erode_test/erode_neon.cc at master · ek9852/erode_test · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
#if defined(__ARM_NEON__) || defined(__aarch64__)

#include "erode_neon.h"
#include <arm_neon.h>
#include <assert.h>

static inline void __attribute__((always_inline))
erode3x3_neon_kernel(uint8_t *in_data, uint8_t *out_data, const int stride,
		const bool parallel_2_rows, const bool left, const bool top, const bool right, const bool bottom)
{
    uint8x8_t p0, p1, p2, p3, p4, p5, p6, p7, p8, p9, p10, p11;

	// | 1 | .... | 8 | 8 |
    uint8_t         *in_ptr   = left ? in_data : (right ? in_data - 8 : in_data - 1);
    uint8x16_t mid_data = vld1q_u8(in_ptr);
    uint8x16_t top_data = top ? mid_data : vld1q_u8(in_ptr - stride);
    uint8x16_t bot_data = (bottom && !parallel_2_rows) ? mid_data : vld1q_u8(in_ptr + stride);
    uint8x16_t bot_2_data;

    if (parallel_2_rows) bot_2_data = bottom ? bot_data : vld1q_u8(in_ptr + 2 * stride);

    if (right) {
        p0 = vget_high_u8(top_data);
        p1 = vext_u8(vget_high_u8(top_data), vrev64_u8(vget_high_u8(top_data)), 1);
        p2 = vext_u8(vget_low_u8(top_data), vget_high_u8(top_data), 7);

        p3 = vget_high_u8(mid_data);
        p4 = vext_u8(vget_high_u8(mid_data), vrev64_u8(vget_high_u8(mid_data)), 1);
        p5 = vext_u8(vget_low_u8(mid_data), vget_high_u8(mid_data), 7);

        p6 = vget_high_u8(bot_data);
        p7 = vext_u8(vget_high_u8(bot_data), vrev64_u8(vget_high_u8(bot_data)), 1);
        p8 = vext_u8(vget_low_u8(bot_data), vget_high_u8(bot_data), 7);

        if (parallel_2_rows) {
            p9 = vget_high_u8(bot_2_data);
            p10 = vext_u8(vget_high_u8(bot_2_data), vrev64_u8(vget_high_u8(bot_2_data)), 1);
            p11 = vext_u8(vget_low_u8(bot_2_data), vget_high_u8(bot_2_data), 7);
        }
    } else if (left) {
        p0 = vget_low_u8(top_data);
        p1 = vext_u8(vrev64_u8(vget_low_u8(top_data)), vget_low_u8(top_data), 7);
        p2 = vext_u8(vget_low_u8(top_data), vget_high_u8(top_data), 1);

        p3 = vget_low_u8(mid_data);
        p4 = vext_u8(vrev64_u8(vget_low_u8(mid_data)), vget_low_u8(mid_data), 7);
        p5 = vext_u8(vget_low_u8(mid_data), vget_high_u8(mid_data), 1);

        p6 = vget_low_u8(bot_data);
        p7 = vext_u8(vrev64_u8(vget_low_u8(bot_data)), vget_low_u8(bot_data), 7);
        p8 = vext_u8(vget_low_u8(bot_data), vget_high_u8(bot_data), 1);

        if (parallel_2_rows) {
            p9 = vget_low_u8(bot_2_data);
            p10 = vext_u8(vrev64_u8(vget_low_u8(bot_2_data)), vget_low_u8(bot_2_data), 7);
            p11 = vext_u8(vget_low_u8(bot_2_data), vget_high_u8(bot_2_data), 1);
        }
    } else {
        p0 = vget_low_u8(top_data);
        p1 = vext_u8(vget_low_u8(top_data), vget_high_u8(top_data), 1);
        p2 = vext_u8(vget_low_u8(top_data), vget_high_u8(top_data), 2);

        p3 = vget_low_u8(mid_data);
        p4 = vext_u8(vget_low_u8(mid_data), vget_high_u8(mid_data), 1);
        p5 = vext_u8(vget_low_u8(mid_data), vget_high_u8(mid_data), 2);

        p6 = vget_low_u8(bot_data);
        p7 = vext_u8(vget_low_u8(bot_data), vget_high_u8(bot_data), 1);
        p8 = vext_u8(vget_low_u8(bot_data), vget_high_u8(bot_data), 2);

        if (parallel_2_rows) {
            p9 = vget_low_u8(bot_2_data);
            p10 = vext_u8(vget_low_u8(bot_2_data), vget_high_u8(bot_2_data), 1);
            p11 = vext_u8(vget_low_u8(bot_2_data), vget_high_u8(bot_2_data), 2);
        }
    }
    p3 = vmin_u8(p3, p4);
    p3 = vmin_u8(p3, p5);
    p3 = vmin_u8(p3, p6);
    p3 = vmin_u8(p3, p7);
    p3 = vmin_u8(p3, p8);

    p0 = vmin_u8(p0, p3);
    p0 = vmin_u8(p0, p1);
    p0 = vmin_u8(p0, p2);

    vst1_u8(out_data, p0);

    if (parallel_2_rows) {
        p9 = vmin_u8(p3, p9);
        p9 = vmin_u8(p9, p10);
        p9 = vmin_u8(p9, p11);

        vst1_u8(out_data+stride, p9);
    }
}

void
erode3x3_neon(uint8_t *in_data, uint8_t *out_data, int w, int h)
{
    int i, j;

    assert(w >= 16);
    assert(h > 2);

    // top row
    erode3x3_neon_kernel(in_data, out_data, w, true, true, true, false, false);
    for (i = 8; i < w - 8; i+=8) {
        erode3x3_neon_kernel(in_data+i, out_data+i, w, true, false, true, false, false);
    }
    // move backward last one, so that it align to 8 bytes
    erode3x3_neon_kernel(in_data+w-8, out_data+w-8, w, true, false, true, true, false);
    in_data += 2*w;
    out_data += 2*w;

    // middle row
    for (j = 2; j < h - 2; j+=2) {
        erode3x3_neon_kernel(in_data, out_data, w, true, true, false, false, false);
        for (i = 8; i < w - 8; i+=8) {
            erode3x3_neon_kernel(in_data+i, out_data+i, w, true, false, false, false, false);
        }
        // move backward last one, so that it align to 8 bytes
        erode3x3_neon_kernel(in_data+w-8, out_data+w-8, w, true, false, false, true, false);
        in_data += 2*w;
        out_data += 2*w;
    }

    // left over 2 rows or 1 row
    bool two_rows = (j == (h-2));

    erode3x3_neon_kernel(in_data, out_data, w, two_rows, true, false, false, true);
    for (i = 8; i < w - 8; i+=8) {
        erode3x3_neon_kernel(in_data+i, out_data+i, w, two_rows, false, false, false, true);
    }
    // move backward last one, so that it align to 8 bytes
    erode3x3_neon_kernel(in_data+w-8, out_data+w-8, w, two_rows, false, false, true, true);
}
#endif