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 op_type = 0;
26
layout (constant_id = 1) const int coeff_term = 0;
28
#define shape_constant_id_offset 2
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;
36
layout (binding = 0) uniform unfp sampler3D bottom_blob1_3d;
37
layout (binding = 1) uniform unfp sampler3D bottom_blob2_3d;
38
layout (binding = 2, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
40
layout (binding = 0) readonly buffer bottom_blob1 { sfpvec8 bottom_blob1_data[]; };
41
layout (binding = 1) readonly buffer bottom_blob2 { sfpvec8 bottom_blob2_data[]; };
42
layout (binding = 2) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
45
layout (push_constant) uniform parameter
59
int gx = int(gl_GlobalInvocationID.x);
60
int gy = int(gl_GlobalInvocationID.y);
61
int gz = int(gl_GlobalInvocationID.z);
63
if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
67
afpvec8 v1 = image3d_ld8(bottom_blob1_3d, ivec3(gx, gy, gz));
68
afpvec8 v2 = image3d_ld8(bottom_blob2_3d, ivec3(gx, gy, gz));
70
const int gi = gz * psc(cstep) + gy * psc(w) + gx;
72
afpvec8 v1 = buffer_ld8(bottom_blob1_data, gi);
73
afpvec8 v2 = buffer_ld8(bottom_blob2_data, gi);
82
res[0] = v1[0] * v2[0];
83
res[1] = v1[1] * v2[1];
88
res[0] = v1[0] + v2[0];
89
res[1] = v1[1] + v2[1];
94
res[0] = max(v1[0], v2[0]);
95
res[1] = max(v1[1], v2[1]);
102
res[0] = v1[0] * v2[0];
103
res[1] = v1[1] * v2[1];
108
res[0] = v1[0] * afp(p.coeff0) + v2[0] * afp(p.coeff1);
109
res[1] = v1[1] * afp(p.coeff0) + v2[1] * afp(p.coeff1);
114
res[0] = max(v1[0], v2[0]);
115
res[1] = max(v1[1], v2[1]);
120
image3d_st8(top_blob_3d, ivec3(gx, gy, gz), res);
122
buffer_st8(top_blob_data, gi, res);