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 sum_workspace_3d;
44
layout (binding = 0) readonly buffer bottom_top_blob { sfpvec8 bottom_top_blob_data[]; };
45
layout (binding = 1) writeonly buffer sum_workspace { sfpvec8 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(outw) || gy >= psc(outh) || gz >= psc(outc))
72
int positive_axis = axis < 0 ? psc(dims) + axis : axis;
73
afpvec8 sum_value = afpvec8(afpvec4(0.f), afpvec4(0.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);
86
afpvec4 sum4 = sum_value[0] + sum_value[1];
87
afpvec2 sum2 = sum4.rg + sum4.ba;
88
afp sum1 = sum2.r + sum2.g;
89
sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
91
image3d_st8(sum_workspace_3d, ivec3(0, 0, 0), sum_value);
93
buffer_st8(sum_workspace_data, 0, sum_value);
98
if (psc(dims) == 2 && positive_axis == 0)
100
for (int i = 0; i < psc(h); i++)
103
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, i, 0));
105
int v_offset = i * psc(w) + gx;
106
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
110
afpvec4 sum4 = sum_value[0] + sum_value[1];
111
afpvec2 sum2 = sum4.rg + sum4.ba;
112
afp sum1 = sum2.r + sum2.g;
113
sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
115
image3d_st8(sum_workspace_3d, ivec3(gx, 0, 0), sum_value);
117
buffer_st8(sum_workspace_data, gx, sum_value);
122
if (psc(dims) == 2 && positive_axis == 1)
124
for (int i = 0; i < psc(w); i++)
127
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, gx, 0));
129
int v_offset = gx * psc(w) + i;
130
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
135
image3d_st8(sum_workspace_3d, ivec3(gx, 0, 0), sum_value);
137
buffer_st8(sum_workspace_data, gx, sum_value);
142
if (psc(dims) == 3 && positive_axis == 0)
144
for (int i = 0; i < psc(c); i++)
147
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, gy, i));
149
int v_offset = i * psc(cstep) + gy * psc(w) + gx;
150
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
154
afpvec4 sum4 = sum_value[0] + sum_value[1];
155
afpvec2 sum2 = sum4.rg + sum4.ba;
156
afp sum1 = sum2.r + sum2.g;
157
sum_value = afpvec8(afpvec4(sum1), afpvec4(sum1));
159
image3d_st8(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
161
buffer_st8(sum_workspace_data, gy * psc(w) + gx, sum_value);
166
if (psc(dims) == 3 && positive_axis == 1)
168
for (int i = 0; i < psc(h); i++)
171
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(gx, i, gy));
173
int v_offset = gy * psc(cstep) + i * psc(w) + gx;
174
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
179
image3d_st8(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
181
buffer_st8(sum_workspace_data, gy * psc(w) + gx, sum_value);
186
if (psc(dims) == 3 && positive_axis == 2)
188
for (int i = 0; i < psc(w); i++)
191
afpvec8 v = image3d_ld8(bottom_top_blob_3d, ivec3(i, gx, gy));
193
int v_offset = gy * psc(cstep) + gx * psc(w) + i;
194
afpvec8 v = buffer_ld8(bottom_top_blob_data, v_offset);
199
image3d_st8(sum_workspace_3d, ivec3(gx, gy, 0), sum_value);
201
buffer_st8(sum_workspace_data, gy * psc(h) + gx, sum_value);