shadps4-emu/ext-cryptopp
1
0
Fork 0
mirror of https://github.com/shadps4-emu/ext-cryptopp.git synced 2024-11-27 11:50:29 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity
b7ced67892
ext-cryptopp/simon-simd.cpp

878 lines
32 KiB
C++
Raw Normal View History

Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
// simon-simd.cpp - written and placed in the public domain by Jeffrey Walton
//
// This source file uses intrinsics and built-ins to gain access to
// SSSE3, ARM NEON and ARMv8a, and Power7 Altivec instructions. A separate
// source file is needed because additional CXXFLAGS are required to enable
// the appropriate instructions sets in some build configurations.
#include "pch.h"
#include "config.h"
#include "simon.h"
#include "misc.h"
// Uncomment for benchmarking C++ against SSE or NEON.
// Do so in both simon.cpp and simon-simd.cpp.
// #undef CRYPTOPP_SSSE3_AVAILABLE
// #undef CRYPTOPP_ARM_NEON_AVAILABLE
// Disable NEON/ASIMD for Cortex-A53 and A57. The shifts are too slow and C/C++ is 3 cpb
// faster than NEON/ASIMD. Also see http://github.com/weidai11/cryptopp/issues/367.
#if (defined(__aarch32__) || defined(__aarch64__)) && defined(CRYPTOPP_SLOW_ARMV8_SHIFT)
# undef CRYPTOPP_ARM_NEON_AVAILABLE
#endif
#if (CRYPTOPP_ARM_NEON_AVAILABLE)
# include <arm_neon.h>
#endif
#if (CRYPTOPP_SSSE3_AVAILABLE)
# include <tmmintrin.h>
#endif
// Clang __m128i casts, http://bugs.llvm.org/show_bug.cgi?id=20670
#define M128_CAST(x) ((__m128i *)(void *)(x))
#define CONST_M128_CAST(x) ((const __m128i *)(const void *)(x))
ANONYMOUS_NAMESPACE_BEGIN
using CryptoPP::byte;
using CryptoPP::word32;
using CryptoPP::word64;
using CryptoPP::rotlFixed;
using CryptoPP::rotrFixed;
using CryptoPP::BlockTransformation;
// *************************** ARM NEON ************************** //
#if defined(CRYPTOPP_ARM_NEON_AVAILABLE)
#if defined(CRYPTOPP_LITTLE_ENDIAN)
const word32 s_one[] = {0, 0, 0, 1<<24}; // uint32x4_t
#else
const word32 s_one[] = {0, 0, 0, 1}; // uint32x4_t
#endif
template <class W, class T>
inline W UnpackHigh64(const T& a, const T& b)
{
const uint64x1_t x = vget_high_u64((uint64x2_t)a);
const uint64x1_t y = vget_high_u64((uint64x2_t)b);
return (W)vcombine_u64(x, y);
}
template <class W, class T>
inline W UnpackLow64(const T& a, const T& b)
{
const uint64x1_t x = vget_low_u64((uint64x2_t)a);
const uint64x1_t y = vget_low_u64((uint64x2_t)b);
return (W)vcombine_u64(x, y);
}
template <unsigned int R>
inline uint64x2_t RotateLeft64(const uint64x2_t& val)
{
CRYPTOPP_ASSERT(R < 64);
const uint64x2_t a(vshlq_n_u64(val, R));
const uint64x2_t b(vshrq_n_u64(val, 64 - R));
return vorrq_u64(a, b);
}
template <unsigned int R>
inline uint64x2_t RotateRight64(const uint64x2_t& val)
{
CRYPTOPP_ASSERT(R < 64);
const uint64x2_t a(vshlq_n_u64(val, 64 - R));
const uint64x2_t b(vshrq_n_u64(val, R));
return vorrq_u64(a, b);
}
Add fast rotate-by-8 for Aarch32 and Aarch64
2017-11-29 17:33:34 +00:00
#if defined(__aarch32__) || defined(__aarch64__)
// Faster than two Shifts and an Or. Thanks to Louis Wingers and Bryan Weeks.
template <>
inline uint64x2_t RotateLeft64<8>(const uint64x2_t& val)
{
const uint8_t maskb[16] = { 14,13,12,11, 10,9,8,15, 6,5,4,3, 2,1,0,7 };
const uint8x16_t mask = vld1q_u8(maskb);
return vreinterpretq_u64_u8(
vqtbl1q_u8(vreinterpretq_u8_u64(val), mask));
}
// Faster than two Shifts and an Or. Thanks to Louis Wingers and Bryan Weeks.
template <>
inline uint64x2_t RotateRight64<8>(const uint64x2_t& val)
{
const uint8_t maskb[16] = { 8,15,14,13, 12,11,10,9, 0,7,6,5, 4,3,2,1 };
const uint8x16_t mask = vld1q_u8(maskb);
return vreinterpretq_u64_u8(
vqtbl1q_u8(vreinterpretq_u8_u64(val), mask));
}
#endif
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
inline uint64x2_t Shuffle64(const uint64x2_t& val)
{
Cleanup comments and whitespace
2017-11-29 15:15:41 +00:00
#if defined(CRYPTOPP_LITTLE_ENDIAN)
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
return vreinterpretq_u64_u8(
vrev64q_u8(vreinterpretq_u8_u64(val)));
Cleanup comments and whitespace
2017-11-29 15:15:41 +00:00
#else
return val;
#endif
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
}
Cleanup comments and whitespace
2017-11-29 15:15:41 +00:00
inline uint64x2_t SIMON128_f(const uint64x2_t& val)
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
{
Cleanup comments and whitespace
2017-11-29 15:15:41 +00:00
return veorq_u64(RotateLeft64<2>(val),
vandq_u64(RotateLeft64<1>(val), RotateLeft64<8>(val)));
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
}
inline void SIMON128_Enc_Block(uint8x16_t &block0, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Enc_Blocks then SIMON128_AdvancedProcessBlocks_NEON.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
// The zero block below is a "don't care". It is present so we can vectorize.
uint8x16_t block1 = {0};
uint64x2_t x1 = UnpackLow64<uint64x2_t>(block0, block1);
uint64x2_t y1 = UnpackHigh64<uint64x2_t>(block0, block1);
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
for (size_t i = 0; static_cast<int>(i) < (rounds & ~1)-1; i += 2)
{
const uint64x2_t rk1 = vld1q_dup_u64(subkeys+i);
const uint64x2_t rk2 = vld1q_dup_u64(subkeys+i+1);
y1 = veorq_u64(y1, SIMON128_f(x1));
y1 = veorq_u64(y1, rk1);
x1 = veorq_u64(x1, SIMON128_f(y1));
x1 = veorq_u64(x1, rk2);
}
if (rounds & 1)
{
const uint64x2_t rk = vld1q_dup_u64(subkeys+rounds-1);
y1 = veorq_u64(y1, SIMON128_f(x1));
y1 = veorq_u64(y1, rk);
const uint64x2_t t = x1; x1 = y1; y1 = t;
}
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
block0 = UnpackLow64<uint8x16_t>(x1, y1);
// block1 = UnpackHigh64<uint8x16_t>(x1, y1);
}
inline void SIMON128_Enc_6_Blocks(uint8x16_t &block0, uint8x16_t &block1,
uint8x16_t &block2, uint8x16_t &block3, uint8x16_t &block4,
uint8x16_t &block5, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Enc_Blocks then SIMON128_AdvancedProcessBlocks_NEON.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
uint64x2_t x1 = UnpackLow64<uint64x2_t>(block0, block1);
uint64x2_t y1 = UnpackHigh64<uint64x2_t>(block0, block1);
uint64x2_t x2 = UnpackLow64<uint64x2_t>(block2, block3);
uint64x2_t y2 = UnpackHigh64<uint64x2_t>(block2, block3);
uint64x2_t x3 = UnpackLow64<uint64x2_t>(block4, block5);
uint64x2_t y3 = UnpackHigh64<uint64x2_t>(block4, block5);
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
x2 = Shuffle64(x2);
y2 = Shuffle64(y2);
x3 = Shuffle64(x3);
y3 = Shuffle64(y3);
for (size_t i = 0; static_cast<int>(i) < (rounds & ~1) - 1; i += 2)
{
const uint64x2_t rk1 = vld1q_dup_u64(subkeys+i);
const uint64x2_t rk2 = vld1q_dup_u64(subkeys+i+1);
y1 = veorq_u64(y1, SIMON128_f(x1));
y2 = veorq_u64(y2, SIMON128_f(x2));
y3 = veorq_u64(y3, SIMON128_f(x3));
y1 = veorq_u64(y1, rk1);
y2 = veorq_u64(y2, rk1);
y3 = veorq_u64(y3, rk1);
x1 = veorq_u64(x1, SIMON128_f(y1));
x2 = veorq_u64(x2, SIMON128_f(y2));
x3 = veorq_u64(x3, SIMON128_f(y3));
x1 = veorq_u64(x1, rk2);
x2 = veorq_u64(x2, rk2);
x3 = veorq_u64(x3, rk2);
}
if (rounds & 1)
{
const uint64x2_t rk = vld1q_dup_u64(subkeys + rounds - 1);
y1 = veorq_u64(y1, SIMON128_f(x1));
y2 = veorq_u64(y2, SIMON128_f(x2));
y3 = veorq_u64(y3, SIMON128_f(x3));
y1 = veorq_u64(y1, rk);
y2 = veorq_u64(y2, rk);
y3 = veorq_u64(y3, rk);
const uint64x2_t t1 = x1; x1 = y1; y1 = t1;
const uint64x2_t t2 = x2; x2 = y2; y2 = t2;
const uint64x2_t t3 = x3; x3 = y3; y3 = t3;
}
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
x2 = Shuffle64(x2);
y2 = Shuffle64(y2);
x3 = Shuffle64(x3);
y3 = Shuffle64(y3);
block0 = UnpackLow64<uint8x16_t>(x1, y1);
block1 = UnpackHigh64<uint8x16_t>(x1, y1);
block2 = UnpackLow64<uint8x16_t>(x2, y2);
block3 = UnpackHigh64<uint8x16_t>(x2, y2);
block4 = UnpackLow64<uint8x16_t>(x3, y3);
block5 = UnpackHigh64<uint8x16_t>(x3, y3);
}
inline void SIMON128_Dec_Block(uint8x16_t &block0, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Dec_Blocks then SIMON128_AdvancedProcessBlocks_NEON.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
// The zero block below is a "don't care". It is present so we can vectorize.
uint8x16_t block1 = {0};
uint64x2_t x1 = UnpackLow64<uint64x2_t>(block0, block1);
uint64x2_t y1 = UnpackHigh64<uint64x2_t>(block0, block1);
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
if (rounds & 1)
{
const uint64x2_t t = x1; x1 = y1; y1 = t;
const uint64x2_t rk = vld1q_dup_u64(subkeys + rounds - 1);
y1 = veorq_u64(y1, rk);
y1 = veorq_u64(y1, SIMON128_f(x1));
rounds--;
}
for (size_t i = rounds-2; static_cast<int>(i) >= 0; i -= 2)
{
const uint64x2_t rk1 = vld1q_dup_u64(subkeys+i+1);
const uint64x2_t rk2 = vld1q_dup_u64(subkeys+i);
x1 = veorq_u64(x1, SIMON128_f(y1));
x1 = veorq_u64(x1, rk1);
y1 = veorq_u64(y1, SIMON128_f(x1));
y1 = veorq_u64(y1, rk2);
}
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
block0 = UnpackLow64<uint8x16_t>(x1, y1);
// block1 = UnpackHigh64<uint8x16_t>(x1, y1);
}
inline void SIMON128_Dec_6_Blocks(uint8x16_t &block0, uint8x16_t &block1,
uint8x16_t &block2, uint8x16_t &block3, uint8x16_t &block4,
uint8x16_t &block5, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Dec_Blocks then SIMON128_AdvancedProcessBlocks_NEON.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
uint64x2_t x1 = UnpackLow64<uint64x2_t>(block0, block1);
uint64x2_t y1 = UnpackHigh64<uint64x2_t>(block0, block1);
uint64x2_t x2 = UnpackLow64<uint64x2_t>(block2, block3);
uint64x2_t y2 = UnpackHigh64<uint64x2_t>(block2, block3);
uint64x2_t x3 = UnpackLow64<uint64x2_t>(block4, block5);
uint64x2_t y3 = UnpackHigh64<uint64x2_t>(block5, block5);
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
x2 = Shuffle64(x2);
y2 = Shuffle64(y2);
x3 = Shuffle64(x3);
y3 = Shuffle64(y3);
if (rounds & 1)
{
const uint64x2_t t = x1; x1 = y1; y1 = t;
const uint64x2_t rk = vld1q_dup_u64(subkeys + rounds - 1);
y1 = veorq_u64(y1, rk);
y2 = veorq_u64(y2, rk);
y1 = veorq_u64(y1, SIMON128_f(x1));
y2 = veorq_u64(y2, SIMON128_f(x2));
rounds--;
}
for (size_t i = rounds - 2; static_cast<int>(i) >= 0; i -= 2)
{
const uint64x2_t rk1 = vld1q_dup_u64(subkeys + i + 1);
const uint64x2_t rk2 = vld1q_dup_u64(subkeys + i);
x1 = veorq_u64(x1, SIMON128_f(y1));
x2 = veorq_u64(x2, SIMON128_f(y2));
x3 = veorq_u64(x3, SIMON128_f(y3));
x1 = veorq_u64(x1, rk1);
x2 = veorq_u64(x2, rk1);
x3 = veorq_u64(x3, rk1);
y1 = veorq_u64(y1, SIMON128_f(x1));
y2 = veorq_u64(y2, SIMON128_f(x2));
y3 = veorq_u64(y3, SIMON128_f(x3));
y1 = veorq_u64(y1, rk2);
y2 = veorq_u64(y2, rk2);
y3 = veorq_u64(y3, rk2);
}
x1 = Shuffle64(x1);
y1 = Shuffle64(y1);
x2 = Shuffle64(x2);
y2 = Shuffle64(y2);
x3 = Shuffle64(x3);
y3 = Shuffle64(y3);
block0 = UnpackLow64<uint8x16_t>(x1, y1);
block1 = UnpackHigh64<uint8x16_t>(x1, y1);
block2 = UnpackLow64<uint8x16_t>(x2, y2);
block3 = UnpackHigh64<uint8x16_t>(x2, y2);
block4 = UnpackLow64<uint8x16_t>(x3, y3);
block5 = UnpackHigh64<uint8x16_t>(x3, y3);
}
template <typename F1, typename F6>
size_t SIMON128_AdvancedProcessBlocks_NEON(F1 func1, F6 func6,
const word64 *subKeys, size_t rounds, const byte *inBlocks,
const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
CRYPTOPP_ASSERT(subKeys);
CRYPTOPP_ASSERT(inBlocks);
CRYPTOPP_ASSERT(outBlocks);
CRYPTOPP_ASSERT(length >= 16);
const size_t blockSize = 16;
size_t inIncrement = (flags & (BlockTransformation::BT_InBlockIsCounter|BlockTransformation::BT_DontIncrementInOutPointers)) ? 0 : blockSize;
size_t xorIncrement = xorBlocks ? blockSize : 0;
size_t outIncrement = (flags & BlockTransformation::BT_DontIncrementInOutPointers) ? 0 : blockSize;
if (flags & BlockTransformation::BT_ReverseDirection)
{
inBlocks += length - blockSize;
xorBlocks += length - blockSize;
outBlocks += length - blockSize;
inIncrement = 0-inIncrement;
xorIncrement = 0-xorIncrement;
outIncrement = 0-outIncrement;
}
if (flags & BlockTransformation::BT_AllowParallel)
{
while (length >= 6*blockSize)
{
uint8x16_t block0, block1, block2, block3, block4, block5, temp;
block0 = vld1q_u8(inBlocks);
if (flags & BlockTransformation::BT_InBlockIsCounter)
{
uint32x4_t be = vld1q_u32(s_one);
block1 = (uint8x16_t)vaddq_u32(vreinterpretq_u32_u8(block0), be);
block2 = (uint8x16_t)vaddq_u32(vreinterpretq_u32_u8(block1), be);
block3 = (uint8x16_t)vaddq_u32(vreinterpretq_u32_u8(block2), be);
block4 = (uint8x16_t)vaddq_u32(vreinterpretq_u32_u8(block3), be);
block5 = (uint8x16_t)vaddq_u32(vreinterpretq_u32_u8(block4), be);
temp = (uint8x16_t)vaddq_u32(vreinterpretq_u32_u8(block5), be);
vst1q_u8(const_cast<byte*>(inBlocks), temp);
}
else
{
const int inc = static_cast<int>(inIncrement);
block1 = vld1q_u8(inBlocks+1*inc);
block2 = vld1q_u8(inBlocks+2*inc);
block3 = vld1q_u8(inBlocks+3*inc);
block4 = vld1q_u8(inBlocks+4*inc);
block5 = vld1q_u8(inBlocks+5*inc);
inBlocks += 6*inc;
}
if (flags & BlockTransformation::BT_XorInput)
{
const int inc = static_cast<int>(xorIncrement);
block0 = veorq_u8(block0, vld1q_u8(xorBlocks+0*inc));
block1 = veorq_u8(block1, vld1q_u8(xorBlocks+1*inc));
block2 = veorq_u8(block2, vld1q_u8(xorBlocks+2*inc));
block3 = veorq_u8(block3, vld1q_u8(xorBlocks+3*inc));
block4 = veorq_u8(block4, vld1q_u8(xorBlocks+4*inc));
block5 = veorq_u8(block5, vld1q_u8(xorBlocks+5*inc));
xorBlocks += 6*inc;
}
func6(block0, block1, block2, block3, block4, block5, subKeys, rounds);
if (xorBlocks && !(flags & BlockTransformation::BT_XorInput))
{
const int inc = static_cast<int>(xorIncrement);
block0 = veorq_u8(block0, vld1q_u8(xorBlocks+0*inc));
block1 = veorq_u8(block1, vld1q_u8(xorBlocks+1*inc));
block2 = veorq_u8(block2, vld1q_u8(xorBlocks+2*inc));
block3 = veorq_u8(block3, vld1q_u8(xorBlocks+3*inc));
block4 = veorq_u8(block4, vld1q_u8(xorBlocks+4*inc));
block5 = veorq_u8(block5, vld1q_u8(xorBlocks+5*inc));
xorBlocks += 6*inc;
}
const int inc = static_cast<int>(outIncrement);
vst1q_u8(outBlocks+0*inc, block0);
vst1q_u8(outBlocks+1*inc, block1);
vst1q_u8(outBlocks+2*inc, block2);
vst1q_u8(outBlocks+3*inc, block3);
vst1q_u8(outBlocks+4*inc, block4);
vst1q_u8(outBlocks+5*inc, block5);
outBlocks += 6*inc;
length -= 6*blockSize;
}
}
while (length >= blockSize)
{
uint8x16_t block = vld1q_u8(inBlocks);
if (flags & BlockTransformation::BT_XorInput)
block = veorq_u8(block, vld1q_u8(xorBlocks));
if (flags & BlockTransformation::BT_InBlockIsCounter)
const_cast<byte *>(inBlocks)[15]++;
func1(block, subKeys, rounds);
if (xorBlocks && !(flags & BlockTransformation::BT_XorInput))
block = veorq_u8(block, vld1q_u8(xorBlocks));
vst1q_u8(outBlocks, block);
inBlocks += inIncrement;
outBlocks += outIncrement;
xorBlocks += xorIncrement;
length -= blockSize;
}
return length;
}
#endif // CRYPTOPP_ARM_NEON_AVAILABLE
// ***************************** IA-32 ***************************** //
#if defined(CRYPTOPP_SSSE3_AVAILABLE)
CRYPTOPP_ALIGN_DATA(16)
const word32 s_one[] = {0, 0, 0, 1<<24};
template <unsigned int R>
inline __m128i RotateLeft64(const __m128i& val)
{
CRYPTOPP_ASSERT(R < 64);
Fix compile using SunCC 12.4
2017-11-29 17:10:19 +00:00
const __m128i a = _mm_slli_epi64(val, R);
const __m128i b = _mm_srli_epi64(val, 64-R);
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
return _mm_or_si128(a, b);
}
template <unsigned int R>
inline __m128i RotateRight64(const __m128i& val)
{
CRYPTOPP_ASSERT(R < 64);
Fix compile using SunCC 12.4
2017-11-29 17:10:19 +00:00
const __m128i a = _mm_slli_epi64(val, 64-R);
const __m128i b = _mm_srli_epi64(val, R);
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
return _mm_or_si128(a, b);
}
Add fast rotate-by-8 for Aarch32 and Aarch64
2017-11-29 17:33:34 +00:00
// Faster than two Shifts and an Or. Thanks to Louis Wingers and Bryan Weeks.
Use EPI8 Shuffle rather than Shifts and Or for rotate when R=8 Louis Wingers and Bryan Weeks from the Simon and Speck team offered the suggestion. The change save 0.7 cpb for Speck, and 5 cpb for Simon on x86_64. Speck is now running very close to the Team's time sor SSE4. Simon is still off, but we know the root cause. For Simon, the Team used a fast bit-sliced implementation
2017-11-29 13:53:48 +00:00
template <>
inline __m128i RotateLeft64<8>(const __m128i& val)
{
Cleanup comments and whitespace
2017-11-29 15:15:41 +00:00
const __m128i mask = _mm_set_epi8(14,13,12,11, 10,9,8,15, 6,5,4,3, 2,1,0,7);
Use EPI8 Shuffle rather than Shifts and Or for rotate when R=8 Louis Wingers and Bryan Weeks from the Simon and Speck team offered the suggestion. The change save 0.7 cpb for Speck, and 5 cpb for Simon on x86_64. Speck is now running very close to the Team's time sor SSE4. Simon is still off, but we know the root cause. For Simon, the Team used a fast bit-sliced implementation
2017-11-29 13:53:48 +00:00
return _mm_shuffle_epi8(val, mask);
}
Add fast rotate-by-8 for Aarch32 and Aarch64
2017-11-29 17:33:34 +00:00
// Faster than two Shifts and an Or. Thanks to Louis Wingers and Bryan Weeks.
Use EPI8 Shuffle rather than Shifts and Or for rotate when R=8 Louis Wingers and Bryan Weeks from the Simon and Speck team offered the suggestion. The change save 0.7 cpb for Speck, and 5 cpb for Simon on x86_64. Speck is now running very close to the Team's time sor SSE4. Simon is still off, but we know the root cause. For Simon, the Team used a fast bit-sliced implementation
2017-11-29 13:53:48 +00:00
template <>
inline __m128i RotateRight64<8>(const __m128i& val)
{
Cleanup comments and whitespace
2017-11-29 15:15:41 +00:00
const __m128i mask = _mm_set_epi8(8,15,14,13, 12,11,10,9, 0,7,6,5, 4,3,2,1);
Use EPI8 Shuffle rather than Shifts and Or for rotate when R=8 Louis Wingers and Bryan Weeks from the Simon and Speck team offered the suggestion. The change save 0.7 cpb for Speck, and 5 cpb for Simon on x86_64. Speck is now running very close to the Team's time sor SSE4. Simon is still off, but we know the root cause. For Simon, the Team used a fast bit-sliced implementation
2017-11-29 13:53:48 +00:00
return _mm_shuffle_epi8(val, mask);
}
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
inline __m128i SIMON128_f(const __m128i& v)
{
return _mm_xor_si128(RotateLeft64<2>(v),
_mm_and_si128(RotateLeft64<1>(v), RotateLeft64<8>(v)));
}
inline void SIMON128_Enc_Block(__m128i &block0, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Enc_Blocks then SIMON128_AdvancedProcessBlocks_SSSE3.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
// The zero block below is a "don't care". It is present so we can vectorize.
__m128i block1 = _mm_setzero_si128();
__m128i x1 = _mm_unpacklo_epi64(block0, block1);
__m128i y1 = _mm_unpackhi_epi64(block0, block1);
const __m128i mask = _mm_set_epi8(8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7);
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
for (size_t i = 0; static_cast<int>(i) < (rounds & ~1)-1; i += 2)
{
const __m128i rk1 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys+i)));
const __m128i rk2 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys+i+1)));
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y1 = _mm_xor_si128(y1, rk1);
x1 = _mm_xor_si128(x1, SIMON128_f(y1));
x1 = _mm_xor_si128(x1, rk2);
}
if (rounds & 1)
{
const __m128i rk = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys+rounds-1)));
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y1 = _mm_xor_si128(y1, rk);
const __m128i t = x1; x1 = y1; y1 = t;
}
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
block0 = _mm_unpacklo_epi64(x1, y1);
// block1 = _mm_unpackhi_epi64(x1, y1);
}
inline void SIMON128_Enc_4_Blocks(__m128i &block0, __m128i &block1,
__m128i &block2, __m128i &block3, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Enc_Blocks then SIMON128_AdvancedProcessBlocks_SSSE3.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
__m128i x1 = _mm_unpacklo_epi64(block0, block1);
__m128i y1 = _mm_unpackhi_epi64(block0, block1);
__m128i x2 = _mm_unpacklo_epi64(block2, block3);
__m128i y2 = _mm_unpackhi_epi64(block2, block3);
const __m128i mask = _mm_set_epi8(8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7);
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
x2 = _mm_shuffle_epi8(x2, mask);
y2 = _mm_shuffle_epi8(y2, mask);
for (size_t i = 0; static_cast<int>(i) < (rounds & ~1) - 1; i += 2)
{
const __m128i rk1 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + i)));
const __m128i rk2 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + i + 1)));
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y2 = _mm_xor_si128(y2, SIMON128_f(x2));
y1 = _mm_xor_si128(y1, rk1);
y2 = _mm_xor_si128(y2, rk1);
x1 = _mm_xor_si128(x1, SIMON128_f(y1));
x2 = _mm_xor_si128(x2, SIMON128_f(y2));
x1 = _mm_xor_si128(x1, rk2);
x2 = _mm_xor_si128(x2, rk2);
}
if (rounds & 1)
{
const __m128i rk = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + rounds - 1)));
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y2 = _mm_xor_si128(y2, SIMON128_f(x2));
y1 = _mm_xor_si128(y1, rk);
y2 = _mm_xor_si128(y2, rk);
const __m128i t1 = x1; x1 = y1; y1 = t1;
const __m128i t2 = x2; x2 = y2; y2 = t2;
}
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
x2 = _mm_shuffle_epi8(x2, mask);
y2 = _mm_shuffle_epi8(y2, mask);
block0 = _mm_unpacklo_epi64(x1, y1);
block1 = _mm_unpackhi_epi64(x1, y1);
block2 = _mm_unpacklo_epi64(x2, y2);
block3 = _mm_unpackhi_epi64(x2, y2);
}
inline void SIMON128_Dec_Block(__m128i &block0, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Dec_Blocks then SIMON128_AdvancedProcessBlocks_SSSE3.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
// The zero block below is a "don't care". It is present so we can vectorize.
__m128i block1 = _mm_setzero_si128();
__m128i x1 = _mm_unpacklo_epi64(block0, block1);
__m128i y1 = _mm_unpackhi_epi64(block0, block1);
const __m128i mask = _mm_set_epi8(8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7);
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
if (rounds & 1)
{
const __m128i t = x1; x1 = y1; y1 = t;
const __m128i rk = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + rounds - 1)));
y1 = _mm_xor_si128(y1, rk);
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
rounds--;
}
for (size_t i = rounds-2; static_cast<int>(i) >= 0; i -= 2)
{
const __m128i rk1 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys+i+1)));
const __m128i rk2 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys+i)));
x1 = _mm_xor_si128(x1, SIMON128_f(y1));
x1 = _mm_xor_si128(x1, rk1);
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y1 = _mm_xor_si128(y1, rk2);
}
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
block0 = _mm_unpacklo_epi64(x1, y1);
// block1 = _mm_unpackhi_epi64(x1, y1);
}
inline void SIMON128_Dec_4_Blocks(__m128i &block0, __m128i &block1,
__m128i &block2, __m128i &block3, const word64 *subkeys, unsigned int rounds)
{
// Hack ahead... Rearrange the data for vectorization. It is easier to permute
Update comments
2017-12-01 07:38:19 +00:00
// the data in SIMON128_Dec_Blocks then SIMON128_AdvancedProcessBlocks_SSSE3.
Add NEON and ASIMD intrinsics for SPECK-128 (GH #539) Performance increased by about 200% on a 980 MHz BananaPi dev-board. Throughput went from about 176.6 cpb to about 60.3 cpb.
2017-11-27 05:36:45 +00:00
__m128i x1 = _mm_unpacklo_epi64(block0, block1);
__m128i y1 = _mm_unpackhi_epi64(block0, block1);
__m128i x2 = _mm_unpacklo_epi64(block2, block3);
__m128i y2 = _mm_unpackhi_epi64(block2, block3);
const __m128i mask = _mm_set_epi8(8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7);
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
x2 = _mm_shuffle_epi8(x2, mask);
y2 = _mm_shuffle_epi8(y2, mask);
if (rounds & 1)
{
const __m128i t = x1; x1 = y1; y1 = t;
const __m128i rk = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + rounds - 1)));
y1 = _mm_xor_si128(y1, rk);
y2 = _mm_xor_si128(y2, rk);
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y2 = _mm_xor_si128(y2, SIMON128_f(x2));
rounds--;
}
for (size_t i = rounds - 2; static_cast<int>(i) >= 0; i -= 2)
{
const __m128i rk1 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + i + 1)));
const __m128i rk2 = _mm_castpd_si128(
_mm_loaddup_pd(reinterpret_cast<const double*>(subkeys + i)));
x1 = _mm_xor_si128(x1, SIMON128_f(y1));
x2 = _mm_xor_si128(x2, SIMON128_f(y2));
x1 = _mm_xor_si128(x1, rk1);
x2 = _mm_xor_si128(x2, rk1);
y1 = _mm_xor_si128(y1, SIMON128_f(x1));
y2 = _mm_xor_si128(y2, SIMON128_f(x2));
y1 = _mm_xor_si128(y1, rk2);
y2 = _mm_xor_si128(y2, rk2);
}
x1 = _mm_shuffle_epi8(x1, mask);
y1 = _mm_shuffle_epi8(y1, mask);
x2 = _mm_shuffle_epi8(x2, mask);
y2 = _mm_shuffle_epi8(y2, mask);
block0 = _mm_unpacklo_epi64(x1, y1);
block1 = _mm_unpackhi_epi64(x1, y1);
block2 = _mm_unpacklo_epi64(x2, y2);
block3 = _mm_unpackhi_epi64(x2, y2);
}
template <typename F1, typename F4>
inline size_t SIMON128_AdvancedProcessBlocks_SSSE3(F1 func1, F4 func4,
const word64 *subKeys, size_t rounds, const byte *inBlocks,
const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
CRYPTOPP_ASSERT(subKeys);
CRYPTOPP_ASSERT(inBlocks);
CRYPTOPP_ASSERT(outBlocks);
CRYPTOPP_ASSERT(length >= 16);
const size_t blockSize = 16;
size_t inIncrement = (flags & (BlockTransformation::BT_InBlockIsCounter|BlockTransformation::BT_DontIncrementInOutPointers)) ? 0 : blockSize;
size_t xorIncrement = xorBlocks ? blockSize : 0;
size_t outIncrement = (flags & BlockTransformation::BT_DontIncrementInOutPointers) ? 0 : blockSize;
if (flags & BlockTransformation::BT_ReverseDirection)
{
inBlocks += length - blockSize;
xorBlocks += length - blockSize;
outBlocks += length - blockSize;
inIncrement = 0-inIncrement;
xorIncrement = 0-xorIncrement;
outIncrement = 0-outIncrement;
}
if (flags & BlockTransformation::BT_AllowParallel)
{
while (length >= 4*blockSize)
{
__m128i block0 = _mm_loadu_si128(CONST_M128_CAST(inBlocks)), block1, block2, block3;
if (flags & BlockTransformation::BT_InBlockIsCounter)
{
const __m128i be1 = *CONST_M128_CAST(s_one);
block1 = _mm_add_epi32(block0, be1);
block2 = _mm_add_epi32(block1, be1);
block3 = _mm_add_epi32(block2, be1);
_mm_storeu_si128(M128_CAST(inBlocks), _mm_add_epi32(block3, be1));
}
else
{
inBlocks += inIncrement;
block1 = _mm_loadu_si128(CONST_M128_CAST(inBlocks));
inBlocks += inIncrement;
block2 = _mm_loadu_si128(CONST_M128_CAST(inBlocks));
inBlocks += inIncrement;
block3 = _mm_loadu_si128(CONST_M128_CAST(inBlocks));
inBlocks += inIncrement;
}
if (flags & BlockTransformation::BT_XorInput)
{
// Coverity finding, appears to be false positive. Assert the condition.
CRYPTOPP_ASSERT(xorBlocks);
block0 = _mm_xor_si128(block0, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
block1 = _mm_xor_si128(block1, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
block2 = _mm_xor_si128(block2, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
block3 = _mm_xor_si128(block3, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
}
func4(block0, block1, block2, block3, subKeys, static_cast<unsigned int>(rounds));
if (xorBlocks && !(flags & BlockTransformation::BT_XorInput))
{
block0 = _mm_xor_si128(block0, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
block1 = _mm_xor_si128(block1, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
block2 = _mm_xor_si128(block2, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
block3 = _mm_xor_si128(block3, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
xorBlocks += xorIncrement;
}
_mm_storeu_si128(M128_CAST(outBlocks), block0);
outBlocks += outIncrement;
_mm_storeu_si128(M128_CAST(outBlocks), block1);
outBlocks += outIncrement;
_mm_storeu_si128(M128_CAST(outBlocks), block2);
outBlocks += outIncrement;
_mm_storeu_si128(M128_CAST(outBlocks), block3);
outBlocks += outIncrement;
length -= 4*blockSize;
}
}
while (length >= blockSize)
{
__m128i block = _mm_loadu_si128(CONST_M128_CAST(inBlocks));
if (flags & BlockTransformation::BT_XorInput)
block = _mm_xor_si128(block, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
if (flags & BlockTransformation::BT_InBlockIsCounter)
const_cast<byte *>(inBlocks)[15]++;
func1(block, subKeys, static_cast<unsigned int>(rounds));
if (xorBlocks && !(flags & BlockTransformation::BT_XorInput))
block = _mm_xor_si128(block, _mm_loadu_si128(CONST_M128_CAST(xorBlocks)));
_mm_storeu_si128(M128_CAST(outBlocks), block);
inBlocks += inIncrement;
outBlocks += outIncrement;
xorBlocks += xorIncrement;
length -= blockSize;
}
return length;
}
#endif // CRYPTOPP_SSSE3_AVAILABLE
ANONYMOUS_NAMESPACE_END
///////////////////////////////////////////////////////////////////////
NAMESPACE_BEGIN(CryptoPP)
// *************************** ARM NEON **************************** //
#if (CRYPTOPP_ARM_NEON_AVAILABLE)
size_t SIMON128_Enc_AdvancedProcessBlocks_NEON(const word64* subKeys, size_t rounds,
const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
return SIMON128_AdvancedProcessBlocks_NEON(SIMON128_Enc_Block, SIMON128_Enc_6_Blocks,
subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);
}
size_t SIMON128_Dec_AdvancedProcessBlocks_NEON(const word64* subKeys, size_t rounds,
const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
return SIMON128_AdvancedProcessBlocks_NEON(SIMON128_Dec_Block, SIMON128_Dec_6_Blocks,
subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);
}
#endif // CRYPTOPP_ARM_NEON_AVAILABLE
// ***************************** IA-32 ***************************** //
#if defined(CRYPTOPP_SSSE3_AVAILABLE)
size_t SIMON128_Enc_AdvancedProcessBlocks_SSSE3(const word64* subKeys, size_t rounds,
const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
return SIMON128_AdvancedProcessBlocks_SSSE3(SIMON128_Enc_Block, SIMON128_Enc_4_Blocks,
subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);
}
size_t SIMON128_Dec_AdvancedProcessBlocks_SSSE3(const word64* subKeys, size_t rounds,
const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)
{
return SIMON128_AdvancedProcessBlocks_SSSE3(SIMON128_Dec_Block, SIMON128_Dec_4_Blocks,
subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);
}
#endif // CRYPTOPP_SSSE3_AVAILABLE
NAMESPACE_END
Reference in New Issue Copy Permalink