Home | History | Annotate | Download | only in dsp

Lines Matching refs:BPS

280       ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
281 ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
282 ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
283 ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
286 ref0 = _mm_cvtsi32_si128(*(int*)&ref[0 * BPS]);
287 ref1 = _mm_cvtsi32_si128(*(int*)&ref[1 * BPS]);
288 ref2 = _mm_cvtsi32_si128(*(int*)&ref[2 * BPS]);
289 ref3 = _mm_cvtsi32_si128(*(int*)&ref[3 * BPS]);
309 _mm_storel_epi64((__m128i*)&dst[0 * BPS], ref0);
310 _mm_storel_epi64((__m128i*)&dst[1 * BPS], ref1);
311 _mm_storel_epi64((__m128i*)&dst[2 * BPS], ref2);
312 _mm_storel_epi64((__m128i*)&dst[3 * BPS], ref3);
315 *((int32_t *)&dst[0 * BPS]) = _mm_cvtsi128_si32(ref0);
316 *((int32_t *)&dst[1 * BPS]) = _mm_cvtsi128_si32(ref1);
317 *((int32_t *)&dst[2 * BPS]) = _mm_cvtsi128_si32(ref2);
318 *((int32_t *)&dst[3 * BPS]) = _mm_cvtsi128_si32(ref3);
347 const __m128i src0 = _mm_loadl_epi64((__m128i*)&src[0 * BPS]);
348 const __m128i src1 = _mm_loadl_epi64((__m128i*)&src[1 * BPS]);
349 const __m128i src2 = _mm_loadl_epi64((__m128i*)&src[2 * BPS]);
350 const __m128i src3 = _mm_loadl_epi64((__m128i*)&src[3 * BPS]);
356 const __m128i ref0 = _mm_loadl_epi64((__m128i*)&ref[0 * BPS]);
357 const __m128i ref1 = _mm_loadl_epi64((__m128i*)&ref[1 * BPS]);
358 const __m128i ref2 = _mm_loadl_epi64((__m128i*)&ref[2 * BPS]);
359 const __m128i ref3 = _mm_loadl_epi64((__m128i*)&ref[3 * BPS]);
468 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[BPS * 0]);
469 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[BPS * 1]);
470 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[BPS * 2]);
471 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[BPS * 3]);
472 const __m128i b0 = _mm_loadu_si128((__m128i*)&b[BPS * 0]);
473 const __m128i b1 = _mm_loadu_si128((__m128i*)&b[BPS * 1]);
474 const __m128i b2 = _mm_loadu_si128((__m128i*)&b[BPS * 2]);
475 const __m128i b3 = _mm_loadu_si128((__m128i*)&b[BPS * 3]);
526 a += 4 * BPS;
527 b += 4 * BPS;
554 const __m128i a0 = _mm_loadl_epi64((__m128i*)&a[BPS * 0]);
555 const __m128i a1 = _mm_loadl_epi64((__m128i*)&a[BPS * 1]);
556 const __m128i a2 = _mm_loadl_epi64((__m128i*)&a[BPS * 2]);
557 const __m128i a3 = _mm_loadl_epi64((__m128i*)&a[BPS * 3]);
558 const __m128i b0 = _mm_loadl_epi64((__m128i*)&b[BPS * 0]);
559 const __m128i b1 = _mm_loadl_epi64((__m128i*)&b[BPS * 1]);
560 const __m128i b2 = _mm_loadl_epi64((__m128i*)&b[BPS * 2]);
561 const __m128i b3 = _mm_loadl_epi64((__m128i*)&b[BPS * 3]);
614 const __m128i inA_0 = _mm_loadl_epi64((__m128i*)&inA[BPS * 0]);
615 const __m128i inA_1 = _mm_loadl_epi64((__m128i*)&inA[BPS * 1]);
616 const __m128i inA_2 = _mm_loadl_epi64((__m128i*)&inA[BPS * 2]);
617 const __m128i inA_3 = _mm_loadl_epi64((__m128i*)&inA[BPS * 3]);
618 const __m128i inB_0 = _mm_loadl_epi64((__m128i*)&inB[BPS * 0]);
619 const __m128i inB_1 = _mm_loadl_epi64((__m128i*)&inB[BPS * 1]);
620 const __m128i inB_2 = _mm_loadl_epi64((__m128i*)&inB[BPS * 2]);
621 const __m128i inB_3 = _mm_loadl_epi64((__m128i*)&inB[BPS * 3]);
764 for (y = 0; y < 16 * BPS; y += 4 * BPS) {