Add CHAM128 SSSE3 implementation (PR #670)

CHAM-128(128) from 10.5 cpb to 4.1 cpb. CHAM-128(256) from 12.5 cpb to 4.7 cpb.

1 files changed, 402 insertions, 0 deletions

diff --git a/cham-simd.cpp b/cham-simd.cpp
new file mode 100644
index 00000000..718e2361
--- /dev/null
+++ b/cham-simd.cpp
@@ -0,0 +1,402 @@
+// cham-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 "cham.h"

+#include "misc.h"

+#include "adv-simd.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

+

+#if (CRYPTOPP_SSSE3_AVAILABLE)

+# include <pmmintrin.h>

+# include <tmmintrin.h>

+#endif

+

+ANONYMOUS_NAMESPACE_BEGIN

+

+using CryptoPP::word32;

+

+#if (CRYPTOPP_SSSE3_AVAILABLE)

+

+template <unsigned int R>

+inline __m128i RotateLeft32(const __m128i& val)

+{

+    return _mm_or_si128(

+        _mm_slli_epi32(val, R), _mm_srli_epi32(val, 32-R));

+}

+

+template <unsigned int R>

+inline __m128i RotateRight32(const __m128i& val)

+{

+    return _mm_or_si128(

+        _mm_slli_epi32(val, 32-R), _mm_srli_epi32(val, R));

+}

+

+// Faster than two Shifts and an Or. Thanks to Louis Wingers and Bryan Weeks.

+template <>

+inline __m128i RotateLeft32<8>(const __m128i& val)

+{

+    const __m128i mask = _mm_set_epi8(14,13,12,15, 10,9,8,11, 6,5,4,7, 2,1,0,3);

+    return _mm_shuffle_epi8(val, mask);

+}

+

+// Faster than two Shifts and an Or. Thanks to Louis Wingers and Bryan Weeks.

+template <>

+inline __m128i RotateRight32<8>(const __m128i& val)

+{

+    const __m128i mask = _mm_set_epi8(12,15,14,13, 8,11,10,9, 4,7,6,5, 0,3,2,1);

+    return _mm_shuffle_epi8(val, mask);

+}

+

+template <unsigned int IDX>

+inline __m128i UnpackXMM(__m128i a, __m128i b, __m128i c, __m128i d)

+{

+    // Should not be instantiated

+    CRYPTOPP_ASSERT(0);;

+    return _mm_setzero_si128();

+}

+

+template <>

+inline __m128i UnpackXMM<0>(__m128i a, __m128i b, __m128i c, __m128i d)

+{

+    // The shuffle converts to and from little-endian for SSE. A specialized

+    // CHAM implementation can avoid the shuffle by framing the data for

+    // encryption, decryption and benchmarks. The library cannot take the

+    // speed-up because of the byte oriented API.

+    const __m128i r1 = _mm_unpacklo_epi32(a, b);

+    const __m128i r2 = _mm_unpacklo_epi32(c, d);

+    return _mm_shuffle_epi8(_mm_unpacklo_epi64(r1, r2),

+        _mm_set_epi8(12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3));

+}

+

+template <>

+inline __m128i UnpackXMM<1>(__m128i a, __m128i b, __m128i c, __m128i d)

+{

+    // The shuffle converts to and from little-endian for SSE. A specialized

+    // CHAM implementation can avoid the shuffle by framing the data for

+    // encryption, decryption and benchmarks. The library cannot take the

+    // speed-up because of the byte oriented API.

+    const __m128i r1 = _mm_unpacklo_epi32(a, b);

+    const __m128i r2 = _mm_unpacklo_epi32(c, d);

+    return _mm_shuffle_epi8(_mm_unpackhi_epi64(r1, r2),

+        _mm_set_epi8(12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3));

+}

+

+template <>

+inline __m128i UnpackXMM<2>(__m128i a, __m128i b, __m128i c, __m128i d)

+{

+    // The shuffle converts to and from little-endian for SSE. A specialized

+    // CHAM implementation can avoid the shuffle by framing the data for

+    // encryption, decryption and benchmarks. The library cannot take the

+    // speed-up because of the byte oriented API.

+    const __m128i r1 = _mm_unpackhi_epi32(a, b);

+    const __m128i r2 = _mm_unpackhi_epi32(c, d);

+    return _mm_shuffle_epi8(_mm_unpacklo_epi64(r1, r2),

+        _mm_set_epi8(12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3));

+}

+

+template <>

+inline __m128i UnpackXMM<3>(__m128i a, __m128i b, __m128i c, __m128i d)

+{

+    // The shuffle converts to and from little-endian for SSE. A specialized

+    // CHAM implementation can avoid the shuffle by framing the data for

+    // encryption, decryption and benchmarks. The library cannot take the

+    // speed-up because of the byte oriented API.

+    const __m128i r1 = _mm_unpackhi_epi32(a, b);

+    const __m128i r2 = _mm_unpackhi_epi32(c, d);

+    return _mm_shuffle_epi8(_mm_unpackhi_epi64(r1, r2),

+        _mm_set_epi8(12,13,14,15, 8,9,10,11, 4,5,6,7, 0,1,2,3));

+}

+

+template <unsigned int IDX>

+inline __m128i UnpackXMM(__m128i v)

+{

+    return UnpackXMM<IDX>(v, v, v, v);

+}

+

+template <unsigned int IDX>

+inline __m128i RepackXMM(__m128i a, __m128i b, __m128i c, __m128i d)

+{

+    return UnpackXMM<IDX>(a, b, c, d);

+}

+#endif

+

+template <unsigned int IDX>

+inline __m128i RepackXMM(__m128i v)

+{

+    return RepackXMM<IDX>(v, v, v, v);

+}

+

+inline void GCC_NO_UBSAN CHAM128_Enc_Block(__m128i &block0,

+    const word32 *subkeys, unsigned int rounds)

+{

+    // Rearrange the data for vectorization. UnpackXMM includes a

+    // little-endian swap for SSE. Thanks to Peter Cordes for help

+    // with packing and unpacking.

+    // [A1 A2 A3 A4][B1 B2 B3 B4] ... => [A1 B1 C1 D1][A2 B2 C2 D2] ...

+    __m128i a = UnpackXMM<0>(block0);

+    __m128i b = UnpackXMM<1>(block0);

+    __m128i c = UnpackXMM<2>(block0);

+    __m128i d = UnpackXMM<3>(block0);

+

+    __m128i counter = _mm_set_epi32(0,0,0,0);

+    __m128i increment = _mm_set_epi32(1,1,1,1);

+

+    const unsigned int MASK = (rounds == 80 ? 7 : 15);

+    for (int i=0; i<static_cast<int>(rounds); i+=4)

+    {

+        __m128i t1, t2, k, k1, k2;

+

+        k = _mm_loadu_si128((const __m128i*) &subkeys[i & MASK]);

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(3,2,1,0, 3,2,1,0, 3,2,1,0, 3,2,1,0));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(7,6,5,4, 7,6,5,4, 7,6,5,4, 7,6,5,4));

+

+        t1 = _mm_xor_si128(a, counter);

+        t2 = _mm_xor_si128(RotateLeft32<1>(b), k1);

+        a = RotateLeft32<8>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+

+        t1 = _mm_xor_si128(b, counter);

+        t2 = _mm_xor_si128(RotateLeft32<8>(c), k2);

+        b = RotateLeft32<1>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(11,10,9,8, 11,10,9,8, 11,10,9,8, 11,10,9,8));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(15,14,13,12, 15,14,13,12, 15,14,13,12, 15,14,13,12));

+

+        t1 = _mm_xor_si128(c, counter);

+        t2 = _mm_xor_si128(RotateLeft32<1>(d), k1);

+        c = RotateLeft32<8>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+

+        t1 = _mm_xor_si128(d, counter);

+        t2 = _mm_xor_si128(RotateLeft32<8>(a), k2);

+        d = RotateLeft32<1>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+    }

+

+    // Repack

+    // [A1 B1 C1 D1][A2 B2 C2 D2] ... => [A1 A2 A3 A4][B1 B2 B3 B4] ...

+    block0 = RepackXMM<0>(a,b,c,d);

+}

+

+inline void GCC_NO_UBSAN CHAM128_Dec_Block(__m128i &block0,

+    const word32 *subkeys, unsigned int rounds)

+{

+    // Rearrange the data for vectorization. UnpackXMM includes a

+    // little-endian swap for SSE. Thanks to Peter Cordes for help

+    // with packing and unpacking.

+    // [A1 A2 A3 A4][B1 B2 B3 B4] ... => [A1 B1 C1 D1][A2 B2 C2 D2] ...

+    __m128i a = UnpackXMM<0>(block0);

+    __m128i b = UnpackXMM<1>(block0);

+    __m128i c = UnpackXMM<2>(block0);

+    __m128i d = UnpackXMM<3>(block0);

+

+    __m128i counter = _mm_set_epi32(rounds-1,rounds-1,rounds-1,rounds-1);

+    __m128i decrement = _mm_set_epi32(1,1,1,1);

+

+    const unsigned int MASK = (rounds == 80 ? 7 : 15);

+    for (int i = static_cast<int>(rounds)-1; i >= 0; i-=4)

+    {

+        __m128i t1, t2, k, k1, k2;

+

+        k = _mm_loadu_si128((const __m128i*) &subkeys[(i-3) & MASK]);

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(15,14,13,12, 15,14,13,12, 15,14,13,12, 15,14,13,12));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(11,10,9,8, 11,10,9,8, 11,10,9,8, 11,10,9,8));

+

+        // Odd round

+        t1 = RotateRight32<1>(d);

+        t2 = _mm_xor_si128(RotateLeft32<8>(a), k1);

+        d = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+

+        // Even round

+        t1 = RotateRight32<8>(c);

+        t2 = _mm_xor_si128(RotateLeft32<1>(d), k2);

+        c = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(7,6,5,4, 7,6,5,4, 7,6,5,4, 7,6,5,4));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(3,2,1,0, 3,2,1,0, 3,2,1,0, 3,2,1,0));

+

+        // Odd round

+        t1 = RotateRight32<1>(b);

+        t2 = _mm_xor_si128(RotateLeft32<8>(c), k1);

+        b = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+

+        // Even round

+        t1 = RotateRight32<8>(a);

+        t2 = _mm_xor_si128(RotateLeft32<1>(b), k2);

+        a = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+    }

+

+    // Repack

+    // [A1 B1 C1 D1][A2 B2 C2 D2] ... => [A1 A2 A3 A4][B1 B2 B3 B4] ...

+    block0 = RepackXMM<0>(a,b,c,d);

+}

+

+inline void GCC_NO_UBSAN CHAM128_Enc_4_Blocks(__m128i &block0, __m128i &block1,

+    __m128i &block2, __m128i &block3, const word32 *subkeys, unsigned int rounds)

+{

+    // Rearrange the data for vectorization. UnpackXMM includes a

+    // little-endian swap for SSE. Thanks to Peter Cordes for help

+    // with packing and unpacking.

+    // [A1 A2 A3 A4][B1 B2 B3 B4] ... => [A1 B1 C1 D1][A2 B2 C2 D2] ...

+    __m128i a = UnpackXMM<0>(block0, block1, block2, block3);

+    __m128i b = UnpackXMM<1>(block0, block1, block2, block3);

+    __m128i c = UnpackXMM<2>(block0, block1, block2, block3);

+    __m128i d = UnpackXMM<3>(block0, block1, block2, block3);

+

+    __m128i counter = _mm_set_epi32(0,0,0,0);

+    __m128i increment = _mm_set_epi32(1,1,1,1);

+

+    const unsigned int MASK = (rounds == 80 ? 7 : 15);

+    for (int i=0; i<static_cast<int>(rounds); i+=4)

+    {

+        __m128i t1, t2, k, k1, k2;

+

+        k = _mm_loadu_si128((const __m128i*) &subkeys[i & MASK]);

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(3,2,1,0, 3,2,1,0, 3,2,1,0, 3,2,1,0));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(7,6,5,4, 7,6,5,4, 7,6,5,4, 7,6,5,4));

+

+        t1 = _mm_xor_si128(a, counter);

+        t2 = _mm_xor_si128(RotateLeft32<1>(b), k1);

+        a = RotateLeft32<8>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+

+        t1 = _mm_xor_si128(b, counter);

+        t2 = _mm_xor_si128(RotateLeft32<8>(c), k2);

+        b = RotateLeft32<1>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(11,10,9,8, 11,10,9,8, 11,10,9,8, 11,10,9,8));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(15,14,13,12, 15,14,13,12, 15,14,13,12, 15,14,13,12));

+

+        t1 = _mm_xor_si128(c, counter);

+        t2 = _mm_xor_si128(RotateLeft32<1>(d), k1);

+        c = RotateLeft32<8>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+

+        t1 = _mm_xor_si128(d, counter);

+        t2 = _mm_xor_si128(RotateLeft32<8>(a), k2);

+        d = RotateLeft32<1>(_mm_add_epi32(t1, t2));

+

+        counter = _mm_add_epi32(counter, increment);

+    }

+

+    // Repack

+    // [A1 B1 C1 D1][A2 B2 C2 D2] ... => [A1 A2 A3 A4][B1 B2 B3 B4] ...

+    block0 = RepackXMM<0>(a,b,c,d);

+    block1 = RepackXMM<1>(a,b,c,d);

+    block2 = RepackXMM<2>(a,b,c,d);

+    block3 = RepackXMM<3>(a,b,c,d);

+}

+

+inline void GCC_NO_UBSAN CHAM128_Dec_4_Blocks(__m128i &block0, __m128i &block1,

+    __m128i &block2, __m128i &block3, const word32 *subkeys, unsigned int rounds)

+{

+    // Rearrange the data for vectorization. UnpackXMM includes a

+    // little-endian swap for SSE. Thanks to Peter Cordes for help

+    // with packing and unpacking.

+    // [A1 A2 A3 A4][B1 B2 B3 B4] ... => [A1 B1 C1 D1][A2 B2 C2 D2] ...

+    __m128i a = UnpackXMM<0>(block0, block1, block2, block3);

+    __m128i b = UnpackXMM<1>(block0, block1, block2, block3);

+    __m128i c = UnpackXMM<2>(block0, block1, block2, block3);

+    __m128i d = UnpackXMM<3>(block0, block1, block2, block3);

+

+    __m128i counter = _mm_set_epi32(rounds-1,rounds-1,rounds-1,rounds-1);

+    __m128i decrement = _mm_set_epi32(1,1,1,1);

+

+    const unsigned int MASK = (rounds == 80 ? 7 : 15);

+    for (int i = static_cast<int>(rounds)-1; i >= 0; i-=4)

+    {

+        __m128i t1, t2, k, k1, k2;

+

+        k = _mm_loadu_si128((const __m128i*) &subkeys[(i-3) & MASK]);

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(15,14,13,12, 15,14,13,12, 15,14,13,12, 15,14,13,12));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(11,10,9,8, 11,10,9,8, 11,10,9,8, 11,10,9,8));

+

+        // Odd round

+        t1 = RotateRight32<1>(d);

+        t2 = _mm_xor_si128(RotateLeft32<8>(a), k1);

+        d = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+

+        // Even round

+        t1 = RotateRight32<8>(c);

+        t2 = _mm_xor_si128(RotateLeft32<1>(d), k2);

+        c = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+

+        k1 = _mm_shuffle_epi8(k, _mm_set_epi8(7,6,5,4, 7,6,5,4, 7,6,5,4, 7,6,5,4));

+        k2 = _mm_shuffle_epi8(k, _mm_set_epi8(3,2,1,0, 3,2,1,0, 3,2,1,0, 3,2,1,0));

+

+        // Odd round

+        t1 = RotateRight32<1>(b);

+        t2 = _mm_xor_si128(RotateLeft32<8>(c), k1);

+        b = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+

+        // Even round

+        t1 = RotateRight32<8>(a);

+        t2 = _mm_xor_si128(RotateLeft32<1>(b), k2);

+        a = _mm_xor_si128(_mm_sub_epi32(t1, t2), counter);

+

+        counter = _mm_sub_epi32(counter, decrement);

+    }

+

+    // Repack

+    // [A1 B1 C1 D1][A2 B2 C2 D2] ... => [A1 A2 A3 A4][B1 B2 B3 B4] ...

+    block0 = RepackXMM<0>(a,b,c,d);

+    block1 = RepackXMM<1>(a,b,c,d);

+    block2 = RepackXMM<2>(a,b,c,d);

+    block3 = RepackXMM<3>(a,b,c,d);

+}

+

+ANONYMOUS_NAMESPACE_END

+

+NAMESPACE_BEGIN(CryptoPP)

+

+#if defined(CRYPTOPP_SSSE3_AVAILABLE)

+size_t CHAM128_Enc_AdvancedProcessBlocks_SSSE3(const word32* subKeys, size_t rounds,

+    const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)

+{

+    return AdvancedProcessBlocks128_4x1_SSE(CHAM128_Enc_Block, CHAM128_Enc_4_Blocks,

+        subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);

+}

+

+size_t CHAM128_Dec_AdvancedProcessBlocks_SSSE3(const word32* subKeys, size_t rounds,

+    const byte *inBlocks, const byte *xorBlocks, byte *outBlocks, size_t length, word32 flags)

+{

+    return AdvancedProcessBlocks128_4x1_SSE(CHAM128_Dec_Block, CHAM128_Dec_4_Blocks,

+        subKeys, rounds, inBlocks, xorBlocks, outBlocks, length, flags);

+}

+#endif // CRYPTOPP_SSSE3_AVAILABLE

+

+NAMESPACE_END