xref: /freebsd/contrib/llvm-project/llvm/lib/Support/BLAKE3/blake3_neon.c (revision 81ad626541db97eb356e2c1d4a20eb2a26a766ab)
1 #include "blake3_impl.h"
2 
3 #if BLAKE3_USE_NEON
4 
5 #include <arm_neon.h>
6 
7 #ifdef __ARM_BIG_ENDIAN
8 #error "This implementation only supports little-endian ARM."
9 // It might be that all we need for big-endian support here is to get the loads
10 // and stores right, but step zero would be finding a way to test it in CI.
11 #endif
12 
loadu_128(const uint8_t src[16])13 INLINE uint32x4_t loadu_128(const uint8_t src[16]) {
14   // vld1q_u32 has alignment requirements. Don't use it.
15   uint32x4_t x;
16   memcpy(&x, src, 16);
17   return x;
18 }
19 
storeu_128(uint32x4_t src,uint8_t dest[16])20 INLINE void storeu_128(uint32x4_t src, uint8_t dest[16]) {
21   // vst1q_u32 has alignment requirements. Don't use it.
22   memcpy(dest, &src, 16);
23 }
24 
add_128(uint32x4_t a,uint32x4_t b)25 INLINE uint32x4_t add_128(uint32x4_t a, uint32x4_t b) {
26   return vaddq_u32(a, b);
27 }
28 
xor_128(uint32x4_t a,uint32x4_t b)29 INLINE uint32x4_t xor_128(uint32x4_t a, uint32x4_t b) {
30   return veorq_u32(a, b);
31 }
32 
set1_128(uint32_t x)33 INLINE uint32x4_t set1_128(uint32_t x) { return vld1q_dup_u32(&x); }
34 
set4(uint32_t a,uint32_t b,uint32_t c,uint32_t d)35 INLINE uint32x4_t set4(uint32_t a, uint32_t b, uint32_t c, uint32_t d) {
36   uint32_t array[4] = {a, b, c, d};
37   return vld1q_u32(array);
38 }
39 
rot16_128(uint32x4_t x)40 INLINE uint32x4_t rot16_128(uint32x4_t x) {
41   return vorrq_u32(vshrq_n_u32(x, 16), vshlq_n_u32(x, 32 - 16));
42 }
43 
rot12_128(uint32x4_t x)44 INLINE uint32x4_t rot12_128(uint32x4_t x) {
45   return vorrq_u32(vshrq_n_u32(x, 12), vshlq_n_u32(x, 32 - 12));
46 }
47 
rot8_128(uint32x4_t x)48 INLINE uint32x4_t rot8_128(uint32x4_t x) {
49   return vorrq_u32(vshrq_n_u32(x, 8), vshlq_n_u32(x, 32 - 8));
50 }
51 
rot7_128(uint32x4_t x)52 INLINE uint32x4_t rot7_128(uint32x4_t x) {
53   return vorrq_u32(vshrq_n_u32(x, 7), vshlq_n_u32(x, 32 - 7));
54 }
55 
56 // TODO: compress_neon
57 
58 // TODO: hash2_neon
59 
60 /*
61  * ----------------------------------------------------------------------------
62  * hash4_neon
63  * ----------------------------------------------------------------------------
64  */
65 
round_fn4(uint32x4_t v[16],uint32x4_t m[16],size_t r)66 INLINE void round_fn4(uint32x4_t v[16], uint32x4_t m[16], size_t r) {
67   v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][0]]);
68   v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][2]]);
69   v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][4]]);
70   v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][6]]);
71   v[0] = add_128(v[0], v[4]);
72   v[1] = add_128(v[1], v[5]);
73   v[2] = add_128(v[2], v[6]);
74   v[3] = add_128(v[3], v[7]);
75   v[12] = xor_128(v[12], v[0]);
76   v[13] = xor_128(v[13], v[1]);
77   v[14] = xor_128(v[14], v[2]);
78   v[15] = xor_128(v[15], v[3]);
79   v[12] = rot16_128(v[12]);
80   v[13] = rot16_128(v[13]);
81   v[14] = rot16_128(v[14]);
82   v[15] = rot16_128(v[15]);
83   v[8] = add_128(v[8], v[12]);
84   v[9] = add_128(v[9], v[13]);
85   v[10] = add_128(v[10], v[14]);
86   v[11] = add_128(v[11], v[15]);
87   v[4] = xor_128(v[4], v[8]);
88   v[5] = xor_128(v[5], v[9]);
89   v[6] = xor_128(v[6], v[10]);
90   v[7] = xor_128(v[7], v[11]);
91   v[4] = rot12_128(v[4]);
92   v[5] = rot12_128(v[5]);
93   v[6] = rot12_128(v[6]);
94   v[7] = rot12_128(v[7]);
95   v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][1]]);
96   v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][3]]);
97   v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][5]]);
98   v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][7]]);
99   v[0] = add_128(v[0], v[4]);
100   v[1] = add_128(v[1], v[5]);
101   v[2] = add_128(v[2], v[6]);
102   v[3] = add_128(v[3], v[7]);
103   v[12] = xor_128(v[12], v[0]);
104   v[13] = xor_128(v[13], v[1]);
105   v[14] = xor_128(v[14], v[2]);
106   v[15] = xor_128(v[15], v[3]);
107   v[12] = rot8_128(v[12]);
108   v[13] = rot8_128(v[13]);
109   v[14] = rot8_128(v[14]);
110   v[15] = rot8_128(v[15]);
111   v[8] = add_128(v[8], v[12]);
112   v[9] = add_128(v[9], v[13]);
113   v[10] = add_128(v[10], v[14]);
114   v[11] = add_128(v[11], v[15]);
115   v[4] = xor_128(v[4], v[8]);
116   v[5] = xor_128(v[5], v[9]);
117   v[6] = xor_128(v[6], v[10]);
118   v[7] = xor_128(v[7], v[11]);
119   v[4] = rot7_128(v[4]);
120   v[5] = rot7_128(v[5]);
121   v[6] = rot7_128(v[6]);
122   v[7] = rot7_128(v[7]);
123 
124   v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][8]]);
125   v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][10]]);
126   v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][12]]);
127   v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][14]]);
128   v[0] = add_128(v[0], v[5]);
129   v[1] = add_128(v[1], v[6]);
130   v[2] = add_128(v[2], v[7]);
131   v[3] = add_128(v[3], v[4]);
132   v[15] = xor_128(v[15], v[0]);
133   v[12] = xor_128(v[12], v[1]);
134   v[13] = xor_128(v[13], v[2]);
135   v[14] = xor_128(v[14], v[3]);
136   v[15] = rot16_128(v[15]);
137   v[12] = rot16_128(v[12]);
138   v[13] = rot16_128(v[13]);
139   v[14] = rot16_128(v[14]);
140   v[10] = add_128(v[10], v[15]);
141   v[11] = add_128(v[11], v[12]);
142   v[8] = add_128(v[8], v[13]);
143   v[9] = add_128(v[9], v[14]);
144   v[5] = xor_128(v[5], v[10]);
145   v[6] = xor_128(v[6], v[11]);
146   v[7] = xor_128(v[7], v[8]);
147   v[4] = xor_128(v[4], v[9]);
148   v[5] = rot12_128(v[5]);
149   v[6] = rot12_128(v[6]);
150   v[7] = rot12_128(v[7]);
151   v[4] = rot12_128(v[4]);
152   v[0] = add_128(v[0], m[(size_t)MSG_SCHEDULE[r][9]]);
153   v[1] = add_128(v[1], m[(size_t)MSG_SCHEDULE[r][11]]);
154   v[2] = add_128(v[2], m[(size_t)MSG_SCHEDULE[r][13]]);
155   v[3] = add_128(v[3], m[(size_t)MSG_SCHEDULE[r][15]]);
156   v[0] = add_128(v[0], v[5]);
157   v[1] = add_128(v[1], v[6]);
158   v[2] = add_128(v[2], v[7]);
159   v[3] = add_128(v[3], v[4]);
160   v[15] = xor_128(v[15], v[0]);
161   v[12] = xor_128(v[12], v[1]);
162   v[13] = xor_128(v[13], v[2]);
163   v[14] = xor_128(v[14], v[3]);
164   v[15] = rot8_128(v[15]);
165   v[12] = rot8_128(v[12]);
166   v[13] = rot8_128(v[13]);
167   v[14] = rot8_128(v[14]);
168   v[10] = add_128(v[10], v[15]);
169   v[11] = add_128(v[11], v[12]);
170   v[8] = add_128(v[8], v[13]);
171   v[9] = add_128(v[9], v[14]);
172   v[5] = xor_128(v[5], v[10]);
173   v[6] = xor_128(v[6], v[11]);
174   v[7] = xor_128(v[7], v[8]);
175   v[4] = xor_128(v[4], v[9]);
176   v[5] = rot7_128(v[5]);
177   v[6] = rot7_128(v[6]);
178   v[7] = rot7_128(v[7]);
179   v[4] = rot7_128(v[4]);
180 }
181 
transpose_vecs_128(uint32x4_t vecs[4])182 INLINE void transpose_vecs_128(uint32x4_t vecs[4]) {
183   // Individually transpose the four 2x2 sub-matrices in each corner.
184   uint32x4x2_t rows01 = vtrnq_u32(vecs[0], vecs[1]);
185   uint32x4x2_t rows23 = vtrnq_u32(vecs[2], vecs[3]);
186 
187   // Swap the top-right and bottom-left 2x2s (which just got transposed).
188   vecs[0] =
189       vcombine_u32(vget_low_u32(rows01.val[0]), vget_low_u32(rows23.val[0]));
190   vecs[1] =
191       vcombine_u32(vget_low_u32(rows01.val[1]), vget_low_u32(rows23.val[1]));
192   vecs[2] =
193       vcombine_u32(vget_high_u32(rows01.val[0]), vget_high_u32(rows23.val[0]));
194   vecs[3] =
195       vcombine_u32(vget_high_u32(rows01.val[1]), vget_high_u32(rows23.val[1]));
196 }
197 
transpose_msg_vecs4(const uint8_t * const * inputs,size_t block_offset,uint32x4_t out[16])198 INLINE void transpose_msg_vecs4(const uint8_t *const *inputs,
199                                 size_t block_offset, uint32x4_t out[16]) {
200   out[0] = loadu_128(&inputs[0][block_offset + 0 * sizeof(uint32x4_t)]);
201   out[1] = loadu_128(&inputs[1][block_offset + 0 * sizeof(uint32x4_t)]);
202   out[2] = loadu_128(&inputs[2][block_offset + 0 * sizeof(uint32x4_t)]);
203   out[3] = loadu_128(&inputs[3][block_offset + 0 * sizeof(uint32x4_t)]);
204   out[4] = loadu_128(&inputs[0][block_offset + 1 * sizeof(uint32x4_t)]);
205   out[5] = loadu_128(&inputs[1][block_offset + 1 * sizeof(uint32x4_t)]);
206   out[6] = loadu_128(&inputs[2][block_offset + 1 * sizeof(uint32x4_t)]);
207   out[7] = loadu_128(&inputs[3][block_offset + 1 * sizeof(uint32x4_t)]);
208   out[8] = loadu_128(&inputs[0][block_offset + 2 * sizeof(uint32x4_t)]);
209   out[9] = loadu_128(&inputs[1][block_offset + 2 * sizeof(uint32x4_t)]);
210   out[10] = loadu_128(&inputs[2][block_offset + 2 * sizeof(uint32x4_t)]);
211   out[11] = loadu_128(&inputs[3][block_offset + 2 * sizeof(uint32x4_t)]);
212   out[12] = loadu_128(&inputs[0][block_offset + 3 * sizeof(uint32x4_t)]);
213   out[13] = loadu_128(&inputs[1][block_offset + 3 * sizeof(uint32x4_t)]);
214   out[14] = loadu_128(&inputs[2][block_offset + 3 * sizeof(uint32x4_t)]);
215   out[15] = loadu_128(&inputs[3][block_offset + 3 * sizeof(uint32x4_t)]);
216   transpose_vecs_128(&out[0]);
217   transpose_vecs_128(&out[4]);
218   transpose_vecs_128(&out[8]);
219   transpose_vecs_128(&out[12]);
220 }
221 
load_counters4(uint64_t counter,bool increment_counter,uint32x4_t * out_low,uint32x4_t * out_high)222 INLINE void load_counters4(uint64_t counter, bool increment_counter,
223                            uint32x4_t *out_low, uint32x4_t *out_high) {
224   uint64_t mask = (increment_counter ? ~0 : 0);
225   *out_low = set4(
226       counter_low(counter + (mask & 0)), counter_low(counter + (mask & 1)),
227       counter_low(counter + (mask & 2)), counter_low(counter + (mask & 3)));
228   *out_high = set4(
229       counter_high(counter + (mask & 0)), counter_high(counter + (mask & 1)),
230       counter_high(counter + (mask & 2)), counter_high(counter + (mask & 3)));
231 }
232 
233 static
blake3_hash4_neon(const uint8_t * const * inputs,size_t blocks,const uint32_t key[8],uint64_t counter,bool increment_counter,uint8_t flags,uint8_t flags_start,uint8_t flags_end,uint8_t * out)234 void blake3_hash4_neon(const uint8_t *const *inputs, size_t blocks,
235                        const uint32_t key[8], uint64_t counter,
236                        bool increment_counter, uint8_t flags,
237                        uint8_t flags_start, uint8_t flags_end, uint8_t *out) {
238   uint32x4_t h_vecs[8] = {
239       set1_128(key[0]), set1_128(key[1]), set1_128(key[2]), set1_128(key[3]),
240       set1_128(key[4]), set1_128(key[5]), set1_128(key[6]), set1_128(key[7]),
241   };
242   uint32x4_t counter_low_vec, counter_high_vec;
243   load_counters4(counter, increment_counter, &counter_low_vec,
244                  &counter_high_vec);
245   uint8_t block_flags = flags | flags_start;
246 
247   for (size_t block = 0; block < blocks; block++) {
248     if (block + 1 == blocks) {
249       block_flags |= flags_end;
250     }
251     uint32x4_t block_len_vec = set1_128(BLAKE3_BLOCK_LEN);
252     uint32x4_t block_flags_vec = set1_128(block_flags);
253     uint32x4_t msg_vecs[16];
254     transpose_msg_vecs4(inputs, block * BLAKE3_BLOCK_LEN, msg_vecs);
255 
256     uint32x4_t v[16] = {
257         h_vecs[0],       h_vecs[1],        h_vecs[2],       h_vecs[3],
258         h_vecs[4],       h_vecs[5],        h_vecs[6],       h_vecs[7],
259         set1_128(IV[0]), set1_128(IV[1]),  set1_128(IV[2]), set1_128(IV[3]),
260         counter_low_vec, counter_high_vec, block_len_vec,   block_flags_vec,
261     };
262     round_fn4(v, msg_vecs, 0);
263     round_fn4(v, msg_vecs, 1);
264     round_fn4(v, msg_vecs, 2);
265     round_fn4(v, msg_vecs, 3);
266     round_fn4(v, msg_vecs, 4);
267     round_fn4(v, msg_vecs, 5);
268     round_fn4(v, msg_vecs, 6);
269     h_vecs[0] = xor_128(v[0], v[8]);
270     h_vecs[1] = xor_128(v[1], v[9]);
271     h_vecs[2] = xor_128(v[2], v[10]);
272     h_vecs[3] = xor_128(v[3], v[11]);
273     h_vecs[4] = xor_128(v[4], v[12]);
274     h_vecs[5] = xor_128(v[5], v[13]);
275     h_vecs[6] = xor_128(v[6], v[14]);
276     h_vecs[7] = xor_128(v[7], v[15]);
277 
278     block_flags = flags;
279   }
280 
281   transpose_vecs_128(&h_vecs[0]);
282   transpose_vecs_128(&h_vecs[4]);
283   // The first four vecs now contain the first half of each output, and the
284   // second four vecs contain the second half of each output.
285   storeu_128(h_vecs[0], &out[0 * sizeof(uint32x4_t)]);
286   storeu_128(h_vecs[4], &out[1 * sizeof(uint32x4_t)]);
287   storeu_128(h_vecs[1], &out[2 * sizeof(uint32x4_t)]);
288   storeu_128(h_vecs[5], &out[3 * sizeof(uint32x4_t)]);
289   storeu_128(h_vecs[2], &out[4 * sizeof(uint32x4_t)]);
290   storeu_128(h_vecs[6], &out[5 * sizeof(uint32x4_t)]);
291   storeu_128(h_vecs[3], &out[6 * sizeof(uint32x4_t)]);
292   storeu_128(h_vecs[7], &out[7 * sizeof(uint32x4_t)]);
293 }
294 
295 /*
296  * ----------------------------------------------------------------------------
297  * hash_many_neon
298  * ----------------------------------------------------------------------------
299  */
300 
301 void blake3_compress_in_place_portable(uint32_t cv[8],
302                                        const uint8_t block[BLAKE3_BLOCK_LEN],
303                                        uint8_t block_len, uint64_t counter,
304                                        uint8_t flags);
305 
hash_one_neon(const uint8_t * input,size_t blocks,const uint32_t key[8],uint64_t counter,uint8_t flags,uint8_t flags_start,uint8_t flags_end,uint8_t out[BLAKE3_OUT_LEN])306 INLINE void hash_one_neon(const uint8_t *input, size_t blocks,
307                           const uint32_t key[8], uint64_t counter,
308                           uint8_t flags, uint8_t flags_start, uint8_t flags_end,
309                           uint8_t out[BLAKE3_OUT_LEN]) {
310   uint32_t cv[8];
311   memcpy(cv, key, BLAKE3_KEY_LEN);
312   uint8_t block_flags = flags | flags_start;
313   while (blocks > 0) {
314     if (blocks == 1) {
315       block_flags |= flags_end;
316     }
317     // TODO: Implement compress_neon. However note that according to
318     // https://github.com/BLAKE2/BLAKE2/commit/7965d3e6e1b4193438b8d3a656787587d2579227,
319     // compress_neon might not be any faster than compress_portable.
320     blake3_compress_in_place_portable(cv, input, BLAKE3_BLOCK_LEN, counter,
321                                       block_flags);
322     input = &input[BLAKE3_BLOCK_LEN];
323     blocks -= 1;
324     block_flags = flags;
325   }
326   memcpy(out, cv, BLAKE3_OUT_LEN);
327 }
328 
blake3_hash_many_neon(const uint8_t * const * inputs,size_t num_inputs,size_t blocks,const uint32_t key[8],uint64_t counter,bool increment_counter,uint8_t flags,uint8_t flags_start,uint8_t flags_end,uint8_t * out)329 void blake3_hash_many_neon(const uint8_t *const *inputs, size_t num_inputs,
330                            size_t blocks, const uint32_t key[8],
331                            uint64_t counter, bool increment_counter,
332                            uint8_t flags, uint8_t flags_start,
333                            uint8_t flags_end, uint8_t *out) {
334   while (num_inputs >= 4) {
335     blake3_hash4_neon(inputs, blocks, key, counter, increment_counter, flags,
336                       flags_start, flags_end, out);
337     if (increment_counter) {
338       counter += 4;
339     }
340     inputs += 4;
341     num_inputs -= 4;
342     out = &out[4 * BLAKE3_OUT_LEN];
343   }
344   while (num_inputs > 0) {
345     hash_one_neon(inputs[0], blocks, key, counter, flags, flags_start,
346                   flags_end, out);
347     if (increment_counter) {
348       counter += 1;
349     }
350     inputs += 1;
351     num_inputs -= 1;
352     out = &out[BLAKE3_OUT_LEN];
353   }
354 }
355 
356 #endif // BLAKE3_USE_NEON
357