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
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, imfmtc1) 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 { sfp bottom_blob_data[]; };53
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };54
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };55
layout (binding = 3) readonly buffer bias_blob { sfpvec4 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 >= psc(outh) || gz >= 1)
87
sum = image3d_ld4(bias_blob, ivec3(gx, 0, 0));
89
sum = buffer_ld4(bias_data, gx);
98
for (int i = 0; i < psc(w) / 8; i++)
101
v[0].r = image3d_ld1(bottom_blob, ivec3(i * 8 + 0, gy, 0));
102
v[0].g = image3d_ld1(bottom_blob, ivec3(i * 8 + 1, gy, 0));
103
v[0].b = image3d_ld1(bottom_blob, ivec3(i * 8 + 2, gy, 0));
104
v[0].a = image3d_ld1(bottom_blob, ivec3(i * 8 + 3, gy, 0));
105
v[1].r = image3d_ld1(bottom_blob, ivec3(i * 8 + 4, gy, 0));
106
v[1].g = image3d_ld1(bottom_blob, ivec3(i * 8 + 5, gy, 0));
107
v[1].b = image3d_ld1(bottom_blob, ivec3(i * 8 + 6, gy, 0));
108
v[1].a = image3d_ld1(bottom_blob, ivec3(i * 8 + 7, gy, 0));
110
afpvec8 k0 = image3d_ld8(weight_blob, ivec3(i * 4 + 0, gx, 0));
111
afpvec8 k1 = image3d_ld8(weight_blob, ivec3(i * 4 + 1, gx, 0));
112
afpvec8 k2 = image3d_ld8(weight_blob, ivec3(i * 4 + 2, gx, 0));
113
afpvec8 k3 = image3d_ld8(weight_blob, ivec3(i * 4 + 3, gx, 0));
116
sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
117
sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
118
sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
119
sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
122
int v_offset = gy * psc(w);
123
int w_offset = gx * psc(w) / 8;
125
for (int i = 0; i < psc(w) / 8; i++)
128
v[0].r = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 0);
129
v[0].g = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 1);
130
v[0].b = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 2);
131
v[0].a = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 3);
132
v[1].r = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 4);
133
v[1].g = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 5);
134
v[1].b = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 6);
135
v[1].a = buffer_ld1(bottom_blob_data, v_offset + i * 8 + 7);
137
afpvec8 k0 = buffer_ld8(weight_data, (w_offset + i) * 4 + 0);
138
afpvec8 k1 = buffer_ld8(weight_data, (w_offset + i) * 4 + 1);
139
afpvec8 k2 = buffer_ld8(weight_data, (w_offset + i) * 4 + 2);
140
afpvec8 k3 = buffer_ld8(weight_data, (w_offset + i) * 4 + 3);
143
sum.r += dot(v[0], k0[0]) + dot(v[1], k0[1]);
144
sum.g += dot(v[0], k1[0]) + dot(v[1], k1[1]);
145
sum.b += dot(v[0], k2[0]) + dot(v[1], k2[1]);
146
sum.a += dot(v[0], k3[0]) + dot(v[1], k3[1]);
150
sum = activation_afpvec4(sum, activation_type, activation_param_0, activation_param_1);
153
image3d_st1(top_blob, ivec3(gx * 4 + 0, gy, 0), sum.r);
154
image3d_st1(top_blob, ivec3(gx * 4 + 1, gy, 0), sum.g);
155
image3d_st1(top_blob, ivec3(gx * 4 + 2, gy, 0), sum.b);
156
image3d_st1(top_blob, ivec3(gx * 4 + 3, gy, 0), sum.a);
158
const int gi = gy * psc(outw) + gx * 4;
159
buffer_st1(top_blob_data, gi + 0, sum.r);
160
buffer_st1(top_blob_data, gi + 1, sum.g);
161
buffer_st1(top_blob_data, gi + 2, sum.b);
162
buffer_st1(top_blob_data, gi + 3, sum.a);