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/gl/kernels/pad.h"
17
18 #include <algorithm>
19 #include <any>
20 #include <cstdint>
21 #include <cstring>
22 #include <memory>
23 #include <string>
24 #include <utility>
25 #include <vector>
26
27 #include "absl/memory/memory.h"
28 #include "tensorflow/lite/delegates/gpu/common/operations.h"
29 #include "tensorflow/lite/delegates/gpu/common/status.h"
30 #include "tensorflow/lite/delegates/gpu/common/types.h"
31 #include "tensorflow/lite/delegates/gpu/gl/variable.h"
32
33 namespace tflite {
34 namespace gpu {
35 namespace gl {
36 namespace {
37
38 class Pad : public NodeShader {
39 public:
GenerateCode(const GenerationContext & ctx,GeneratedCode * generated_code) const40 absl::Status GenerateCode(const GenerationContext& ctx,
41 GeneratedCode* generated_code) const final {
42 const auto& attr = std::any_cast<const PadAttributes&>(ctx.op_attr);
43
44 if (attr.type != PaddingContentType::ZEROS &&
45 attr.type != PaddingContentType::REFLECT) {
46 return absl::UnimplementedError(
47 "Only ZERO and REFLECT padding types are supported.");
48 }
49 if (attr.appended.h < 0 || attr.appended.w < 0 || attr.appended.c < 0 ||
50 attr.prepended.h < 0 || attr.prepended.w < 0 || attr.prepended.c < 0) {
51 return absl::UnimplementedError("Negative padding is not supported.");
52 }
53 if (attr.appended.b != 0 || attr.prepended.b != 0) {
54 return absl::UnimplementedError("Padding for BATCH is not supported.");
55 }
56 std::vector<Variable> parameters = {
57 {"input_data_0_h", static_cast<int>(ctx.input_shapes[0][1])},
58 {"input_data_0_w", static_cast<int>(ctx.input_shapes[0][2])},
59 {"input_data_0_c", static_cast<int>(ctx.input_shapes[0][3])},
60 {"prepended",
61 int4(attr.prepended.w, attr.prepended.h, attr.prepended.c, 0)},
62 };
63 std::string source;
64 if (attr.type == PaddingContentType::REFLECT) {
65 source = R"(
66 int src_x = gid.x - $prepended.x$;
67 src_x = abs(src_x);
68 src_x = $input_data_0_w$ - 1 - abs(src_x - $input_data_0_w$ + 1);
69
70 int src_y = gid.y - $prepended.y$;
71 src_y = abs(src_y);
72 src_y = $input_data_0_h$ - 1 - abs(src_y - $input_data_0_h$ + 1);
73 )";
74 if (attr.prepended.c == 0 && attr.appended.c == 0) {
75 // optimized case
76 source += " value_0 = $input_data_0[src_x, src_y, gid.z]$;\n";
77 } else {
78 source += R"(
79 int start_channel = gid.z * 4;
80 for (int i = 0; i < 4; ++i) {
81 int channel = start_channel + i;
82 int src_z = channel - $prepended.z$;
83 src_z = abs(src_z);
84 src_z = $input_data_0_c$ - 1 - abs(src_z - $input_data_0_c$ + 1);
85 // We need additional clamp for z, so that we use alignment for channels
86 // and can proceed extra channels that can lead to reading out of
87 // resource.
88 src_z = clamp(src_z, 0, $input_data_0_c$ - 1);
89 value_0[i] = $input_data_0[src_x, src_y, src_z / 4]$[src_z % 4];
90 }
91 )";
92 }
93 } else {
94 source = R"(
95 int src_x = gid.x - $prepended.x$;
96 int src_y = gid.y - $prepended.y$;
97 if (src_x >= 0 && src_x < $input_data_0_w$ && src_y >= 0 && src_y < $input_data_0_h$) {
98 )";
99 if (attr.prepended.c == 0 && attr.appended.c == 0) {
100 // optimized case
101 source += " value_0 = $input_data_0[src_x, src_y, gid.z]$;\n";
102 } else if (attr.prepended.c % 4 == 0) {
103 parameters.push_back(
104 {"src_slices",
105 DivideRoundUp(static_cast<int>(ctx.input_shapes[0][3]), 4)});
106 source += R"(
107 int src_z = gid.z - $prepended.z$ / 4;
108 if (src_z >= 0 && src_z < $src_slices$) {
109 value_0 = $input_data_0[src_x, src_y, src_z]$;
110 }
111 )";
112 } else {
113 source += R"(
114 int start_channel = gid.z * 4;
115 for (int i = 0; i < 4; ++i) {
116 int channel = start_channel + i;
117 int src_z = channel - $prepended.z$;
118 if (src_z >= 0 && src_z < $input_data_0_c$) {
119 value_0[i] = $input_data_0[src_x, src_y, src_z / 4]$[src_z % 4];
120 }
121 }
122 )";
123 }
124 source += " }\n";
125 }
126 *generated_code = {
127 /*parameters=*/std::move(parameters),
128 /*objects=*/{},
129 /*shared_variables=*/{},
130 /*workload=*/uint3(),
131 /*workgroup=*/uint3(),
132 /*source_code=*/std::move(source),
133 /*input=*/IOStructure::ONLY_DEFINITIONS,
134 /*output=*/IOStructure::AUTO,
135 };
136 return absl::OkStatus();
137 }
138 };
139
140 } // namespace
141
NewPadNodeShader()142 std::unique_ptr<NodeShader> NewPadNodeShader() {
143 return std::make_unique<Pad>();
144 }
145
146 } // namespace gl
147 } // namespace gpu
148 } // namespace tflite
149