
add eax, 256
movzx ecx, bpl
shr ebp, 8
xor ebx, DWORD PTR [r12+rcx*4+1024]
movzx ecx, sil
shr esi, 8
xor r9d, DWORD PTR [r12+rcx*4+1024]
add r12, 2048
movzx ecx, r10b
shr r10d, 8
add r10d, 256
mov r11d, DWORD PTR [r12+rax*4]
xor r11d, DWORD PTR [r12+rcx*4]
xor r11d, r9d
movzx ecx, sil
mov r10d, DWORD PTR [r12+r10*4]
shr esi, 8
add esi, 256
xor r10d, DWORD PTR [r12+rcx*4]
movzx ecx, bpl
xor r10d, ebx
shr ebp, 8
add ebp, 256
movd xmm1, r11d
mov r9d, DWORD PTR [r12+rcx*4]
xor r9d, DWORD PTR [r12+rsi*4]
mov eax, DWORD PTR [r12+rbp*4]
xor r9d, edi
movq rdi, xmm8
movzx ecx, r14b
movd xmm0, r10d
movd xmm2, r9d
punpckldq xmm2, xmm1
movq xmm1, r8
xor eax, DWORD PTR [r12+rcx*4]
xor eax, r15d
movd xmm3, eax
movq rax, xmm7
punpckldq xmm3, xmm0
movq xmm0, r13
punpcklqdq xmm1, xmm0
punpckldq xmm3, xmm2
pxor xmm3, xmm1
movq r9, xmm3
mov r10d, r9d
and r10d, 2097136
movdqa xmm0, xmm3
pxor xmm0, xmm4
movdqu XMMWORD PTR [rdx+rdi], xmm0
psrldq xmm0, 11
movq rcx, xmm0
movzx ecx, cl
mov cl, BYTE PTR [rcx+rax]
mov BYTE PTR [rdi+rdx+11], cl
mov rbx, QWORD PTR [r10+rdi]
mov rcx, r9
lea r9, QWORD PTR [r10+rdi]
mov r11, QWORD PTR [r9+8]
mov rax, rbx
movdqa xmm4, xmm3
mul rcx
movq rcx, xmm6
add r8, rdx
add r13, rax
movq rax, xmm5
xor rax, r13
mov QWORD PTR [r9], r8
xor rax, r8
xor r8, rbx
mov QWORD PTR [r9+8], rax
movq rax, xmm9
mov rdx, r8
xor r13, r11
and edx, 2097136
mov QWORD PTR [rsp+64], rdx
sub eax, 1
jne cnv1_main_loop_rto_soft_aes_sandybridge
movaps xmm6, XMMWORD PTR [rsp]
movaps xmm7, XMMWORD PTR [rsp+16]
movaps xmm8, XMMWORD PTR [rsp+32]
movaps xmm9, XMMWORD PTR [rsp+48]
add rsp, 72
pop r15
pop r14
pop r13
pop r12
pop rdi
pop rsi
pop rbp
pop rbx

4
src/version.h
View file

@ -36,14 +36,14 @@
#define APP_DESC "XMRigCC CPU miner"
#define APP_COPYRIGHT "Copyright (C) 2017- BenDr0id"
#endif
#define APP_VERSION "1.8.12b (based on XMRig)"
#define APP_VERSION "1.8.13 (based on XMRig)"
#define APP_DOMAIN ""
#define APP_SITE "https://github.com/Bendr0id/xmrigCC"
#define APP_KIND "cpu"
#define APP_VER_MAJOR 1
#define APP_VER_MINOR 8
#define APP_VER_BUILD 12
#define APP_VER_BUILD 13
#define APP_VER_REV 0
#ifndef NDEBUG