Lines Matching refs:__m128i
30 __m128i xmm_crc0 = _mm_loadu_si128((__m128i *)s->crc0 + 0);\
31 __m128i xmm_crc1 = _mm_loadu_si128((__m128i *)s->crc0 + 1);\
32 __m128i xmm_crc2 = _mm_loadu_si128((__m128i *)s->crc0 + 2);\
33 __m128i xmm_crc3 = _mm_loadu_si128((__m128i *)s->crc0 + 3);\
34 __m128i xmm_crc_part = _mm_loadu_si128((__m128i *)s->crc0 + 4);
37 _mm_storeu_si128((__m128i *)s->crc0 + 0, xmm_crc0);\
38 _mm_storeu_si128((__m128i *)s->crc0 + 1, xmm_crc1);\
39 _mm_storeu_si128((__m128i *)s->crc0 + 2, xmm_crc2);\
40 _mm_storeu_si128((__m128i *)s->crc0 + 3, xmm_crc3);\
41 _mm_storeu_si128((__m128i *)s->crc0 + 4, xmm_crc_part);\
59 __m128i *xmm_crc0, __m128i *xmm_crc1,
60 __m128i *xmm_crc2, __m128i *xmm_crc3)
62 const __m128i xmm_fold4 = _mm_set_epi32(
66 __m128i x_tmp3;
85 __m128i *xmm_crc0, __m128i *xmm_crc1,
86 __m128i *xmm_crc2, __m128i *xmm_crc3)
88 const __m128i xmm_fold4 = _mm_set_epi32(
92 __m128i x_tmp3, x_tmp2;
119 __m128i *xmm_crc0, __m128i *xmm_crc1,
120 __m128i *xmm_crc2, __m128i *xmm_crc3)
122 const __m128i xmm_fold4 = _mm_set_epi32(
126 __m128i x_tmp3;
159 __m128i *xmm_crc0, __m128i *xmm_crc1,
160 __m128i *xmm_crc2, __m128i *xmm_crc3)
162 const __m128i xmm_fold4 = _mm_set_epi32(
166 __m128i x_tmp0, x_tmp1, x_tmp2, x_tmp3;
225 __m128i *xmm_crc0, __m128i *xmm_crc1,
226 __m128i *xmm_crc2, __m128i *xmm_crc3,
227 __m128i *xmm_crc_part)
230 const __m128i xmm_fold4 = _mm_set_epi32(
233 const __m128i xmm_mask3 = _mm_set1_epi32(0x80808080);
235 __m128i xmm_shl, xmm_shr, xmm_tmp1, xmm_tmp2, xmm_tmp3;
236 __m128i xmm_a0_0, xmm_a0_1;
239 xmm_shl = _mm_load_si128((__m128i *)pshufb_shf_table + (len - 1));
278 __m128i xmm_t0, xmm_t1, xmm_t2, xmm_t3;
290 xmm_crc_part = _mm_loadu_si128((__m128i *)src);
291 _mm_storeu_si128((__m128i *)dst, xmm_crc_part);
302 xmm_t0 = _mm_load_si128((__m128i *)src);
303 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
304 xmm_t2 = _mm_load_si128((__m128i *)src + 2);
305 xmm_t3 = _mm_load_si128((__m128i *)src + 3);
309 _mm_storeu_si128((__m128i *)dst, xmm_t0);
310 _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
311 _mm_storeu_si128((__m128i *)dst + 2, xmm_t2);
312 _mm_storeu_si128((__m128i *)dst + 3, xmm_t3);
329 xmm_t0 = _mm_load_si128((__m128i *)src);
330 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
331 xmm_t2 = _mm_load_si128((__m128i *)src + 2);
335 _mm_storeu_si128((__m128i *)dst, xmm_t0);
336 _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
337 _mm_storeu_si128((__m128i *)dst + 2, xmm_t2);
351 xmm_t0 = _mm_load_si128((__m128i *)src);
352 xmm_t1 = _mm_load_si128((__m128i *)src + 1);
356 _mm_storeu_si128((__m128i *)dst, xmm_t0);
357 _mm_storeu_si128((__m128i *)dst + 1, xmm_t1);
370 xmm_t0 = _mm_load_si128((__m128i *)src);
374 _mm_storeu_si128((__m128i *)dst, xmm_t0);
406 _mm_storeu_si128((__m128i *)dst, xmm_crc_part);
432 const __m128i xmm_mask = _mm_load_si128((__m128i *)crc_mask);
433 const __m128i xmm_mask2 = _mm_load_si128((__m128i *)crc_mask2);
436 __m128i x_tmp0, x_tmp1, x_tmp2, crc_fold;
438 __m128i xmm_crc0 = _mm_loadu_si128((__m128i *)s->crc0 + 0);
439 __m128i xmm_crc1 = _mm_loadu_si128((__m128i *)s->crc0 + 1);
440 __m128i xmm_crc2 = _mm_loadu_si128((__m128i *)s->crc0 + 2);
441 __m128i xmm_crc3 = _mm_loadu_si128((__m128i *)s->crc0 + 3);
446 crc_fold = _mm_load_si128((__m128i *)crc_k);
466 crc_fold = _mm_load_si128((__m128i *)crc_k + 1);
484 crc_fold = _mm_load_si128((__m128i *)crc_k + 2);