Hugging Face's logo

camenduru
/
ncnn

Model card Files
xet
Community
ncnn
File size: 8,909 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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
 
// 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"

#define LOCAL_MEMORY_UNROLL_INCH 8

layout (constant_id = 0) const int bias_term = 0;
layout (constant_id = 1) const int activation_type = 0;
layout (constant_id = 2) const float activation_param_0 = 0;
layout (constant_id = 3) const float activation_param_1 = 0;

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

layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
layout (constant_id = shape_constant_id_offset + 6) const int outc = 0;
layout (constant_id = shape_constant_id_offset + 7) 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[]; };
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
#endif

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

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

#if NCNN_shader_local_memory
shared lfpvec4 tmp_v[8][LOCAL_MEMORY_UNROLL_INCH][4];
shared lfpvec4 tmp_k[8][LOCAL_MEMORY_UNROLL_INCH][4];
#endif

void main()
{
    int gx = int(gl_GlobalInvocationID.x) * 4;
    int gy = int(gl_GlobalInvocationID.y);

#if NCNN_image_shader
    if (gx >= psc(outw) * psc(outh) || gy >= psc(outc))
        return;
#else
#if !NCNN_shader_local_memory
    if (gx >= psc(outcstep) || gy >= psc(outc))
        return;
#endif
#endif

    afpvec4 sum0;
    afpvec4 sum1;
    afpvec4 sum2;
    afpvec4 sum3;

    if (bias_term == 1)
    {
#if NCNN_image_shader
        afpvec4 b = image3d_ld4(bias_blob, ivec3(gy, 0, 0));
#else
        afpvec4 b = buffer_ld4(bias_data, gy);
#endif
        sum0 = b;
        sum1 = b;
        sum2 = b;
        sum3 = b;
    }
    else
    {
        sum0 = afpvec4(0.f);
        sum1 = afpvec4(0.f);
        sum2 = afpvec4(0.f);
        sum3 = afpvec4(0.f);
    }

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

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

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

        afpmat4 k = afpmat4(
            image3d_ld4(weight_blob, ivec3(0, z, gy)),
            image3d_ld4(weight_blob, ivec3(1, z, gy)),
            image3d_ld4(weight_blob, ivec3(2, z, gy)),
            image3d_ld4(weight_blob, ivec3(3, z, gy))
        );

        sum0 += v0 * k;
        sum1 += v1 * k;
        sum2 += v2 * k;
        sum3 += v3 * k;
    }
#else
    int w_offset = gy * psc(c) * 4;
    int v_offset = gx;

#if NCNN_shader_local_memory
    const int lx = int(gl_LocalInvocationID.x);
    const int ly = int(gl_LocalInvocationID.y);

    int z = 0;
    for (; z + (LOCAL_MEMORY_UNROLL_INCH - 1) < psc(c); z += LOCAL_MEMORY_UNROLL_INCH)
    {
        if (ly < 4)
        {
            for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
            {
                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_blob_data[v_offset + z4 * psc(cstep) + ly]);
            }
        }

        if (lx < 4)
        {
            for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
            {
                tmp_k[ly][z4][lx] = sfp2lfpvec4(weight_data[w_offset + z4 * 4 + lx]);
            }
        }

        barrier();

        for (int z4 = 0; z4 < LOCAL_MEMORY_UNROLL_INCH; z4++)
        {
            afpvec4 v0 = lfp2afpvec4(tmp_v[lx][z4][0]);
            afpvec4 v1 = lfp2afpvec4(tmp_v[lx][z4][1]);
            afpvec4 v2 = lfp2afpvec4(tmp_v[lx][z4][2]);
            afpvec4 v3 = lfp2afpvec4(tmp_v[lx][z4][3]);

            afpvec4 k0 = lfp2afpvec4(tmp_k[ly][z4][0]);
            afpvec4 k1 = lfp2afpvec4(tmp_k[ly][z4][1]);
            afpvec4 k2 = lfp2afpvec4(tmp_k[ly][z4][2]);
            afpvec4 k3 = lfp2afpvec4(tmp_k[ly][z4][3]);

            afpmat4 k = afpmat4(k0, k1, k2, k3);

            sum0 += v0 * k;
            sum1 += v1 * k;
            sum2 += v2 * k;
            sum3 += v3 * k;
        }

        v_offset += LOCAL_MEMORY_UNROLL_INCH * psc(cstep);
        w_offset += LOCAL_MEMORY_UNROLL_INCH * 4;

        barrier();
    }

    if (z < psc(c))
    {
        const int remain = psc(c) - z;

        if (ly < 4)
        {
            for (int z4 = 0; z4 < remain; z4++)
            {
                tmp_v[lx][z4][ly] = sfp2lfpvec4(bottom_blob_data[v_offset + z4 * psc(cstep) + ly]);
            }
        }

        if (lx < 4)
        {
            for (int z4 = 0; z4 < remain; z4++)
            {
                tmp_k[ly][z4][lx] = sfp2lfpvec4(weight_data[w_offset + z4 * 4 + lx]);
            }
        }

        barrier();

        for (int z4 = 0; z4 < remain; z4++)
        {
            afpvec4 v0 = lfp2afpvec4(tmp_v[lx][z4][0]);
            afpvec4 v1 = lfp2afpvec4(tmp_v[lx][z4][1]);
            afpvec4 v2 = lfp2afpvec4(tmp_v[lx][z4][2]);
            afpvec4 v3 = lfp2afpvec4(tmp_v[lx][z4][3]);

            afpvec4 k0 = lfp2afpvec4(tmp_k[ly][z4][0]);
            afpvec4 k1 = lfp2afpvec4(tmp_k[ly][z4][1]);
            afpvec4 k2 = lfp2afpvec4(tmp_k[ly][z4][2]);
            afpvec4 k3 = lfp2afpvec4(tmp_k[ly][z4][3]);

            afpmat4 k = afpmat4(k0, k1, k2, k3);

            sum0 += v0 * k;
            sum1 += v1 * k;
            sum2 += v2 * k;
            sum3 += v3 * k;
        }
    }
#else
    for (int z = 0; z < psc(c); z++)
    {
        afpvec4 v0 = buffer_ld4(bottom_blob_data, v_offset + 0);
        afpvec4 v1 = buffer_ld4(bottom_blob_data, v_offset + 1);
        afpvec4 v2 = buffer_ld4(bottom_blob_data, v_offset + 2);
        afpvec4 v3 = buffer_ld4(bottom_blob_data, v_offset + 3);

        afpmat4 k = afpmat4(
            buffer_ld4(weight_data, w_offset + 0),
            buffer_ld4(weight_data, w_offset + 1),
            buffer_ld4(weight_data, w_offset + 2),
            buffer_ld4(weight_data, w_offset + 3)
        );

        sum0 += v0 * k;
        sum1 += v1 * k;
        sum2 += v2 * k;
        sum3 += v3 * k;

        w_offset += 4;
        v_offset += psc(cstep);
    }
#endif
#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);

#if NCNN_image_shader
    image3d_st4(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);
    image3d_st4(top_blob, ivec3(sx4.g, sy4.g, gy), sum1);
    image3d_st4(top_blob, ivec3(sx4.b, sy4.b, gy), sum2);
    image3d_st4(top_blob, ivec3(sx4.a, sy4.a, gy), sum3);
#else
#if NCNN_shader_local_memory
    if (gx >= psc(outcstep) || gy >= psc(outc))
        return;
#endif

    int gi = gy * psc(outcstep) + gx;

    buffer_st4(top_blob_data, gi + 0, sum0);
    if (gx + 1 < psc(outcstep)) buffer_st4(top_blob_data, gi + 1, sum1);
    if (gx + 2 < psc(outcstep)) buffer_st4(top_blob_data, gi + 2, sum2);
    if (gx + 3 < psc(outcstep)) buffer_st4(top_blob_data, gi + 3, sum3);
#endif
}