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
#endif
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
22
#endif
23

24
#define LOCAL_MEMORY_UNROLL_INCH 8
25

26
layout (constant_id = 0) const int maxk = 1;
27

28
#define shape_constant_id_offset 1
29
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
30
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
31
layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
32
layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
33

34
layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
35
layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
36

37
#if NCNN_image_shader
38
layout (binding = 0) uniform unfp sampler3D bottom_blob;
39
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D col_blob;
40
layout (binding = 2) uniform unfp sampler3D weight_blob;
41
#else
42
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
43
layout (binding = 1) writeonly buffer col_blob { sfp col_blob_data[]; };
44
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
45
#endif
46

47
layout (push_constant) uniform parameter
48
{
49
    int w;
50
    int h;
51
    int c;
52
    int cstep;
53

54
    int outw;
55
    int outh;
56
} p;
57

58
#if NCNN_shader_local_memory
59
shared lfpvec4 tmp_v[8][LOCAL_MEMORY_UNROLL_INCH][4];
60
shared lfpvec4 tmp_k[8][LOCAL_MEMORY_UNROLL_INCH];
61
#endif
62

63
void main()
64
{
65
    int gx = int(gl_GlobalInvocationID.x) * 4;
66
    int gy = int(gl_GlobalInvocationID.y);
67

68
#if !NCNN_shader_local_memory
69
    if (gx >= psc(outw) || gy >= psc(outh))
70
        return;
71
#endif
72

73
    afp sum0 = afp(0.f);
74
    afp sum1 = afp(0.f);
75
    afp sum2 = afp(0.f);
76
    afp sum3 = afp(0.f);
77

78
#if NCNN_image_shader
79
    ivec4 gx4 = gx + ivec4(0, 1, 2, 3);
80

81
    ivec4 sy4 = gx4 / psc(w);
82
    ivec4 sx4 = gx4 % psc(w);
83

84
    for (int z = 0; z < psc(c); z++)
85
    {
86
        afpvec4 v0 = image3d_ld4(bottom_blob, ivec3(sx4.r, sy4.r, z));
87
        afpvec4 v1 = image3d_ld4(bottom_blob, ivec3(sx4.g, sy4.g, z));
88
        afpvec4 v2 = image3d_ld4(bottom_blob, ivec3(sx4.b, sy4.b, z));
89
        afpvec4 v3 = image3d_ld4(bottom_blob, ivec3(sx4.a, sy4.a, z));
90

91
        afpvec4 k = image3d_ld4(weight_blob, ivec3(z, gy, 0));
92

93
        sum0 += dot(v0, k);
94
        sum1 += dot(v1, k);
95
        sum2 += dot(v2, k);
96
        sum3 += dot(v3, k);
97
    }
98
#else
99
    int v_offset = gx;
100
    int w_offset = gy * psc(c);
101

102
#if NCNN_shader_local_memory
103
    const int lx = int(gl_LocalInvocationID.x);
104
    const int ly = int(gl_LocalInvocationID.y);
105

106
    int z = 0;
107
    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < psc(c); z += LOCAL_MEMORY_UNROLL_INCH)
108
    {
109
        if (ly < 4)
110
        {
111
            for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
112
            {
113
                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_blob_data[v_offset + z4 * psc(cstep) + ly]);
114
            }
115
        }
116

117
        if (lx == 0)
118
        {
119
            for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
120
            {
121
                tmp_k[ly][z4] = sfp2lfpvec4(weight_data[w_offset + z4]);
122
            }
123
        }
124

125
        barrier();
126

127
        for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
128
        {
129
            afpvec4 v0 = lfp2afpvec4(tmp_v[lx][z4][0]);
130
            afpvec4 v1 = lfp2afpvec4(tmp_v[lx][z4][1]);
131
            afpvec4 v2 = lfp2afpvec4(tmp_v[lx][z4][2]);
132
            afpvec4 v3 = lfp2afpvec4(tmp_v[lx][z4][3]);
133

134
            afpvec4 k = lfp2afpvec4(tmp_k[ly][z4]);
135

136
            sum0 += dot(v0, k);
137
            sum1 += dot(v1, k);
138
            sum2 += dot(v2, k);
139
            sum3 += dot(v3, k);
140
        }
141

142
        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(cstep);
143
        w_offset += LOCAL_MEMORY_UNROLL_INCH;
144

145
        barrier();
146
    }
147

148
    if (z < psc(c))
149
    {
150
        const int remain = psc(c) - z;
151

152
        if (ly < 4)
153
        {
154
            for (int z4 = 0; z4 < remain; z4++)
155
            {
156
                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_blob_data[v_offset + z4 * psc(cstep) + ly]);
157
            }
158
        }
159

160
        if (lx == 0)
161
        {
162
            for (int z4 = 0; z4 < remain; z4++)
163
            {
164
                tmp_k[ly][z4] = sfp2lfpvec4(weight_data[w_offset + z4]);
165
            }
166
        }
167

168
        barrier();
169

170
        for (int z4 = 0; z4 < remain; z4++)
171
        {
172
            afpvec4 v0 = lfp2afpvec4(tmp_v[lx][z4][0]);
173
            afpvec4 v1 = lfp2afpvec4(tmp_v[lx][z4][1]);
174
            afpvec4 v2 = lfp2afpvec4(tmp_v[lx][z4][2]);
175
            afpvec4 v3 = lfp2afpvec4(tmp_v[lx][z4][3]);
176

177
            afpvec4 k = lfp2afpvec4(tmp_k[ly][z4]);
178

179
            sum0 += dot(v0, k);
180
            sum1 += dot(v1, k);
181
            sum2 += dot(v2, k);
182
            sum3 += dot(v3, k);
183
        }
184
    }
185
#else
186
    for (int z = 0; z < psc(c); z++)
187
    {
188
        afpvec4 v0 = buffer_ld4(bottom_blob_data, v_offset + 0);
189
        afpvec4 v1 = buffer_ld4(bottom_blob_data, v_offset + 1);
190
        afpvec4 v2 = buffer_ld4(bottom_blob_data, v_offset + 2);
191
        afpvec4 v3 = buffer_ld4(bottom_blob_data, v_offset + 3);
192

193
        afpvec4 k = buffer_ld4(weight_data, w_offset);
194

195
        sum0 += dot(v0, k);
196
        sum1 += dot(v1, k);
197
        sum2 += dot(v2, k);
198
        sum3 += dot(v3, k);
199

200
        v_offset += psc(cstep);
201
        w_offset += 1;
202
    }
203
#endif
204
#endif
205

206
#if NCNN_shader_local_memory
207
    if (gx >= psc(outw) || gy >= psc(outh))
208
        return;
209
#endif
210

211
#if NCNN_image_shader
212
    image3d_st1(col_blob, ivec3(gx4.r, gy, 0), sum0);
213
    image3d_st1(col_blob, ivec3(gx4.g, gy, 0), sum1);
214
    image3d_st1(col_blob, ivec3(gx4.b, gy, 0), sum2);
215
    image3d_st1(col_blob, ivec3(gx4.a, gy, 0), sum3);
216
#else
217
    const int gi = gy * psc(outw) + gx;
218

219
    buffer_st1(col_blob_data, gi, sum0);
220
    if (gx + 1 < psc(outw)) buffer_st1(col_blob_data, gi + 1, sum1);
221
    if (gx + 2 < psc(outw)) buffer_st1(col_blob_data, gi + 2, sum2);
222
    if (gx + 3 < psc(outw)) buffer_st1(col_blob_data, gi + 3, sum3);
223
#endif
224
}
225