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/concat_xy.h"
17
18 #include <map>
19 #include <string>
20 #include <vector>
21
22 #include "tensorflow/lite/delegates/gpu/common/operations.h"
23 #include "tensorflow/lite/delegates/gpu/common/task/work_group_picking.h"
24 #include "tensorflow/lite/delegates/gpu/common/types.h"
25
26 namespace tflite {
27 namespace gpu {
28 namespace {
GetConcatKernelCode(const OperationDef & op_def,const ConcatAttributes & attr)29 std::string GetConcatKernelCode(const OperationDef& op_def,
30 const ConcatAttributes& attr) {
31 std::vector<std::string> tensor_names(op_def.src_tensors.size());
32 for (int i = 0; i < op_def.src_tensors.size(); ++i) {
33 tensor_names[i] = "src_tensor_" + std::to_string(i);
34 }
35
36 std::map<Axis, std::string> axis_to_selector = {
37 {Axis::WIDTH, "Width"}, {Axis::HEIGHT, "Height"},
38 {Axis::DEPTH, "Depth"}, {Axis::CHANNELS, "Channels"},
39 {Axis::BATCH, "Batch"},
40 };
41 std::map<Axis, std::string> axis_to_coord = {
42 {Axis::WIDTH, "X"}, {Axis::HEIGHT, "Y"}, {Axis::DEPTH, "D"},
43 {Axis::CHANNELS, "S"}, {Axis::BATCH, "B"},
44 };
45
46 std::vector<std::string> src_coords;
47 std::vector<std::string> dst_coords;
48 for (auto axis :
49 {Axis::WIDTH, Axis::HEIGHT, Axis::DEPTH, Axis::CHANNELS, Axis::BATCH}) {
50 if (op_def.src_tensors[0].HasAxis(axis) && axis != Axis::BATCH) {
51 if (axis == attr.axis) {
52 src_coords.push_back("coord");
53 } else {
54 src_coords.push_back(axis_to_coord[axis]);
55 }
56 }
57 if (op_def.dst_tensors[0].HasAxis(axis)) {
58 dst_coords.push_back(axis_to_coord[axis]);
59 }
60 }
61 std::string src_coord = src_coords[0];
62 for (int i = 1; i < src_coords.size(); ++i) {
63 src_coord += ", " + src_coords[i];
64 }
65 std::string dst_coord = dst_coords[0];
66 for (int i = 1; i < dst_coords.size(); ++i) {
67 dst_coord += ", " + dst_coords[i];
68 }
69
70 std::string c;
71 c += "MAIN_FUNCTION($0) {\n";
72 if (op_def.dst_tensors[0].HasAxis(Axis::BATCH)) {
73 c += " int linear_id_0 = GLOBAL_ID_0;\n";
74 c += " int X = linear_id_0 / args.dst_tensor.Batch();\n";
75 c += " int B = linear_id_0 % args.dst_tensor.Batch();\n";
76 } else {
77 c += " int X = GLOBAL_ID_0;\n";
78 }
79 if (op_def.dst_tensors[0].HasAxis(Axis::DEPTH)) {
80 c += " int linear_id_1 = GLOBAL_ID_1;\n";
81 c += " int Y = linear_id_1 / args.dst_tensor.Depth();\n";
82 c += " int D = linear_id_1 % args.dst_tensor.Depth();\n";
83 } else {
84 c += " int Y = GLOBAL_ID_1;\n";
85 }
86 c += " int S = GLOBAL_ID_2;\n";
87 c += " if (X >= args.dst_tensor.Width() || Y >= args.dst_tensor.Height() || "
88 "S >= args.dst_tensor.Slices()) { \n";
89 c += " return; \n";
90 c += " } \n";
91 c += " FLT4 result = INIT_FLT4(0.0f);\n";
92 c += " int coord = " + axis_to_coord[attr.axis] + ";\n";
93 for (int i = 0; i < op_def.src_tensors.size(); ++i) {
94 const std::string field =
95 "args." + tensor_names[i] + "." + axis_to_selector[attr.axis] + "()";
96 c += " if (coord >= 0 && coord < " + field + ") { \n";
97 if (op_def.src_tensors[i].HasAxis(Axis::BATCH)) {
98 if (attr.axis == Axis::BATCH) {
99 c += " args." + tensor_names[i] + ".SetBatchRef(coord);\n";
100 } else {
101 c += " args." + tensor_names[i] + ".SetBatchRef(B);\n";
102 }
103 }
104 c += " result = args." + tensor_names[i] + ".Read(" + src_coord + ");\n";
105 c += " } \n";
106 c += " coord -= " + field + ";\n";
107 }
108 c += " args.dst_tensor.Write(result, " + dst_coord + ");\n";
109 c += "}\n";
110 return c;
111 }
112 } // namespace
113
CreateConcatXY(const OperationDef & definition,const ConcatAttributes & attr)114 GPUOperation CreateConcatXY(const OperationDef& definition,
115 const ConcatAttributes& attr) {
116 GPUOperation op(definition);
117 for (int i = 0; i < definition.src_tensors.size(); ++i) {
118 const std::string name = "src_tensor_" + std::to_string(i);
119 op.AddSrcTensor(name, definition.src_tensors[i]);
120 }
121 op.AddDstTensor("dst_tensor", definition.dst_tensors[0]);
122 op.code_ = GetConcatKernelCode(definition, attr);
123 op.tensor_to_grid_ = TensorToGrid::kWBToX_HDToY_SToZ;
124 return op;
125 }
126
127 } // namespace gpu
128 } // namespace tflite
129