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_top_blob_3d;
42
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D max_workspace_3d;
44
layout (binding = 0) readonly buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
45
layout (binding = 1) writeonly buffer max_workspace { sfpvec8 max_workspace_data[]; };
48
layout (push_constant) uniform parameter
65
int gx = int(gl_GlobalInvocationID.x);
66
int gy = int(gl_GlobalInvocationID.y);
67
int gz = int(gl_GlobalInvocationID.z);
69
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
72
int positive_axis = axis < 0 ? psc(dims) + axis : axis;
73
afpvec8 max_value = afpvec8(afpvec4(-99999999.f), afpvec4(-99999999.f));
75
if (psc(dims) == 1) // positive_axis == 0
77
for (int i = 0; i < psc(w); i++)
80
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, 0, 0));
82
afpvec8 v = buffer_ld8(bottom_top_blob_data, i);
84
max_value[0] = max(max_value[0], v[0]);
85
max_value[1] = max(max_value[1], v[1]);
87
afpvec4 max4 = max(max_value[0], max_value[1]);
88
afpvec2 max2 = max(max4.rg, max4.ba);
89
afp max1 = max(max2.r, max2.g);
90
max_value = afpvec8(afpvec4(max1), afpvec4(max1));
92
image3d_st8(max_workspace_3d, ivec3(0, 0, 0), max_value);
94
buffer_st8(max_workspace_data, 0, max_value);
99
if (psc(dims) == 2 && positive_axis == 0)
101
for (int i = 0; i < psc(h); i++)
104
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, i, 0));
106
int v_offset = i * psc(w) + gx;
107
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
109
max_value[0] = max(max_value[0], v[0]);
110
max_value[1] = max(max_value[1], v[1]);
112
afpvec4 max4 = max(max_value[0], max_value[1]);
113
afpvec2 max2 = max(max4.rg, max4.ba);
114
afp max1 = max(max2.r, max2.g);
115
max_value = afpvec8(afpvec4(max1), afpvec4(max1));
117
image3d_st8(max_workspace_3d, ivec3(gx, 0, 0), max_value);
119
buffer_st8(max_workspace_data, gx, max_value);
124
if (psc(dims) == 2 && positive_axis == 1)
126
for (int i = 0; i < psc(w); i++)
129
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, gx, 0));
131
int v_offset = gx * psc(w) + i;
132
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
134
max_value[0] = max(max_value[0], v[0]);
135
max_value[1] = max(max_value[1], v[1]);
138
image3d_st8(max_workspace_3d, ivec3(gx, 0, 0), max_value);
140
buffer_st8(max_workspace_data, gx, max_value);
145
if (psc(dims) == 3 && positive_axis == 0)
147
for (int i = 0; i < psc(c); i++)
150
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, gy, i));
152
int v_offset = i * psc(cstep) + gy * psc(w) + gx;
153
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
155
max_value[0] = max(max_value[0], v[0]);
156
max_value[1] = max(max_value[1], v[1]);
158
afpvec4 max4 = max(max_value[0], max_value[1]);
159
afpvec2 max2 = max(max4.rg, max4.ba);
160
afp max1 = max(max2.r, max2.g);
161
max_value = afpvec8(afpvec4(max1), afpvec4(max1));
163
image3d_st8(max_workspace_3d, ivec3(gx, gy, 0), max_value);
165
buffer_st8(max_workspace_data, gy * psc(w) + gx, max_value);
170
if (psc(dims) == 3 && positive_axis == 1)
172
for (int i = 0; i < psc(h); i++)
175
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, i, gy));
177
int v_offset = gy * psc(cstep) + i * psc(w) + gx;
178
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
180
max_value[0] = max(max_value[0], v[0]);
181
max_value[1] = max(max_value[1], v[1]);
184
image3d_st8(max_workspace_3d, ivec3(gx, gy, 0), max_value);
186
buffer_st8(max_workspace_data, gy * psc(w) + gx, max_value);
191
if (psc(dims) == 3 && positive_axis == 2)
193
for (int i = 0; i < psc(w); i++)
196
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, gx, gy));
198
int v_offset = gy * psc(cstep) + gx * psc(w) + i;
199
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
201
max_value[0] = max(max_value[0], v[0]);
202
max_value[1] = max(max_value[1], v[1]);
205
image3d_st8(max_workspace_3d, ivec3(gx, gy, 0), max_value);
207
buffer_st8(max_workspace_data, gy * psc(h) + gx, max_value);