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#define shape_constant_id_offset 0
25layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
26layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
27layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
28layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
29layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
30
31layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
32layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
33layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
34layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
35layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
36
37#if NCNN_image_shader
38layout (binding = 0) uniform unfp sampler3D bottom_blob;
39layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
40#else
41layout (binding = 0) readonly buffer bottom_blob { sfpvec4 bottom_blob_data[]; };
42layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
43#endif
44layout (binding = 2) readonly buffer alpha_blob { sfpvec4 alpha_blob_data[]; };
45layout (binding = 3) readonly buffer xofs_blob { int xofs_blob_data[]; };
46layout (binding = 4) readonly buffer beta_blob { sfpvec4 beta_blob_data[]; };
47layout (binding = 5) readonly buffer yofs_blob { int yofs_blob_data[]; };
48
49layout (push_constant) uniform parameter
50{
51    int dims;
52    int w;
53    int h;
54    int c;
55    int cstep;
56
57    int outdims;
58    int outw;
59    int outh;
60    int outc;
61    int outcstep;
62} p;
63
64void main()
65{
66    int gx = int(gl_GlobalInvocationID.x);
67    int gy = int(gl_GlobalInvocationID.y);
68    int gz = int(gl_GlobalInvocationID.z);
69
70    if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
71        return;
72
73    if (psc(dims) == 2)
74    {
75        int sx = xofs_blob_data[gx];
76
77#if NCNN_image_shader
78        afpvec4 b0 = image3d_ld4(bottom_blob, ivec3(sx - 1, gy, gz));
79        afpvec4 b1 = image3d_ld4(bottom_blob, ivec3(sx + 0, gy, gz));
80        afpvec4 b2 = image3d_ld4(bottom_blob, ivec3(sx + 1, gy, gz));
81        afpvec4 b3 = image3d_ld4(bottom_blob, ivec3(sx + 2, gy, gz));
82
83        afpmat4 b0123 = afpmat4(b0, b1, b2, b3);
84#else
85        int v_offset_1 = gz * psc(cstep) + gy * psc(w) + sx;
86
87        afpvec4 b0 = buffer_ld4(bottom_blob_data, v_offset_1 - 1);
88        afpvec4 b1 = buffer_ld4(bottom_blob_data, v_offset_1 + 0);
89        afpvec4 b2 = buffer_ld4(bottom_blob_data, v_offset_1 + 1);
90        afpvec4 b3 = buffer_ld4(bottom_blob_data, v_offset_1 + 2);
91
92        afpmat4 b0123 = afpmat4(b0, b1, b2, b3);
93#endif
94
95        afpvec4 alpha = buffer_ld4(alpha_blob_data, gx);
96
97        afpvec4 v = b0123 * alpha;
98
99#if NCNN_image_shader
100        image3d_st4(top_blob, ivec3(gx, gy, gz), v);
101#else
102        const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
103
104        buffer_st4(top_blob_data, gi, v);
105#endif
106        return;
107    }
108
109    int sx = xofs_blob_data[gx];
110    int sy = yofs_blob_data[gy];
111
112#if NCNN_image_shader
113    afpvec4 a0 = image3d_ld4(bottom_blob, ivec3(sx - 1, sy - 1, gz));
114    afpvec4 a1 = image3d_ld4(bottom_blob, ivec3(sx + 0, sy - 1, gz));
115    afpvec4 a2 = image3d_ld4(bottom_blob, ivec3(sx + 1, sy - 1, gz));
116    afpvec4 a3 = image3d_ld4(bottom_blob, ivec3(sx + 2, sy - 1, gz));
117
118    afpmat4 a0123 = afpmat4(a0, a1, a2, a3);
119
120    afpvec4 b0 = image3d_ld4(bottom_blob, ivec3(sx - 1, sy + 0, gz));
121    afpvec4 b1 = image3d_ld4(bottom_blob, ivec3(sx + 0, sy + 0, gz));
122    afpvec4 b2 = image3d_ld4(bottom_blob, ivec3(sx + 1, sy + 0, gz));
123    afpvec4 b3 = image3d_ld4(bottom_blob, ivec3(sx + 2, sy + 0, gz));
124
125    afpmat4 b0123 = afpmat4(b0, b1, b2, b3);
126
127    afpvec4 c0 = image3d_ld4(bottom_blob, ivec3(sx - 1, sy + 1, gz));
128    afpvec4 c1 = image3d_ld4(bottom_blob, ivec3(sx + 0, sy + 1, gz));
129    afpvec4 c2 = image3d_ld4(bottom_blob, ivec3(sx + 1, sy + 1, gz));
130    afpvec4 c3 = image3d_ld4(bottom_blob, ivec3(sx + 2, sy + 1, gz));
131
132    afpmat4 c0123 = afpmat4(c0, c1, c2, c3);
133
134    afpvec4 d0 = image3d_ld4(bottom_blob, ivec3(sx - 1, sy + 2, gz));
135    afpvec4 d1 = image3d_ld4(bottom_blob, ivec3(sx + 0, sy + 2, gz));
136    afpvec4 d2 = image3d_ld4(bottom_blob, ivec3(sx + 1, sy + 2, gz));
137    afpvec4 d3 = image3d_ld4(bottom_blob, ivec3(sx + 2, sy + 2, gz));
138
139    afpmat4 d0123 = afpmat4(d0, d1, d2, d3);
140#else
141    int v_offset_0 = gz * psc(cstep) + (sy - 1) * psc(w) + sx;
142    int v_offset_1 = gz * psc(cstep) + (sy + 0) * psc(w) + sx;
143    int v_offset_2 = gz * psc(cstep) + (sy + 1) * psc(w) + sx;
144    int v_offset_3 = gz * psc(cstep) + (sy + 2) * psc(w) + sx;
145
146    afpvec4 a0 = buffer_ld4(bottom_blob_data, v_offset_0 - 1);
147    afpvec4 a1 = buffer_ld4(bottom_blob_data, v_offset_0 + 0);
148    afpvec4 a2 = buffer_ld4(bottom_blob_data, v_offset_0 + 1);
149    afpvec4 a3 = buffer_ld4(bottom_blob_data, v_offset_0 + 2);
150
151    afpmat4 a0123 = afpmat4(a0, a1, a2, a3);
152
153    afpvec4 b0 = buffer_ld4(bottom_blob_data, v_offset_1 - 1);
154    afpvec4 b1 = buffer_ld4(bottom_blob_data, v_offset_1 + 0);
155    afpvec4 b2 = buffer_ld4(bottom_blob_data, v_offset_1 + 1);
156    afpvec4 b3 = buffer_ld4(bottom_blob_data, v_offset_1 + 2);
157
158    afpmat4 b0123 = afpmat4(b0, b1, b2, b3);
159
160    afpvec4 c0 = buffer_ld4(bottom_blob_data, v_offset_2 - 1);
161    afpvec4 c1 = buffer_ld4(bottom_blob_data, v_offset_2 + 0);
162    afpvec4 c2 = buffer_ld4(bottom_blob_data, v_offset_2 + 1);
163    afpvec4 c3 = buffer_ld4(bottom_blob_data, v_offset_2 + 2);
164
165    afpmat4 c0123 = afpmat4(c0, c1, c2, c3);
166
167    afpvec4 d0 = buffer_ld4(bottom_blob_data, v_offset_3 - 1);
168    afpvec4 d1 = buffer_ld4(bottom_blob_data, v_offset_3 + 0);
169    afpvec4 d2 = buffer_ld4(bottom_blob_data, v_offset_3 + 1);
170    afpvec4 d3 = buffer_ld4(bottom_blob_data, v_offset_3 + 2);
171
172    afpmat4 d0123 = afpmat4(d0, d1, d2, d3);
173#endif
174
175    afpvec4 alpha = buffer_ld4(alpha_blob_data, gx);
176
177    afpvec4 a = a0123 * alpha;
178    afpvec4 b = b0123 * alpha;
179    afpvec4 c = c0123 * alpha;
180    afpvec4 d = d0123 * alpha;
181
182    afpmat4 abcd = afpmat4(a, b, c, d);
183
184    afpvec4 beta = buffer_ld4(beta_blob_data, gy);
185
186    afpvec4 v = abcd * beta;
187
188#if NCNN_image_shader
189    image3d_st4(top_blob, ivec3(gx, gy, gz), v);
190#else
191    const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
192
193    buffer_st4(top_blob_data, gi, v);
194#endif
195}
196