1 // Based on public domain code written in 2012 by Samuel Neves
2
3 extern const byte blake2s_sigma[10][16];
4
5 // Initialization vector.
6 static __m128i blake2s_IV_0_3, blake2s_IV_4_7;
7
8 #ifdef _WIN_64
9 // Constants for cyclic rotation. Used in 64-bit mode in mm_rotr_epi32 macro.
10 static __m128i crotr8, crotr16;
11 #endif
12
blake2s_init_sse()13 static void blake2s_init_sse()
14 {
15 // We cannot initialize these 128 bit variables in place when declaring
16 // them globally, because global scope initialization is performed before
17 // our SSE check and it would make code incompatible with older non-SSE2
18 // CPUs. Also we cannot initialize them as static inside of function
19 // using these variables, because SSE static initialization is not thread
20 // safe: first thread starts initialization and sets "init done" flag even
21 // if it is not done yet, second thread can attempt to access half-init
22 // SSE data. So we moved init code here.
23
24 blake2s_IV_0_3 = _mm_setr_epi32( 0x6A09E667, 0xBB67AE85, 0x3C6EF372, 0xA54FF53A );
25 blake2s_IV_4_7 = _mm_setr_epi32( 0x510E527F, 0x9B05688C, 0x1F83D9AB, 0x5BE0CD19 );
26
27 #ifdef _WIN_64
28 crotr8 = _mm_set_epi8( 12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1 );
29 crotr16 = _mm_set_epi8( 13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2 );
30 #endif
31 }
32
33
34 #define LOAD(p) _mm_load_si128( (__m128i *)(p) )
35 #define STORE(p,r) _mm_store_si128((__m128i *)(p), r)
36
37 #ifdef _WIN_32
38 // 32-bit mode has less SSE2 registers and in MSVC2008 it is more efficient
39 // to not use _mm_shuffle_epi8 here.
40 #define mm_rotr_epi32(r, c) ( \
41 _mm_xor_si128(_mm_srli_epi32( (r), c ),_mm_slli_epi32( (r), 32-c )) )
42 #else
43 #define mm_rotr_epi32(r, c) ( \
44 c==8 ? _mm_shuffle_epi8(r,crotr8) \
45 : c==16 ? _mm_shuffle_epi8(r,crotr16) \
46 : _mm_xor_si128(_mm_srli_epi32( (r), c ),_mm_slli_epi32( (r), 32-c )) )
47 #endif
48
49
50 #define G1(row1,row2,row3,row4,buf) \
51 row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
52 row4 = _mm_xor_si128( row4, row1 ); \
53 row4 = mm_rotr_epi32(row4, 16); \
54 row3 = _mm_add_epi32( row3, row4 ); \
55 row2 = _mm_xor_si128( row2, row3 ); \
56 row2 = mm_rotr_epi32(row2, 12);
57
58 #define G2(row1,row2,row3,row4,buf) \
59 row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
60 row4 = _mm_xor_si128( row4, row1 ); \
61 row4 = mm_rotr_epi32(row4, 8); \
62 row3 = _mm_add_epi32( row3, row4 ); \
63 row2 = _mm_xor_si128( row2, row3 ); \
64 row2 = mm_rotr_epi32(row2, 7);
65
66 #define DIAGONALIZE(row1,row2,row3,row4) \
67 row4 = _mm_shuffle_epi32( row4, _MM_SHUFFLE(2,1,0,3) ); \
68 row3 = _mm_shuffle_epi32( row3, _MM_SHUFFLE(1,0,3,2) ); \
69 row2 = _mm_shuffle_epi32( row2, _MM_SHUFFLE(0,3,2,1) );
70
71 #define UNDIAGONALIZE(row1,row2,row3,row4) \
72 row4 = _mm_shuffle_epi32( row4, _MM_SHUFFLE(0,3,2,1) ); \
73 row3 = _mm_shuffle_epi32( row3, _MM_SHUFFLE(1,0,3,2) ); \
74 row2 = _mm_shuffle_epi32( row2, _MM_SHUFFLE(2,1,0,3) );
75
76 #ifdef _WIN_64
77 // MSVC 2008 in x64 mode expands _mm_set_epi32 to store to stack and load
78 // from stack operations, which are slower than this code.
79 #define _mm_set_epi32(i3,i2,i1,i0) \
80 _mm_unpacklo_epi32(_mm_unpacklo_epi32(_mm_cvtsi32_si128(i0),_mm_cvtsi32_si128(i2)), \
81 _mm_unpacklo_epi32(_mm_cvtsi32_si128(i1),_mm_cvtsi32_si128(i3)))
82 #endif
83
84 // Original BLAKE2 SSE4.1 message loading code was a little slower in x86 mode
85 // and about the same in x64 mode in our test. Perhaps depends on compiler.
86 // We also tried _mm_i32gather_epi32 and _mm256_i32gather_epi32 AVX2 gather
87 // instructions here, but they did not show any speed gain on i7-6700K.
88 #define SSE_ROUND(m,row,r) \
89 { \
90 __m128i buf; \
91 buf=_mm_set_epi32(m[blake2s_sigma[r][6]],m[blake2s_sigma[r][4]],m[blake2s_sigma[r][2]],m[blake2s_sigma[r][0]]); \
92 G1(row[0],row[1],row[2],row[3],buf); \
93 buf=_mm_set_epi32(m[blake2s_sigma[r][7]],m[blake2s_sigma[r][5]],m[blake2s_sigma[r][3]],m[blake2s_sigma[r][1]]); \
94 G2(row[0],row[1],row[2],row[3],buf); \
95 DIAGONALIZE(row[0],row[1],row[2],row[3]); \
96 buf=_mm_set_epi32(m[blake2s_sigma[r][14]],m[blake2s_sigma[r][12]],m[blake2s_sigma[r][10]],m[blake2s_sigma[r][8]]); \
97 G1(row[0],row[1],row[2],row[3],buf); \
98 buf=_mm_set_epi32(m[blake2s_sigma[r][15]],m[blake2s_sigma[r][13]],m[blake2s_sigma[r][11]],m[blake2s_sigma[r][9]]); \
99 G2(row[0],row[1],row[2],row[3],buf); \
100 UNDIAGONALIZE(row[0],row[1],row[2],row[3]); \
101 }
102
103
blake2s_compress_sse(blake2s_state * S,const byte block[BLAKE2S_BLOCKBYTES])104 static int blake2s_compress_sse( blake2s_state *S, const byte block[BLAKE2S_BLOCKBYTES] )
105 {
106 __m128i row[4];
107 __m128i ff0, ff1;
108
109 const uint32 *m = ( uint32 * )block;
110
111 row[0] = ff0 = LOAD( &S->h[0] );
112 row[1] = ff1 = LOAD( &S->h[4] );
113
114 row[2] = blake2s_IV_0_3;
115 row[3] = _mm_xor_si128( blake2s_IV_4_7, LOAD( &S->t[0] ) );
116 SSE_ROUND( m, row, 0 );
117 SSE_ROUND( m, row, 1 );
118 SSE_ROUND( m, row, 2 );
119 SSE_ROUND( m, row, 3 );
120 SSE_ROUND( m, row, 4 );
121 SSE_ROUND( m, row, 5 );
122 SSE_ROUND( m, row, 6 );
123 SSE_ROUND( m, row, 7 );
124 SSE_ROUND( m, row, 8 );
125 SSE_ROUND( m, row, 9 );
126 STORE( &S->h[0], _mm_xor_si128( ff0, _mm_xor_si128( row[0], row[2] ) ) );
127 STORE( &S->h[4], _mm_xor_si128( ff1, _mm_xor_si128( row[1], row[3] ) ) );
128 return 0;
129 }
130