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 activation_type = 0;
36
layout (constant_id = 8) const float activation_param_0 = 0;
37
layout (constant_id = 9) const float activation_param_1 = 0;
39
#define shape_constant_id_offset 10
40
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
41
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
42
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
43
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
44
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
46
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
47
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
48
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
49
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
50
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
53
layout (binding = 0) uniform unfp sampler3D bottom_blob;
54
layout (binding = 1, imfmtc4) writeonly uniform unfp image3D top_blob;
55
layout (binding = 2) uniform unfp sampler3D weight_blob;
56
layout (binding = 3) uniform unfp sampler3D bias_blob;
58
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
59
layout (binding = 1) writeonly buffer top_blob { sfpvec8 top_blob_data[]; };
60
layout (binding = 2) readonly buffer weight_blob { sfpvec8 weight_data[]; };
61
layout (binding = 3) readonly buffer bias_blob { sfpvec8 bias_data[]; };
64
layout (push_constant) uniform parameter
81
int gx = int(gl_GlobalInvocationID.x) * 2;
82
int gy = int(gl_GlobalInvocationID.y) * 2;
83
int gz = int(gl_GlobalInvocationID.z) * 2;
85
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
88
const ivec2 gx2 = gx + ivec2(0, 1);
89
const ivec2 gy2 = gy + ivec2(0, 1);
90
const ivec2 gz2 = gz + ivec2(0, 1);
104
sum0 = image3d_ld8(bias_blob, ivec3(gz2.x, 0, 0));
105
sum4 = image3d_ld8(bias_blob, ivec3(gz2.y, 0, 0));
107
sum0 = buffer_ld8(bias_data, gz2.x);
108
sum4 = buffer_ld8(bias_data, gz2.y);
119
sum0 = afpvec8(afpvec4(0.f), afpvec4(0.f));
120
sum1 = afpvec8(afpvec4(0.f), afpvec4(0.f));
121
sum2 = afpvec8(afpvec4(0.f), afpvec4(0.f));
122
sum3 = afpvec8(afpvec4(0.f), afpvec4(0.f));
123
sum4 = afpvec8(afpvec4(0.f), afpvec4(0.f));
124
sum5 = afpvec8(afpvec4(0.f), afpvec4(0.f));
125
sum6 = afpvec8(afpvec4(0.f), afpvec4(0.f));
126
sum7 = afpvec8(afpvec4(0.f), afpvec4(0.f));
130
for (int z = 0; z < psc(c); z++)
132
ivec2 sy = gy2 * stride_h;
135
for (int y = 0; y < kernel_h; y++)
137
ivec2 sx = gx2 * stride_w;
139
for (int x = 0; x < kernel_w; x++)
141
afp v0 = image3d_ld1(bottom_blob, ivec3(sx.x, sy.x, z));
142
afp v1 = image3d_ld1(bottom_blob, ivec3(sx.y, sy.x, z));
143
afp v2 = image3d_ld1(bottom_blob, ivec3(sx.x, sy.y, z));
144
afp v3 = image3d_ld1(bottom_blob, ivec3(sx.y, sy.y, z));
146
afpvec8 k0 = image3d_ld8(weight_blob, ivec3(wx, z, gz2.x));
147
afpvec8 k1 = image3d_ld8(weight_blob, ivec3(wx, z, gz2.y));
150
sum0[0] += v0 * k0[0];
151
sum0[1] += v0 * k0[1];
152
sum1[0] += v1 * k0[0];
153
sum1[1] += v1 * k0[1];
154
sum2[0] += v2 * k0[0];
155
sum2[1] += v2 * k0[1];
156
sum3[0] += v3 * k0[0];
157
sum3[1] += v3 * k0[1];
158
sum4[0] += v0 * k1[0];
159
sum4[1] += v0 * k1[1];
160
sum5[0] += v1 * k1[0];
161
sum5[1] += v1 * k1[1];
162
sum6[0] += v2 * k1[0];
163
sum6[1] += v2 * k1[1];
164
sum7[0] += v3 * k1[0];
165
sum7[1] += v3 * k1[1];
175
ivec2 w_offset = gz2 * psc(c) * kernel_w * kernel_h;
177
for (int z = 0; z < psc(c); z++)
180
v_offset.rg = z * psc(cstep) + gy2.x * stride_h * psc(w) + gx2 * stride_w;
181
v_offset.ba = z * psc(cstep) + gy2.y * stride_h * psc(w) + gx2 * stride_w;
183
for (int y = 0; y < kernel_h; y++)
185
for (int x = 0; x < kernel_w; x++)
187
afp v0 = buffer_ld1(bottom_blob_data, v_offset.r + x * dilation_w);
188
afp v1 = buffer_ld1(bottom_blob_data, v_offset.g + x * dilation_w);
189
afp v2 = buffer_ld1(bottom_blob_data, v_offset.b + x * dilation_w);
190
afp v3 = buffer_ld1(bottom_blob_data, v_offset.a + x * dilation_w);
192
afpvec8 k0 = buffer_ld8(weight_data, w_offset.x + x);
193
afpvec8 k1 = buffer_ld8(weight_data, w_offset.y + x);
196
sum0[0] += v0 * k0[0];
197
sum0[1] += v0 * k0[1];
198
sum1[0] += v1 * k0[0];
199
sum1[1] += v1 * k0[1];
200
sum2[0] += v2 * k0[0];
201
sum2[1] += v2 * k0[1];
202
sum3[0] += v3 * k0[0];
203
sum3[1] += v3 * k0[1];
204
sum4[0] += v0 * k1[0];
205
sum4[1] += v0 * k1[1];
206
sum5[0] += v1 * k1[0];
207
sum5[1] += v1 * k1[1];
208
sum6[0] += v2 * k1[0];
209
sum6[1] += v2 * k1[1];
210
sum7[0] += v3 * k1[0];
211
sum7[1] += v3 * k1[1];
214
v_offset += dilation_h * psc(w);
215
w_offset += kernel_w;
220
sum0 = activation_afpvec8(sum0, activation_type, activation_param_0, activation_param_1);
221
sum1 = activation_afpvec8(sum1, activation_type, activation_param_0, activation_param_1);
222
sum2 = activation_afpvec8(sum2, activation_type, activation_param_0, activation_param_1);
223
sum3 = activation_afpvec8(sum3, activation_type, activation_param_0, activation_param_1);
224
sum4 = activation_afpvec8(sum4, activation_type, activation_param_0, activation_param_1);
225
sum5 = activation_afpvec8(sum5, activation_type, activation_param_0, activation_param_1);
226
sum6 = activation_afpvec8(sum6, activation_type, activation_param_0, activation_param_1);
227
sum7 = activation_afpvec8(sum7, activation_type, activation_param_0, activation_param_1);
230
image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.x), sum0);
231
image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.x), sum1);
232
image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.x), sum2);
233
image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.x), sum3);
234
image3d_st8(top_blob, ivec3(gx2.x, gy2.x, gz2.y), sum4);
235
image3d_st8(top_blob, ivec3(gx2.y, gy2.x, gz2.y), sum5);
236
image3d_st8(top_blob, ivec3(gx2.x, gy2.y, gz2.y), sum6);
237
image3d_st8(top_blob, ivec3(gx2.y, gy2.y, gz2.y), sum7);
239
const ivec2 gi = gz2 * psc(outcstep) + gy * psc(outw) + gx;
241
buffer_st8(top_blob_data, gi.x, sum0);
242
if (gx + 1 < psc(outw)) buffer_st8(top_blob_data, gi.x + 1, sum1);
243
if (gy + 1 < psc(outh)) buffer_st8(top_blob_data, gi.x + psc(outw), sum2);
244
if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st8(top_blob_data, gi.x + psc(outw) + 1, sum3);
245
if (gz + 1 < psc(outc))
247
buffer_st8(top_blob_data, gi.y, sum4);
248
if (gx + 1 < psc(outw)) buffer_st8(top_blob_data, gi.y + 1, sum5);
249
if (gy + 1 < psc(outh)) buffer_st8(top_blob_data, gi.y + psc(outw), sum6);
250
if (gy + 1 < psc(outh) && gx + 1 < psc(outw)) buffer_st8(top_blob_data, gi.y + psc(outw) + 1, sum7);