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 { float 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 == 0 only
79
mat2x4 sum = mat2x4(0.f);
82
ivec4 z4 = gz * 8 + ivec4(0, 1, 2, 3);
85
for (int z = 0; z < local_size; z++)
87
sum[0].r += texelFetch(square_workspace, ivec3(gx, gy, z4.r), 0).r;
88
sum[0].g += texelFetch(square_workspace, ivec3(gx, gy, z4.g), 0).r;
89
sum[0].b += texelFetch(square_workspace, ivec3(gx, gy, z4.b), 0).r;
90
sum[0].a += texelFetch(square_workspace, ivec3(gx, gy, z4.a), 0).r;
91
sum[1].r += texelFetch(square_workspace, ivec3(gx, gy, zz4.r), 0).r;
92
sum[1].g += texelFetch(square_workspace, ivec3(gx, gy, zz4.g), 0).r;
93
sum[1].b += texelFetch(square_workspace, ivec3(gx, gy, zz4.b), 0).r;
94
sum[1].a += texelFetch(square_workspace, ivec3(gx, gy, zz4.a), 0).r;
100
ivec4 z4 = gz * 8 + ivec4(0, 1, 2, 3);
102
ivec4 v_offset = z4 * psc(cstep) + gy * psc(w) + gx;
103
ivec4 vv_offset = zz4 * psc(cstep) + gy * psc(w) + gx;
105
for (int z = 0; z < local_size; z++)
107
sum[0].r += square_workspace_data[v_offset.r];
108
sum[0].g += square_workspace_data[v_offset.g];
109
sum[0].b += square_workspace_data[v_offset.b];
110
sum[0].a += square_workspace_data[v_offset.a];
111
sum[1].r += square_workspace_data[vv_offset.r];
112
sum[1].g += square_workspace_data[vv_offset.g];
113
sum[1].b += square_workspace_data[vv_offset.b];
114
sum[1].a += square_workspace_data[vv_offset.a];
116
v_offset += psc(cstep);
117
vv_offset += psc(cstep);
121
const float alpha_div_size = alpha / local_size;
123
scale[0] = afpvec4(pow(bias_constant + alpha_div_size * sum[0], vec4(-beta)));
124
scale[1] = afpvec4(pow(bias_constant + alpha_div_size * sum[1], vec4(-beta)));
127
afpvec8 v = image3d_ld8(bottom_blob, ivec3(gx, gy, gz));
129
int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
131
afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
138
image3d_st8(top_blob, ivec3(gx, gy, gz), v);
140
buffer_st8(bottom_top_blob_data, gi, v);