Lines Matching defs:v_r1
312 __m128i v_r1 = _mm_packs_epi32(_mm_loadu_si128((__m128i const *)(row1 + x)),
317 __m128i v_2r1 = _mm_adds_epu16(v_r1, v_r1), v_4r1 = _mm_adds_epu16(v_2r1, v_2r1);
319 __m128i v_dst10 = _mm_slli_epi16(_mm_adds_epu16(v_r1, v_r2), 2);
323 v_r1 = _mm_packs_epi32(_mm_loadu_si128((__m128i const *)(row1 + x + 8)),
328 v_2r1 = _mm_adds_epu16(v_r1, v_r1), v_4r1 = _mm_adds_epu16(v_2r1, v_2r1);
330 __m128i v_dst11 = _mm_slli_epi16(_mm_adds_epu16(v_r1, v_r2), 2);
342 __m128i v_r1 = _mm_packs_epi32(_mm_loadu_si128((__m128i const *)(row1 + x)),
347 __m128i v_2r1 = _mm_adds_epu16(v_r1, v_r1), v_4r1 = _mm_adds_epu16(v_2r1, v_2r1);
349 __m128i v_dst1 = _mm_slli_epi16(_mm_adds_epu16(v_r1, v_r2), 2);
375 v_r1 = _mm_loadu_si128((__m128i const *)(row1 + x)),
377 __m128i v_2r1 = _mm_slli_epi32(v_r1, 1), v_4r1 = _mm_slli_epi32(v_r1, 2);
379 __m128i v_dst10 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
382 v_r1 = _mm_loadu_si128((__m128i const *)(row1 + x + 4));
384 v_2r1 = _mm_slli_epi32(v_r1, 1);
385 v_4r1 = _mm_slli_epi32(v_r1, 2);
387 __m128i v_dst11 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
400 v_r1 = _mm_loadu_si128((__m128i const *)(row1 + x)),
402 __m128i v_2r1 = _mm_slli_epi32(v_r1, 1), v_4r1 = _mm_slli_epi32(v_r1, 2);
405 __m128i v_dst1 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
435 v_r1 = _mm_loadu_si128((__m128i const *)(row1 + x)),
437 __m128i v_2r1 = _mm_slli_epi32(v_r1, 1), v_4r1 = _mm_slli_epi32(v_r1, 2);
439 __m128i v_dst10 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
442 v_r1 = _mm_loadu_si128((__m128i const *)(row1 + x + 4));
444 v_2r1 = _mm_slli_epi32(v_r1, 1);
445 v_4r1 = _mm_slli_epi32(v_r1, 2);
447 __m128i v_dst11 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
460 v_r1 = _mm_loadu_si128((__m128i const *)(row1 + x)),
462 __m128i v_2r1 = _mm_slli_epi32(v_r1, 1), v_4r1 = _mm_slli_epi32(v_r1, 2);
465 __m128i v_dst1 = _mm_slli_epi32(_mm_add_epi32(v_r1, v_r2), 2);
500 __m128 v_r1 = _mm_loadu_ps(row1 + x);
503 _mm_storeu_ps(dst1 + x, _mm_mul_ps(v_scale4, _mm_add_ps(v_r1, v_r2)));
504 _mm_storeu_ps(dst0 + x, _mm_mul_ps(v_scale, _mm_add_ps(_mm_add_ps(v_r0, _mm_mul_ps(v_6, v_r1)), v_r2)));
507 v_r1 = _mm_loadu_ps(row1 + x + 4);
510 _mm_storeu_ps(dst1 + x + 4, _mm_mul_ps(v_scale4, _mm_add_ps(v_r1, v_r2)));
511 _mm_storeu_ps(dst0 + x + 4, _mm_mul_ps(v_scale, _mm_add_ps(_mm_add_ps(v_r0, _mm_mul_ps(v_6, v_r1)), v_r2)));
533 uint16x8_t v_r1 = vcombine_u16(vqmovn_u32(vld1q_u32(row1 + x)), vqmovn_u32(vld1q_u32(row1 + x + 4)));
539 v_r1 = vaddq_u16(vaddq_u16(v_r1, v_r2), v_r3);
540 uint16x8_t v_dst0 = vaddq_u16(v_r0, vshlq_n_u16(v_r1, 2));
543 v_r1 = vcombine_u16(vqmovn_u32(vld1q_u32(row1 + x + 8)), vqmovn_u32(vld1q_u32(row1 + x + 12)));
549 v_r1 = vaddq_u16(vaddq_u16(v_r1, v_r2), v_r3);
550 uint16x8_t v_dst1 = vaddq_u16(v_r0, vshlq_n_u16(v_r1, 2));
638 float32x4_t v_r1 = vld1q_f32(row1 + x);
644 v_r1 = vaddq_f32(vaddq_f32(v_r1, v_r2), v_r3);
645 vst1q_f32(dst + x, vmulq_f32(vmlaq_f32(v_r0, v_4, v_r1), v_scale));
648 v_r1 = vld1q_f32(row1 + x + 4);
654 v_r1 = vaddq_f32(vaddq_f32(v_r1, v_r2), v_r3);
655 vst1q_f32(dst + x + 4, vmulq_f32(vmlaq_f32(v_r0, v_4, v_r1), v_scale));
674 uint16x8_t v_r1 = vcombine_u16(vqmovn_u32(vld1q_u32(row1 + x)), vqmovn_u32(vld1q_u32(row1 + x + 4)));
677 uint16x8_t v_2r1 = vaddq_u16(v_r1, v_r1), v_4r1 = vaddq_u16(v_2r1, v_2r1);
679 uint16x8_t v_dst10 = vshlq_n_u16(vaddq_u16(v_r1, v_r2), 2);
682 v_r1 = vcombine_u16(vqmovn_u32(vld1q_u32(row1 + x + 8)), vqmovn_u32(vld1q_u32(row1 + x + 12)));
685 v_2r1 = vaddq_u16(v_r1, v_r1), v_4r1 = vaddq_u16(v_2r1, v_2r1);
687 uint16x8_t v_dst11 = vshlq_n_u16(vaddq_u16(v_r1, v_r2), 2);
698 uint16x8_t v_r1 = vcombine_u16(vqmovn_u32(vld1q_u32(row1 + x)), vqmovn_u32(vld1q_u32(row1 + x + 4)));
701 uint16x8_t v_2r1 = vaddq_u16(v_r1, v_r1), v_4r1 = vaddq_u16(v_2r1, v_2r1);
703 uint16x8_t v_dst1 = vshlq_n_u16(vaddq_u16(v_r1, v_r2), 2);
724 uint32x4_t v_r0 = vld1q_u32(row0 + x), v_r1 = vld1q_u32(row1 + x), v_r2 = vld1q_u32(row2 + x);
725 uint32x4_t v_2r1 = vshlq_n_u32(v_r1, 1), v_4r1 = vshlq_n_u32(v_r1, 2);
727 uint32x4_t v_dst10 = vshlq_n_u32(vaddq_u32(v_r1, v_r2), 2);
730 v_r1 = vld1q_u32(row1 + x + 4);
732 v_2r1 = vshlq_n_u32(v_r1, 1);
733 v_4r1 = vshlq_n_u32(v_r1, 2);
735 v_r1, v_r2), 2);
745 uint32x4_t v_r0 = vld1q_u32(row0 + x), v_r1 = vld1q_u32(row1 + x), v_r2 = vld1q_u32(row2 + x);
746 uint32x4_t v_2r1 = vshlq_n_u32(v_r1, 1), v_4r1 = vshlq_n_u32(v_r1, 2);
749 uint32x4_t v_dst1 = vshlq_n_u32(vaddq_u32(v_r1, v_r2), 2);
770 int32x4_t v_r0 = vld1q_s32(row0 + x), v_r1 = vld1q_s32(row1 + x), v_r2 = vld1q_s32(row2 + x);
771 int32x4_t v_2r1 = vshlq_n_s32(v_r1, 1), v_4r1 = vshlq_n_s32(v_r1, 2);
773 int32x4_t v_dst10 = vshlq_n_s32(vaddq_s32(v_r1, v_r2), 2);
776 v_r1 = vld1q_s32(row1 + x + 4);
778 v_2r1 = vshlq_n_s32(v_r1, 1);
779 v_4r1 = vshlq_n_s32(v_r1, 2);
781 int32x4_t v_dst11 = vshlq_n_s32(vaddq_s32(v_r1, v_r2), 2);
791 int32x4_t v_r0 = vld1q_s32(row0 + x), v_r1 = vld1q_s32(row1 + x), v_r2 = vld1q_s32(row2 + x);
792 int32x4_t v_2r1 = vshlq_n_s32(v_r1, 1), v_4r1 = vshlq_n_s32(v_r1, 2);
795 int32x4_t v_dst1 = vshlq_n_s32(vaddq_s32(v_r1, v_r2), 2);
817 float32x4_t v_r1 = vld1q_f32(row1 + x);
820 vst1q_f32(dst1 + x, vmulq_f32(v_scale4, vaddq_f32(v_r1, v_r2)));
821 vst1q_f32(dst0 + x, vmulq_f32(v_scale, vaddq_f32(vmlaq_f32(v_r0, v_6, v_r1), v_r2)));
824 v_r1 = vld1q_f32(row1 + x + 4);
827 vst1q_f32(dst1 + x + 4, vmulq_f32(v_scale4, vaddq_f32(v_r1, v_r2)));
828 vst1q_f32(dst0 + x + 4, vmulq_f32(v_scale, vaddq_f32(vmlaq_f32(v_r0, v_6, v_r1), v_r2)));
