1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2018 THL A29 Limited, a Tencent company. All rights reserved.
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
8
// https://opensource.org/licenses/BSD-3-Clause
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.
18
#extension GL_EXT_shader_16bit_storage: require
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
24
layout (constant_id = 0) const int type = 1;
25
layout (constant_id = 1) const float value = 0;
26
layout (constant_id = 2) const int per_channel_pad = 0;
28
#define shape_constant_id_offset 3
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;
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;
42
layout (binding = 0) uniform unfp sampler3D bottom_blob;
43
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
44
layout (binding = 2) uniform unfp sampler3D per_channel_pad_blob;
46
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
47
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
48
layout (binding = 2) readonly buffer per_channel_pad_blob { sfp per_channel_pad_blob_data[]; };
51
layout (push_constant) uniform parameter
72
int gx = int(gl_GlobalInvocationID.x);
73
int gy = int(gl_GlobalInvocationID.y);
74
int gz = int(gl_GlobalInvocationID.z);
76
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
85
if (x >= 0 && x < psc(w))
88
image3d_cp1(top_blob, ivec3(gx, 0, 0), bottom_blob, ivec3(x, 0, 0));
90
buffer_cp1(top_blob_data, gx, bottom_blob_data, x);
97
image3d_st1(top_blob, ivec3(gx, 0, 0), v);
99
buffer_st1(top_blob_data, gx, v);
105
x = clamp(x, 0, psc(w) - 1);
108
image3d_cp1(top_blob, ivec3(gx, 0, 0), bottom_blob, ivec3(x, 0, 0));
110
buffer_cp1(top_blob_data, gx, bottom_blob_data, x);
116
// NOTE psc(X) get zeros on nvidia
117
// TODO only enable this workaround for some nvidia driver
118
x = (p.w - 1) - abs(x - (p.w - 1));
119
// x = (psc(w) - 1) - abs(x - (psc(w) - 1));
122
image3d_cp1(top_blob, ivec3(gx, 0, 0), bottom_blob, ivec3(x, 0, 0));
124
buffer_cp1(top_blob_data, gx, bottom_blob_data, x);
128
else if (psc(dims) == 2)
130
const int gi = gy * psc(outw) + gx;
137
if (x >= 0 && x < psc(w) && y >= 0 && y < psc(h))
140
image3d_cp1(top_blob, ivec3(gx, gy, 0), bottom_blob, ivec3(x, y, 0));
142
int v_offset = y * psc(w) + x;
143
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
150
image3d_st1(top_blob, ivec3(gx, gy, 0), v);
152
buffer_st1(top_blob_data, gi, v);
158
x = clamp(x, 0, psc(w) - 1);
159
y = clamp(y, 0, psc(h) - 1);
162
image3d_cp1(top_blob, ivec3(gx, gy, 0), bottom_blob, ivec3(x, y, 0));
164
int v_offset = y * psc(w) + x;
165
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
172
// NOTE psc(X) get zeros on nvidia
173
// TODO only enable this workaround for some nvidia driver
174
x = (p.w - 1) - abs(x - (p.w - 1));
175
y = (p.h - 1) - abs(y - (p.h - 1));
176
// x = (psc(w) - 1) - abs(x - (psc(w) - 1));
177
// y = (psc(h) - 1) - abs(y - (psc(h) - 1));
180
image3d_cp1(top_blob, ivec3(gx, gy, 0), bottom_blob, ivec3(x, y, 0));
182
int v_offset = y * psc(w) + x;
183
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
187
else // if (psc(dims) == 3)
189
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
193
int z = gz - p.front;
197
if (x >= 0 && x < psc(w) && y >= 0 && y < psc(h) && z >= 0 && z < psc(c))
200
image3d_cp1(top_blob, ivec3(gx, gy, gz), bottom_blob, ivec3(x, y, z));
202
int v_offset = z * psc(cstep) + y * psc(w) + x;
203
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
206
else if (per_channel_pad == 1)
209
image3d_cp1(top_blob, ivec3(gx, gy, gz), per_channel_pad_blob, ivec3(gz, 0, 0));
211
buffer_cp1(top_blob_data, gi, per_channel_pad_blob_data, gz);
218
image3d_st1(top_blob, ivec3(gx, gy, gz), v);
220
buffer_st1(top_blob_data, gi, v);
226
x = clamp(x, 0, psc(w) - 1);
227
y = clamp(y, 0, psc(h) - 1);
228
z = clamp(z, 0, psc(c) - 1);
231
image3d_cp1(top_blob, ivec3(gx, gy, gz), bottom_blob, ivec3(x, y, z));
233
int v_offset = z * psc(cstep) + y * psc(w) + x;
234
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
242
// NOTE psc(X) get zeros on nvidia
243
// TODO only enable this workaround for some nvidia driver
244
x = (p.w - 1) - abs(x - (p.w - 1));
245
y = (p.h - 1) - abs(y - (p.h - 1));
246
z = (p.c - 1) - abs(z - (p.c - 1));
247
// x = (psc(w) - 1) - abs(x - (psc(w) - 1));
248
// y = (psc(h) - 1) - abs(y - (psc(h) - 1));
249
// z = (psc(c) - 1) - abs(z - (psc(c) - 1));
252
image3d_cp1(top_blob, ivec3(gx, gy, gz), bottom_blob, ivec3(x, y, z));
254
int v_offset = z * psc(cstep) + y * psc(w) + x;
255
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);