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 w = 0;
35
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
36
layout (constant_id = shape_constant_id_offset + 2) const int c = 0;
37
layout (constant_id = shape_constant_id_offset + 3) const int cstep = 0;
39
layout (constant_id = shape_constant_id_offset + 4) const int outw = 0;
40
layout (constant_id = shape_constant_id_offset + 5) const int outh = 0;
41
layout (constant_id = shape_constant_id_offset + 6) const int outc = 0;
42
layout (constant_id = shape_constant_id_offset + 7) const int outcstep = 0;
45
layout (binding = 0) uniform unfp sampler3D bottom_blob;
46
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
47
layout (binding = 2) uniform unfp sampler3D weight_blob;
48
layout (binding = 3) uniform unfp sampler3D bias_blob;
50
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
51
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
52
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
53
layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
56
layout (push_constant) uniform parameter
71
int gx = int(gl_GlobalInvocationID.x) * 4;
72
int gy = int(gl_GlobalInvocationID.y);
75
if (gx >= psc(outw) * psc(outh) || gy >= psc(outc))
78
if (gx >= psc(outcstep) || gy >= psc(outc))
90
afpvec8 b = image3d_ld8(bias_blob, ivec3(gy, 0, 0));
92
afpvec8 b = buffer_ld8(bias_data, gy);
101
sum0 = afpvec8(afpvec4(0.f), afpvec4(0.f));
102
sum1 = afpvec8(afpvec4(0.f), afpvec4(0.f));
103
sum2 = afpvec8(afpvec4(0.f), afpvec4(0.f));
104
sum3 = afpvec8(afpvec4(0.f), afpvec4(0.f));
108
ivec4 gx4 = gx + ivec4(0, 1, 2, 3);
110
ivec4 sy4 = gx4 / psc(w);
111
ivec4 sx4 = gx4 % psc(w);
113
for (int z = 0; z < psc(c); z++)
115
afp v0 = image3d_ld1(bottom_blob, ivec3(sx4.r, sy4.r, z));
116
afp v1 = image3d_ld1(bottom_blob, ivec3(sx4.g, sy4.g, z));
117
afp v2 = image3d_ld1(bottom_blob, ivec3(sx4.b, sy4.b, z));
118
afp v3 = image3d_ld1(bottom_blob, ivec3(sx4.a, sy4.a, z));
120
afpvec8 k = image3d_ld8(weight_blob, ivec3(0, z, gy));
123
sum0[0] += v0 * k[0];
124
sum0[1] += v0 * k[1];
126
sum1[0] += v1 * k[0];
127
sum1[1] += v1 * k[1];
129
sum2[0] += v2 * k[0];
130
sum2[1] += v2 * k[1];
132
sum3[0] += v3 * k[0];
133
sum3[1] += v3 * k[1];
136
int w_offset = gy * psc(c);
139
for (int z = 0; z < psc(c); z++)
141
afp v0 = buffer_ld1(bottom_blob_data, v_offset + 0);
142
afp v1 = buffer_ld1(bottom_blob_data, v_offset + 1);
143
afp v2 = buffer_ld1(bottom_blob_data, v_offset + 2);
144
afp v3 = buffer_ld1(bottom_blob_data, v_offset + 3);
146
afpvec8 k = buffer_ld8(weight_data, w_offset);
149
sum0[0] += v0 * k[0];
150
sum0[1] += v0 * k[1];
152
sum1[0] += v1 * k[0];
153
sum1[1] += v1 * k[1];
155
sum2[0] += v2 * k[0];
156
sum2[1] += v2 * k[1];
158
sum3[0] += v3 * k[0];
159
sum3[1] += v3 * k[1];
162
v_offset += psc(cstep);
166
sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
167
sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
168
sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
169
sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
172
image3d_st8(top_blob, ivec3(sx4.r, sy4.r, gy), sum0);
173
image3d_st8(top_blob, ivec3(sx4.g, sy4.g, gy), sum1);
174
image3d_st8(top_blob, ivec3(sx4.b, sy4.b, gy), sum2);
175
image3d_st8(top_blob, ivec3(sx4.a, sy4.a, gy), sum3);
177
int gi = gy * psc(outcstep) + gx;
179
buffer_st8(top_blob_data, gi + 0, sum0);
180
if (gx + 1 < psc(outcstep)) buffer_st8(top_blob_data, gi + 1, sum1);
181
if (gx + 2 < psc(outcstep)) buffer_st8(top_blob_data, gi + 2, sum2);
182
if (gx + 3 < psc(outcstep)) buffer_st8(top_blob_data, gi + 3, sum3);