1
// Tencent is pleased to support the open source community by making ncnn available.
3
// Copyright (C) 2019 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
layout (constant_id = 0) const int resize_type = 0;
25
layout (constant_id = 1) const int align_corner = 0;
27
#define shape_constant_id_offset 2
28
layout (constant_id = shape_constant_id_offset + 0) const int dims = 0;
29
layout (constant_id = shape_constant_id_offset + 1) const int w = 0;
30
layout (constant_id = shape_constant_id_offset + 2) const int h = 0;
31
layout (constant_id = shape_constant_id_offset + 3) const int c = 0;
32
layout (constant_id = shape_constant_id_offset + 4) const int cstep = 0;
34
layout (constant_id = shape_constant_id_offset + 5) const int outdims = 0;
35
layout (constant_id = shape_constant_id_offset + 6) const int outw = 0;
36
layout (constant_id = shape_constant_id_offset + 7) const int outh = 0;
37
layout (constant_id = shape_constant_id_offset + 8) const int outc = 0;
38
layout (constant_id = shape_constant_id_offset + 9) const int outcstep = 0;
41
layout (binding = 0) uniform unfp sampler3D bottom_blob;
42
layout (binding = 1, imfmtc1) writeonly uniform unfp image3D top_blob;
44
layout (binding = 0) readonly buffer bottom_blob { sfp bottom_blob_data[]; };
45
layout (binding = 1) writeonly buffer top_blob { sfp top_blob_data[]; };
48
layout (push_constant) uniform parameter
68
int gx = int(gl_GlobalInvocationID.x);
69
int gy = int(gl_GlobalInvocationID.y);
70
int gz = int(gl_GlobalInvocationID.z);
72
if (gx >= psc(outw) || gy >= psc(outh) || gz >= psc(outc))
78
image3d_cp1(top_blob, ivec3(gx, gy, gz), bottom_blob, ivec3(gz, 0, 0));
80
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
82
buffer_cp1(top_blob_data, gi, bottom_blob_data, gz);
90
if (resize_type == 1) // nearest
92
int sx = min(int(floor(afp(gx) * afp(p.scale_x))), psc(w) - 1);
95
image3d_cp1(top_blob, ivec3(gx, gy, gz), bottom_blob, ivec3(sx, gy, gz));
97
int v_offset = gz * psc(cstep) + gy * psc(w) + sx;
99
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
101
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
104
if (resize_type == 2) // bilinear
107
if (align_corner == 1)
109
fx = afp(gx) * afp(p.scale_x);
113
fx = (afp(gx) + afp(0.5f)) * afp(p.scale_x) - afp(0.5f);
116
int sx = int(floor(fx));
120
int sx_max = psc(w) - 2;
127
else if (sx > sx_max)
134
afp a0 = image3d_ld1(bottom_blob, ivec3(sx, gy, gz));
135
afp a1 = image3d_ld1(bottom_blob, ivec3(sx + 1, gy, gz));
137
int v_offset_0 = gz * psc(cstep) + gy * psc(w) + sx;
139
afp a0 = buffer_ld1(bottom_blob_data, v_offset_0);
140
afp a1 = buffer_ld1(bottom_blob_data, v_offset_0 + 1);
143
afp res = a0 * (afp(1.f) - fx) + a1 * fx;
146
image3d_st1(top_blob, ivec3(gx, gy, gz), res);
148
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
150
buffer_st1(top_blob_data, gi, res);
157
if (resize_type == 1) // nearest
159
afpvec2 gxy = afpvec2(gx, gy);
160
ivec2 sxy_max = ivec2(psc(w) - 1, psc(h) - 1);
161
ivec2 sxy = min(ivec2(floor(gxy * afpvec2(p.scale_x, p.scale_y))), sxy_max);
167
image3d_cp1(top_blob, ivec3(gx, gy, gz), bottom_blob, ivec3(sx, sy, gz));
169
int v_offset = gz * psc(cstep) + sy * psc(w) + sx;
171
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
173
buffer_cp1(top_blob_data, gi, bottom_blob_data, v_offset);
176
if (resize_type == 2) // bilinear
178
afpvec2 gxy = afpvec2(gx, gy);
180
if (align_corner == 1)
182
fxy = gxy * afpvec2(p.scale_x, p.scale_y);
186
fxy = (gxy + afp(0.5f)) * afpvec2(p.scale_x, p.scale_y) - afp(0.5f);
189
ivec2 sxy = ivec2(floor(fxy));
193
ivec2 sxy_max = ivec2(psc(w) - 2, psc(h) - 2);
195
bvec2 underflow = lessThan(sxy, ivec2(0));
196
bvec2 overflow = greaterThan(sxy, sxy_max);
198
sxy = clamp(sxy, ivec2(0), sxy_max);
200
fxy = mix(fxy, afpvec2(0.f), underflow);
201
fxy = mix(fxy, afpvec2(1.f), overflow);
207
afp a0 = image3d_ld1(bottom_blob, ivec3(sx, sy, gz));
208
afp a1 = image3d_ld1(bottom_blob, ivec3(sx + 1, sy, gz));
209
afp b0 = image3d_ld1(bottom_blob, ivec3(sx, sy + 1, gz));
210
afp b1 = image3d_ld1(bottom_blob, ivec3(sx + 1, sy + 1, gz));
212
int v_offset_0 = gz * psc(cstep) + sy * psc(w) + sx;
213
int v_offset_1 = gz * psc(cstep) + (sy + 1) * psc(w) + sx;
215
afp a0 = buffer_ld1(bottom_blob_data, v_offset_0);
216
afp a1 = buffer_ld1(bottom_blob_data, v_offset_0 + 1);
217
afp b0 = buffer_ld1(bottom_blob_data, v_offset_1);
218
afp b1 = buffer_ld1(bottom_blob_data, v_offset_1 + 1);
224
afpvec2 ab = afpvec2(a0, b0) * (afp(1.f) - fx) + afpvec2(a1, b1) * fx;
226
afp res = ab.r * (afp(1.f) - fy) + ab.g * fy;
229
image3d_st1(top_blob, ivec3(gx, gy, gz), res);
231
const int gi = gz * psc(outcstep) + gy * psc(outw) + gx;
233
buffer_st1(top_blob_data, gi, res);