Hugging Face's logo

camenduru
/
ncnn

Model card Files
xet
Community
ncnn
File size: 10,202 Bytes
be903e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
 
// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2019 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#version 450

#if NCNN_fp16_storage
#extension GL_EXT_shader_16bit_storage: require
#endif
#if NCNN_fp16_arithmetic
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
#endif

#extension GL_GOOGLE_include_directive: enable
#include "vulkan_activation.comp"

layout (constant_id = 0) const int kernel_w = 1;
layout (constant_id = 1) const int kernel_h = 1;
layout (constant_id = 2) const int dilation_w = 1;
layout (constant_id = 3) const int dilation_h = 1;
layout (constant_id = 4) const int stride_w = 1;
layout (constant_id = 5) const int stride_h = 1;
layout (constant_id = 6) const int bias_term = 0;
layout (constant_id = 7) const int activation_type = 0;
layout (constant_id = 8) const float activation_param_0 = 0;
layout (constant_id = 9) const float activation_param_1 = 0;

#define shape_constant_id_offset 10
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;

layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;

#if NCNN_image_shader
layout (binding = 0) uniform unfp sampler3D bottom_blob;
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
layout (binding = 2) uniform unfp sampler3D weight_blob;
layout (binding = 3) uniform unfp sampler3D bias_blob;
#else
layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
#else
layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
#endif
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
#endif

layout (push_constant) uniform parameter
{
    int dims;
    int w;
    int h;
    int c;
    int cstep;

    int outdims;
    int outw;
    int outh;
    int outc;
    int outcstep;
} p;

void main()
{
    int gx = int(gl_GlobalInvocationID.x) * 2;
    int gy = int(gl_GlobalInvocationID.y) * 2;
    int gz = int(gl_GlobalInvocationID.z) * 2;

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

    const ivec2 gx2 = gx + ivec2(0, 1);
    const ivec2 gy2 = gy + ivec2(0, 1);
    const ivec2 gz2 = gz + ivec2(0, 1);

    afpvec4 sum0;
    afpvec4 sum1;
    afpvec4 sum2;
    afpvec4 sum3;
    afpvec4 sum4;
    afpvec4 sum5;
    afpvec4 sum6;
    afpvec4 sum7;

    if (bias_term == 1)
    {
#if NCNN_image_shader
        sum0 = image3d_ld4(bias_blob, ivec3(gz2.x, 0, 0));
        sum4 = image3d_ld4(bias_blob, ivec3(gz2.y, 0, 0));
#else
        sum0 = buffer_ld4(bias_data, gz2.x);
        sum4 = buffer_ld4(bias_data, gz2.y);
#endif
        sum1 = sum0;
        sum2 = sum0;
        sum3 = sum0;
        sum5 = sum4;
        sum6 = sum4;
        sum7 = sum4;
    }
    else
    {
        sum0 = afpvec4(0.f);
        sum1 = afpvec4(0.f);
        sum2 = afpvec4(0.f);
        sum3 = afpvec4(0.f);
        sum4 = afpvec4(0.f);
        sum5 = afpvec4(0.f);
        sum6 = afpvec4(0.f);
        sum7 = afpvec4(0.f);
    }

#if NCNN_image_shader
    for (int z = 0; z < psc(c); z++)
    {
        ivec2 sy = gy2 * stride_h;
        int wx = 0;

        for (int y = 0; y < kernel_h; y++)
        {
            ivec2 sx = gx2 * stride_w;

            for (int x = 0; x < kernel_w; x++)
            {
                afpvec4 v0 = image3d_ld4(bottom_blob, ivec3(sx.x, sy.x, z));
                afpvec4 v1 = image3d_ld4(bottom_blob, ivec3(sx.y, sy.x, z));
                afpvec4 v2 = image3d_ld4(bottom_blob, ivec3(sx.x, sy.y, z));
                afpvec4 v3 = image3d_ld4(bottom_blob, ivec3(sx.y, sy.y, z));

                afpmat4 k0 = afpmat4(
                    image3d_ld4(weight_blob, ivec3(wx + 0, z, gz2.x)),
                    image3d_ld4(weight_blob, ivec3(wx + 1, z, gz2.x)),
                    image3d_ld4(weight_blob, ivec3(wx + 2, z, gz2.x)),
                    image3d_ld4(weight_blob, ivec3(wx + 3, z, gz2.x))
                );
                afpmat4 k1 = afpmat4(
                    image3d_ld4(weight_blob, ivec3(wx + 0, z, gz2.y)),
                    image3d_ld4(weight_blob, ivec3(wx + 1, z, gz2.y)),
                    image3d_ld4(weight_blob, ivec3(wx + 2, z, gz2.y)),
                    image3d_ld4(weight_blob, ivec3(wx + 3, z, gz2.y))
                );

                sum0 += v0 * k0;
                sum1 += v1 * k0;
                sum2 += v2 * k0;
                sum3 += v3 * k0;
                sum4 += v0 * k1;
                sum5 += v1 * k1;
                sum6 += v2 * k1;
                sum7 += v3 * k1;

                sx += dilation_w;
                wx += 4;
            }

            sy += dilation_h;
        }
    }
#else
    ivec2 w_offset = gz2 * psc(c) * kernel_w * kernel_h;

    for (int z = 0; z < psc(c); z++)
    {
        ivec4 v_offset;
        v_offset.rg = z * psc(cstep) + gy2.x * stride_h * psc(w) + gx2 * stride_w;
        v_offset.ba = z * psc(cstep) + gy2.y * stride_h * psc(w) + gx2 * stride_w;

        for (int y = 0; y < kernel_h; y++)
        {
            for (int x = 0; x < kernel_w; x++)
            {
                afpvec4 v0 = buffer_ld4(bottom_blob_data, v_offset.r + x * dilation_w);
                afpvec4 v1 = buffer_ld4(bottom_blob_data, v_offset.g + x * dilation_w);
                afpvec4 v2 = buffer_ld4(bottom_blob_data, v_offset.b + x * dilation_w);
                afpvec4 v3 = buffer_ld4(bottom_blob_data, v_offset.a + x * dilation_w);

#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
                // GL_EXT_shader_16bit_storage does not define f16mat4 type :(
                afpmat4 k0 = afpmat4(
                    buffer_ld4(weight_data, (w_offset.x + x) * 4 + 0),
                    buffer_ld4(weight_data, (w_offset.x + x) * 4 + 1),
                    buffer_ld4(weight_data, (w_offset.x + x) * 4 + 2),
                    buffer_ld4(weight_data, (w_offset.x + x) * 4 + 3)
                );
                afpmat4 k1 = afpmat4(
                    buffer_ld4(weight_data, (w_offset.y + x) * 4 + 0),
                    buffer_ld4(weight_data, (w_offset.y + x) * 4 + 1),
                    buffer_ld4(weight_data, (w_offset.y + x) * 4 + 2),
                    buffer_ld4(weight_data, (w_offset.y + x) * 4 + 3)
                );
#else
                afpmat4 k0 = sfp2afpmat4(weight_data[w_offset.x + x]);
                afpmat4 k1 = sfp2afpmat4(weight_data[w_offset.y + x]);
#endif

                sum0 += v0 * k0;
                sum1 += v1 * k0;
                sum2 += v2 * k0;
                sum3 += v3 * k0;
                sum4 += v0 * k1;
                sum5 += v1 * k1;
                sum6 += v2 * k1;
                sum7 += v3 * k1;
            }

            v_offset += dilation_h * psc(w);
            w_offset += kernel_w;
        }
    }
#endif

    sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
    sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
    sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
    sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
    sum4 = activation_afpvec4(sum4, activation_type, activation_param_0, activation_param_1);
    sum5 = activation_afpvec4(sum5, activation_type, activation_param_0, activation_param_1);
    sum6 = activation_afpvec4(sum6, activation_type, activation_param_0, activation_param_1);
    sum7 = activation_afpvec4(sum7, activation_type, activation_param_0, activation_param_1);

#if NCNN_image_shader
    image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
    image3d_st4(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
    image3d_st4(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
    image3d_st4(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
    image3d_st4(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
#else
    const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;

    buffer_st4(top_blob_data, gi.x, sum0);
    if (gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi.x + 1, sum1);
    if (gy + 1 < psc(outh)) buffer_st4(top_blob_data, gi.x + psc(outw), sum2);
    if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi.x + psc(outw) + 1, sum3);
    if (gz + 1 < psc(outc))
    {
        buffer_st4(top_blob_data, gi.y, sum4);
        if (gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi.y + 1, sum5);
        if (gy + 1 < psc(outh)) buffer_st4(top_blob_data, gi.y + psc(outw), sum6);
        if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi.y + psc(outw) + 1, sum7);
    }
#endif
}