ncnn

softmax_reduce_sum_pack4.comp
198 строк · 6.0 Кб
Перенос по словам
1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2019 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 axis = 0;
25

26
#define shape_constant_id_offset 1
27
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
28
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
29
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
30
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
31
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
32

33
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
34
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
35
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
36
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
37
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
38

39
#if NCNN_image_shader
40
layout (binding = 0) uniform unfp sampler3D bottom_top_blob_3d;
41
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D sum_workspace_3d;
42
#else
43
layout (binding = 0) readonly buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
44
layout (binding = 1) writeonly buffer sum_workspace { sfpvec4 sum_workspace_data[]; };
45
#endif
46

47
layout (push_constant) uniform parameter
48
{
49
    int dims;
50
    int w;
51
    int h;
52
    int c;
53
    int cstep;
54

55
    int outdims;
56
    int outw;
57
    int outh;
58
    int outc;
59
    int outcstep;
60
} p;
61

62
void main()
63
{
64
    int gx = int(gl_GlobalInvocationID.x);
65
    int gy = int(gl_GlobalInvocationID.y);
66
    int gz = int(gl_GlobalInvocationID.z);
67

68
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
69
        return;
70

71
    int positive_axis = axis < 0 ? psc(dims) + axis : axis;
72
    afpvec4 sum_value = afpvec4(0.f);
73

74
    if (psc(dims) == 1) // positive_axis == 0
75
    {
76
        for (int i = 0; i < psc(w); i++)
77
        {
78
#if NCNN_image_shader
79
            afpvec4 v = image3d_ld4(bottom_top_blob_3d, ivec3(i, 0, 0));
80
#else
81
            afpvec4 v = buffer_ld4(bottom_top_blob_data, i);
82
#endif
83
            sum_value += v;
84
        }
85
        afpvec2 sum2 = sum_value.rg + sum_value.ba;
86
        sum_value = afpvec4(sum2.r + sum2.g);
87
#if NCNN_image_shader
88
        image3d_st4(sum_workspace_3d, ivec3(0, 0, 0), sum_value);
89
#else
90
        buffer_st4(sum_workspace_data, 0, sum_value);
91
#endif
92
        return;
93
    }
94

95
    if (psc(dims) == 2 && positive_axis == 0)
96
    {
97
        for (int i = 0; i < psc(h); i++)
98
        {
99
#if NCNN_image_shader
100
            afpvec4 v = image3d_ld4(bottom_top_blob_3d, ivec3(gx, i, 0));
101
#else
102
            int v_offset = i * psc(w) + gx;
103
            afpvec4 v = buffer_ld4(bottom_top_blob_data, v_offset);
104
#endif
105
            sum_value += v;
106
        }
107
        afpvec2 sum2 = sum_value.rg + sum_value.ba;
108
        sum_value = afpvec4(sum2.r + sum2.g);
109
#if NCNN_image_shader
110
        image3d_st4(sum_workspace_3d, ivec3(gx, 0, 0), sum_value);
111
#else
112
        buffer_st4(sum_workspace_data, gx, sum_value);
113
#endif
114
        return;
115
    }
116

117
    if (psc(dims) == 2 && positive_axis == 1)
118
    {
119
        for (int i = 0; i < psc(w); i++)
120
        {
121
#if NCNN_image_shader
122
            afpvec4 v = image3d_ld4(bottom_top_blob_3d, ivec3(i, gx, 0));
123
#else
124
            int v_offset = gx * psc(w) + i;
125
            afpvec4 v = buffer_ld4(bottom_top_blob_data, v_offset);
126
#endif
127
            sum_value += v;
128
        }
129
#if NCNN_image_shader
130
        image3d_st4(sum_workspace_3d, ivec3(gx, 0, 0), sum_value);
131
#else
132
        buffer_st4(sum_workspace_data, gx, sum_value);
133
#endif
134
        return;
135
    }
136

137
    if (psc(dims) == 3 && positive_axis == 0)
138
    {
139
        for (int i = 0; i < psc(c); i++)
140
        {
141
#if NCNN_image_shader
142
            afpvec4 v = image3d_ld4(bottom_top_blob_3d, ivec3(gx, gy, i));
143
#else
144
            int v_offset = i * psc(cstep) + gy * psc(w) + gx;
145
            afpvec4 v = buffer_ld4(bottom_top_blob_data, v_offset);
146
#endif
147
            sum_value += v;
148
        }
149
        afpvec2 sum2 = sum_value.rg + sum_value.ba;
150
        sum_value = afpvec4(sum2.r + sum2.g);
151
#if NCNN_image_shader
152
        image3d_st4(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
153
#else
154
        buffer_st4(sum_workspace_data, gy * psc(w) + gx, sum_value);
155
#endif
156
        return;
157
    }
158

159
    if (psc(dims) == 3 && positive_axis == 1)
160
    {
161
        for (int i = 0; i < psc(h); i++)
162
        {
163
#if NCNN_image_shader
164
            afpvec4 v = image3d_ld4(bottom_top_blob_3d, ivec3(gx, i, gy));
165
#else
166
            int v_offset = gy * psc(cstep) + i * psc(w) + gx;
167
            afpvec4 v = buffer_ld4(bottom_top_blob_data, v_offset);
168
#endif
169
            sum_value += v;
170
        }
171
#if NCNN_image_shader
172
        image3d_st4(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
173
#else
174
        buffer_st4(sum_workspace_data, gy * psc(w) + gx, sum_value);
175
#endif
176
        return;
177
    }
178

179
    if (psc(dims) == 3 && positive_axis == 2)
180
    {
181
        for (int i = 0; i < psc(w); i++)
182
        {
183
#if NCNN_image_shader
184
            afpvec4 v = image3d_ld4(bottom_top_blob_3d, ivec3(i, gx, gy));
185
#else
186
            int v_offset = gy * psc(cstep) + gx * psc(w) + i;
187
            afpvec4 v = buffer_ld4(bottom_top_blob_data, v_offset);
188
#endif
189
            sum_value += v;
190
        }
191
#if NCNN_image_shader
192
        image3d_st4(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
193
#else
194
        buffer_st4(sum_workspace_data, gy * psc(h) + gx, sum_value);
195
#endif
196
        return;
197
    }
198
}
199
ncnn

Использование cookies
