ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2023 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
#endif
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
22
#endif
23

24
layout (constant_id = 0) const int op_type = 0;
25

26
#define shape_constant_id_offset 1
27
layout (constant_id = shape_constant_id_offset + 0) const int adims = 0;
28
layout (constant_id = shape_constant_id_offset + 1) const int aw = 0;
29
layout (constant_id = shape_constant_id_offset + 2) const int ah = 0;
30
layout (constant_id = shape_constant_id_offset + 3) const int ad = 0;
31
layout (constant_id = shape_constant_id_offset + 4) const int ac = 0;
32
layout (constant_id = shape_constant_id_offset + 5) const int acstep = 0;
33

34
layout (constant_id = shape_constant_id_offset + 6) const int bdims = 0;
35
layout (constant_id = shape_constant_id_offset + 7) const int bw = 0;
36
layout (constant_id = shape_constant_id_offset + 8) const int bh = 0;
37
layout (constant_id = shape_constant_id_offset + 9) const int bd = 0;
38
layout (constant_id = shape_constant_id_offset + 10) const int bc = 0;
39
layout (constant_id = shape_constant_id_offset + 11) const int bcstep = 0;
40

41
layout (constant_id = shape_constant_id_offset + 12) const int outdims = 0;
42
layout (constant_id = shape_constant_id_offset + 13) const int outw = 0;
43
layout (constant_id = shape_constant_id_offset + 14) const int outh = 0;
44
layout (constant_id = shape_constant_id_offset + 15) const int outd = 0;
45
layout (constant_id = shape_constant_id_offset + 16) const int outc = 0;
46
layout (constant_id = shape_constant_id_offset + 17) const int outcstep = 0;
47

48
#if NCNN_image_shader
49
layout (binding = 0) uniform unfp sampler3D a_blob_3d;
50
layout (binding = 1) uniform unfp sampler3D b_blob_3d;
51
layout (binding = 2, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
52
#else
53
layout (binding = 0) readonly buffer a_blob { sfpvec4 a_blob_data[]; };
54
layout (binding = 1) readonly buffer b_blob { sfp b_blob_data[]; };
55
layout (binding = 2) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
56
#endif
57

58
layout (push_constant) uniform parameter
59
{
60
    int adims;
61
    int aw;
62
    int ah;
63
    int ad;
64
    int ac;
65
    int acstep;
66

67
    int bdims;
68
    int bw;
69
    int bh;
70
    int bd;
71
    int bc;
72
    int bcstep;
73

74
    int outdims;
75
    int outw;
76
    int outh;
77
    int outd;
78
    int outc;
79
    int outcstep;
80
} p;
81

82
void main()
83
{
84
    int gx = int(gl_GlobalInvocationID.x);
85
    int gy = int(gl_GlobalInvocationID.y);
86
    int gz = int(gl_GlobalInvocationID.z);
87

88
    if (gx >= psc(outw) || gy >= psc(outh) * psc(outd) || gz >= psc(outc))
89
        return;
90

91
    int yd = gy / psc(outh);
92
    int yh = gy % psc(outh);
93

94
    // explicit broadcast
95
    int ax = min(gx, psc(aw) - 1);
96
    int ay = min(yd, psc(ad) - 1) * psc(ah) + min(yh, psc(ah) - 1);
97
    int az = min(gz, psc(ac) - 1);
98

99
    int bx = min(gx, psc(bw) - 1);
100
    int by = min(yd, psc(bd) - 1) * psc(bh) + min(yh, psc(bh) - 1);
101
    int bz = min(gz, psc(bc) - 1);
102

103
#if NCNN_image_shader
104
    afpvec4 v1 = image3d_ld4(a_blob_3d, ivec3(ax, ay, az));
105
    afpvec4 v2 = afpvec4(image3d_ld1(b_blob_3d, ivec3(bx, by, bz)));
106
#else
107
    int ai = az * psc(acstep) + ay * psc(aw) + ax;
108
    int bi = bz * psc(bcstep) + by * psc(bw) + bx;
109

110
    afpvec4 v1 = buffer_ld4(a_blob_data, ai);
111
    afpvec4 v2 = afpvec4(buffer_ld1(b_blob_data, bi));
112
#endif
113

114
    afpvec4 res;
115

116
    if (op_type == 0) res = v1 + v2;
117
    if (op_type == 1) res = v1 - v2;
118
    if (op_type == 2) res = v1 * v2;
119
    if (op_type == 3) res = v1 / v2;
120
    if (op_type == 4) res = max(v1, v2);
121
    if (op_type == 5) res = min(v1, v2);
122
    if (op_type == 6) res = pow(v1, v2);
123
    if (op_type == 7) res = v2 - v1;
124
    if (op_type == 8) res = v2 / v1;
125
    if (op_type == 9) res = pow(v2, v1);
126
#if NCNN_moltenvk
127
    if (op_type == 10) res = afpvec4(atan(vec4(v1), vec4(v2)));
128
    if (op_type == 11) res = afpvec4(atan(vec4(v2), vec4(v1)));
129
#else
130
    if (op_type == 10) res = atan(v1, v2);
131
    if (op_type == 11) res = atan(v2, v1);
132
#endif
133

134
#if NCNN_image_shader
135
    image3d_st4(top_blob_3d, ivec3(gx, gy, gz), res);
136
#else
137
    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
138
    buffer_st4(top_blob_data, gi, res);
139
#endif
140
}
141