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
20
#if NCNN_fp16_arithmetic
21
#extension GL_EXT_shader_explicit_arithmetic_types_float16: require
24
#extension GL_GOOGLE_include_directive: enable
25
#include "vulkan_activation.comp"
27
layout (constant_id = 0) const int bias_term = 0;
28
layout (constant_id = 1) const int activation_type = 0;
29
layout (constant_id = 2) const float activation_param_0 = 0;
30
layout (constant_id = 3) const float activation_param_1 = 0;
32
#define shape_constant_id_offset 4
33
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
34
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
35
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
36
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
37
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
39
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
40
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
41
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
42
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
43
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
46
layout (binding = 0) uniform unfp sampler3D bottom_blob;
47
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
48
layout (binding = 2) uniform unfp sampler3D weight_blob;
49
layout (binding = 3) uniform unfp sampler3D bias_blob;
51
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
52
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
53
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
54
// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
55
layout (binding = 2) readonly buffer weight_blob { sfpvec4 weight_data[]; };
57
layout (binding = 2) readonly buffer weight_blob { sfpmat4 weight_data[]; };
59
layout (binding = 3) readonly buffer bias_blob { sfpvec4 bias_data[]; };
62
layout (push_constant) uniform parameter
79
int gx = int(gl_GlobalInvocationID.x);
80
int gy = int(gl_GlobalInvocationID.y);
81
int gz = int(gl_GlobalInvocationID.z);
83
if (gx * 4 >= psc(outw) || gy >= psc(outh) || gz >= 1)
91
sum = image3d_ld4(bias_blob, ivec3(gx, 0, 0));
93
sum = buffer_ld4(bias_data, gx);
102
for (int i = 0; i < psc(w) / 4; i++)
105
v.r = image3d_ld1(bottom_blob, ivec3(i * 4 + 0, gy, 0));
106
v.g = image3d_ld1(bottom_blob, ivec3(i * 4 + 1, gy, 0));
107
v.b = image3d_ld1(bottom_blob, ivec3(i * 4 + 2, gy, 0));
108
v.a = image3d_ld1(bottom_blob, ivec3(i * 4 + 3, gy, 0));
111
image3d_ld4(weight_blob, ivec3(i * 4 + 0, gx, 0)),
112
image3d_ld4(weight_blob, ivec3(i * 4 + 1, gx, 0)),
113
image3d_ld4(weight_blob, ivec3(i * 4 + 2, gx, 0)),
114
image3d_ld4(weight_blob, ivec3(i * 4 + 3, gx, 0))
120
int v_offset = gy * psc(w);
121
int w_offset = gx * psc(w) / 4;
123
for (int i = 0; i < psc(w) / 4; i++)
126
v.r = buffer_ld1(bottom_blob_data, v_offset + i * 4 + 0);
127
v.g = buffer_ld1(bottom_blob_data, v_offset + i * 4 + 1);
128
v.b = buffer_ld1(bottom_blob_data, v_offset + i * 4 + 2);
129
v.a = buffer_ld1(bottom_blob_data, v_offset + i * 4 + 3);
131
#if NCNN_fp16_packed || (NCNN_fp16_storage && !NCNN_fp16_arithmetic)
132
// GL_EXT_shader_16bit_storage does not define f16mat4 type :(
134
buffer_ld4(weight_data, (w_offset + i) * 4 + 0),
135
buffer_ld4(weight_data, (w_offset + i) * 4 + 1),
136
buffer_ld4(weight_data, (w_offset + i) * 4 + 2),
137
buffer_ld4(weight_data, (w_offset + i) * 4 + 3)
140
afpmat4 k = afpmat4(weight_data[w_offset + i]);
147
sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
150
image3d_st1(top_blob, ivec3(gx * 4 + 0, gy, 0), sum.r);
151
image3d_st1(top_blob, ivec3(gx * 4 + 1, gy, 0), sum.g);
152
image3d_st1(top_blob, ivec3(gx * 4 + 2, gy, 0), sum.b);
153
image3d_st1(top_blob, ivec3(gx * 4 + 3, gy, 0), sum.a);
155
const int gi = gy * psc(outw) + gx * 4;
156
buffer_st1(top_blob_data, gi + 0, sum.r);
157
buffer_st1(top_blob_data, gi + 1, sum.g);
158
buffer_st1(top_blob_data, gi + 2, sum.b);
159
buffer_st1(top_blob_data, gi + 3, sum.a);