1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2019 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 region_type = 0;
25
layout (constant_id = 1) const int local_size = 0;
26
layout (constant_id = 2) const float alpha = 0;
27
layout (constant_id = 3) const float beta = 0;
28
layout (constant_id = 4) const float bias_constant = 0;
30
#define shape_constant_id_offset 5
31
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
32
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
33
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
34
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
35
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
37
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
38
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
39
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
40
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
41
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
44
layout (binding = 0) uniform highp sampler3D square_workspace;
45
layout (binding = 1) uniform unfp sampler3D bottom_blob;
46
layout (binding = 2, imfmtc4) writeonly uniform unfp image3D top_blob;
48
layout (binding = 0) readonly buffer square_workspace { float square_workspace_data[]; };
49
layout (binding = 1) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
52
layout (push_constant) uniform parameter
69
int gx = int(gl_GlobalInvocationID.x);
70
int gy = int(gl_GlobalInvocationID.y);
71
int gz = int(gl_GlobalInvocationID.z);
73
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
76
// support region_type == 0 only
81
ivec4 z4 = gz * 4 + ivec4(0, 1, 2, 3);
83
for (int z = 0; z < local_size; z++)
85
sum.r += texelFetch(square_workspace, ivec3(gx, gy, z4.r), 0).r;
86
sum.g += texelFetch(square_workspace, ivec3(gx, gy, z4.g), 0).r;
87
sum.b += texelFetch(square_workspace, ivec3(gx, gy, z4.b), 0).r;
88
sum.a += texelFetch(square_workspace, ivec3(gx, gy, z4.a), 0).r;
93
ivec4 z4 = gz * 4 + ivec4(0, 1, 2, 3);
94
ivec4 v_offset = z4 * psc(cstep) + gy * psc(w) + gx;
96
for (int z = 0; z < local_size; z++)
98
sum.r += square_workspace_data[v_offset.r];
99
sum.g += square_workspace_data[v_offset.g];
100
sum.b += square_workspace_data[v_offset.b];
101
sum.a += square_workspace_data[v_offset.a];
103
v_offset += psc(cstep);
107
const float alpha_div_size = alpha / local_size;
108
afpvec4 scale = afpvec4(pow(bias_constant + alpha_div_size * sum, vec4(-beta)));
111
afpvec4 v = image3d_ld4(bottom_blob, ivec3(gx, gy, gz));
113
int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
115
afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
121
image3d_st4(top_blob, ivec3(gx, gy, gz), v);
123
buffer_st4(bottom_top_blob_data, gi, v);