ncnn

Форк
0
/
packing_pack1to8_fp32_to_fp16.comp 
215 строк · 7.9 Кб
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 storage_type_from = 0;
26
layout (constant_id = 1) const int storage_type_to = 0;
27

28
#define shape_constant_id_offset 2
29
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
30
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
31
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
32
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
33
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
34

35
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
36
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
37
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
38
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
39
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
40

41
layout (binding = 0) readonly buffer bottom_blob_fp32 { float bottom_blob_fp32_data[]; };
42
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
43
#if NCNN_image_shader
44
layout (binding = 2) uniform highp sampler3D bottom_blob_3d_fp32;
45
layout (binding = 3, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
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
    afpvec8 v;
73

74
    if (psc(dims) == 1)
75
    {
76
        if (storage_type_from == 0)
77
        {
78
            ivec4 x4 = ivec4(gx * 8) + ivec4(0, 1, 2, 3);
79

80
            ivec4 v_offset = x4;
81
            ivec4 vv_offset = x4 + 4;
82

83
            v[0].r = afp(bottom_blob_fp32_data[v_offset.r]);
84
            v[0].g = afp(bottom_blob_fp32_data[v_offset.g]);
85
            v[0].b = afp(bottom_blob_fp32_data[v_offset.b]);
86
            v[0].a = afp(bottom_blob_fp32_data[v_offset.a]);
87
            v[1].r = afp(bottom_blob_fp32_data[vv_offset.r]);
88
            v[1].g = afp(bottom_blob_fp32_data[vv_offset.g]);
89
            v[1].b = afp(bottom_blob_fp32_data[vv_offset.b]);
90
            v[1].a = afp(bottom_blob_fp32_data[vv_offset.a]);
91
        }
92
#if NCNN_image_shader
93
        if (storage_type_from == 1)
94
        {
95
            int x4 = gx * 8;
96

97
            v[0].r = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 0, 0, 0), 0).r);
98
            v[0].g = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 1, 0, 0), 0).r);
99
            v[0].b = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 2, 0, 0), 0).r);
100
            v[0].a = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 3, 0, 0), 0).r);
101
            v[1].r = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 4, 0, 0), 0).r);
102
            v[1].g = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 5, 0, 0), 0).r);
103
            v[1].b = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 6, 0, 0), 0).r);
104
            v[1].a = afp(texelFetch(bottom_blob_3d_fp32, ivec3(x4 + 7, 0, 0), 0).r);
105
        }
106
#endif
107

108
        if (storage_type_to == 0)
109
        {
110
            int gi = gx;
111

112
            buffer_st8(top_blob_data, gi, v);
113
        }
114
#if NCNN_image_shader
115
        if (storage_type_to == 1)
116
        {
117
            image3d_st8(top_blob_3d, ivec3(gx, 0, 0), v);
118
        }
119
#endif
120
    }
121
    else if (psc(dims) == 2)
122
    {
123
        if (storage_type_from == 0)
124
        {
125
            ivec4 y4 = ivec4(gy * 8) + ivec4(0, 1, 2, 3);
126

127
            ivec4 v_offset = y4 * psc(w) + gx;
128
            ivec4 vv_offset = (y4 + 4) * psc(w) + gx;
129

130
            v[0].r = afp(bottom_blob_fp32_data[v_offset.r]);
131
            v[0].g = afp(bottom_blob_fp32_data[v_offset.g]);
132
            v[0].b = afp(bottom_blob_fp32_data[v_offset.b]);
133
            v[0].a = afp(bottom_blob_fp32_data[v_offset.a]);
134
            v[1].r = afp(bottom_blob_fp32_data[vv_offset.r]);
135
            v[1].g = afp(bottom_blob_fp32_data[vv_offset.g]);
136
            v[1].b = afp(bottom_blob_fp32_data[vv_offset.b]);
137
            v[1].a = afp(bottom_blob_fp32_data[vv_offset.a]);
138
        }
139
#if NCNN_image_shader
140
        if (storage_type_from == 1)
141
        {
142
            int y4 = gy * 8;
143

144
            v[0].r = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 0, 0), 0).r);
145
            v[0].g = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 1, 0), 0).r);
146
            v[0].b = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 2, 0), 0).r);
147
            v[0].a = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 3, 0), 0).r);
148
            v[1].r = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 4, 0), 0).r);
149
            v[1].g = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 5, 0), 0).r);
150
            v[1].b = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 6, 0), 0).r);
151
            v[1].a = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, y4 + 7, 0), 0).r);
152
        }
153
#endif
154

155
        if (storage_type_to == 0)
156
        {
157
            int gi = gy * psc(outw) + gx;
158

159
            buffer_st8(top_blob_data, gi, v);
160
        }
161
#if NCNN_image_shader
162
        if (storage_type_to == 1)
163
        {
164
            image3d_st8(top_blob_3d, ivec3(gx, gy, 0), v);
165
        }
166
#endif
167
    }
168
    else // if (psc(dims) == 3)
169
    {
170
        if (storage_type_from == 0)
171
        {
172
            ivec4 z4 = ivec4(gz * 8) + ivec4(0, 1, 2, 3);
173

174
            ivec4 v_offset = z4 * psc(cstep) + ivec4(gy * psc(w) + gx);
175
            ivec4 vv_offset = (z4 + 4) * psc(cstep) + ivec4(gy * psc(w) + gx);
176

177
            v[0].r = afp(bottom_blob_fp32_data[v_offset.r]);
178
            v[0].g = afp(bottom_blob_fp32_data[v_offset.g]);
179
            v[0].b = afp(bottom_blob_fp32_data[v_offset.b]);
180
            v[0].a = afp(bottom_blob_fp32_data[v_offset.a]);
181
            v[1].r = afp(bottom_blob_fp32_data[vv_offset.r]);
182
            v[1].g = afp(bottom_blob_fp32_data[vv_offset.g]);
183
            v[1].b = afp(bottom_blob_fp32_data[vv_offset.b]);
184
            v[1].a = afp(bottom_blob_fp32_data[vv_offset.a]);
185
        }
186
#if NCNN_image_shader
187
        if (storage_type_from == 1)
188
        {
189
            int z4 = gz * 8;
190

191
            v[0].r = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 0), 0).r);
192
            v[0].g = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 1), 0).r);
193
            v[0].b = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 2), 0).r);
194
            v[0].a = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 3), 0).r);
195
            v[1].r = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 4), 0).r);
196
            v[1].g = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 5), 0).r);
197
            v[1].b = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 6), 0).r);
198
            v[1].a = afp(texelFetch(bottom_blob_3d_fp32, ivec3(gx, gy, z4 + 7), 0).r);
199
        }
200
#endif
201

202
        if (storage_type_to == 0)
203
        {
204
            int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
205

206
            buffer_st8(top_blob_data, gi, v);
207
        }
208
#if NCNN_image_shader
209
        if (storage_type_to == 1)
210
        {
211
            image3d_st8(top_blob_3d, ivec3(gx, gy, gz), v);
212
        }
213
#endif
214
    }
215
}
216

Использование cookies

Мы используем файлы cookie в соответствии с Политикой конфиденциальности и Политикой использования cookies.

Нажимая кнопку «Принимаю», Вы даете АО «СберТех» согласие на обработку Ваших персональных данных в целях совершенствования нашего веб-сайта и Сервиса GitVerse, а также повышения удобства их использования.

Запретить использование cookies Вы можете самостоятельно в настройках Вашего браузера.