Lines Matching full:v
276 INLINE void round_fn(__m128i v[16], __m128i m[16], size_t r) { in round_fn()
277 v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][0]]); in round_fn()
278 v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][2]]); in round_fn()
279 v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][4]]); in round_fn()
280 v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][6]]); in round_fn()
281 v[0] = addv(v[0], v[4]); in round_fn()
282 v[1] = addv(v[1], v[5]); in round_fn()
283 v[2] = addv(v[2], v[6]); in round_fn()
284 v[3] = addv(v[3], v[7]); in round_fn()
285 v[12] = xorv(v[12], v[0]); in round_fn()
286 v[13] = xorv(v[13], v[1]); in round_fn()
287 v[14] = xorv(v[14], v[2]); in round_fn()
288 v[15] = xorv(v[15], v[3]); in round_fn()
289 v[12] = rot16(v[12]); in round_fn()
290 v[13] = rot16(v[13]); in round_fn()
291 v[14] = rot16(v[14]); in round_fn()
292 v[15] = rot16(v[15]); in round_fn()
293 v[8] = addv(v[8], v[12]); in round_fn()
294 v[9] = addv(v[9], v[13]); in round_fn()
295 v[10] = addv(v[10], v[14]); in round_fn()
296 v[11] = addv(v[11], v[15]); in round_fn()
297 v[4] = xorv(v[4], v[8]); in round_fn()
298 v[5] = xorv(v[5], v[9]); in round_fn()
299 v[6] = xorv(v[6], v[10]); in round_fn()
300 v[7] = xorv(v[7], v[11]); in round_fn()
301 v[4] = rot12(v[4]); in round_fn()
302 v[5] = rot12(v[5]); in round_fn()
303 v[6] = rot12(v[6]); in round_fn()
304 v[7] = rot12(v[7]); in round_fn()
305 v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][1]]); in round_fn()
306 v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][3]]); in round_fn()
307 v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][5]]); in round_fn()
308 v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][7]]); in round_fn()
309 v[0] = addv(v[0], v[4]); in round_fn()
310 v[1] = addv(v[1], v[5]); in round_fn()
311 v[2] = addv(v[2], v[6]); in round_fn()
312 v[3] = addv(v[3], v[7]); in round_fn()
313 v[12] = xorv(v[12], v[0]); in round_fn()
314 v[13] = xorv(v[13], v[1]); in round_fn()
315 v[14] = xorv(v[14], v[2]); in round_fn()
316 v[15] = xorv(v[15], v[3]); in round_fn()
317 v[12] = rot8(v[12]); in round_fn()
318 v[13] = rot8(v[13]); in round_fn()
319 v[14] = rot8(v[14]); in round_fn()
320 v[15] = rot8(v[15]); in round_fn()
321 v[8] = addv(v[8], v[12]); in round_fn()
322 v[9] = addv(v[9], v[13]); in round_fn()
323 v[10] = addv(v[10], v[14]); in round_fn()
324 v[11] = addv(v[11], v[15]); in round_fn()
325 v[4] = xorv(v[4], v[8]); in round_fn()
326 v[5] = xorv(v[5], v[9]); in round_fn()
327 v[6] = xorv(v[6], v[10]); in round_fn()
328 v[7] = xorv(v[7], v[11]); in round_fn()
329 v[4] = rot7(v[4]); in round_fn()
330 v[5] = rot7(v[5]); in round_fn()
331 v[6] = rot7(v[6]); in round_fn()
332 v[7] = rot7(v[7]); in round_fn()
334 v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][8]]); in round_fn()
335 v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][10]]); in round_fn()
336 v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][12]]); in round_fn()
337 v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][14]]); in round_fn()
338 v[0] = addv(v[0], v[5]); in round_fn()
339 v[1] = addv(v[1], v[6]); in round_fn()
340 v[2] = addv(v[2], v[7]); in round_fn()
341 v[3] = addv(v[3], v[4]); in round_fn()
342 v[15] = xorv(v[15], v[0]); in round_fn()
343 v[12] = xorv(v[12], v[1]); in round_fn()
344 v[13] = xorv(v[13], v[2]); in round_fn()
345 v[14] = xorv(v[14], v[3]); in round_fn()
346 v[15] = rot16(v[15]); in round_fn()
347 v[12] = rot16(v[12]); in round_fn()
348 v[13] = rot16(v[13]); in round_fn()
349 v[14] = rot16(v[14]); in round_fn()
350 v[10] = addv(v[10], v[15]); in round_fn()
351 v[11] = addv(v[11], v[12]); in round_fn()
352 v[8] = addv(v[8], v[13]); in round_fn()
353 v[9] = addv(v[9], v[14]); in round_fn()
354 v[5] = xorv(v[5], v[10]); in round_fn()
355 v[6] = xorv(v[6], v[11]); in round_fn()
356 v[7] = xorv(v[7], v[8]); in round_fn()
357 v[4] = xorv(v[4], v[9]); in round_fn()
358 v[5] = rot12(v[5]); in round_fn()
359 v[6] = rot12(v[6]); in round_fn()
360 v[7] = rot12(v[7]); in round_fn()
361 v[4] = rot12(v[4]); in round_fn()
362 v[0] = addv(v[0], m[(size_t)MSG_SCHEDULE[r][9]]); in round_fn()
363 v[1] = addv(v[1], m[(size_t)MSG_SCHEDULE[r][11]]); in round_fn()
364 v[2] = addv(v[2], m[(size_t)MSG_SCHEDULE[r][13]]); in round_fn()
365 v[3] = addv(v[3], m[(size_t)MSG_SCHEDULE[r][15]]); in round_fn()
366 v[0] = addv(v[0], v[5]); in round_fn()
367 v[1] = addv(v[1], v[6]); in round_fn()
368 v[2] = addv(v[2], v[7]); in round_fn()
369 v[3] = addv(v[3], v[4]); in round_fn()
370 v[15] = xorv(v[15], v[0]); in round_fn()
371 v[12] = xorv(v[12], v[1]); in round_fn()
372 v[13] = xorv(v[13], v[2]); in round_fn()
373 v[14] = xorv(v[14], v[3]); in round_fn()
374 v[15] = rot8(v[15]); in round_fn()
375 v[12] = rot8(v[12]); in round_fn()
376 v[13] = rot8(v[13]); in round_fn()
377 v[14] = rot8(v[14]); in round_fn()
378 v[10] = addv(v[10], v[15]); in round_fn()
379 v[11] = addv(v[11], v[12]); in round_fn()
380 v[8] = addv(v[8], v[13]); in round_fn()
381 v[9] = addv(v[9], v[14]); in round_fn()
382 v[5] = xorv(v[5], v[10]); in round_fn()
383 v[6] = xorv(v[6], v[11]); in round_fn()
384 v[7] = xorv(v[7], v[8]); in round_fn()
385 v[4] = xorv(v[4], v[9]); in round_fn()
386 v[5] = rot7(v[5]); in round_fn()
387 v[6] = rot7(v[6]); in round_fn()
388 v[7] = rot7(v[7]); in round_fn()
389 v[4] = rot7(v[4]); in round_fn()
476 __m128i v[16] = { in blake3_hash4_sse41() local
482 round_fn(v, msg_vecs, 0); in blake3_hash4_sse41()
483 round_fn(v, msg_vecs, 1); in blake3_hash4_sse41()
484 round_fn(v, msg_vecs, 2); in blake3_hash4_sse41()
485 round_fn(v, msg_vecs, 3); in blake3_hash4_sse41()
486 round_fn(v, msg_vecs, 4); in blake3_hash4_sse41()
487 round_fn(v, msg_vecs, 5); in blake3_hash4_sse41()
488 round_fn(v, msg_vecs, 6); in blake3_hash4_sse41()
489 h_vecs[0] = xorv(v[0], v[8]); in blake3_hash4_sse41()
490 h_vecs[1] = xorv(v[1], v[9]); in blake3_hash4_sse41()
491 h_vecs[2] = xorv(v[2], v[10]); in blake3_hash4_sse41()
492 h_vecs[3] = xorv(v[3], v[11]); in blake3_hash4_sse41()
493 h_vecs[4] = xorv(v[4], v[12]); in blake3_hash4_sse41()
494 h_vecs[5] = xorv(v[5], v[13]); in blake3_hash4_sse41()
495 h_vecs[6] = xorv(v[6], v[14]); in blake3_hash4_sse41()
496 h_vecs[7] = xorv(v[7], v[15]); in blake3_hash4_sse41()