Lines Matching full:bps
248 ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
249 ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
250 ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
251 ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
254 ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]);
255 ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]);
256 ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]);
257 ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]);
277 _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
278 _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
279 _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
280 _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
283 *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0);
284 *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1);
285 *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2);
286 *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3);
309 const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]);
310 const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]);
311 const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]);
312 const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]);
318 const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
319 const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
320 const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
321 const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
446 const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]);
447 const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]);
448 const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]);
449 const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]);
450 const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]);
451 const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]);
452 const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]);
453 const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]);
507 const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]);
508 const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]);
509 const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]);
510 const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]);
511 const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]);
512 const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]);
513 const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]);
514 const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]);
673 for (y = 0; y < 16 * BPS; y += 4 * BPS) {
