ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2022 THL A29 Limited, a Tencent company. All rights reserved.
4
//
5
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6
// in compliance with the License. You may obtain a copy of the License at
7
//
8
// https://opensource.org/licenses/BSD-3-Clause
9
//
10
// Unless required by applicable law or agreed to in writing, software distributed
11
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
13
// specific language governing permissions and limitations under the License.
14

15
#version 450
16

17
#if NCNN_fp16_storage
18
#extension GL_EXT_shader_16bit_storage: require
19
struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
20
#endif
21
#if NCNN_fp16_arithmetic
22
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
23
#endif
24

25
#define shape_constant_id_offset 0
26
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
27

28
layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
29
layout (constant_id = shape_constant_id_offset + 2) const int outh = 0;
30

31
#if NCNN_image_shader
32
layout (binding = 0) uniform unfp sampler3D bottom_blob;
33
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
34
layout (binding = 2) uniform unfp sampler3D weight_blob;
35
#else
36
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
37
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
38
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
39
#endif
40

41
layout (push_constant) uniform parameter
42
{
43
    int w;
44

45
    int outw;
46
    int outh;
47
} p;
48

49
void main()
50
{
51
    int gx = int(gl_GlobalInvocationID.x);
52
    int gy = int(gl_GlobalInvocationID.y);
53
    int gz = int(gl_GlobalInvocationID.z);
54

55
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= 1)
56
        return;
57

58
    const int start = gx * 8;
59
    const int end = min(gx * 8 + 8, psc(w));
60

61
    afpvec8 sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
62

63
#if NCNN_image_shader
64
    int wx = start * 8;
65

66
    for (int i = start; i < end; i++)
67
    {
68
        afpvec4 v = image3d_ld4(bottom_blob, ivec3(i, 0, 0));
69

70
        afpvec4 k0 = image3d_ld4(weight_blob, ivec3(wx + 0, gy, 0));
71
        afpvec4 k1 = image3d_ld4(weight_blob, ivec3(wx + 1, gy, 0));
72
        afpvec4 k2 = image3d_ld4(weight_blob, ivec3(wx + 2, gy, 0));
73
        afpvec4 k3 = image3d_ld4(weight_blob, ivec3(wx + 3, gy, 0));
74
        afpvec4 k4 = image3d_ld4(weight_blob, ivec3(wx + 4, gy, 0));
75
        afpvec4 k5 = image3d_ld4(weight_blob, ivec3(wx + 5, gy, 0));
76
        afpvec4 k6 = image3d_ld4(weight_blob, ivec3(wx + 6, gy, 0));
77
        afpvec4 k7 = image3d_ld4(weight_blob, ivec3(wx + 7, gy, 0));
78

79
        // sum += v * k;
80
        sum[0].r += dot(v, k0);
81
        sum[0].g += dot(v, k1);
82
        sum[0].b += dot(v, k2);
83
        sum[0].a += dot(v, k3);
84
        sum[1].r += dot(v, k4);
85
        sum[1].g += dot(v, k5);
86
        sum[1].b += dot(v, k6);
87
        sum[1].a += dot(v, k7);
88

89
        wx += 8;
90
    }
91
#else
92
    int w_offset = (gy * psc(w) + start) * 8;
93

94
    for (int i = start; i < end; i++)
95
    {
96
        afpvec4 v = buffer_ld4(bottom_blob_data, i);
97

98
        afpvec4 k0 = buffer_ld4(weight_data, w_offset + 0);
99
        afpvec4 k1 = buffer_ld4(weight_data, w_offset + 1);
100
        afpvec4 k2 = buffer_ld4(weight_data, w_offset + 2);
101
        afpvec4 k3 = buffer_ld4(weight_data, w_offset + 3);
102
        afpvec4 k4 = buffer_ld4(weight_data, w_offset + 4);
103
        afpvec4 k5 = buffer_ld4(weight_data, w_offset + 5);
104
        afpvec4 k6 = buffer_ld4(weight_data, w_offset + 6);
105
        afpvec4 k7 = buffer_ld4(weight_data, w_offset + 7);
106

107
        // sum += v * k;
108
        sum[0].r += dot(v, k0);
109
        sum[0].g += dot(v, k1);
110
        sum[0].b += dot(v, k2);
111
        sum[0].a += dot(v, k3);
112
        sum[1].r += dot(v, k4);
113
        sum[1].g += dot(v, k5);
114
        sum[1].b += dot(v, k6);
115
        sum[1].a += dot(v, k7);
116

117
        w_offset += 8;
118
    }
119
#endif
120

121
#if NCNN_image_shader
122
    image3d_st8(top_blob, ivec3(gx, gy, 0), sum);
123
#else
124
    const int gi = gy * psc(outw) + gx;
125
    buffer_st8(top_blob_data, gi, sum);
126
#endif
127
}
128