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 kernel_w = 1;
29
layout (constant_id = 1) const int kernel_h = 1;
30
layout (constant_id = 2) const int dilation_w = 1;
31
layout (constant_id = 3) const int dilation_h = 1;
32
layout (constant_id = 4) const int stride_w = 1;
33
layout (constant_id = 5) const int stride_h = 1;
34
layout (constant_id = 6) const int bias_term = 0;
35
layout (constant_id = 7) const int group = 1;
36
layout (constant_id = 8) const int activation_type = 0;
37
layout (constant_id = 9) const float activation_param_0 = 0;
38
layout (constant_id = 10) const float activation_param_1 = 0;
40
#define shape_constant_id_offset 11
41
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
42
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
43
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
44
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
45
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
47
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
48
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
49
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
50
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
51
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
54
layout (binding = 0) uniform unfp sampler3D bottom_blob;
55
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
56
layout (binding = 2) uniform unfp sampler3D weight_blob;
57
layout (binding = 3) uniform unfp sampler3D bias_blob;
59
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
60
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
61
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
62
layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
65
layout (push_constant) uniform parameter
82
int gx = int(gl_GlobalInvocationID.x);
83
int gy = int(gl_GlobalInvocationID.y);
84
int gz = int(gl_GlobalInvocationID.z);
86
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
94
sum = image3d_ld8(bias_blob, ivec3(gz, 0, 0));
96
sum = buffer_ld8(bias_data, gz);
101
sum = afpvec8(afpvec4(0.f), afpvec4(0.f));
104
// depth-wise convolution
106
int sy = gy * stride_h;
109
for (int y = 0; y < kernel_h; y++)
111
int sx = gx * stride_w;
113
for (int x = 0; x < kernel_w; x++)
115
afpvec8 v = image3d_ld8(bottom_blob, ivec3(sx, sy, gz));
117
afpvec8 k = image3d_ld8(weight_blob, ivec3(wx, gz, 0));
120
sum[0] += v[0] * k[0];
121
sum[1] += v[1] * k[1];
130
int w_offset = gz * kernel_w * kernel_h;
131
int v_offset = gz * psc(cstep) + gy * stride_h * psc(w) + gx * stride_w;
133
for (int y = 0; y < kernel_h; y++)
135
for (int x = 0; x < kernel_w; x++)
137
afpvec8 v = buffer_ld8(bottom_blob_data, v_offset + x * dilation_w);
139
afpvec8 k = buffer_ld8(weight_data, w_offset + x);
142
sum[0] += v[0] * k[0];
143
sum[1] += v[1] * k[1];
146
v_offset += dilation_h * psc(w);
147
w_offset += kernel_w;
151
sum = activation_afpvec8(sum, activation_type, activation_param_0, activation_param_1);
154
image3d_st8(top_blob, ivec3(gx, gy, gz), sum);
156
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
158
buffer_st8(top_blob_data, gi, sum);