1 | /*M/////////////////////////////////////////////////////////////////////////////////////// |
2 | // |
3 | // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. |
4 | // |
5 | // By downloading, copying, installing or using the software you agree to this license. |
6 | // If you do not agree to this license, do not download, install, |
7 | // copy or use the software. |
8 | // |
9 | // |
10 | // License Agreement |
11 | // For Open Source Computer Vision Library |
12 | // |
13 | // Copyright (C) 2000-2008, Intel Corporation, all rights reserved. |
14 | // Copyright (C) 2009, Willow Garage Inc., all rights reserved. |
15 | // Copyright (C) 2014-2015, Itseez Inc., all rights reserved. |
16 | // Third party copyrights are property of their respective owners. |
17 | // |
18 | // Redistribution and use in source and binary forms, with or without modification, |
19 | // are permitted provided that the following conditions are met: |
20 | // |
21 | // * Redistribution's of source code must retain the above copyright notice, |
22 | // this list of conditions and the following disclaimer. |
23 | // |
24 | // * Redistribution's in binary form must reproduce the above copyright notice, |
25 | // this list of conditions and the following disclaimer in the documentation |
26 | // and/or other materials provided with the distribution. |
27 | // |
28 | // * The name of the copyright holders may not be used to endorse or promote products |
29 | // derived from this software without specific prior written permission. |
30 | // |
31 | // This software is provided by the copyright holders and contributors "as is" and |
32 | // any express or implied warranties, including, but not limited to, the implied |
33 | // warranties of merchantability and fitness for a particular purpose are disclaimed. |
34 | // In no event shall the Intel Corporation or contributors be liable for any direct, |
35 | // indirect, incidental, special, exemplary, or consequential damages |
36 | // (including, but not limited to, procurement of substitute goods or services; |
37 | // loss of use, data, or profits; or business interruption) however caused |
38 | // and on any theory of liability, whether in contract, strict liability, |
39 | // or tort (including negligence or otherwise) arising in any way out of |
40 | // the use of this software, even if advised of the possibility of such damage. |
41 | // |
42 | //M*/ |
43 | |
44 | #include "precomp.hpp" |
45 | #undef CV_FORCE_SIMD128_CPP // expected AVX implementation only |
46 | #include "opencv2/core/hal/intrin.hpp" |
47 | #include "corner.hpp" |
48 | |
49 | namespace cv |
50 | { |
51 | |
52 | // load three 8-packed float vector and deinterleave |
53 | // probably it's better to write down somewhere else |
54 | static void load_deinterleave(const float* ptr, __m256& a, __m256& b, __m256& c) |
55 | { |
56 | __m256 s0 = _mm256_loadu_ps(p: ptr); // a0, b0, c0, a1, b1, c1, a2, b2, |
57 | __m256 s1 = _mm256_loadu_ps(p: ptr + 8); // c2, a3, b3, c3, a4, b4, c4, a5, |
58 | __m256 s2 = _mm256_loadu_ps(p: ptr + 16); // b5, c5, a6, b6, c6, a7, b7, c7, |
59 | __m256 s3 = _mm256_permute2f128_ps(s1, s2, 0x21); // a4, b4, c4, a5, b5, c5, a6, b6, |
60 | __m256 s4 = _mm256_permute2f128_ps(s2, s2, 0x33); // c6, a7, b7, c7, c6, a7, b7, c7, |
61 | |
62 | __m256 v00 = _mm256_unpacklo_ps(a: s0, b: s3); // a0, a4, b0, b4, b1, b5, c1, c5, |
63 | __m256 v01 = _mm256_unpackhi_ps(a: s0, b: s3); // c0, c4, a1, a5, a2, a6, b2, b6, |
64 | __m256 v02 = _mm256_unpacklo_ps(a: s1, b: s4); // c2, c6, a3, a7, x, x, x, x, |
65 | __m256 v03 = _mm256_unpackhi_ps(a: s1, b: s4); // b3, b7, c3, c7, x, x, x, x, |
66 | __m256 v04 = _mm256_permute2f128_ps(v02, v03, 0x20); // c2, c6, a3, a7, b3, b7, c3, c7, |
67 | __m256 v05 = _mm256_permute2f128_ps(v01, v03, 0x21); // a2, a6, b2, b6, b3, b7, c3, c7, |
68 | |
69 | __m256 v10 = _mm256_unpacklo_ps(a: v00, b: v05); // a0, a2, a4, a6, b1, b3, b5, b7, |
70 | __m256 v11 = _mm256_unpackhi_ps(a: v00, b: v05); // b0, b2, b4, b6, c1, c3, c5, c7, |
71 | __m256 v12 = _mm256_unpacklo_ps(a: v01, b: v04); // c0, c2, c4, c6, x, x, x, x, |
72 | __m256 v13 = _mm256_unpackhi_ps(a: v01, b: v04); // a1, a3, a5, a7, x, x, x, x, |
73 | __m256 v14 = _mm256_permute2f128_ps(v11, v12, 0x20); // b0, b2, b4, b6, c0, c2, c4, c6, |
74 | __m256 v15 = _mm256_permute2f128_ps(v10, v11, 0x31); // b1, b3, b5, b7, c1, c3, c5, c7, |
75 | |
76 | __m256 v20 = _mm256_unpacklo_ps(a: v14, b: v15); // b0, b1, b2, b3, c0, c1, c2, c3, |
77 | __m256 v21 = _mm256_unpackhi_ps(a: v14, b: v15); // b4, b5, b6, b7, c4, c5, c6, c7, |
78 | __m256 v22 = _mm256_unpacklo_ps(a: v10, b: v13); // a0, a1, a2, a3, x, x, x, x, |
79 | __m256 v23 = _mm256_unpackhi_ps(a: v10, b: v13); // a4, a5, a6, a7, x, x, x, x, |
80 | |
81 | a = _mm256_permute2f128_ps(v22, v23, 0x20); // a0, a1, a2, a3, a4, a5, a6, a7, |
82 | b = _mm256_permute2f128_ps(v20, v21, 0x20); // b0, b1, b2, b3, b4, b5, b6, b7, |
83 | c = _mm256_permute2f128_ps(v20, v21, 0x31); // c0, c1, c2, c3, c4, c5, c6, c7, |
84 | } |
85 | |
86 | // realign four 3-packed vector to three 4-packed vector |
87 | static void v_pack4x3to3x4(const __m128i& s0, const __m128i& s1, const __m128i& s2, const __m128i& s3, __m128i& d0, __m128i& d1, __m128i& d2) |
88 | { |
89 | d0 = _mm_or_si128(a: s0, _mm_slli_si128(s1, 12)); |
90 | d1 = _mm_or_si128(_mm_srli_si128(s1, 4), _mm_slli_si128(s2, 8)); |
91 | d2 = _mm_or_si128(_mm_srli_si128(s2, 8), _mm_slli_si128(s3, 4)); |
92 | } |
93 | |
94 | // separate high and low 128 bit and cast to __m128i |
95 | static void v_separate_lo_hi(const __m256& src, __m128i& lo, __m128i& hi) |
96 | { |
97 | lo = _mm_castps_si128(a: _mm256_castps256_ps128(a: src)); |
98 | hi = _mm_castps_si128(_mm256_extractf128_ps(src, 1)); |
99 | } |
100 | |
101 | // interleave three 8-float vector and store |
102 | static void store_interleave(float* ptr, const __m256& a, const __m256& b, const __m256& c) |
103 | { |
104 | __m128i a0, a1, b0, b1, c0, c1; |
105 | v_separate_lo_hi(src: a, lo&: a0, hi&: a1); |
106 | v_separate_lo_hi(src: b, lo&: b0, hi&: b1); |
107 | v_separate_lo_hi(src: c, lo&: c0, hi&: c1); |
108 | |
109 | v_uint32x4 z = v_setzero_u32(); |
110 | v_uint32x4 u0, u1, u2, u3; |
111 | v_transpose4x4(a0: v_uint32x4(a0), a1: v_uint32x4(b0), a2: v_uint32x4(c0), a3: z, b0&: u0, b1&: u1, b2&: u2, b3&: u3); |
112 | v_pack4x3to3x4(s0: u0.val, s1: u1.val, s2: u2.val, s3: u3.val, d0&: a0, d1&: b0, d2&: c0); |
113 | v_transpose4x4(a0: v_uint32x4(a1), a1: v_uint32x4(b1), a2: v_uint32x4(c1), a3: z, b0&: u0, b1&: u1, b2&: u2, b3&: u3); |
114 | v_pack4x3to3x4(s0: u0.val, s1: u1.val, s2: u2.val, s3: u3.val, d0&: a1, d1&: b1, d2&: c1); |
115 | |
116 | #if !defined(__GNUC__) || defined(__INTEL_COMPILER) |
117 | _mm256_storeu_ps(ptr, _mm256_castsi256_ps(_mm256_setr_m128i(a0, b0))); |
118 | _mm256_storeu_ps(ptr + 8, _mm256_castsi256_ps(_mm256_setr_m128i(c0, a1))); |
119 | _mm256_storeu_ps(ptr + 16, _mm256_castsi256_ps(_mm256_setr_m128i(b1, c1))); |
120 | #else |
121 | // GCC: workaround for missing AVX intrinsic: "_mm256_setr_m128()" |
122 | _mm256_storeu_ps(p: ptr, _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_castsi128_ps(a0)), _mm_castsi128_ps(b0), 1)); |
123 | _mm256_storeu_ps(p: ptr + 8, _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_castsi128_ps(c0)), _mm_castsi128_ps(a1), 1)); |
124 | _mm256_storeu_ps(p: ptr + 16, _mm256_insertf128_ps(_mm256_castps128_ps256(_mm_castsi128_ps(b1)), _mm_castsi128_ps(c1), 1)); |
125 | #endif |
126 | } |
127 | |
128 | int calcMinEigenValLine_AVX(const float* cov, float* dst, int width) |
129 | { |
130 | int j = 0; |
131 | __m256 half = _mm256_set1_ps(w: 0.5f); |
132 | for (; j <= width - 8; j += 8) |
133 | { |
134 | __m256 v_a, v_b, v_c, v_t; |
135 | load_deinterleave(ptr: cov + j * 3, a&: v_a, b&: v_b, c&: v_c); |
136 | v_a = _mm256_mul_ps(a: v_a, b: half); |
137 | v_c = _mm256_mul_ps(a: v_c, b: half); |
138 | v_t = _mm256_sub_ps(a: v_a, b: v_c); |
139 | v_t = _mm256_add_ps(a: _mm256_mul_ps(a: v_b, b: v_b), b: _mm256_mul_ps(a: v_t, b: v_t)); |
140 | _mm256_storeu_ps(p: dst + j, a: _mm256_sub_ps(a: _mm256_add_ps(a: v_a, b: v_c), b: _mm256_sqrt_ps(a: v_t))); |
141 | } |
142 | return j; |
143 | } |
144 | |
145 | int calcHarrisLine_AVX(const float* cov, float* dst, double k, int width) |
146 | { |
147 | int j = 0; |
148 | __m256 v_k = _mm256_set1_ps(w: (float)k); |
149 | |
150 | for (; j <= width - 8; j += 8) |
151 | { |
152 | __m256 v_a, v_b, v_c; |
153 | load_deinterleave(ptr: cov + j * 3, a&: v_a, b&: v_b, c&: v_c); |
154 | |
155 | __m256 v_ac_bb = _mm256_sub_ps(a: _mm256_mul_ps(a: v_a, b: v_c), b: _mm256_mul_ps(a: v_b, b: v_b)); |
156 | __m256 v_ac = _mm256_add_ps(a: v_a, b: v_c); |
157 | __m256 v_dst = _mm256_sub_ps(a: v_ac_bb, b: _mm256_mul_ps(a: v_k, b: _mm256_mul_ps(a: v_ac, b: v_ac))); |
158 | _mm256_storeu_ps(p: dst + j, a: v_dst); |
159 | } |
160 | return j; |
161 | } |
162 | |
163 | int cornerEigenValsVecsLine_AVX(const float* dxdata, const float* dydata, float* cov_data, int width) |
164 | { |
165 | int j = 0; |
166 | for (; j <= width - 8; j += 8) |
167 | { |
168 | __m256 v_dx = _mm256_loadu_ps(p: dxdata + j); |
169 | __m256 v_dy = _mm256_loadu_ps(p: dydata + j); |
170 | |
171 | __m256 v_dst0, v_dst1, v_dst2; |
172 | v_dst0 = _mm256_mul_ps(a: v_dx, b: v_dx); |
173 | v_dst1 = _mm256_mul_ps(a: v_dx, b: v_dy); |
174 | v_dst2 = _mm256_mul_ps(a: v_dy, b: v_dy); |
175 | |
176 | store_interleave(ptr: cov_data + j * 3, a: v_dst0, b: v_dst1, c: v_dst2); |
177 | } |
178 | return j; |
179 | } |
180 | |
181 | } |
182 | /* End of file */ |
183 | |