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 axis = 0;
27
#define shape_constant_id_offset 1
28
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
29
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
30
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
31
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
32
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
34
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
35
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
36
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
37
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
38
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
41
layout (binding = 0) uniform unfp sampler3D bottom_blob_3d;
42
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
43
layout (binding = 2) uniform unfp sampler3D max_workspace_3d;
45
layout (binding = 0) buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
46
layout (binding = 1) readonly buffer max_workspace { sfpvec8 max_workspace_data[]; };
49
layout (push_constant) uniform parameter
66
int gx = int(gl_GlobalInvocationID.x);
67
int gy = int(gl_GlobalInvocationID.y);
68
int gz = int(gl_GlobalInvocationID.z);
70
if (gx >= psc(w) || gy >= psc(h) || gz >= psc(c))
73
int positive_axis = axis < 0 ? psc(dims) + axis : axis;
78
if (psc(dims) == 1) // positive_axis == 0
80
v = image3d_ld8(bottom_blob_3d, ivec3(gx, 0, 0));
81
max_value = image3d_ld8(max_workspace_3d, ivec3(0, 0, 0));
83
else if (psc(dims) == 2 && positive_axis == 0)
85
v = image3d_ld8(bottom_blob_3d, ivec3(gx, gy, 0));
86
max_value = image3d_ld8(max_workspace_3d, ivec3(gx, 0, 0));
88
else if (psc(dims) == 2 && positive_axis == 1)
90
v = image3d_ld8(bottom_blob_3d, ivec3(gx, gy, 0));
91
max_value = image3d_ld8(max_workspace_3d, ivec3(gy, 0, 0));
93
else if (psc(dims) == 3 && positive_axis == 0)
95
v = image3d_ld8(bottom_blob_3d, ivec3(gx, gy, gz));
96
max_value = image3d_ld8(max_workspace_3d, ivec3(gx, gy, 0));
98
else if (psc(dims) == 3 && positive_axis == 1)
100
v = image3d_ld8(bottom_blob_3d, ivec3(gx, gy, gz));
101
max_value = image3d_ld8(max_workspace_3d, ivec3(gx, gz, 0));
103
else if (psc(dims) == 3 && positive_axis == 2)
105
v = image3d_ld8(bottom_blob_3d, ivec3(gx, gy, gz));
106
max_value = image3d_ld8(max_workspace_3d, ivec3(gy, gz, 0));
109
const int gi = gz * psc(cstep) + gy * psc(w) + gx;
111
afpvec8 v = buffer_ld8(bottom_top_blob_data, gi);
115
if (psc(dims) == 1) // positive_axis == 0
117
max_value = buffer_ld8(max_workspace_data, 0);
119
else if (psc(dims) == 2 && positive_axis == 0)
121
max_value = buffer_ld8(max_workspace_data, gx);
123
else if (psc(dims) == 2 && positive_axis == 1)
125
max_value = buffer_ld8(max_workspace_data, gy);
127
else if (psc(dims) == 3 && positive_axis == 0)
129
max_value = buffer_ld8(max_workspace_data, gy * psc(w) + gx);
131
else if (psc(dims) == 3 && positive_axis == 1)
133
max_value = buffer_ld8(max_workspace_data, gz * psc(w) + gx);
135
else if (psc(dims) == 3 && positive_axis == 2)
137
max_value = buffer_ld8(max_workspace_data, gz * psc(h) + gy);
140
#if NCNN_fp16_packed || NCNN_fp16_storage
141
// NOTE reduce max may produce (X, undef, X, undef) on nvidia fp16p/fp16s
142
// TODO only enable this workaround for some nvidia driver
143
if (positive_axis == 0)
145
max_value = afpvec8(afpvec4(max_value[0].r), afpvec4(max_value[0].r));
150
v[0] = exp(v[0] - max_value[0]);
151
v[1] = exp(v[1] - max_value[1]);
154
image3d_st8(top_blob_3d, ivec3(gx, gy, gz), v);
156
buffer_st8(bottom_top_blob_data, gi, v);