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
layout (constant_id = 0) const int region_type = 0;
26
layout (constant_id = 1) const int local_size = 0;
27
layout (constant_id = 2) const float alpha = 0;
28
layout (constant_id = 3) const float beta = 0;
29
layout (constant_id = 4) const float bias_constant = 0;
31
#define shape_constant_id_offset 5
32
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
33
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
34
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
35
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
36
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
38
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
39
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
40
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
41
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
42
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
45
layout (binding = 0) uniform highp sampler3D square_workspace;
46
layout (binding = 1) uniform unfp sampler3D bottom_blob;
47
layout (binding = 2, imfmtc4) writeonly uniform unfp image3D top_blob;
49
layout (binding = 0) readonly buffer square_workspace { mat2x4 square_workspace_data[]; };
50
layout (binding = 1) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
53
layout (push_constant) uniform parameter
70
int gx = int(gl_GlobalInvocationID.x);
71
int gy = int(gl_GlobalInvocationID.y);
72
int gz = int(gl_GlobalInvocationID.z);
74
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
77
// support region_type == 1 only
79
mat2x4 sum = mat2x4(0.f);
82
for (int y = 0; y < local_size; y++)
84
for (int x = 0; x < local_size; x++)
86
sum[0] += texelFetch(square_workspace, ivec3((gx + x) * 2, gy + y, gz), 0);
87
sum[1] += texelFetch(square_workspace, ivec3((gx + x) * 2 + 1, gy + y, gz), 0);
91
int v_offset = gz * psc(cstep) + gy * psc(w) + gx;
93
for (int y = 0; y < local_size; y++)
95
for (int x = 0; x < local_size; x++)
97
sum += square_workspace_data[v_offset + x];
104
const float alpha_div_size = alpha / (local_size * local_size);
106
scale[0] = afpvec4(pow(bias_constant + alpha_div_size * sum[0], vec4(-beta)));
107
scale[1] = afpvec4(pow(bias_constant + alpha_div_size * sum[1], vec4(-beta)));
110
afpvec8 v = image3d_ld8(bottom_blob, ivec3(gx, gy, gz));
112
int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
114
afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
121
image3d_st8(top_blob, ivec3(gx, gy, gz), v);
123
buffer_st8(bottom_top_blob_data, gi, v);