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 #ifndef TENSORFLOW_LITE_DELEGATES_GPU_COMMON_TASKS_CONVOLUTION_TRANSPOSED_THIN_H_
17 #define TENSORFLOW_LITE_DELEGATES_GPU_COMMON_TASKS_CONVOLUTION_TRANSPOSED_THIN_H_
18
19 #include <vector>
20
21 #include "tensorflow/lite/delegates/gpu/common/data_type.h"
22 #include "tensorflow/lite/delegates/gpu/common/operations.h"
23 #include "tensorflow/lite/delegates/gpu/common/shape.h"
24 #include "tensorflow/lite/delegates/gpu/common/status.h"
25 #include "tensorflow/lite/delegates/gpu/common/task/buffer_desc.h"
26 #include "tensorflow/lite/delegates/gpu/common/task/gpu_operation.h"
27 #include "tensorflow/lite/delegates/gpu/common/task/tensor_desc.h"
28 #include "tensorflow/lite/delegates/gpu/common/task/texture2d_desc.h"
29 #include "tensorflow/lite/delegates/gpu/common/tensor.h"
30 #include "tensorflow/lite/delegates/gpu/common/types.h"
31
32 namespace tflite {
33 namespace gpu {
34
35 class ConvolutionTransposedThin : public GPUOperation {
36 public:
37 ConvolutionTransposedThin() = default;
38 int3 GetGridSize() const override;
39
40 // Move only
41 ConvolutionTransposedThin(ConvolutionTransposedThin&& operation);
42 ConvolutionTransposedThin& operator=(ConvolutionTransposedThin&& operation);
43 ConvolutionTransposedThin(const ConvolutionTransposedThin&) = delete;
44 ConvolutionTransposedThin& operator=(const ConvolutionTransposedThin&) =
45 delete;
46
47 private:
48 friend ConvolutionTransposedThin CreateConvolutionTransposedThin(
49 const GpuInfo& gpu_info, const OperationDef& definition,
50 const ConvolutionTransposedAttributes& attr);
51 ConvolutionTransposedThin(const OperationDef& definition,
52 const ConvolutionTransposedAttributes& attr,
53 const GpuInfo& gpu_info);
54 template <DataType T>
55 void UploadData(const tflite::gpu::Tensor<OHWI, T>& weights,
56 const tflite::gpu::Tensor<Linear, T>& biases);
57
58 template <DataType S, typename T>
59 void RearrangeWeightsData(const tflite::gpu::Tensor<OHWI, S>& weights,
60 absl::Span<T> dst);
61 std::string GenerateConvolutionTransposedCode(const OperationDef& op_def,
62 int src_depth, int dst_channels,
63 const int2& kernel_size);
64 };
65
66 template <DataType T>
UploadData(const tflite::gpu::Tensor<OHWI,T> & weights,const tflite::gpu::Tensor<Linear,T> & biases)67 void ConvolutionTransposedThin::UploadData(
68 const tflite::gpu::Tensor<OHWI, T>& weights,
69 const tflite::gpu::Tensor<Linear, T>& biases) {
70 const int src_depth = DivideRoundUp(weights.shape.i, 4);
71 const int flt4_count =
72 weights.shape.w * weights.shape.h * src_depth * weights.shape.o;
73
74 const bool f32_weights = definition_.precision == CalculationsPrecision::F32;
75 const int flt4_size = f32_weights ? sizeof(float4) : sizeof(half4);
76
77 BufferDescriptor desc;
78 desc.element_type = f32_weights ? DataType::FLOAT32 : DataType::FLOAT16;
79 desc.element_size = 4;
80 desc.memory_type = MemoryType::CONSTANT;
81 desc.size = flt4_size * (flt4_count + 1);
82 desc.data.resize(desc.size);
83
84 if (f32_weights) {
85 float4* gpu_data = reinterpret_cast<float4*>(desc.data.data());
86 RearrangeWeightsData(weights, absl::MakeSpan(gpu_data, flt4_count));
87 float4 bias_value(0.0f);
88 for (int i = 0; i < weights.shape.o; ++i) {
89 bias_value[i] = biases.data[i];
90 }
91 gpu_data[flt4_count] = bias_value;
92 } else {
93 half4* gpu_data = reinterpret_cast<half4*>(desc.data.data());
94 RearrangeWeightsData(weights, absl::MakeSpan(gpu_data, flt4_count));
95 half4 bias_value(0.0f);
96 for (int i = 0; i < weights.shape.o; ++i) {
97 bias_value[i] = biases.data[i];
98 }
99 gpu_data[flt4_count] = bias_value;
100 }
101
102 args_.AddObject("weights",
103 absl::make_unique<BufferDescriptor>(std::move(desc)));
104 }
105
106 template <DataType S, typename T>
RearrangeWeightsData(const tflite::gpu::Tensor<OHWI,S> & weights,absl::Span<T> dst)107 void ConvolutionTransposedThin::RearrangeWeightsData(
108 const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
109 const int src_depth = DivideRoundUp(weights.shape.i, 4);
110 const int kernel_x = weights.shape.w;
111 const int kernel_y = weights.shape.h;
112
113 int counter = 0;
114 for (int s = 0; s < src_depth; ++s) {
115 for (int y = 0; y < kernel_y; ++y) {
116 for (int x = 0; x < kernel_x; ++x) {
117 std::vector<T> filters(weights.shape.o);
118 for (int j = 0; j < weights.shape.o; ++j) {
119 for (int i = 0; i < 4; ++i) {
120 const int s_ch = s * 4 + i;
121 const int d_ch = j;
122 if (s_ch < weights.shape.i && d_ch < weights.shape.o) {
123 const int f_index = weights.shape.LinearIndex({d_ch, y, x, s_ch});
124 filters[j][i] = weights.data[f_index];
125 } else {
126 filters[j][i] = 0.0f;
127 }
128 }
129 }
130 for (int j = 0; j < weights.shape.o; ++j) {
131 dst[counter++] = filters[j];
132 }
133 }
134 }
135 }
136 }
137
138 bool IsConvolutionTransposedThinSupported(
139 const ConvolutionTransposedAttributes& attr);
140
141 ConvolutionTransposedThin CreateConvolutionTransposedThin(
142 const GpuInfo& gpu_info, const OperationDef& definition,
143 const ConvolutionTransposedAttributes& attr);
144
145 } // namespace gpu
146 } // namespace tflite
147
148 #endif // TENSORFLOW_LITE_DELEGATES_GPU_COMMON_TASKS_CONVOLUTION_TRANSPOSED_THIN_H_
149