From 940f1e79143f8ad7f34b267bb013f73d6c79700d Mon Sep 17 00:00:00 2001 From: Ilya Lavrenov Date: Mon, 12 Jan 2015 10:59:29 +0300 Subject: [PATCH] interliving / deinterliving --- modules/imgproc/src/color.cpp | 90 +++++++++++++++++++++++++++++++++++ 1 file changed, 90 insertions(+) diff --git a/modules/imgproc/src/color.cpp b/modules/imgproc/src/color.cpp index b5ac47328..9bbb7b6ad 100644 --- a/modules/imgproc/src/color.cpp +++ b/modules/imgproc/src/color.cpp @@ -183,6 +183,96 @@ static IppStatus sts = ippInit(); v_b1 = _mm_packus_epi16(_mm_srli_epi16(layer1_chunk4, 8), _mm_srli_epi16(layer1_chunk5, 8)); \ } +#define _MM_DEINTERLIV_EPI16(layer0_chunk0, layer0_chunk1, layer0_chunk2, \ + layer0_chunk3, layer0_chunk4, layer0_chunk5) \ + { \ + __m128i layer1_chunk0 = _mm_unpacklo_epi16(layer0_chunk0, layer0_chunk3); \ + __m128i layer1_chunk1 = _mm_unpackhi_epi16(layer0_chunk0, layer0_chunk3); \ + __m128i layer1_chunk2 = _mm_unpacklo_epi16(layer0_chunk1, layer0_chunk4); \ + __m128i layer1_chunk3 = _mm_unpackhi_epi16(layer0_chunk1, layer0_chunk4); \ + __m128i layer1_chunk4 = _mm_unpacklo_epi16(layer0_chunk2, layer0_chunk5); \ + __m128i layer1_chunk5 = _mm_unpackhi_epi16(layer0_chunk2, layer0_chunk5); \ + \ + __m128i layer2_chunk0 = _mm_unpacklo_epi16(layer1_chunk0, layer1_chunk3); \ + __m128i layer2_chunk1 = _mm_unpackhi_epi16(layer1_chunk0, layer1_chunk3); \ + __m128i layer2_chunk2 = _mm_unpacklo_epi16(layer1_chunk1, layer1_chunk4); \ + __m128i layer2_chunk3 = _mm_unpackhi_epi16(layer1_chunk1, layer1_chunk4); \ + __m128i layer2_chunk4 = _mm_unpacklo_epi16(layer1_chunk2, layer1_chunk5); \ + __m128i layer2_chunk5 = _mm_unpackhi_epi16(layer1_chunk2, layer1_chunk5); \ + \ + __m128i layer3_chunk0 = _mm_unpacklo_epi16(layer2_chunk0, layer2_chunk3); \ + __m128i layer3_chunk1 = _mm_unpackhi_epi16(layer2_chunk0, layer2_chunk3); \ + __m128i layer3_chunk2 = _mm_unpacklo_epi16(layer2_chunk1, layer2_chunk4); \ + __m128i layer3_chunk3 = _mm_unpackhi_epi16(layer2_chunk1, layer2_chunk4); \ + __m128i layer3_chunk4 = _mm_unpacklo_epi16(layer2_chunk2, layer2_chunk5); \ + __m128i layer3_chunk5 = _mm_unpackhi_epi16(layer2_chunk2, layer2_chunk5); \ + \ + layer0_chunk0 = _mm_unpacklo_epi16(layer3_chunk0, layer3_chunk3); \ + layer0_chunk1 = _mm_unpackhi_epi16(layer3_chunk0, layer3_chunk3); \ + layer0_chunk2 = _mm_unpacklo_epi16(layer3_chunk1, layer3_chunk4); \ + layer0_chunk3 = _mm_unpackhi_epi16(layer3_chunk1, layer3_chunk4); \ + layer0_chunk4 = _mm_unpacklo_epi16(layer3_chunk2, layer3_chunk5); \ + layer0_chunk5 = _mm_unpackhi_epi16(layer3_chunk2, layer3_chunk5); \ + } + +#define _MM_INTERLIV_EPI16(v_r0, v_r1, v_g0, v_g1, v_b0, v_b1) \ + { \ + __m128i v_mask = _mm_set1_epi32(0x0000ffff); \ + \ + __m128i layer3_chunk0 = _mm_packus_epi32(_mm_and_si128(v_r0, v_mask), _mm_and_si128(v_r1, v_mask)); \ + __m128i layer3_chunk3 = _mm_packus_epi32(_mm_srli_epi32(v_r0, 16), _mm_srli_epi32(v_r1, 16)); \ + __m128i layer3_chunk1 = _mm_packus_epi32(_mm_and_si128(v_g0, v_mask), _mm_and_si128(v_g1, v_mask)); \ + __m128i layer3_chunk4 = _mm_packus_epi32(_mm_srli_epi32(v_g0, 16), _mm_srli_epi32(v_g1, 16)); \ + __m128i layer3_chunk2 = _mm_packus_epi32(_mm_and_si128(v_b0, v_mask), _mm_and_si128(v_b1, v_mask)); \ + __m128i layer3_chunk5 = _mm_packus_epi32(_mm_srli_epi32(v_b0, 16), _mm_srli_epi32(v_b1, 16)); \ + \ + __m128i layer2_chunk0 = _mm_packus_epi32(_mm_and_si128(layer3_chunk0, v_mask), _mm_and_si128(layer3_chunk1, v_mask)); \ + __m128i layer2_chunk3 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk0, 16), _mm_srli_epi32(layer3_chunk1, 16)); \ + __m128i layer2_chunk1 = _mm_packus_epi32(_mm_and_si128(layer3_chunk2, v_mask), _mm_and_si128(layer3_chunk3, v_mask)); \ + __m128i layer2_chunk4 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk2, 16), _mm_srli_epi32(layer3_chunk3, 16)); \ + __m128i layer2_chunk2 = _mm_packus_epi32(_mm_and_si128(layer3_chunk4, v_mask), _mm_and_si128(layer3_chunk5, v_mask)); \ + __m128i layer2_chunk5 = _mm_packus_epi32(_mm_srli_epi32(layer3_chunk4, 16), _mm_srli_epi32(layer3_chunk5, 16)); \ + \ + __m128i layer1_chunk0 = _mm_packus_epi32(_mm_and_si128(layer2_chunk0, v_mask), _mm_and_si128(layer2_chunk1, v_mask)); \ + __m128i layer1_chunk3 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk0, 16), _mm_srli_epi32(layer2_chunk1, 16)); \ + __m128i layer1_chunk1 = _mm_packus_epi32(_mm_and_si128(layer2_chunk2, v_mask), _mm_and_si128(layer2_chunk3, v_mask)); \ + __m128i layer1_chunk4 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk2, 16), _mm_srli_epi32(layer2_chunk3, 16)); \ + __m128i layer1_chunk2 = _mm_packus_epi32(_mm_and_si128(layer2_chunk4, v_mask), _mm_and_si128(layer2_chunk5, v_mask)); \ + __m128i layer1_chunk5 = _mm_packus_epi32(_mm_srli_epi32(layer2_chunk4, 16), _mm_srli_epi32(layer2_chunk5, 16)); \ + \ + v_r0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk0, v_mask), _mm_and_si128(layer1_chunk1, v_mask)); \ + v_g1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk0, 16), _mm_srli_epi32(layer1_chunk1, 16)); \ + v_r1 = _mm_packus_epi32(_mm_and_si128(layer1_chunk2, v_mask), _mm_and_si128(layer1_chunk3, v_mask)); \ + v_b0 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk2, 16), _mm_srli_epi32(layer1_chunk3, 16)); \ + v_g0 = _mm_packus_epi32(_mm_and_si128(layer1_chunk4, v_mask), _mm_and_si128(layer1_chunk5, v_mask)); \ + v_b1 = _mm_packus_epi32(_mm_srli_epi32(layer1_chunk4, 16), _mm_srli_epi32(layer1_chunk5, 16)); \ + } + +#define _MM_DEINTERLIV_PS(layer0_chunk0, layer0_chunk1, layer0_chunk2, \ + layer0_chunk3, layer0_chunk4, layer0_chunk5) \ + { \ + __m128 layer1_chunk0 = _mm_unpacklo_ps(layer0_chunk0, layer0_chunk3); \ + __m128 layer1_chunk1 = _mm_unpackhi_ps(layer0_chunk0, layer0_chunk3); \ + __m128 layer1_chunk2 = _mm_unpacklo_ps(layer0_chunk1, layer0_chunk4); \ + __m128 layer1_chunk3 = _mm_unpackhi_ps(layer0_chunk1, layer0_chunk4); \ + __m128 layer1_chunk4 = _mm_unpacklo_ps(layer0_chunk2, layer0_chunk5); \ + __m128 layer1_chunk5 = _mm_unpackhi_ps(layer0_chunk2, layer0_chunk5); \ + \ + __m128 layer2_chunk0 = _mm_unpacklo_ps(layer1_chunk0, layer1_chunk3); \ + __m128 layer2_chunk1 = _mm_unpackhi_ps(layer1_chunk0, layer1_chunk3); \ + __m128 layer2_chunk2 = _mm_unpacklo_ps(layer1_chunk1, layer1_chunk4); \ + __m128 layer2_chunk3 = _mm_unpackhi_ps(layer1_chunk1, layer1_chunk4); \ + __m128 layer2_chunk4 = _mm_unpacklo_ps(layer1_chunk2, layer1_chunk5); \ + __m128 layer2_chunk5 = _mm_unpackhi_ps(layer1_chunk2, layer1_chunk5); \ + \ + layer0_chunk0 = _mm_unpacklo_ps(layer2_chunk0, layer2_chunk3); \ + layer0_chunk1 = _mm_unpackhi_ps(layer2_chunk0, layer2_chunk3); \ + layer0_chunk2 = _mm_unpacklo_ps(layer2_chunk1, layer2_chunk4); \ + layer0_chunk3 = _mm_unpackhi_ps(layer2_chunk1, layer2_chunk4); \ + layer0_chunk4 = _mm_unpacklo_ps(layer2_chunk2, layer2_chunk5); \ + layer0_chunk5 = _mm_unpackhi_ps(layer2_chunk2, layer2_chunk5); \ + } + #endif namespace cv