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
layout (constant_id = 0) const int batch = 1;
26

27
#define shape_constant_id_offset 1
28
layout (constant_id = shape_constant_id_offset + 0) const int c = 0;
29
layout (constant_id = shape_constant_id_offset + 1) const int cstep = 0;
30

31
layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
32
layout (constant_id = shape_constant_id_offset + 3) const int outc = 0;
33
layout (constant_id = shape_constant_id_offset + 4) const int outcstep = 0;
34

35
#if NCNN_image_shader
36
layout (binding = 0) uniform unfp sampler3D bottom_tm_blob;
37
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_tm_blob;
38
layout (binding = 2) uniform unfp sampler3D weight_tm_blob;
39
#else
40
layout (binding = 0) readonly buffer bottom_tm_blob { sfpvec8 bottom_tm_blob_data[]; };
41
layout (binding = 1) writeonly buffer top_tm_blob { sfpvec4 top_tm_blob_data[]; };
42
layout (binding = 2) readonly buffer weight_tm_blob { sfpvec8 weight_tm_data[]; };
43
#endif
44

45
layout (push_constant) uniform parameter
46
{
47
    int c;
48
    int cstep;
49

50
    int outw;
51
    int outc;
52
    int outcstep;
53
} p;
54

55
void main()
56
{
57
    int gx = int(gl_GlobalInvocationID.x) * 4;
58
    int gy = int(gl_GlobalInvocationID.y);
59
    int gz = int(gl_GlobalInvocationID.z);
60

61
    if (gx >= psc(outw) || gy >= psc(outc) || gz >= batch)
62
        return;
63

64
    afpvec4 sum0 = afpvec4(0.f);
65
    afpvec4 sum1 = afpvec4(0.f);
66
    afpvec4 sum2 = afpvec4(0.f);
67
    afpvec4 sum3 = afpvec4(0.f);
68

69
#if NCNN_image_shader
70
    for (int z = 0; z < psc(c); z++)
71
    {
72
        afpvec8 v0 = image3d_ld8(bottom_tm_blob, ivec3(gx + 0, z, gz));
73
        afpvec8 v1 = image3d_ld8(bottom_tm_blob, ivec3(gx + 1, z, gz));
74
        afpvec8 v2 = image3d_ld8(bottom_tm_blob, ivec3(gx + 2, z, gz));
75
        afpvec8 v3 = image3d_ld8(bottom_tm_blob, ivec3(gx + 3, z, gz));
76

77
        afpvec8 k0 = image3d_ld8(weight_tm_blob, ivec3(z * 4 + 0, gy, gz));
78
        afpvec8 k1 = image3d_ld8(weight_tm_blob, ivec3(z * 4 + 1, gy, gz));
79
        afpvec8 k2 = image3d_ld8(weight_tm_blob, ivec3(z * 4 + 2, gy, gz));
80
        afpvec8 k3 = image3d_ld8(weight_tm_blob, ivec3(z * 4 + 3, gy, gz));
81

82
        // sum += v * k;
83
        sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
84
        sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
85
        sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
86
        sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
87

88
        sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
89
        sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
90
        sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
91
        sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
92

93
        sum2.r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);
94
        sum2.g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);
95
        sum2.b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);
96
        sum2.a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);
97

98
        sum3.r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);
99
        sum3.g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);
100
        sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
101
        sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
102
    }
103
#else
104
    int v_offset = gz * psc(cstep) + gx;
105
    int w_offset = (gz * psc(c) * psc(outc) + gy * psc(c)) * 4;
106

107
    for (int z = 0; z < psc(c); z++)
108
    {
109
        afpvec8 v0 = buffer_ld8(bottom_tm_blob_data, v_offset + 0);
110
        afpvec8 v1 = buffer_ld8(bottom_tm_blob_data, v_offset + 1);
111
        afpvec8 v2 = buffer_ld8(bottom_tm_blob_data, v_offset + 2);
112
        afpvec8 v3 = buffer_ld8(bottom_tm_blob_data, v_offset + 3);
113

114
        afpvec8 k0 = buffer_ld8(weight_tm_data, w_offset + 0);
115
        afpvec8 k1 = buffer_ld8(weight_tm_data, w_offset + 1);
116
        afpvec8 k2 = buffer_ld8(weight_tm_data, w_offset + 2);
117
        afpvec8 k3 = buffer_ld8(weight_tm_data, w_offset + 3);
118

119
        // sum += v * k;
120
        sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
121
        sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
122
        sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
123
        sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
124

125
        sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
126
        sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
127
        sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
128
        sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
129

130
        sum2.r += dot(v2[0], k0[0]) + dot(v2[1], k0[1]);
131
        sum2.g += dot(v2[0], k1[0]) + dot(v2[1], k1[1]);
132
        sum2.b += dot(v2[0], k2[0]) + dot(v2[1], k2[1]);
133
        sum2.a += dot(v2[0], k3[0]) + dot(v2[1], k3[1]);
134

135
        sum3.r += dot(v3[0], k0[0]) + dot(v3[1], k0[1]);
136
        sum3.g += dot(v3[0], k1[0]) + dot(v3[1], k1[1]);
137
        sum3.b += dot(v3[0], k2[0]) + dot(v3[1], k2[1]);
138
        sum3.a += dot(v3[0], k3[0]) + dot(v3[1], k3[1]);
139

140
        v_offset += psc(outw);
141
        w_offset += 4;
142
    }
143
#endif
144

145
#if NCNN_image_shader
146
    image3d_st4(top_tm_blob, ivec3(gx + 0, gy, gz), sum0);
147
    image3d_st4(top_tm_blob, ivec3(gx + 1, gy, gz), sum1);
148
    image3d_st4(top_tm_blob, ivec3(gx + 2, gy, gz), sum2);
149
    image3d_st4(top_tm_blob, ivec3(gx + 3, gy, gz), sum3);
150
#else
151
    int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
152

153
    buffer_st4(top_tm_blob_data, gi + 0, sum0);
154
    if (gx + 1 < psc(outw)) buffer_st4(top_tm_blob_data, gi + 1, sum1);
155
    if (gx + 2 < psc(outw)) buffer_st4(top_tm_blob_data, gi + 2, sum2);
156
    if (gx + 3 < psc(outw)) buffer_st4(top_tm_blob_data, gi + 3, sum3);
157
#endif
158
}
159