1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2019 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
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
24
layout (constant_id = 0) const int axis = 0;
26
#define shape_constant_id_offset 1
27
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
28
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
29
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
30
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
31
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
33
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
34
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
35
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
36
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
37
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
40
layout (binding = 0) uniform unfp sampler3D bottom_blob_3d;
41
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob_3d;
42
layout (binding = 2) uniform unfp sampler3D sum_workspace_3d;
44
layout (binding = 0) buffer bottom_top_blob { sfpvec4 bottom_top_blob_data[]; };
45
layout (binding = 1) readonly buffer sum_workspace { sfpvec4 sum_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(w) || gy >= psc(h) || gz >= psc(c))
72
int positive_axis = axis < 0 ? psc(dims) + axis : axis;
77
if (psc(dims) == 1) // positive_axis == 0
79
v = image3d_ld4(bottom_blob_3d, ivec3(gx, 0, 0));
80
sum = image3d_ld4(sum_workspace_3d, ivec3(0, 0, 0));
82
else if (psc(dims) == 2 && positive_axis == 0)
84
v = image3d_ld4(bottom_blob_3d, ivec3(gx, gy, 0));
85
sum = image3d_ld4(sum_workspace_3d, ivec3(gx, 0, 0));
87
else if (psc(dims) == 2 && positive_axis == 1)
89
v = image3d_ld4(bottom_blob_3d, ivec3(gx, gy, 0));
90
sum = image3d_ld4(sum_workspace_3d, ivec3(gy, 0, 0));
92
else if (psc(dims) == 3 && positive_axis == 0)
94
v = image3d_ld4(bottom_blob_3d, ivec3(gx, gy, gz));
95
sum = image3d_ld4(sum_workspace_3d, ivec3(gx, gy, 0));
97
else if (psc(dims) == 3 && positive_axis == 1)
99
v = image3d_ld4(bottom_blob_3d, ivec3(gx, gy, gz));
100
sum = image3d_ld4(sum_workspace_3d, ivec3(gx, gz, 0));
102
else if (psc(dims) == 3 && positive_axis == 2)
104
v = image3d_ld4(bottom_blob_3d, ivec3(gx, gy, gz));
105
sum = image3d_ld4(sum_workspace_3d, ivec3(gy, gz, 0));
108
const int gi = gz * psc(cstep) + gy * psc(w) + gx;
110
afpvec4 v = buffer_ld4(bottom_top_blob_data, gi);
114
if (psc(dims) == 1) // positive_axis == 0
116
sum = buffer_ld4(sum_workspace_data, 0);
118
else if (psc(dims) == 2 && positive_axis == 0)
120
sum = buffer_ld4(sum_workspace_data, gx);
122
else if (psc(dims) == 2 && positive_axis == 1)
124
sum = buffer_ld4(sum_workspace_data, gy);
126
else if (psc(dims) == 3 && positive_axis == 0)
128
sum = buffer_ld4(sum_workspace_data, gy * psc(w) + gx);
130
else if (psc(dims) == 3 && positive_axis == 1)
132
sum = buffer_ld4(sum_workspace_data, gz * psc(w) + gx);
134
else if (psc(dims) == 3 && positive_axis == 2)
136
sum = buffer_ld4(sum_workspace_data, gz * psc(h) + gy);
139
#if NCNN_fp16_packed || NCNN_fp16_storage
140
// NOTE reduce max may produce (X, undef, X, undef) on nvidia fp16p/fp16s
141
// TODO only enable this workaround for some nvidia driver
142
if (positive_axis == 0)
144
sum = afpvec4(sum.r);
152
image3d_st4(top_blob_3d, ivec3(gx, gy, gz), v);
154
buffer_st4(bottom_top_blob_data, gi, v);