1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2023 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 dilation_w = 1;
30
layout (constant_id = 2) const int stride_w = 1;
31
layout (constant_id = 3) const int bias_term = 0;
32
layout (constant_id = 4) const int activation_type = 0;
33
layout (constant_id = 5) const float activation_param_0 = 0;
34
layout (constant_id = 6) const float activation_param_1 = 0;
36
#define shape_constant_id_offset 7
37
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
38
layout (constant_id = shape_constant_id_offset + 1) const int h = 0;
40
layout (constant_id = shape_constant_id_offset + 2) const int outw = 0;
41
layout (constant_id = shape_constant_id_offset + 3) const int outh = 0;
44
layout (binding = 0) uniform unfp sampler3D bottom_blob;
45
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
46
layout (binding = 2) uniform unfp sampler3D weight_blob;
47
layout (binding = 3) uniform unfp sampler3D bias_blob;
49
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
50
layout (binding = 1) writeonly buffer top_blob { sfpvec4 top_blob_data[]; };
51
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
52
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
55
layout (push_constant) uniform parameter
66
int gx = int(gl_GlobalInvocationID.x) * 2;
67
int gy = int(gl_GlobalInvocationID.y) * 2;
69
if (gx >= psc(outw) || gy >= psc(outh))
72
const ivec2 gx2 = gx + ivec2(0, 1);
73
const ivec2 gy2 = gy + ivec2(0, 1);
75
afpvec4 sum0 = afpvec4(0.0f);
76
afpvec4 sum1 = afpvec4(0.0f);
77
afpvec4 sum2 = afpvec4(0.0f);
78
afpvec4 sum3 = afpvec4(0.0f);
83
sum0 = image3d_ld4(bias_blob, ivec3(gy2.x, 0, 0));
84
sum2 = image3d_ld4(bias_blob, ivec3(gy2.y, 0, 0));
86
sum0 = buffer_ld4(bias_data, gy2.x);
87
sum2 = buffer_ld4(bias_data, gy2.y);
95
ivec2 v_offset = gx2 * stride_w;
97
for (int y = 0; y < psc(h); y++)
101
for (int x = 0; x < kernel_w; x++)
103
afpvec8 v0 = image3d_ld8(bottom_blob, ivec3(v_offset.x + x * dilation_w, y, 0));
104
afpvec8 v1 = image3d_ld8(bottom_blob, ivec3(v_offset.y + x * dilation_w, y, 0));
106
afpvec8 k0 = image3d_ld8(weight_blob, ivec3(wx + 0, y, gy2.x));
107
afpvec8 k1 = image3d_ld8(weight_blob, ivec3(wx + 1, y, gy2.x));
108
afpvec8 k2 = image3d_ld8(weight_blob, ivec3(wx + 2, y, gy2.x));
109
afpvec8 k3 = image3d_ld8(weight_blob, ivec3(wx + 3, y, gy2.x));
110
afpvec8 k4 = image3d_ld8(weight_blob, ivec3(wx + 0, y, gy2.y));
111
afpvec8 k5 = image3d_ld8(weight_blob, ivec3(wx + 1, y, gy2.y));
112
afpvec8 k6 = image3d_ld8(weight_blob, ivec3(wx + 2, y, gy2.y));
113
afpvec8 k7 = image3d_ld8(weight_blob, ivec3(wx + 3, y, gy2.y));
115
sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
116
sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
117
sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
118
sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
120
sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
121
sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
122
sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
123
sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
125
sum2.r += dot(v0[0], k4[0]) + dot(v0[1], k4[1]);
126
sum2.g += dot(v0[0], k5[0]) + dot(v0[1], k5[1]);
127
sum2.b += dot(v0[0], k6[0]) + dot(v0[1], k6[1]);
128
sum2.a += dot(v0[0], k7[0]) + dot(v0[1], k7[1]);
130
sum3.r += dot(v1[0], k4[0]) + dot(v1[1], k4[1]);
131
sum3.g += dot(v1[0], k5[0]) + dot(v1[1], k5[1]);
132
sum3.b += dot(v1[0], k6[0]) + dot(v1[1], k6[1]);
133
sum3.a += dot(v1[0], k7[0]) + dot(v1[1], k7[1]);
141
ivec2 v_offset = gx2 * stride_w;
142
ivec2 w_offset = gy2 * psc(h) * kernel_w;
144
for (int y = 0; y < psc(h); y++)
146
for (int x = 0; x < kernel_w; x++)
148
afpvec8 v0 = buffer_ld8(bottom_blob_data, v_offset.x + x * dilation_w);
149
afpvec8 v1 = buffer_ld8(bottom_blob_data, v_offset.y + x * dilation_w);
151
afpvec8 k0 = buffer_ld8(weight_data, (w_offset.x + x) * 4 + 0);
152
afpvec8 k1 = buffer_ld8(weight_data, (w_offset.x + x) * 4 + 1);
153
afpvec8 k2 = buffer_ld8(weight_data, (w_offset.x + x) * 4 + 2);
154
afpvec8 k3 = buffer_ld8(weight_data, (w_offset.x + x) * 4 + 3);
155
afpvec8 k4 = buffer_ld8(weight_data, (w_offset.y + x) * 4 + 0);
156
afpvec8 k5 = buffer_ld8(weight_data, (w_offset.y + x) * 4 + 1);
157
afpvec8 k6 = buffer_ld8(weight_data, (w_offset.y + x) * 4 + 2);
158
afpvec8 k7 = buffer_ld8(weight_data, (w_offset.y + x) * 4 + 3);
160
sum0.r += dot(v0[0], k0[0]) + dot(v0[1], k0[1]);
161
sum0.g += dot(v0[0], k1[0]) + dot(v0[1], k1[1]);
162
sum0.b += dot(v0[0], k2[0]) + dot(v0[1], k2[1]);
163
sum0.a += dot(v0[0], k3[0]) + dot(v0[1], k3[1]);
165
sum1.r += dot(v1[0], k0[0]) + dot(v1[1], k0[1]);
166
sum1.g += dot(v1[0], k1[0]) + dot(v1[1], k1[1]);
167
sum1.b += dot(v1[0], k2[0]) + dot(v1[1], k2[1]);
168
sum1.a += dot(v1[0], k3[0]) + dot(v1[1], k3[1]);
170
sum2.r += dot(v0[0], k4[0]) + dot(v0[1], k4[1]);
171
sum2.g += dot(v0[0], k5[0]) + dot(v0[1], k5[1]);
172
sum2.b += dot(v0[0], k6[0]) + dot(v0[1], k6[1]);
173
sum2.a += dot(v0[0], k7[0]) + dot(v0[1], k7[1]);
175
sum3.r += dot(v1[0], k4[0]) + dot(v1[1], k4[1]);
176
sum3.g += dot(v1[0], k5[0]) + dot(v1[1], k5[1]);
177
sum3.b += dot(v1[0], k6[0]) + dot(v1[1], k6[1]);
178
sum3.a += dot(v1[0], k7[0]) + dot(v1[1], k7[1]);
181
w_offset += kernel_w;
186
sum0 = activation_afpvec4(sum0, activation_type, activation_param_0, activation_param_1);
187
sum1 = activation_afpvec4(sum1, activation_type, activation_param_0, activation_param_1);
188
sum2 = activation_afpvec4(sum2, activation_type, activation_param_0, activation_param_1);
189
sum3 = activation_afpvec4(sum3, activation_type, activation_param_0, activation_param_1);
193
image3d_st4(top_blob, ivec3(gx2.x, gy2.x, 0), sum0);
194
image3d_st4(top_blob, ivec3(gx2.y, gy2.x, 0), sum1);
195
image3d_st4(top_blob, ivec3(gx2.x, gy2.y, 0), sum2);
196
image3d_st4(top_blob, ivec3(gx2.y, gy2.y, 0), sum3);
200
const int gi = gy * psc(outw) + gx;
202
buffer_st4(top_blob_data, gi, sum0);
203
if (gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi + 1, sum1);
204
if (gy + 1 < psc(outh)) buffer_st4(top_blob_data, gi + psc(outw), sum2);
205
if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st4(top_blob_data, gi + psc(outw) + 1, sum3);