1 // blake2_simd.cpp - written and placed in the public domain by
2 // Samuel Neves, Jeffrey Walton, Uri Blumenthal
3 // and Marcel Raad.
4 //
5 // This source file uses intrinsics to gain access to ARMv7a/ARMv8a
6 // NEON, Power7 and SSE4.1 instructions. A separate source file is
7 // needed because additional CXXFLAGS are required to enable the
8 // appropriate instructions sets in some build configurations.
9
10 // The BLAKE2b and BLAKE2s numbers are consistent with the BLAKE2 team's
11 // numbers. However, we have an Altivec implementation of BLAKE2s,
12 // and a POWER8 implementation of BLAKE2b (BLAKE2 team is missing them).
13 // Altivec code is about 2x faster than C++ when using GCC 5.0 or
14 // above. The POWER8 code is about 2.5x faster than C++ when using GCC 5.0
15 // or above. If you use GCC 4.0 (PowerMac) or GCC 4.8 (GCC Compile Farm)
16 // then the PowerPC code will be slower than C++. Be sure to use GCC 5.0
17 // or above for PowerPC builds or disable Altivec for BLAKE2b and BLAKE2s
18 // if using the old compilers.
19
20 #include "pch.h"
21 #include "config.h"
22 #include "misc.h"
23 #include "blake2.h"
24
25 // Uncomment for benchmarking C++ against SSE2 or NEON.
26 // Do so in both blake2.cpp and blake2_simd.cpp.
27 // #undef CRYPTOPP_SSE41_AVAILABLE
28 // #undef CRYPTOPP_ARM_NEON_AVAILABLE
29 // #undef CRYPTOPP_ALTIVEC_AVAILABLE
30
31 // Disable NEON/ASIMD for Cortex-A53 and A57. The shifts are too slow and C/C++ is about
32 // 3 cpb faster than NEON/ASIMD. Also see http://github.com/weidai11/cryptopp/issues/367.
33 #if (defined(__aarch32__) || defined(__aarch64__)) && defined(CRYPTOPP_SLOW_ARMV8_SHIFT)
34 # undef CRYPTOPP_ARM_NEON_AVAILABLE
35 #endif
36
37 // BLAKE2s bug on AIX 7.1 (POWER7) with XLC 12.01
38 // https://github.com/weidai11/cryptopp/issues/743
39 #if defined(__xlC__) && (__xlC__ < 0x0d01)
40 # define CRYPTOPP_DISABLE_ALTIVEC 1
41 # undef CRYPTOPP_POWER7_AVAILABLE
42 # undef CRYPTOPP_ALTIVEC_AVAILABLE
43 #endif
44
45 #if defined(__XOP__)
46 # include <ammintrin.h>
47 # if defined(__GNUC__)
48 # include <x86intrin.h>
49 # endif
50 #endif
51
52 #if (CRYPTOPP_SSE41_AVAILABLE)
53 # include <emmintrin.h>
54 # include <tmmintrin.h>
55 # include <smmintrin.h>
56 #endif
57
58 #if (CRYPTOPP_ARM_NEON_HEADER)
59 # include <arm_neon.h>
60 #endif
61
62 #if (CRYPTOPP_ARM_ACLE_HEADER)
63 # include <stdint.h>
64 # include <arm_acle.h>
65 #endif
66
67 #if (CRYPTOPP_ALTIVEC_AVAILABLE)
68 # include "ppc_simd.h"
69 #endif
70
71 #if defined(CRYPTOPP_GCC_DIAGNOSTIC_AVAILABLE)
72 /* Ignore "warning: vec_lvsl is deprecated..." */
73 # pragma GCC diagnostic ignored "-Wdeprecated"
74 #endif
75
76 // Squash MS LNK4221 and libtool warnings
77 extern const char BLAKE2S_SIMD_FNAME[] = __FILE__;
78
79 NAMESPACE_BEGIN(CryptoPP)
80
81 // Exported by blake2.cpp
82 extern const word32 BLAKE2S_IV[8];
83 extern const word64 BLAKE2B_IV[8];
84
85 #if CRYPTOPP_SSE41_AVAILABLE
86
87 #define LOADU(p) _mm_loadu_si128((const __m128i *)(const void*)(p))
88 #define STOREU(p,r) _mm_storeu_si128((__m128i *)(void*)(p), r)
89 #define TOF(reg) _mm_castsi128_ps((reg))
90 #define TOI(reg) _mm_castps_si128((reg))
91
BLAKE2_Compress32_SSE4(const byte * input,BLAKE2s_State & state)92 void BLAKE2_Compress32_SSE4(const byte* input, BLAKE2s_State& state)
93 {
94 #define BLAKE2S_LOAD_MSG_0_1(buf) \
95 buf = TOI(_mm_shuffle_ps(TOF(m0), TOF(m1), _MM_SHUFFLE(2,0,2,0)));
96
97 #define BLAKE2S_LOAD_MSG_0_2(buf) \
98 buf = TOI(_mm_shuffle_ps(TOF(m0), TOF(m1), _MM_SHUFFLE(3,1,3,1)));
99
100 #define BLAKE2S_LOAD_MSG_0_3(buf) \
101 t0 = _mm_shuffle_epi32(m2, _MM_SHUFFLE(3,2,0,1)); \
102 t1 = _mm_shuffle_epi32(m3, _MM_SHUFFLE(0,1,3,2)); \
103 buf = _mm_blend_epi16(t0, t1, 0xC3);
104
105 #define BLAKE2S_LOAD_MSG_0_4(buf) \
106 t0 = _mm_blend_epi16(t0, t1, 0x3C); \
107 buf = _mm_shuffle_epi32(t0, _MM_SHUFFLE(2,3,0,1));
108
109 #define BLAKE2S_LOAD_MSG_1_1(buf) \
110 t0 = _mm_blend_epi16(m1, m2, 0x0C); \
111 t1 = _mm_slli_si128(m3, 4); \
112 t2 = _mm_blend_epi16(t0, t1, 0xF0); \
113 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,1,0,3));
114
115 #define BLAKE2S_LOAD_MSG_1_2(buf) \
116 t0 = _mm_shuffle_epi32(m2,_MM_SHUFFLE(0,0,2,0)); \
117 t1 = _mm_blend_epi16(m1,m3,0xC0); \
118 t2 = _mm_blend_epi16(t0, t1, 0xF0); \
119 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,3,0,1));
120
121 #define BLAKE2S_LOAD_MSG_1_3(buf) \
122 t0 = _mm_slli_si128(m1, 4); \
123 t1 = _mm_blend_epi16(m2, t0, 0x30); \
124 t2 = _mm_blend_epi16(m0, t1, 0xF0); \
125 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(3,0,1,2));
126
127 #define BLAKE2S_LOAD_MSG_1_4(buf) \
128 t0 = _mm_unpackhi_epi32(m0,m1); \
129 t1 = _mm_slli_si128(m3, 4); \
130 t2 = _mm_blend_epi16(t0, t1, 0x0C); \
131 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(3,0,1,2));
132
133 #define BLAKE2S_LOAD_MSG_2_1(buf) \
134 t0 = _mm_unpackhi_epi32(m2,m3); \
135 t1 = _mm_blend_epi16(m3,m1,0x0C); \
136 t2 = _mm_blend_epi16(t0, t1, 0x0F); \
137 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(3,1,0,2));
138
139 #define BLAKE2S_LOAD_MSG_2_2(buf) \
140 t0 = _mm_unpacklo_epi32(m2,m0); \
141 t1 = _mm_blend_epi16(t0, m0, 0xF0); \
142 t2 = _mm_slli_si128(m3, 8); \
143 buf = _mm_blend_epi16(t1, t2, 0xC0);
144
145 #define BLAKE2S_LOAD_MSG_2_3(buf) \
146 t0 = _mm_blend_epi16(m0, m2, 0x3C); \
147 t1 = _mm_srli_si128(m1, 12); \
148 t2 = _mm_blend_epi16(t0,t1,0x03); \
149 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(0,3,2,1));
150
151 #define BLAKE2S_LOAD_MSG_2_4(buf) \
152 t0 = _mm_slli_si128(m3, 4); \
153 t1 = _mm_blend_epi16(m0, m1, 0x33); \
154 t2 = _mm_blend_epi16(t1, t0, 0xC0); \
155 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(1,2,3,0));
156
157 #define BLAKE2S_LOAD_MSG_3_1(buf) \
158 t0 = _mm_unpackhi_epi32(m0,m1); \
159 t1 = _mm_unpackhi_epi32(t0, m2); \
160 t2 = _mm_blend_epi16(t1, m3, 0x0C); \
161 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(3,1,0,2));
162
163 #define BLAKE2S_LOAD_MSG_3_2(buf) \
164 t0 = _mm_slli_si128(m2, 8); \
165 t1 = _mm_blend_epi16(m3,m0,0x0C); \
166 t2 = _mm_blend_epi16(t1, t0, 0xC0); \
167 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,0,1,3));
168
169 #define BLAKE2S_LOAD_MSG_3_3(buf) \
170 t0 = _mm_blend_epi16(m0,m1,0x0F); \
171 t1 = _mm_blend_epi16(t0, m3, 0xC0); \
172 buf = _mm_shuffle_epi32(t1, _MM_SHUFFLE(0,1,2,3));
173
174 #define BLAKE2S_LOAD_MSG_3_4(buf) \
175 t0 = _mm_alignr_epi8(m0, m1, 4); \
176 buf = _mm_blend_epi16(t0, m2, 0x33);
177
178 #define BLAKE2S_LOAD_MSG_4_1(buf) \
179 t0 = _mm_unpacklo_epi64(m1,m2); \
180 t1 = _mm_unpackhi_epi64(m0,m2); \
181 t2 = _mm_blend_epi16(t0,t1,0x33); \
182 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,0,1,3));
183
184 #define BLAKE2S_LOAD_MSG_4_2(buf) \
185 t0 = _mm_unpackhi_epi64(m1,m3); \
186 t1 = _mm_unpacklo_epi64(m0,m1); \
187 buf = _mm_blend_epi16(t0,t1,0x33);
188
189 #define BLAKE2S_LOAD_MSG_4_3(buf) \
190 t0 = _mm_unpackhi_epi64(m3,m1); \
191 t1 = _mm_unpackhi_epi64(m2,m0); \
192 t2 = _mm_blend_epi16(t1,t0,0x33); \
193 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,1,0,3));
194
195 #define BLAKE2S_LOAD_MSG_4_4(buf) \
196 t0 = _mm_blend_epi16(m0,m2,0x03); \
197 t1 = _mm_slli_si128(t0, 8); \
198 t2 = _mm_blend_epi16(t1,m3,0x0F); \
199 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,0,3,1));
200
201 #define BLAKE2S_LOAD_MSG_5_1(buf) \
202 t0 = _mm_unpackhi_epi32(m0,m1); \
203 t1 = _mm_unpacklo_epi32(m0,m2); \
204 buf = _mm_unpacklo_epi64(t0,t1);
205
206 #define BLAKE2S_LOAD_MSG_5_2(buf) \
207 t0 = _mm_srli_si128(m2, 4); \
208 t1 = _mm_blend_epi16(m0,m3,0x03); \
209 buf = _mm_blend_epi16(t1,t0,0x3C);
210
211 #define BLAKE2S_LOAD_MSG_5_3(buf) \
212 t0 = _mm_blend_epi16(m1,m0,0x0C); \
213 t1 = _mm_srli_si128(m3, 4); \
214 t2 = _mm_blend_epi16(t0,t1,0x30); \
215 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,3,0,1));
216
217 #define BLAKE2S_LOAD_MSG_5_4(buf) \
218 t0 = _mm_unpacklo_epi64(m2,m1); \
219 t1 = _mm_shuffle_epi32(m3, _MM_SHUFFLE(2,0,1,0)); \
220 t2 = _mm_srli_si128(t0, 4); \
221 buf = _mm_blend_epi16(t1,t2,0x33);
222
223 #define BLAKE2S_LOAD_MSG_6_1(buf) \
224 t0 = _mm_slli_si128(m1, 12); \
225 t1 = _mm_blend_epi16(m0,m3,0x33); \
226 buf = _mm_blend_epi16(t1,t0,0xC0);
227
228 #define BLAKE2S_LOAD_MSG_6_2(buf) \
229 t0 = _mm_blend_epi16(m3,m2,0x30); \
230 t1 = _mm_srli_si128(m1, 4); \
231 t2 = _mm_blend_epi16(t0,t1,0x03); \
232 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,1,3,0));
233
234 #define BLAKE2S_LOAD_MSG_6_3(buf) \
235 t0 = _mm_unpacklo_epi64(m0,m2); \
236 t1 = _mm_srli_si128(m1, 4); \
237 t2 = _mm_blend_epi16(t0,t1,0x0C); \
238 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(3,1,0,2));
239
240 #define BLAKE2S_LOAD_MSG_6_4(buf) \
241 t0 = _mm_unpackhi_epi32(m1,m2); \
242 t1 = _mm_unpackhi_epi64(m0,t0); \
243 buf = _mm_shuffle_epi32(t1, _MM_SHUFFLE(0,1,2,3));
244
245 #define BLAKE2S_LOAD_MSG_7_1(buf) \
246 t0 = _mm_unpackhi_epi32(m0,m1); \
247 t1 = _mm_blend_epi16(t0,m3,0x0F); \
248 buf = _mm_shuffle_epi32(t1,_MM_SHUFFLE(2,0,3,1));
249
250 #define BLAKE2S_LOAD_MSG_7_2(buf) \
251 t0 = _mm_blend_epi16(m2,m3,0x30); \
252 t1 = _mm_srli_si128(m0,4); \
253 t2 = _mm_blend_epi16(t0,t1,0x03); \
254 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(1,0,2,3));
255
256 #define BLAKE2S_LOAD_MSG_7_3(buf) \
257 t0 = _mm_unpackhi_epi64(m0,m3); \
258 t1 = _mm_unpacklo_epi64(m1,m2); \
259 t2 = _mm_blend_epi16(t0,t1,0x3C); \
260 buf = _mm_shuffle_epi32(t2,_MM_SHUFFLE(2,3,1,0));
261
262 #define BLAKE2S_LOAD_MSG_7_4(buf) \
263 t0 = _mm_unpacklo_epi32(m0,m1); \
264 t1 = _mm_unpackhi_epi32(m1,m2); \
265 t2 = _mm_unpacklo_epi64(t0,t1); \
266 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(2,1,0,3));
267
268 #define BLAKE2S_LOAD_MSG_8_1(buf) \
269 t0 = _mm_unpackhi_epi32(m1,m3); \
270 t1 = _mm_unpacklo_epi64(t0,m0); \
271 t2 = _mm_blend_epi16(t1,m2,0xC0); \
272 buf = _mm_shufflehi_epi16(t2,_MM_SHUFFLE(1,0,3,2));
273
274 #define BLAKE2S_LOAD_MSG_8_2(buf) \
275 t0 = _mm_unpackhi_epi32(m0,m3); \
276 t1 = _mm_blend_epi16(m2,t0,0xF0); \
277 buf = _mm_shuffle_epi32(t1,_MM_SHUFFLE(0,2,1,3));
278
279 #define BLAKE2S_LOAD_MSG_8_3(buf) \
280 t0 = _mm_unpacklo_epi64(m0,m3); \
281 t1 = _mm_srli_si128(m2,8); \
282 t2 = _mm_blend_epi16(t0,t1,0x03); \
283 buf = _mm_shuffle_epi32(t2, _MM_SHUFFLE(1,3,2,0));
284
285 #define BLAKE2S_LOAD_MSG_8_4(buf) \
286 t0 = _mm_blend_epi16(m1,m0,0x30); \
287 buf = _mm_shuffle_epi32(t0,_MM_SHUFFLE(0,3,2,1));
288
289 #define BLAKE2S_LOAD_MSG_9_1(buf) \
290 t0 = _mm_blend_epi16(m0,m2,0x03); \
291 t1 = _mm_blend_epi16(m1,m2,0x30); \
292 t2 = _mm_blend_epi16(t1,t0,0x0F); \
293 buf = _mm_shuffle_epi32(t2,_MM_SHUFFLE(1,3,0,2));
294
295 #define BLAKE2S_LOAD_MSG_9_2(buf) \
296 t0 = _mm_slli_si128(m0,4); \
297 t1 = _mm_blend_epi16(m1,t0,0xC0); \
298 buf = _mm_shuffle_epi32(t1,_MM_SHUFFLE(1,2,0,3));
299
300 #define BLAKE2S_LOAD_MSG_9_3(buf) \
301 t0 = _mm_unpackhi_epi32(m0,m3); \
302 t1 = _mm_unpacklo_epi32(m2,m3); \
303 t2 = _mm_unpackhi_epi64(t0,t1); \
304 buf = _mm_shuffle_epi32(t2,_MM_SHUFFLE(0,2,1,3));
305
306 #define BLAKE2S_LOAD_MSG_9_4(buf) \
307 t0 = _mm_blend_epi16(m3,m2,0xC0); \
308 t1 = _mm_unpacklo_epi32(m0,m3); \
309 t2 = _mm_blend_epi16(t0,t1,0x0F); \
310 buf = _mm_shuffle_epi32(t2,_MM_SHUFFLE(1,2,3,0));
311
312 #ifdef __XOP__
313 # define MM_ROTI_EPI32(r, c) \
314 _mm_roti_epi32(r, c)
315 #else
316 # define MM_ROTI_EPI32(r, c) ( \
317 (8==-(c)) ? _mm_shuffle_epi8(r,r8) \
318 : (16==-(c)) ? _mm_shuffle_epi8(r,r16) \
319 : _mm_xor_si128(_mm_srli_epi32((r), -(c)), \
320 _mm_slli_epi32((r), 32-(-(c)))))
321 #endif
322
323 #define BLAKE2S_G1(row1,row2,row3,row4,buf) \
324 row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
325 row4 = _mm_xor_si128( row4, row1 ); \
326 row4 = MM_ROTI_EPI32(row4, -16); \
327 row3 = _mm_add_epi32( row3, row4 ); \
328 row2 = _mm_xor_si128( row2, row3 ); \
329 row2 = MM_ROTI_EPI32(row2, -12);
330
331 #define BLAKE2S_G2(row1,row2,row3,row4,buf) \
332 row1 = _mm_add_epi32( _mm_add_epi32( row1, buf), row2 ); \
333 row4 = _mm_xor_si128( row4, row1 ); \
334 row4 = MM_ROTI_EPI32(row4, -8); \
335 row3 = _mm_add_epi32( row3, row4 ); \
336 row2 = _mm_xor_si128( row2, row3 ); \
337 row2 = MM_ROTI_EPI32(row2, -7);
338
339 #define DIAGONALIZE(row1,row2,row3,row4) \
340 row1 = _mm_shuffle_epi32( row1, _MM_SHUFFLE(2,1,0,3) ); \
341 row4 = _mm_shuffle_epi32( row4, _MM_SHUFFLE(1,0,3,2) ); \
342 row3 = _mm_shuffle_epi32( row3, _MM_SHUFFLE(0,3,2,1) );
343
344 #define UNDIAGONALIZE(row1,row2,row3,row4) \
345 row1 = _mm_shuffle_epi32( row1, _MM_SHUFFLE(0,3,2,1) ); \
346 row4 = _mm_shuffle_epi32( row4, _MM_SHUFFLE(1,0,3,2) ); \
347 row3 = _mm_shuffle_epi32( row3, _MM_SHUFFLE(2,1,0,3) );
348
349 #define BLAKE2S_ROUND(r) \
350 BLAKE2S_LOAD_MSG_ ##r ##_1(buf1); \
351 BLAKE2S_G1(row1,row2,row3,row4,buf1); \
352 BLAKE2S_LOAD_MSG_ ##r ##_2(buf2); \
353 BLAKE2S_G2(row1,row2,row3,row4,buf2); \
354 DIAGONALIZE(row1,row2,row3,row4); \
355 BLAKE2S_LOAD_MSG_ ##r ##_3(buf3); \
356 BLAKE2S_G1(row1,row2,row3,row4,buf3); \
357 BLAKE2S_LOAD_MSG_ ##r ##_4(buf4); \
358 BLAKE2S_G2(row1,row2,row3,row4,buf4); \
359 UNDIAGONALIZE(row1,row2,row3,row4);
360
361 __m128i row1, row2, row3, row4;
362 __m128i buf1, buf2, buf3, buf4;
363 __m128i t0, t1, t2, ff0, ff1;
364
365 const __m128i r8 = _mm_set_epi8(12, 15, 14, 13, 8, 11, 10, 9, 4, 7, 6, 5, 0, 3, 2, 1);
366 const __m128i r16 = _mm_set_epi8(13, 12, 15, 14, 9, 8, 11, 10, 5, 4, 7, 6, 1, 0, 3, 2);
367
368 const __m128i m0 = LOADU(input + 00);
369 const __m128i m1 = LOADU(input + 16);
370 const __m128i m2 = LOADU(input + 32);
371 const __m128i m3 = LOADU(input + 48);
372
373 row1 = ff0 = LOADU(state.h()+0);
374 row2 = ff1 = LOADU(state.h()+4);
375 row3 = LOADU(BLAKE2S_IV+0);
376 row4 = _mm_xor_si128(LOADU(BLAKE2S_IV+4), LOADU(state.t()+0));
377
378 BLAKE2S_ROUND(0);
379 BLAKE2S_ROUND(1);
380 BLAKE2S_ROUND(2);
381 BLAKE2S_ROUND(3);
382 BLAKE2S_ROUND(4);
383 BLAKE2S_ROUND(5);
384 BLAKE2S_ROUND(6);
385 BLAKE2S_ROUND(7);
386 BLAKE2S_ROUND(8);
387 BLAKE2S_ROUND(9);
388
389 STOREU(state.h()+0, _mm_xor_si128(ff0, _mm_xor_si128(row1, row3)));
390 STOREU(state.h()+4, _mm_xor_si128(ff1, _mm_xor_si128(row2, row4)));
391 }
392 #endif // CRYPTOPP_SSE41_AVAILABLE
393
394 #if CRYPTOPP_ARM_NEON_AVAILABLE
BLAKE2_Compress32_NEON(const byte * input,BLAKE2s_State & state)395 void BLAKE2_Compress32_NEON(const byte* input, BLAKE2s_State& state)
396 {
397 #define BLAKE2S_LOAD_MSG_0_1(buf) \
398 do { uint32x2_t t0, t1; \
399 t0 = vzip_u32(vget_low_u32(m0), vget_high_u32(m0)).val[0]; \
400 t1 = vzip_u32(vget_low_u32(m1), vget_high_u32(m1)).val[0]; \
401 buf = vcombine_u32(t0, t1); } while(0)
402
403 #define BLAKE2S_LOAD_MSG_0_2(buf) \
404 do { uint32x2_t t0, t1; \
405 t0 = vzip_u32(vget_low_u32(m0), vget_high_u32(m0)).val[1]; \
406 t1 = vzip_u32(vget_low_u32(m1), vget_high_u32(m1)).val[1]; \
407 buf = vcombine_u32(t0, t1); } while(0)
408
409 #define BLAKE2S_LOAD_MSG_0_3(buf) \
410 do { uint32x2_t t0, t1; \
411 t0 = vzip_u32(vget_low_u32(m2), vget_high_u32(m2)).val[0]; \
412 t1 = vzip_u32(vget_low_u32(m3), vget_high_u32(m3)).val[0]; \
413 buf = vcombine_u32(t0, t1); } while(0)
414
415 #define BLAKE2S_LOAD_MSG_0_4(buf) \
416 do { uint32x2_t t0, t1; \
417 t0 = vzip_u32(vget_low_u32(m2), vget_high_u32(m2)).val[1]; \
418 t1 = vzip_u32(vget_low_u32(m3), vget_high_u32(m3)).val[1]; \
419 buf = vcombine_u32(t0, t1); } while(0)
420
421 #define BLAKE2S_LOAD_MSG_1_1(buf) \
422 do { uint32x2_t t0, t1; \
423 t0 = vzip_u32(vget_high_u32(m3), vget_low_u32(m1)).val[0]; \
424 t1 = vzip_u32(vget_low_u32(m2), vget_low_u32(m3)).val[1]; \
425 buf = vcombine_u32(t0, t1); } while(0)
426
427 #define BLAKE2S_LOAD_MSG_1_2(buf) \
428 do { uint32x2_t t0, t1; \
429 t0 = vzip_u32(vget_high_u32(m2), vget_low_u32(m2)).val[0]; \
430 t1 = vext_u32(vget_high_u32(m3), vget_high_u32(m1), 1); \
431 buf = vcombine_u32(t0, t1); } while(0)
432
433 #define BLAKE2S_LOAD_MSG_1_3(buf) \
434 do { uint32x2_t t0, t1; \
435 t0 = vext_u32(vget_low_u32(m0), vget_low_u32(m0), 1); \
436 t1 = vzip_u32(vget_high_u32(m2), vget_low_u32(m1)).val[1]; \
437 buf = vcombine_u32(t0, t1); } while(0)
438
439 #define BLAKE2S_LOAD_MSG_1_4(buf) \
440 do { uint32x2_t t0, t1; \
441 t0 = vzip_u32(vget_low_u32(m3), vget_high_u32(m0)).val[0]; \
442 t1 = vzip_u32(vget_high_u32(m1), vget_high_u32(m0)).val[1]; \
443 buf = vcombine_u32(t0, t1); } while(0)
444
445 #define BLAKE2S_LOAD_MSG_2_1(buf) \
446 do { uint32x2_t t0, t1; \
447 t0 = vext_u32(vget_high_u32(m2), vget_low_u32(m3), 1); \
448 t1 = vzip_u32(vget_low_u32(m1), vget_high_u32(m3)).val[1]; \
449 buf = vcombine_u32(t0, t1); } while(0)
450
451 #define BLAKE2S_LOAD_MSG_2_2(buf) \
452 do { uint32x2_t t0, t1; \
453 t0 = vzip_u32(vget_low_u32(m2), vget_low_u32(m0)).val[0]; \
454 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m0), vget_low_u32(m3)); \
455 buf = vcombine_u32(t0, t1); } while(0)
456
457 #define BLAKE2S_LOAD_MSG_2_3(buf) \
458 do { uint32x2_t t0, t1; \
459 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m2), vget_high_u32(m0)); \
460 t1 = vzip_u32(vget_high_u32(m1), vget_low_u32(m2)).val[1]; \
461 buf = vcombine_u32(t0, t1); } while(0)
462
463 #define BLAKE2S_LOAD_MSG_2_4(buf) \
464 do { uint32x2_t t0, t1; \
465 t0 = vzip_u32(vget_high_u32(m3), vget_high_u32(m1)).val[0]; \
466 t1 = vext_u32(vget_low_u32(m0), vget_low_u32(m1), 1); \
467 buf = vcombine_u32(t0, t1); } while(0)
468
469 #define BLAKE2S_LOAD_MSG_3_1(buf) \
470 do { uint32x2_t t0, t1; \
471 t0 = vzip_u32(vget_high_u32(m1), vget_high_u32(m0)).val[1]; \
472 t1 = vzip_u32(vget_low_u32(m3), vget_high_u32(m2)).val[1]; \
473 buf = vcombine_u32(t0, t1); } while(0)
474
475 #define BLAKE2S_LOAD_MSG_3_2(buf) \
476 do { uint32x2_t t0, t1; \
477 t0 = vzip_u32(vget_low_u32(m2), vget_low_u32(m0)).val[1]; \
478 t1 = vzip_u32(vget_low_u32(m3), vget_high_u32(m3)).val[0]; \
479 buf = vcombine_u32(t0, t1); } while(0)
480
481 #define BLAKE2S_LOAD_MSG_3_3(buf) \
482 do { uint32x2_t t0, t1; \
483 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m0), vget_low_u32(m1)); \
484 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m1), vget_high_u32(m3)); \
485 buf = vcombine_u32(t0, t1); } while(0)
486
487 #define BLAKE2S_LOAD_MSG_3_4(buf) \
488 do { uint32x2_t t0, t1; \
489 t0 = vzip_u32(vget_high_u32(m1), vget_high_u32(m2)).val[0]; \
490 t1 = vzip_u32(vget_low_u32(m0), vget_low_u32(m2)).val[0]; \
491 buf = vcombine_u32(t0, t1); } while(0)
492
493 #define BLAKE2S_LOAD_MSG_4_1(buf) \
494 do { uint32x2_t t0, t1; \
495 t0 = vzip_u32(vget_low_u32(m2), vget_low_u32(m1)).val[1]; \
496 t1 = vzip_u32((vget_high_u32(m0)), vget_high_u32(m2)).val[0]; \
497 buf = vcombine_u32(t0, t1); } while(0)
498
499 #define BLAKE2S_LOAD_MSG_4_2(buf) \
500 do { uint32x2_t t0, t1; \
501 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m0), vget_high_u32(m1)); \
502 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m1), vget_high_u32(m3)); \
503 buf = vcombine_u32(t0, t1); } while(0)
504
505 #define BLAKE2S_LOAD_MSG_4_3(buf) \
506 do { uint32x2_t t0, t1; \
507 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m3), vget_high_u32(m2)); \
508 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m1), vget_high_u32(m0)); \
509 buf = vcombine_u32(t0, t1); } while(0)
510
511 #define BLAKE2S_LOAD_MSG_4_4(buf) \
512 do { uint32x2_t t0, t1; \
513 t0 = vext_u32(vget_low_u32(m0), vget_low_u32(m3), 1); \
514 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m2), vget_low_u32(m3)); \
515 buf = vcombine_u32(t0, t1); } while(0)
516
517 #define BLAKE2S_LOAD_MSG_5_1(buf) \
518 do { uint32x2_t t0, t1; \
519 t0 = vzip_u32((vget_high_u32(m0)), vget_high_u32(m1)).val[0]; \
520 t1 = vzip_u32(vget_low_u32(m0), vget_low_u32(m2)).val[0]; \
521 buf = vcombine_u32(t0, t1); } while(0)
522
523 #define BLAKE2S_LOAD_MSG_5_2(buf) \
524 do { uint32x2_t t0, t1; \
525 t0 = vzip_u32(vget_low_u32(m3), vget_high_u32(m2)).val[0]; \
526 t1 = vzip_u32(vget_high_u32(m2), vget_high_u32(m0)).val[1]; \
527 buf = vcombine_u32(t0, t1); } while(0)
528
529 #define BLAKE2S_LOAD_MSG_5_3(buf) \
530 do { uint32x2_t t0, t1; \
531 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m1), vget_high_u32(m1)); \
532 t1 = vzip_u32(vget_high_u32(m3), vget_low_u32(m0)).val[1]; \
533 buf = vcombine_u32(t0, t1); } while(0)
534
535 #define BLAKE2S_LOAD_MSG_5_4(buf) \
536 do { uint32x2_t t0, t1; \
537 t0 = vzip_u32(vget_low_u32(m3), vget_low_u32(m1)).val[1]; \
538 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m3), vget_low_u32(m2)); \
539 buf = vcombine_u32(t0, t1); } while(0)
540
541 #define BLAKE2S_LOAD_MSG_6_1(buf) \
542 do { uint32x2_t t0, t1; \
543 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m3), vget_low_u32(m0)); \
544 t1 = vzip_u32(vget_high_u32(m3), vget_low_u32(m1)).val[0]; \
545 buf = vcombine_u32(t0, t1); } while(0)
546
547 #define BLAKE2S_LOAD_MSG_6_2(buf) \
548 do { uint32x2_t t0, t1; \
549 t0 = vzip_u32(vget_low_u32(m1), vget_high_u32(m3)).val[1]; \
550 t1 = vext_u32(vget_low_u32(m3), vget_high_u32(m2), 1); \
551 buf = vcombine_u32(t0, t1); } while(0)
552
553 #define BLAKE2S_LOAD_MSG_6_3(buf) \
554 do { uint32x2_t t0, t1; \
555 t0 = vzip_u32(vget_low_u32(m0), vget_high_u32(m1)).val[0]; \
556 t1 = vext_u32(vget_low_u32(m2), vget_low_u32(m2), 1); \
557 buf = vcombine_u32(t0, t1); } while(0)
558
559 #define BLAKE2S_LOAD_MSG_6_4(buf) \
560 do { uint32x2_t t0, t1; \
561 t0 = vzip_u32(vget_high_u32(m1), vget_high_u32(m0)).val[1]; \
562 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m0), vget_high_u32(m2)); \
563 buf = vcombine_u32(t0, t1); } while(0)
564
565 #define BLAKE2S_LOAD_MSG_7_1(buf) \
566 do { uint32x2_t t0, t1; \
567 t0 = vzip_u32(vget_low_u32(m3), vget_high_u32(m1)).val[1]; \
568 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m3), vget_high_u32(m0)); \
569 buf = vcombine_u32(t0, t1); } while(0)
570
571 #define BLAKE2S_LOAD_MSG_7_2(buf) \
572 do { uint32x2_t t0, t1; \
573 t0 = vext_u32(vget_high_u32(m2), vget_high_u32(m3), 1); \
574 t1 = vzip_u32(vget_low_u32(m0), vget_low_u32(m2)).val[1]; \
575 buf = vcombine_u32(t0, t1); } while(0)
576
577 #define BLAKE2S_LOAD_MSG_7_3(buf) \
578 do { uint32x2_t t0, t1; \
579 t0 = vzip_u32(vget_low_u32(m1), vget_high_u32(m3)).val[1]; \
580 t1 = vzip_u32(vget_low_u32(m2), vget_high_u32(m0)).val[0]; \
581 buf = vcombine_u32(t0, t1); } while(0)
582
583 #define BLAKE2S_LOAD_MSG_7_4(buf) \
584 do { uint32x2_t t0, t1; \
585 t0 = vzip_u32(vget_low_u32(m0), vget_low_u32(m1)).val[0]; \
586 t1 = vzip_u32(vget_high_u32(m1), vget_high_u32(m2)).val[0]; \
587 buf = vcombine_u32(t0, t1); } while(0)
588
589 #define BLAKE2S_LOAD_MSG_8_1(buf) \
590 do { uint32x2_t t0, t1; \
591 t0 = vzip_u32(vget_high_u32(m1), vget_high_u32(m3)).val[0]; \
592 t1 = vext_u32(vget_high_u32(m2), vget_low_u32(m0), 1); \
593 buf = vcombine_u32(t0, t1); } while(0)
594
595 #define BLAKE2S_LOAD_MSG_8_2(buf) \
596 do { uint32x2_t t0, t1; \
597 t0 = vzip_u32(vget_high_u32(m3), vget_low_u32(m2)).val[1]; \
598 t1 = vext_u32(vget_high_u32(m0), vget_low_u32(m2), 1); \
599 buf = vcombine_u32(t0, t1); } while(0)
600
601 #define BLAKE2S_LOAD_MSG_8_3(buf) \
602 do { uint32x2_t t0, t1; \
603 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m3), vget_low_u32(m3)); \
604 t1 = vext_u32(vget_low_u32(m0), vget_high_u32(m2), 1); \
605 buf = vcombine_u32(t0, t1); } while(0)
606
607 #define BLAKE2S_LOAD_MSG_8_4(buf) \
608 do { uint32x2_t t0, t1; \
609 t0 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m0), vget_high_u32(m1)); \
610 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_low_u32(m1), vget_low_u32(m1)); \
611 buf = vcombine_u32(t0, t1); } while(0)
612
613 #define BLAKE2S_LOAD_MSG_9_1(buf) \
614 do { uint32x2_t t0, t1; \
615 t0 = vzip_u32(vget_high_u32(m2), vget_low_u32(m2)).val[0]; \
616 t1 = vzip_u32(vget_high_u32(m1), vget_low_u32(m0)).val[1]; \
617 buf = vcombine_u32(t0, t1); } while(0)
618
619 #define BLAKE2S_LOAD_MSG_9_2(buf) \
620 do { uint32x2_t t0, t1; \
621 t0 = vzip_u32((vget_high_u32(m0)), vget_low_u32(m1)).val[0]; \
622 t1 = vbsl_u32(vcreate_u32(0xFFFFFFFF), vget_high_u32(m1), vget_low_u32(m1)); \
623 buf = vcombine_u32(t0, t1); } while(0)
624
625 #define BLAKE2S_LOAD_MSG_9_3(buf) \
626 do { uint32x2_t t0, t1; \
627 t0 = vzip_u32(vget_high_u32(m3), vget_low_u32(m2)).val[1]; \
628 t1 = vzip_u32((vget_high_u32(m0)), vget_low_u32(m3)).val[1]; \
629 buf = vcombine_u32(t0, t1); } while(0)
630
631 #define BLAKE2S_LOAD_MSG_9_4(buf) \
632 do { uint32x2_t t0, t1; \
633 t0 = vext_u32(vget_high_u32(m2), vget_high_u32(m3), 1); \
634 t1 = vzip_u32(vget_low_u32(m3), vget_low_u32(m0)).val[0]; \
635 buf = vcombine_u32(t0, t1); } while(0)
636
637 #define vrorq_n_u32_16(x) vreinterpretq_u32_u16(vrev32q_u16(vreinterpretq_u16_u32(x)))
638
639 #define vrorq_n_u32_8(x) vsriq_n_u32(vshlq_n_u32((x), 24), (x), 8)
640
641 #define vrorq_n_u32(x, c) vsriq_n_u32(vshlq_n_u32((x), 32-(c)), (x), (c))
642
643 #define BLAKE2S_G1(row1,row2,row3,row4,buf) \
644 do { \
645 row1 = vaddq_u32(vaddq_u32(row1, buf), row2); row4 = veorq_u32(row4, row1); \
646 row4 = vrorq_n_u32_16(row4); row3 = vaddq_u32(row3, row4); \
647 row2 = veorq_u32(row2, row3); row2 = vrorq_n_u32(row2, 12); \
648 } while(0)
649
650 #define BLAKE2S_G2(row1,row2,row3,row4,buf) \
651 do { \
652 row1 = vaddq_u32(vaddq_u32(row1, buf), row2); row4 = veorq_u32(row4, row1); \
653 row4 = vrorq_n_u32_8(row4); row3 = vaddq_u32(row3, row4); \
654 row2 = veorq_u32(row2, row3); row2 = vrorq_n_u32(row2, 7); \
655 } while(0)
656
657 #define BLAKE2S_DIAGONALIZE(row1,row2,row3,row4) \
658 do { \
659 row4 = vextq_u32(row4, row4, 3); row3 = vextq_u32(row3, row3, 2); row2 = vextq_u32(row2, row2, 1); \
660 } while(0)
661
662 #define BLAKE2S_UNDIAGONALIZE(row1,row2,row3,row4) \
663 do { \
664 row4 = vextq_u32(row4, row4, 1); \
665 row3 = vextq_u32(row3, row3, 2); \
666 row2 = vextq_u32(row2, row2, 3); \
667 } while(0)
668
669 #define BLAKE2S_ROUND(r) \
670 do { \
671 uint32x4_t buf1, buf2, buf3, buf4; \
672 BLAKE2S_LOAD_MSG_ ##r ##_1(buf1); \
673 BLAKE2S_G1(row1,row2,row3,row4,buf1); \
674 BLAKE2S_LOAD_MSG_ ##r ##_2(buf2); \
675 BLAKE2S_G2(row1,row2,row3,row4,buf2); \
676 BLAKE2S_DIAGONALIZE(row1,row2,row3,row4); \
677 BLAKE2S_LOAD_MSG_ ##r ##_3(buf3); \
678 BLAKE2S_G1(row1,row2,row3,row4,buf3); \
679 BLAKE2S_LOAD_MSG_ ##r ##_4(buf4); \
680 BLAKE2S_G2(row1,row2,row3,row4,buf4); \
681 BLAKE2S_UNDIAGONALIZE(row1,row2,row3,row4); \
682 } while(0)
683
684 const uint32x4_t m0 = vreinterpretq_u32_u8(vld1q_u8(input + 00));
685 const uint32x4_t m1 = vreinterpretq_u32_u8(vld1q_u8(input + 16));
686 const uint32x4_t m2 = vreinterpretq_u32_u8(vld1q_u8(input + 32));
687 const uint32x4_t m3 = vreinterpretq_u32_u8(vld1q_u8(input + 48));
688
689 uint32x4_t row1, row2, row3, row4;
690
691 const uint32x4_t f0 = row1 = vld1q_u32(state.h()+0);
692 const uint32x4_t f1 = row2 = vld1q_u32(state.h()+4);
693 row3 = vld1q_u32(BLAKE2S_IV+0);
694 row4 = veorq_u32(vld1q_u32(BLAKE2S_IV+4), vld1q_u32(state.t()+0));
695
696 BLAKE2S_ROUND(0);
697 BLAKE2S_ROUND(1);
698 BLAKE2S_ROUND(2);
699 BLAKE2S_ROUND(3);
700 BLAKE2S_ROUND(4);
701 BLAKE2S_ROUND(5);
702 BLAKE2S_ROUND(6);
703 BLAKE2S_ROUND(7);
704 BLAKE2S_ROUND(8);
705 BLAKE2S_ROUND(9);
706
707 vst1q_u32(state.h()+0, veorq_u32(f0, veorq_u32(row1, row3)));
708 vst1q_u32(state.h()+4, veorq_u32(f1, veorq_u32(row2, row4)));
709 }
710 #endif // CRYPTOPP_ARM_NEON_AVAILABLE
711
712 #if (CRYPTOPP_ALTIVEC_AVAILABLE)
713
714 template <class T>
VecLoad32(const T * p)715 inline uint32x4_p VecLoad32(const T* p)
716 {
717 return VecLoad(p);
718 }
719
720 template <class T>
VecLoad32LE(const T * p,const uint8x16_p le_mask)721 inline uint32x4_p VecLoad32LE(const T* p, const uint8x16_p le_mask)
722 {
723 #if defined(CRYPTOPP_BIG_ENDIAN)
724 const uint32x4_p v = VecLoad(p);
725 return VecPermute(v, v, le_mask);
726 #else
727 CRYPTOPP_UNUSED(le_mask);
728 return VecLoad(p);
729 #endif
730 }
731
732 template <class T>
VecStore32(T * p,const uint32x4_p x)733 inline void VecStore32(T* p, const uint32x4_p x)
734 {
735 VecStore(x, p);
736 }
737
738 template <class T>
VecStore32LE(T * p,const uint32x4_p x,const uint8x16_p le_mask)739 inline void VecStore32LE(T* p, const uint32x4_p x, const uint8x16_p le_mask)
740 {
741 #if defined(CRYPTOPP_BIG_ENDIAN)
742 const uint32x4_p v = VecPermute(x, x, le_mask);
743 VecStore(v, p);
744 #else
745 CRYPTOPP_UNUSED(le_mask);
746 VecStore(x, p);
747 #endif
748 }
749
750 template <unsigned int E1, unsigned int E2>
VectorSet32(const uint32x4_p a,const uint32x4_p b)751 inline uint32x4_p VectorSet32(const uint32x4_p a, const uint32x4_p b)
752 {
753 // Re-index. I'd like to use something like Z=Y*4 and then
754 // VecShiftLeftOctet<Z>(b) but it crashes early Red Hat
755 // GCC compilers.
756 enum {X=E1&3, Y=E2&3};
757
758 // Don't care element
759 const unsigned int DC = 31;
760
761 // Element 0 combinations
762 if (X == 0 && Y == 0)
763 {
764 const uint8x16_p mask = {0,1,2,3, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
765 return VecPermute(a, b, mask);
766 }
767 else if (X == 0 && Y == 1)
768 {
769 const uint8x16_p mask = {0,1,2,3, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
770 return VecPermute(a, VecShiftLeftOctet<4>(b), mask);
771 }
772 else if (X == 0 && Y == 2)
773 {
774 const uint8x16_p mask = {0,1,2,3, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
775 return VecPermute(a, VecShiftLeftOctet<8>(b), mask);
776 }
777 else if (X == 0 && Y == 3)
778 {
779 const uint8x16_p mask = {0,1,2,3, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
780 return VecPermute(a, VecShiftLeftOctet<12>(b), mask);
781 }
782
783 // Element 1 combinations
784 else if (X == 1 && Y == 0)
785 {
786 const uint8x16_p mask = {4,5,6,7, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
787 return VecPermute(a, b, mask);
788 }
789 else if (X == 1 && Y == 1)
790 {
791 const uint8x16_p mask = {4,5,6,7, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
792 return VecPermute(a, VecShiftLeftOctet<4>(b), mask);
793 }
794 else if (X == 1 && Y == 2)
795 {
796 const uint8x16_p mask = {4,5,6,7, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
797 return VecPermute(a, VecShiftLeftOctet<8>(b), mask);
798 }
799 else if (X == 1 && Y == 3)
800 {
801 const uint8x16_p mask = {4,5,6,7, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
802 return VecPermute(a, VecShiftLeftOctet<12>(b), mask);
803 }
804
805 // Element 2 combinations
806 else if (X == 2 && Y == 0)
807 {
808 const uint8x16_p mask = {8,9,10,11, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
809 return VecPermute(a, b, mask);
810 }
811 else if (X == 2 && Y == 1)
812 {
813 const uint8x16_p mask = {8,9,10,11, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
814 return VecPermute(a, VecShiftLeftOctet<4>(b), mask);
815 }
816 else if (X == 2 && Y == 2)
817 {
818 const uint8x16_p mask = {8,9,10,11, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
819 return VecPermute(a, VecShiftLeftOctet<8>(b), mask);
820 }
821 else if (X == 2 && Y == 3)
822 {
823 const uint8x16_p mask = {8,9,10,11, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
824 return VecPermute(a, VecShiftLeftOctet<12>(b), mask);
825 }
826
827 // Element 3 combinations
828 else if (X == 3 && Y == 0)
829 {
830 const uint8x16_p mask = {12,13,14,15, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
831 return VecPermute(a, b, mask);
832 }
833 else if (X == 3 && Y == 1)
834 {
835 const uint8x16_p mask = {12,13,14,15, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
836 return VecPermute(a, VecShiftLeftOctet<4>(b), mask);
837 }
838 else if (X == 3 && Y == 2)
839 {
840 const uint8x16_p mask = {12,13,14,15, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
841 return VecPermute(a, VecShiftLeftOctet<8>(b), mask);
842 }
843 else if (X == 3 && Y == 3)
844 {
845 const uint8x16_p mask = {12,13,14,15, 16,17,18,19, DC,DC,DC,DC, DC,DC,DC,DC};
846 return VecPermute(a, VecShiftLeftOctet<12>(b), mask);
847 }
848
849 // Quiet IBM XLC warning
850 return VecXor(a, a);
851 }
852
853 template <unsigned int E1, unsigned int E2, unsigned int E3, unsigned int E4>
VectorSet32(const uint32x4_p a,const uint32x4_p b,const uint32x4_p c,const uint32x4_p d)854 inline uint32x4_p VectorSet32(const uint32x4_p a, const uint32x4_p b,
855 const uint32x4_p c, const uint32x4_p d)
856 {
857 // Re-index
858 enum {W=E1&3, X=E2&3, Y=E3&3, Z=E4&3};
859
860 const uint32x4_p t0 = VectorSet32<W,X>(a, b);
861 const uint32x4_p t1 = VectorSet32<Y,Z>(c, d);
862
863 // PowerPC follows SSE2's implementation, and this is _mm_set_epi32.
864 const uint8x16_p mask = {20,21,22,23, 16,17,18,19, 4,5,6,7, 0,1,2,3};
865 return VecPermute(t0, t1, mask);
866 }
867
868 template<>
VectorSet32(const uint32x4_p a,const uint32x4_p b,const uint32x4_p c,const uint32x4_p d)869 uint32x4_p VectorSet32<2,0,2,0>(const uint32x4_p a, const uint32x4_p b,
870 const uint32x4_p c, const uint32x4_p d)
871 {
872 // a=b, c=d, mask is {2,0, 2,0}
873 CRYPTOPP_UNUSED(b); CRYPTOPP_UNUSED(d);
874 const uint8x16_p mask = {16,17,18,19, 24,25,26,27, 0,1,2,3, 8,9,10,11};
875 return VecPermute(a, c, mask);
876 }
877
878 template<>
VectorSet32(const uint32x4_p a,const uint32x4_p b,const uint32x4_p c,const uint32x4_p d)879 uint32x4_p VectorSet32<3,1,3,1>(const uint32x4_p a, const uint32x4_p b,
880 const uint32x4_p c, const uint32x4_p d)
881 {
882 // a=b, c=d, mask is {3,1, 3,1}
883 CRYPTOPP_UNUSED(b); CRYPTOPP_UNUSED(d);
884 const uint8x16_p mask = {20,21,22,23, 28,29,30,31, 4,5,6,7, 12,13,14,15};
885 return VecPermute(a, c, mask);
886 }
887
BLAKE2_Compress32_ALTIVEC(const byte * input,BLAKE2s_State & state)888 void BLAKE2_Compress32_ALTIVEC(const byte* input, BLAKE2s_State& state)
889 {
890 # define m1 m0
891 # define m2 m0
892 # define m3 m0
893
894 # define m5 m4
895 # define m6 m4
896 # define m7 m4
897
898 # define m9 m8
899 # define m10 m8
900 # define m11 m8
901
902 # define m13 m12
903 # define m14 m12
904 # define m15 m12
905
906 // #define BLAKE2S_LOAD_MSG_0_1(buf) buf = VectorSet32<6,4,2,0>(m6,m4,m2,m0);
907 #define BLAKE2S_LOAD_MSG_0_1(buf) buf = VectorSet32<2,0,2,0>(m6,m4,m2,m0);
908 // #define BLAKE2S_LOAD_MSG_0_2(buf) buf = VectorSet32<7,5,3,1>(m7,m5,m3,m1);
909 #define BLAKE2S_LOAD_MSG_0_2(buf) buf = VectorSet32<3,1,3,1>(m7,m5,m3,m1);
910 // #define BLAKE2S_LOAD_MSG_0_3(buf) buf = VectorSet32<14,12,10,8>(m14,m12,m10,m8);
911 #define BLAKE2S_LOAD_MSG_0_3(buf) buf = VectorSet32<2,0,2,0>(m14,m12,m10,m8);
912 // #define BLAKE2S_LOAD_MSG_0_4(buf) buf = VectorSet32<15,13,11,9>(m15,m13,m11,m9);
913 #define BLAKE2S_LOAD_MSG_0_4(buf) buf = VectorSet32<3,1,3,1>(m15,m13,m11,m9);
914
915 #define BLAKE2S_LOAD_MSG_1_1(buf) buf = VectorSet32<13,9,4,14>(m13,m9,m4,m14);
916 #define BLAKE2S_LOAD_MSG_1_2(buf) buf = VectorSet32<6,15,8,10>(m6,m15,m8,m10)
917 #define BLAKE2S_LOAD_MSG_1_3(buf) buf = VectorSet32<5,11,0,1>(m5,m11,m0,m1)
918 #define BLAKE2S_LOAD_MSG_1_4(buf) buf = VectorSet32<3,7,2,12>(m3,m7,m2,m12)
919
920 #define BLAKE2S_LOAD_MSG_2_1(buf) buf = VectorSet32<15,5,12,11>(m15,m5,m12,m11)
921 #define BLAKE2S_LOAD_MSG_2_2(buf) buf = VectorSet32<13,2,0,8>(m13,m2,m0,m8)
922 #define BLAKE2S_LOAD_MSG_2_3(buf) buf = VectorSet32<9,7,3,10>(m9,m7,m3,m10)
923 #define BLAKE2S_LOAD_MSG_2_4(buf) buf = VectorSet32<4,1,6,14>(m4,m1,m6,m14)
924
925 #define BLAKE2S_LOAD_MSG_3_1(buf) buf = VectorSet32<11,13,3,7>(m11,m13,m3,m7)
926 #define BLAKE2S_LOAD_MSG_3_2(buf) buf = VectorSet32<14,12,1,9>(m14,m12,m1,m9)
927 #define BLAKE2S_LOAD_MSG_3_3(buf) buf = VectorSet32<15,4,5,2>(m15,m4,m5,m2)
928 #define BLAKE2S_LOAD_MSG_3_4(buf) buf = VectorSet32<8,0,10,6>(m8,m0,m10,m6)
929
930 #define BLAKE2S_LOAD_MSG_4_1(buf) buf = VectorSet32<10,2,5,9>(m10,m2,m5,m9)
931 #define BLAKE2S_LOAD_MSG_4_2(buf) buf = VectorSet32<15,4,7,0>(m15,m4,m7,m0)
932 #define BLAKE2S_LOAD_MSG_4_3(buf) buf = VectorSet32<3,6,11,14>(m3,m6,m11,m14)
933 #define BLAKE2S_LOAD_MSG_4_4(buf) buf = VectorSet32<13,8,12,1>(m13,m8,m12,m1)
934
935 #define BLAKE2S_LOAD_MSG_5_1(buf) buf = VectorSet32<8,0,6,2>(m8,m0,m6,m2)
936 #define BLAKE2S_LOAD_MSG_5_2(buf) buf = VectorSet32<3,11,10,12>(m3,m11,m10,m12)
937 #define BLAKE2S_LOAD_MSG_5_3(buf) buf = VectorSet32<1,15,7,4>(m1,m15,m7,m4)
938 #define BLAKE2S_LOAD_MSG_5_4(buf) buf = VectorSet32<9,14,5,13>(m9,m14,m5,m13)
939
940 #define BLAKE2S_LOAD_MSG_6_1(buf) buf = VectorSet32<4,14,1,12>(m4,m14,m1,m12)
941 #define BLAKE2S_LOAD_MSG_6_2(buf) buf = VectorSet32<10,13,15,5>(m10,m13,m15,m5)
942 #define BLAKE2S_LOAD_MSG_6_3(buf) buf = VectorSet32<8,9,6,0>(m8,m9,m6,m0)
943 #define BLAKE2S_LOAD_MSG_6_4(buf) buf = VectorSet32<11,2,3,7>(m11,m2,m3,m7)
944
945 #define BLAKE2S_LOAD_MSG_7_1(buf) buf = VectorSet32<3,12,7,13>(m3,m12,m7,m13)
946 #define BLAKE2S_LOAD_MSG_7_2(buf) buf = VectorSet32<9,1,14,11>(m9,m1,m14,m11)
947 #define BLAKE2S_LOAD_MSG_7_3(buf) buf = VectorSet32<2,8,15,5>(m2,m8,m15,m5)
948 #define BLAKE2S_LOAD_MSG_7_4(buf) buf = VectorSet32<10,6,4,0>(m10,m6,m4,m0)
949
950 #define BLAKE2S_LOAD_MSG_8_1(buf) buf = VectorSet32<0,11,14,6>(m0,m11,m14,m6)
951 #define BLAKE2S_LOAD_MSG_8_2(buf) buf = VectorSet32<8,3,9,15>(m8,m3,m9,m15)
952 #define BLAKE2S_LOAD_MSG_8_3(buf) buf = VectorSet32<10,1,13,12>(m10,m1,m13,m12)
953 #define BLAKE2S_LOAD_MSG_8_4(buf) buf = VectorSet32<5,4,7,2>(m5,m4,m7,m2)
954
955 #define BLAKE2S_LOAD_MSG_9_1(buf) buf = VectorSet32<1,7,8,10>(m1,m7,m8,m10)
956 #define BLAKE2S_LOAD_MSG_9_2(buf) buf = VectorSet32<5,6,4,2>(m5,m6,m4,m2)
957 #define BLAKE2S_LOAD_MSG_9_3(buf) buf = VectorSet32<13,3,9,15>(m13,m3,m9,m15)
958 #define BLAKE2S_LOAD_MSG_9_4(buf) buf = VectorSet32<0,12,14,11>(m0,m12,m14,m11)
959
960 #define vec_ror_16(x) VecRotateRight<16>(x)
961 #define vec_ror_12(x) VecRotateRight<12>(x)
962 #define vec_ror_8(x) VecRotateRight<8>(x)
963 #define vec_ror_7(x) VecRotateRight<7>(x)
964
965 #define BLAKE2S_G1(row1,row2,row3,row4,buf) \
966 row1 = VecAdd(VecAdd(row1, buf), row2); \
967 row4 = VecXor(row4, row1); \
968 row4 = vec_ror_16(row4); \
969 row3 = VecAdd(row3, row4); \
970 row2 = VecXor(row2, row3); \
971 row2 = vec_ror_12(row2);
972
973 #define BLAKE2S_G2(row1,row2,row3,row4,buf) \
974 row1 = VecAdd(VecAdd(row1, buf), row2); \
975 row4 = VecXor(row4, row1); \
976 row4 = vec_ror_8(row4); \
977 row3 = VecAdd(row3, row4); \
978 row2 = VecXor(row2, row3); \
979 row2 = vec_ror_7(row2);
980
981 const uint8x16_p D2103_MASK = {12,13,14,15, 0,1,2,3, 4,5,6,7, 8,9,10,11};
982 const uint8x16_p D1032_MASK = {8,9,10,11, 12,13,14,15, 0,1,2,3, 4,5,6,7};
983 const uint8x16_p D0321_MASK = {4,5,6,7, 8,9,10,11, 12,13,14,15, 0,1,2,3};
984
985 #define BLAKE2S_DIAGONALIZE(row1,row2,row3,row4) \
986 row4 = VecPermute(row4, row4, D2103_MASK); \
987 row3 = VecPermute(row3, row3, D1032_MASK); \
988 row2 = VecPermute(row2, row2, D0321_MASK);
989
990 #define BLAKE2S_UNDIAGONALIZE(row1,row2,row3,row4) \
991 row4 = VecPermute(row4, row4, D0321_MASK); \
992 row3 = VecPermute(row3, row3, D1032_MASK); \
993 row2 = VecPermute(row2, row2, D2103_MASK);
994
995 #define BLAKE2S_ROUND(r) \
996 BLAKE2S_LOAD_MSG_ ##r ##_1(buf1); \
997 BLAKE2S_G1(row1,row2,row3,row4,buf1); \
998 BLAKE2S_LOAD_MSG_ ##r ##_2(buf2); \
999 BLAKE2S_G2(row1,row2,row3,row4,buf2); \
1000 BLAKE2S_DIAGONALIZE(row1,row2,row3,row4); \
1001 BLAKE2S_LOAD_MSG_ ##r ##_3(buf3); \
1002 BLAKE2S_G1(row1,row2,row3,row4,buf3); \
1003 BLAKE2S_LOAD_MSG_ ##r ##_4(buf4); \
1004 BLAKE2S_G2(row1,row2,row3,row4,buf4); \
1005 BLAKE2S_UNDIAGONALIZE(row1,row2,row3,row4);
1006
1007 // Possibly unaligned user messages
1008 uint32x4_p m0, m4, m8, m12;
1009 // Endian conversion mask
1010 const uint8x16_p le_mask = {3,2,1,0, 7,6,5,4, 11,10,9,8, 15,14,13,12};
1011
1012 #if defined(_ARCH_PWR9)
1013 // POWER9 provides loads for char's and short's
1014 m0 = (uint32x4_p) vec_xl( 0, CONST_V8_CAST( input ));
1015 m4 = (uint32x4_p) vec_xl( 16, CONST_V8_CAST( input ));
1016 m8 = (uint32x4_p) vec_xl( 32, CONST_V8_CAST( input ));
1017 m12 = (uint32x4_p) vec_xl( 48, CONST_V8_CAST( input ));
1018
1019 # if defined(CRYPTOPP_BIG_ENDIAN)
1020 m0 = vec_perm(m0, m0, le_mask);
1021 m4 = vec_perm(m4, m4, le_mask);
1022 m8 = vec_perm(m8, m8, le_mask);
1023 m12 = vec_perm(m12, m12, le_mask);
1024 # endif
1025 #else
1026 // Altivec only provides 16-byte aligned loads
1027 // http://www.nxp.com/docs/en/reference-manual/ALTIVECPEM.pdf
1028 m0 = (uint32x4_p) vec_ld( 0, CONST_V8_CAST( input ));
1029 m4 = (uint32x4_p) vec_ld( 16, CONST_V8_CAST( input ));
1030 m8 = (uint32x4_p) vec_ld( 32, CONST_V8_CAST( input ));
1031 m12 = (uint32x4_p) vec_ld( 48, CONST_V8_CAST( input ));
1032
1033 // Alignment check for load of the message buffer
1034 const uintptr_t addr = (uintptr_t)input;
1035 if (addr%16 == 0)
1036 {
1037 // Already aligned. Perform a little-endian swap as required
1038 # if defined(CRYPTOPP_BIG_ENDIAN)
1039 m0 = vec_perm(m0, m0, le_mask);
1040 m4 = vec_perm(m4, m4, le_mask);
1041 m8 = vec_perm(m8, m8, le_mask);
1042 m12 = vec_perm(m12, m12, le_mask);
1043 # endif
1044 }
1045 else
1046 {
1047 // Not aligned. Fix vectors and perform a little-endian swap as required
1048 // http://mirror.informatimago.com/next/developer.apple.com/
1049 // hardwaredrivers/ve/code_optimization.html
1050 uint32x4_p ex; uint8x16_p perm;
1051 ex = (uint32x4_p) vec_ld(48+15, CONST_V8_CAST( input ));
1052 perm = vec_lvsl(0, CONST_V8_CAST( addr ));
1053
1054 # if defined(CRYPTOPP_BIG_ENDIAN)
1055 // Combine the vector permute with the little-endian swap
1056 perm = vec_perm(perm, perm, le_mask);
1057 # endif
1058
1059 m0 = vec_perm(m0, m4, perm);
1060 m4 = vec_perm(m4, m8, perm);
1061 m8 = vec_perm(m8, m12, perm);
1062 m12 = vec_perm(m12, ex, perm);
1063 }
1064 #endif
1065
1066 uint32x4_p row1, row2, row3, row4;
1067 uint32x4_p buf1, buf2, buf3, buf4;
1068 uint32x4_p ff0, ff1;
1069
1070 row1 = ff0 = VecLoad32LE(state.h()+0, le_mask);
1071 row2 = ff1 = VecLoad32LE(state.h()+4, le_mask);
1072 row3 = VecLoad32(BLAKE2S_IV+0);
1073 row4 = VecXor(VecLoad32(BLAKE2S_IV+4), VecLoad32(state.t()+0));
1074
1075 BLAKE2S_ROUND(0);
1076 BLAKE2S_ROUND(1);
1077 BLAKE2S_ROUND(2);
1078 BLAKE2S_ROUND(3);
1079 BLAKE2S_ROUND(4);
1080 BLAKE2S_ROUND(5);
1081 BLAKE2S_ROUND(6);
1082 BLAKE2S_ROUND(7);
1083 BLAKE2S_ROUND(8);
1084 BLAKE2S_ROUND(9);
1085
1086 VecStore32LE(state.h()+0, VecXor(ff0, VecXor(row1, row3)), le_mask);
1087 VecStore32LE(state.h()+4, VecXor(ff1, VecXor(row2, row4)), le_mask);
1088 }
1089 #endif // CRYPTOPP_ALTIVEC_AVAILABLE
1090
1091 NAMESPACE_END
1092