ncnn

softmax_reduce_sum_pack8.comp
205 строк · 6.4 Кб
Перенос по словам
1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2020 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 axis = 0;
26

27
#define shape_constant_id_offset 1
28
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
29
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
30
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
31
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
32
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
33

34
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
35
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
36
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
37
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
38
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
39

40
#if NCNN_image_shader
41
layout (binding = 0) uniform unfp sampler3D bottom_top_blob_3d;
42
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D sum_workspace_3d;
43
#else
44
layout (binding = 0) readonly buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
45
layout (binding = 1) writeonly buffer sum_workspace { sfpvec8 sum_workspace_data[]; };
46
#endif
47

48
layout (push_constant) uniform parameter
49
{
50
    int dims;
51
    int w;
52
    int h;
53
    int c;
54
    int cstep;
55

56
    int outdims;
57
    int outw;
58
    int outh;
59
    int outc;
60
    int outcstep;
61
} p;
62

63
void main()
64
{
65
    int gx = int(gl_GlobalInvocationID.x);
66
    int gy = int(gl_GlobalInvocationID.y);
67
    int gz = int(gl_GlobalInvocationID.z);
68

69
    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
70
        return;
71

72
    int positive_axis = axis < 0 ? psc(dims) + axis : axis;
73
    afpvec8 sum_value = afpvec8(afpvec4(0.f), afpvec4(0.f));
74

75
    if (psc(dims) == 1) // positive_axis == 0
76
    {
77
        for (int i = 0; i < psc(w); i++)
78
        {
79
#if NCNN_image_shader
80
            afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, 0, 0));
81
#else
82
            afpvec8 v = buffer_ld8(bottom_top_blob_data, i);
83
#endif
84
            sum_value += v;
85
        }
86
        afpvec4 sum4 = sum_value[0] + sum_value[1];
87
        afpvec2 sum2 = sum4.rg + sum4.ba;
88
        afp sum1 = sum2.r + sum2.g;
89
        sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
90
#if NCNN_image_shader
91
        image3d_st8(sum_workspace_3d, ivec3(0, 0, 0), sum_value);
92
#else
93
        buffer_st8(sum_workspace_data, 0, sum_value);
94
#endif
95
        return;
96
    }
97

98
    if (psc(dims) == 2 && positive_axis == 0)
99
    {
100
        for (int i = 0; i < psc(h); i++)
101
        {
102
#if NCNN_image_shader
103
            afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, i, 0));
104
#else
105
            int v_offset = i * psc(w) + gx;
106
            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
107
#endif
108
            sum_value += v;
109
        }
110
        afpvec4 sum4 = sum_value[0] + sum_value[1];
111
        afpvec2 sum2 = sum4.rg + sum4.ba;
112
        afp sum1 = sum2.r + sum2.g;
113
        sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
114
#if NCNN_image_shader
115
        image3d_st8(sum_workspace_3d, ivec3(gx, 0, 0), sum_value);
116
#else
117
        buffer_st8(sum_workspace_data, gx, sum_value);
118
#endif
119
        return;
120
    }
121

122
    if (psc(dims) == 2 && positive_axis == 1)
123
    {
124
        for (int i = 0; i < psc(w); i++)
125
        {
126
#if NCNN_image_shader
127
            afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, gx, 0));
128
#else
129
            int v_offset = gx * psc(w) + i;
130
            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
131
#endif
132
            sum_value += v;
133
        }
134
#if NCNN_image_shader
135
        image3d_st8(sum_workspace_3d, ivec3(gx, 0, 0), sum_value);
136
#else
137
        buffer_st8(sum_workspace_data, gx, sum_value);
138
#endif
139
        return;
140
    }
141

142
    if (psc(dims) == 3 && positive_axis == 0)
143
    {
144
        for (int i = 0; i < psc(c); i++)
145
        {
146
#if NCNN_image_shader
147
            afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, gy, i));
148
#else
149
            int v_offset = i * psc(cstep) + gy * psc(w) + gx;
150
            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
151
#endif
152
            sum_value += v;
153
        }
154
        afpvec4 sum4 = sum_value[0] + sum_value[1];
155
        afpvec2 sum2 = sum4.rg + sum4.ba;
156
        afp sum1 = sum2.r + sum2.g;
157
        sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
158
#if NCNN_image_shader
159
        image3d_st8(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
160
#else
161
        buffer_st8(sum_workspace_data, gy * psc(w) + gx, sum_value);
162
#endif
163
        return;
164
    }
165

166
    if (psc(dims) == 3 && positive_axis == 1)
167
    {
168
        for (int i = 0; i < psc(h); i++)
169
        {
170
#if NCNN_image_shader
171
            afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, i, gy));
172
#else
173
            int v_offset = gy * psc(cstep) + i * psc(w) + gx;
174
            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
175
#endif
176
            sum_value += v;
177
        }
178
#if NCNN_image_shader
179
        image3d_st8(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
180
#else
181
        buffer_st8(sum_workspace_data, gy * psc(w) + gx, sum_value);
182
#endif
183
        return;
184
    }
185

186
    if (psc(dims) == 3 && positive_axis == 2)
187
    {
188
        for (int i = 0; i < psc(w); i++)
189
        {
190
#if NCNN_image_shader
191
            afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, gx, gy));
192
#else
193
            int v_offset = gy * psc(cstep) + gx * psc(w) + i;
194
            afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
195
#endif
196
            sum_value += v;
197
        }
198
#if NCNN_image_shader
199
        image3d_st8(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
200
#else
201
        buffer_st8(sum_workspace_data, gy * psc(h) + gx, sum_value);
202
#endif
203
        return;
204
    }
205
}
206
ncnn

Использование cookies
