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 <memory>
20 #include <string>
21 #include <utility>
22 #include <vector>
23
24 #include "tensorflow/lite/delegates/gpu/common/data_type.h"
25 #include "tensorflow/lite/delegates/gpu/common/operations.h"
26 #include "tensorflow/lite/delegates/gpu/common/shape.h"
27 #include "tensorflow/lite/delegates/gpu/common/status.h"
28 #include "tensorflow/lite/delegates/gpu/common/task/buffer_desc.h"
29 #include "tensorflow/lite/delegates/gpu/common/task/gpu_operation.h"
30 #include "tensorflow/lite/delegates/gpu/common/task/tensor_desc.h"
31 #include "tensorflow/lite/delegates/gpu/common/task/texture2d_desc.h"
32 #include "tensorflow/lite/delegates/gpu/common/tensor.h"
33 #include "tensorflow/lite/delegates/gpu/common/types.h"
34
35 namespace tflite {
36 namespace gpu {
37
38 class ConvolutionTransposedThin : public GPUOperation {
39 public:
40 ConvolutionTransposedThin() = default;
41 int3 GetGridSize() const override;
42
43 // Move only
44 ConvolutionTransposedThin(ConvolutionTransposedThin&& operation);
45 ConvolutionTransposedThin& operator=(ConvolutionTransposedThin&& operation);
46 ConvolutionTransposedThin(const ConvolutionTransposedThin&) = delete;
47 ConvolutionTransposedThin& operator=(const ConvolutionTransposedThin&) =
48 delete;
49
50 private:
51 friend ConvolutionTransposedThin CreateConvolutionTransposedThin(
52 const GpuInfo& gpu_info, const OperationDef& definition,
53 const ConvolutionTransposedAttributes& attr);
54 ConvolutionTransposedThin(const OperationDef& definition,
55 const ConvolutionTransposedAttributes& attr,
56 const GpuInfo& gpu_info);
57 template <DataType T>
58 void UploadData(const tflite::gpu::Tensor<OHWI, T>& weights,
59 const tflite::gpu::Tensor<Linear, T>& biases);
60
61 template <DataType S, typename T>
62 void RearrangeWeightsData(const tflite::gpu::Tensor<OHWI, S>& weights,
63 absl::Span<T> dst);
64 std::string GenerateConvolutionTransposedCode(const OperationDef& op_def,
65 int src_depth, int dst_channels,
66 const int2& kernel_size);
67 };
68
69 template <DataType T>
UploadData(const tflite::gpu::Tensor<OHWI,T> & weights,const tflite::gpu::Tensor<Linear,T> & biases)70 void ConvolutionTransposedThin::UploadData(
71 const tflite::gpu::Tensor<OHWI, T>& weights,
72 const tflite::gpu::Tensor<Linear, T>& biases) {
73 const int src_depth = DivideRoundUp(weights.shape.i, 4);
74 const int flt4_count =
75 weights.shape.w * weights.shape.h * src_depth * weights.shape.o;
76
77 const bool f32_weights = definition_.precision == CalculationsPrecision::F32;
78 const int flt4_size = f32_weights ? sizeof(float4) : sizeof(half4);
79
80 BufferDescriptor desc;
81 desc.element_type = f32_weights ? DataType::FLOAT32 : DataType::FLOAT16;
82 desc.element_size = 4;
83 desc.memory_type = MemoryType::CONSTANT;
84 desc.size = flt4_size * (flt4_count + 1);
85 desc.data.resize(desc.size);
86
87 if (f32_weights) {
88 float4* gpu_data = reinterpret_cast<float4*>(desc.data.data());
89 RearrangeWeightsData(weights, absl::MakeSpan(gpu_data, flt4_count));
90 float4 bias_value(0.0f);
91 for (int i = 0; i < weights.shape.o; ++i) {
92 bias_value[i] = biases.data[i];
93 }
94 gpu_data[flt4_count] = bias_value;
95 } else {
96 half4* gpu_data = reinterpret_cast<half4*>(desc.data.data());
97 RearrangeWeightsData(weights, absl::MakeSpan(gpu_data, flt4_count));
98 half4 bias_value(0.0f);
99 for (int i = 0; i < weights.shape.o; ++i) {
100 bias_value[i] = biases.data[i];
101 }
102 gpu_data[flt4_count] = bias_value;
103 }
104
105 args_.AddObject("weights",
106 std::make_unique<BufferDescriptor>(std::move(desc)));
107 }
108
109 template <DataType S, typename T>
RearrangeWeightsData(const tflite::gpu::Tensor<OHWI,S> & weights,absl::Span<T> dst)110 void ConvolutionTransposedThin::RearrangeWeightsData(
111 const tflite::gpu::Tensor<OHWI, S>& weights, absl::Span<T> dst) {
112 const int src_depth = DivideRoundUp(weights.shape.i, 4);
113 const int kernel_x = weights.shape.w;
114 const int kernel_y = weights.shape.h;
115
116 int counter = 0;
117 for (int s = 0; s < src_depth; ++s) {
118 for (int y = 0; y < kernel_y; ++y) {
119 for (int x = 0; x < kernel_x; ++x) {
120 std::vector<T> filters(weights.shape.o);
121 for (int j = 0; j < weights.shape.o; ++j) {
122 for (int i = 0; i < 4; ++i) {
123 const int s_ch = s * 4 + i;
124 const int d_ch = j;
125 if (s_ch < weights.shape.i && d_ch < weights.shape.o) {
126 const int f_index = weights.shape.LinearIndex({d_ch, y, x, s_ch});
127 filters[j][i] = weights.data[f_index];
128 } else {
129 filters[j][i] = 0.0f;
130 }
131 }
132 }
133 for (int j = 0; j < weights.shape.o; ++j) {
134 dst[counter++] = filters[j];
135 }
136 }
137 }
138 }
139 }
140
141 bool IsConvolutionTransposedThinSupported(
142 const ConvolutionTransposedAttributes& attr);
143
144 ConvolutionTransposedThin CreateConvolutionTransposedThin(
145 const GpuInfo& gpu_info, const OperationDef& definition,
146 const ConvolutionTransposedAttributes& attr);
147
148 } // namespace gpu
149 } // namespace tflite
150
151 #endif // TENSORFLOW_LITE_DELEGATES_GPU_COMMON_TASKS_CONVOLUTION_TRANSPOSED_THIN_H_
152