1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2022 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
#define shape_constant_id_offset 0
26
layout (constant_id = shape_constant_id_offset + 0) const int w = 0;
28
layout (constant_id = shape_constant_id_offset + 1) const int outw = 0;
29
layout (constant_id = shape_constant_id_offset + 2) const int outh = 0;
32
layout (binding = 0) uniform unfp sampler3D bottom_blob;
33
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
34
layout (binding = 2) uniform unfp sampler3D weight_blob;
36
layout (binding = 0) readonly buffer bottom_blob { sfpvec8 bottom_blob_data[]; };
37
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
38
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
41
layout (push_constant) uniform parameter
51
int gx = int(gl_GlobalInvocationID.x);
52
int gy = int(gl_GlobalInvocationID.y);
53
int gz = int(gl_GlobalInvocationID.z);
55
if (gx >= psc(outw) || gy >= psc(outh) || gz >= 1)
58
const int start = gx * 8;
59
const int end = min(gx * 8 + 8, psc(w));
64
for (int i = start; i < end; i++)
66
afpvec8 v = image3d_ld8(bottom_blob, ivec3(i, 0, 0));
68
afpvec8 k = image3d_ld8(weight_blob, ivec3(i, gy, 0));
71
sum += dot(v[0], k[0]) + dot(v[1], k[1]);
74
int w_offset = gy * psc(w);
76
for (int i = start; i < end; i++)
78
afpvec8 v = buffer_ld8(bottom_blob_data, i);
80
afpvec8 k = buffer_ld8(weight_data, w_offset + i);
83
sum += dot(v[0], k[0]) + dot(v[1], k[1]);
88
image3d_st1(top_blob, ivec3(gx, gy, 0), sum);
90
const int gi = gy * psc(outw) + gx;
91
buffer_st1(top_blob_data, gi, sum);