1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2020 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
19
struct sfpvec8 { f16vec4 abcd; f16vec4 efgh; };
21
#if NCNN_fp16_arithmetic
22
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
25
#define FLT_MAX 3.402823466e+38
27
layout (constant_id = 0) const int pooling_type = 0;
28
layout (constant_id = 1) const int kernel_w = 1;
29
layout (constant_id = 2) const int kernel_h = 1;
30
layout (constant_id = 3) const int stride_w = 1;
31
layout (constant_id = 4) const int stride_h = 1;
32
layout (constant_id = 5) const int pad_left = 0;
33
layout (constant_id = 6) const int pad_right = 0;
34
layout (constant_id = 7) const int pad_top = 0;
35
layout (constant_id = 8) const int pad_bottom = 0;
36
layout (constant_id = 9) const int global_pooling = 0;
37
layout (constant_id = 10) const int pad_mode = 0;
38
layout (constant_id = 11) const int avgpool_count_include_pad = 0;
40
#define shape_constant_id_offset 12
41
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
42
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
43
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
44
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
45
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
47
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
48
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
49
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
50
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
51
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
54
layout (binding = 0) uniform unfp sampler3D bottom_blob;
55
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
57
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
58
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
61
layout (push_constant) uniform parameter
81
int gx = int(gl_GlobalInvocationID.x);
82
int gy = int(gl_GlobalInvocationID.y);
83
int gz = int(gl_GlobalInvocationID.z);
85
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
90
if (pooling_type == 0)
92
res = afpvec8(afpvec4(-FLT_MAX), afpvec4(-FLT_MAX));
95
int sx = gx * stride_w;
96
int sy = gy * stride_h;
98
for (int y = 0; y < kernel_h; y++)
100
for (int x = 0; x < kernel_w; x++)
102
afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx + x, sy + y, gz));
103
res[0] = max(res[0], v[0]);
104
res[1] = max(res[1], v[1]);
108
int v_offset = gz * psc(cstep) + gy * stride_h * psc(w) + gx * stride_w;
110
for (int y = 0; y < kernel_h; y++)
112
for (int x = 0; x < kernel_w; x++)
114
afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
115
res[0] = max(res[0], v[0]);
116
res[1] = max(res[1], v[1]);
123
else if (pooling_type == 1 && avgpool_count_include_pad == 0)
125
res = afpvec8(afpvec4(0.f), afpvec4(0.f));
128
int sx = gx * stride_w;
129
int sy = gy * stride_h;
132
for (int y = 0; y < kernel_h; y++)
134
if (sy + y < pad_top)
137
if (sy + y >= psc(h) - pad_bottom - p.htailpad)
140
for (int x = 0; x < kernel_w; x++)
142
if (sx + x < pad_left)
145
if (sx + x >= psc(w) - pad_right - p.wtailpad)
148
afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx + x, sy + y, gz));
155
int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
157
for (int y = 0; y < kernel_h; y++)
159
if (sy + y < pad_top)
165
if (sy + y >= psc(h) - pad_bottom - p.htailpad)
168
for (int x = 0; x < kernel_w; x++)
170
if (sx + x < pad_left)
175
if (sx + x >= psc(w) - pad_right - p.wtailpad)
178
afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
191
else if (pooling_type == 1 && avgpool_count_include_pad == 1)
193
res = afpvec8(afpvec4(0.f), afpvec4(0.f));
196
int sx = gx * stride_w;
197
int sy = gy * stride_h;
199
for (int y = 0; y < kernel_h; y++)
201
for (int x = 0; x < kernel_w; x++)
203
afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx + x, sy + y, gz));
209
int v_offset = gz * psc(cstep) + gy * stride_h * psc(w) + gx * stride_w;
211
for (int y = 0; y < kernel_h; y++)
213
for (int x = 0; x < kernel_w; x++)
215
afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x);
224
afp area = afp(kernel_w * kernel_h);
230
image3d_st8(top_blob, ivec3(gx, gy, gz), res);
232
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
234
buffer_st8(top_blob_data, gi, res);