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
#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, imfmtc4) 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 { sfpvec8 bottom_blob_data[]; };
53
layout (binding = 1) writeonly buffer top_blob { sfpvec8 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 >= psc(outw) || gy >= 1 || 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));
100
for (int i = 0; i < psc(w); i++)
102
afpvec8 v = image3d_ld8(bottom_blob, ivec3(i, 0, 0));
104
afpvec8 k0 = image3d_ld8(weight_blob, ivec3(wx + 0, gx, 0));
105
afpvec8 k1 = image3d_ld8(weight_blob, ivec3(wx + 1, gx, 0));
106
afpvec8 k2 = image3d_ld8(weight_blob, ivec3(wx + 2, gx, 0));
107
afpvec8 k3 = image3d_ld8(weight_blob, ivec3(wx + 3, gx, 0));
108
afpvec8 k4 = image3d_ld8(weight_blob, ivec3(wx + 4, gx, 0));
109
afpvec8 k5 = image3d_ld8(weight_blob, ivec3(wx + 5, gx, 0));
110
afpvec8 k6 = image3d_ld8(weight_blob, ivec3(wx + 6, gx, 0));
111
afpvec8 k7 = image3d_ld8(weight_blob, ivec3(wx + 7, gx, 0));
114
sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
115
sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
116
sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
117
sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
118
sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
119
sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
120
sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
121
sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
126
int w_offset = gx * psc(w);
128
for (int i = 0; i < psc(w); i++)
130
afpvec8 v = buffer_ld8(bottom_blob_data, i);
132
afpvec8 k0 = buffer_ld8(weight_data, (w_offset + i) * 8 + 0);
133
afpvec8 k1 = buffer_ld8(weight_data, (w_offset + i) * 8 + 1);
134
afpvec8 k2 = buffer_ld8(weight_data, (w_offset + i) * 8 + 2);
135
afpvec8 k3 = buffer_ld8(weight_data, (w_offset + i) * 8 + 3);
136
afpvec8 k4 = buffer_ld8(weight_data, (w_offset + i) * 8 + 4);
137
afpvec8 k5 = buffer_ld8(weight_data, (w_offset + i) * 8 + 5);
138
afpvec8 k6 = buffer_ld8(weight_data, (w_offset + i) * 8 + 6);
139
afpvec8 k7 = buffer_ld8(weight_data, (w_offset + i) * 8 + 7);
142
sum[0].r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
143
sum[0].g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
144
sum[0].b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
145
sum[0].a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
146
sum[1].r += dot(v[0], k4[0]) + dot(v[1], k4[1]);
147
sum[1].g += dot(v[0], k5[0]) + dot(v[1], k5[1]);
148
sum[1].b += dot(v[0], k6[0]) + dot(v[1], k6[1]);
149
sum[1].a += dot(v[0], k7[0]) + dot(v[1], k7[1]);
153
sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
156
image3d_st8(top_blob, ivec3(gx, 0, 0), sum);
158
buffer_st8(top_blob_data, gx, sum);