1 /* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
2
3 Licensed under the Apache License, Version 2.0 (the "License");
4 you may not use this file except in compliance with the License.
5 You may obtain a copy of the License at
6
7 http://www.apache.org/licenses/LICENSE-2.0
8
9 Unless required by applicable law or agreed to in writing, software
10 distributed under the License is distributed on an "AS IS" BASIS,
11 WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12 See the License for the specific language governing permissions and
13 limitations under the License.
14 ==============================================================================*/
15
16 #include "tensorflow/lite/delegates/gpu/common/tasks/padding.h"
17
18 #include <string>
19
20 #include "tensorflow/lite/delegates/gpu/common/operations.h"
21 #include "tensorflow/lite/delegates/gpu/common/task/work_group_picking.h"
22
23 namespace tflite {
24 namespace gpu {
25
26 namespace {
GetPaddingCode(const OperationDef & op_def,const PadAttributes & attr,GPUOperation * op)27 std::string GetPaddingCode(const OperationDef& op_def,
28 const PadAttributes& attr, GPUOperation* op) {
29 op->AddSrcTensor("src_tensor", op_def.src_tensors[0]);
30 op->AddDstTensor("dst_tensor", op_def.dst_tensors[0]);
31 op->args_.AddInt("prepended_x", attr.prepended.w);
32 op->args_.AddInt("prepended_y", attr.prepended.h);
33 op->args_.AddInt("prepended_z", attr.prepended.c);
34 op->args_.AddInt("prepended_w", attr.prepended.b);
35
36 const std::string dst_batch =
37 op_def.dst_tensors[0].HasAxis(Axis::BATCH) ? "B" : "0";
38 std::string c;
39 const std::string channels[] = {".x", ".y", ".z", ".w"};
40
41 if (attr.type == PaddingContentType::REFLECT) {
42 c += "int reflect_coord(int x, int size) {\n";
43 c += " int t = abs(x) - size + 1;\n";
44 c += " return size - 1 - abs(t);\n";
45 c += "}\n\n";
46 }
47
48 c += "MAIN_FUNCTION($0) {\n";
49 if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) {
50 c += " int linear_id = GLOBAL_ID_0;\n";
51 c += " int X = linear_id / args.dst_tensor.Batch();\n";
52 c += " int B = linear_id % args.dst_tensor.Batch();\n";
53 c += " args.dst_tensor.SetBatchRef(B);\n";
54 } else {
55 c += " int X = GLOBAL_ID_0;\n";
56 }
57 c += " int Y = GLOBAL_ID_1;\n";
58 c += " int Z = GLOBAL_ID_2;\n";
59 c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || "
60 "Z >= args.dst_tensor.Slices()) { \n";
61 c += " return; \n";
62 c += " } \n";
63 c += " FLT4 result = INIT_FLT4(0.0);\n";
64 c += " int s_x = X - args.prepended_x;\n";
65 c += " int s_y = Y - args.prepended_y;\n";
66 if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) {
67 c += " int s_b = " + dst_batch + " - args.prepended_w;\n";
68 c += " args.src_tensor.SetBatchRef(s_b);\n";
69 }
70 if (attr.type == PaddingContentType::REFLECT) {
71 c += " s_x = reflect_coord(s_x, args.src_tensor.Width());\n";
72 c += " s_y = reflect_coord(s_y, args.src_tensor.Height());\n";
73 if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) {
74 c += " int s_b = reflect_coord(s_b, args.src_tensor.Batch());\n";
75 }
76 if (attr.prepended.c == 0 && attr.appended.c == 0) {
77 // optimized case
78 c += " result = args.src_tensor.Read(s_x, s_y, Z);\n";
79 } else {
80 c += " int start_channel = Z * 4;\n";
81 for (int i = 0; i < 4; ++i) {
82 const auto& s = channels[i];
83 c += " {\n";
84 c += " int channel = start_channel + " + std::to_string(i) + ";\n";
85 c += " int s_z = channel - args.prepended_z;\n";
86 // We need additional clamp for z, so that we use alignment for channels
87 // and can proceed extra channels that can lead to reading out of
88 // resource.
89 c += " s_z = clamp(reflect_coord(s_z, args.src_tensor.Channels()), "
90 "0, "
91 "args.src_tensor.Channels() - "
92 "1);\n";
93 c += " FLT4 t = args.src_tensor.Read(s_x, s_y, s_z / 4);\n";
94 c += " result" + s + " = SELECT_BY_INDEX_FROM_FLT4(t, s_z % 4);\n";
95 c += " }\n";
96 }
97 }
98 } else {
99 c += " bool inside_x = s_x >= 0 && s_x < args.src_tensor.Width();\n";
100 c += " bool inside_y = s_y >= 0 && s_y < args.src_tensor.Height();\n";
101 if (op_def.src_tensors[0].HasAxis(Axis::BATCH)) {
102 c += " inside_y &= (s_b >= 0 && s_b < args.src_tensor.Batch());\n";
103 }
104 c += " if (inside_x && inside_y) {\n";
105 if (attr.prepended.c == 0 && attr.appended.c == 0) {
106 // optimized case
107 c += " result = args.src_tensor.Read(s_x, s_y, Z);\n";
108 } else if (attr.prepended.c % 4 == 0) {
109 c += " int s_z = Z - args.prepended_z / 4;\n";
110 c += " if (s_z >= 0 && s_z < args.src_tensor.Slices()) {\n";
111 c += " result = args.src_tensor.Read(s_x, s_y, s_z);\n";
112 c += " }\n";
113 } else {
114 c += " int start_channel = Z * 4;\n";
115 for (int i = 0; i < 4; ++i) {
116 const auto& s = channels[i];
117 c += " {\n";
118 c += " int channel = start_channel + " + std::to_string(i) + ";\n";
119 c += " int s_z = channel - args.prepended_z;\n";
120 c += " if (s_z >= 0 && s_z < args.src_tensor.Channels()) {\n";
121 c += " FLT4 t = args.src_tensor.Read(s_x, s_y, s_z / 4);\n";
122 c += " result" + s + " = SELECT_BY_INDEX_FROM_FLT4(t, s_z % 4);\n";
123 c += " }\n";
124 c += " }\n";
125 }
126 }
127 c += " }\n";
128 }
129 c += " args.dst_tensor.Write(result, X, Y, Z);\n";
130 c += "}\n";
131
132 return c;
133 }
134
135 } // namespace
136
CreatePadding(const OperationDef & definition,const PadAttributes & attr)137 GPUOperation CreatePadding(const OperationDef& definition,
138 const PadAttributes& attr) {
139 GPUOperation op(definition);
140 op.code_ = GetPaddingCode(definition, attr, &op);
141 op.tensor_to_grid_ = TensorToGrid::kWBToX_HDToY_SToZ;
142 return op;
143 }
144
145 } // namespace gpu
146 } // namespace tflite
147