1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2021 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
#extension GL_GOOGLE_include_directive: enable
26
#include "vulkan_activation.comp"
28
layout (constant_id = 0) const int bias_term = 0;
29
layout (constant_id = 1) const int activation_type = 0;
30
layout (constant_id = 2) const float activation_param_0 = 0;
31
layout (constant_id = 3) const float activation_param_1 = 0;
33
#define shape_constant_id_offset 4
34
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
35
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
36
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
37
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
38
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
40
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
41
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
42
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
43
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
44
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
47
layout (binding = 0) uniform unfp sampler3D bottom_blob;
48
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
49
layout (binding = 2) uniform unfp sampler3D weight_blob;
50
layout (binding = 3) uniform unfp sampler3D bias_blob;
52
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
53
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
54
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
55
layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
58
layout (push_constant) uniform parameter
75
int gx = int(gl_GlobalInvocationID.x);
76
int gy = int(gl_GlobalInvocationID.y);
77
int gz = int(gl_GlobalInvocationID.z);
79
if (gx * 8 >= psc(outw) || gy >= psc(outh) || gz >= 1)
87
sum = image3d_ld8(bias_blob, ivec3(gx, 0, 0));
89
sum = buffer_ld8(bias_data, gx);
94
sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
98
for (int i = 0; i < psc(w); i++)
100
afp v = image3d_ld1(bottom_blob, ivec3(i, gy, 0));
102
afpvec8 k = image3d_ld8(weight_blob, ivec3(i, gx, 0));
109
int v_offset = gy * psc(w);
110
int w_offset = gx * psc(w);
112
for (int i = 0; i < psc(w); i++)
114
afp v = buffer_ld1(bottom_blob_data, v_offset + i);
116
afpvec8 k = buffer_ld8(weight_data, w_offset + i);
124
sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
127
image3d_st1(top_blob, ivec3(gx * 8 + 0, gy, 0), sum[0].r);
128
image3d_st1(top_blob, ivec3(gx * 8 + 1, gy, 0), sum[0].g);
129
image3d_st1(top_blob, ivec3(gx * 8 + 2, gy, 0), sum[0].b);
130
image3d_st1(top_blob, ivec3(gx * 8 + 3, gy, 0), sum[0].a);
131
image3d_st1(top_blob, ivec3(gx * 8 + 4, gy, 0), sum[1].r);
132
image3d_st1(top_blob, ivec3(gx * 8 + 5, gy, 0), sum[1].g);
133
image3d_st1(top_blob, ivec3(gx * 8 + 6, gy, 0), sum[1].b);
134
image3d_st1(top_blob, ivec3(gx * 8 + 7, gy, 0), sum[1].a);
136
const int gi = gy * psc(outw) + gx * 8;
137
buffer_st1(top_blob_data, gi + 0, sum[0].r);
138
buffer_st1(top_blob_data, gi + 1, sum[0].g);
139
buffer_st1(top_blob_data, gi + 2, sum[0].b);
140
buffer_st1(top_blob_data, gi + 3, sum[0].a);
141
buffer_st1(top_blob_data, gi + 4, sum[1].r);
142
buffer_st1(top_blob_data, gi + 5, sum[1].g);
143
buffer_st1(top_blob_data, gi + 6, sum[1].b);
144
buffer_st1(top_blob_data, gi + 7, sum[1].a);