1 /*
2 * Copyright (C) 2018 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "Multinomial.h"
18
19 #include "CpuExecutor.h"
20 #include "CpuOperationUtils.h"
21 #include "HalInterfaces.h"
22 #include "Tracing.h"
23
24 #include "guarded_philox_random.h"
25 #include "philox_random.h"
26 #include "simple_philox.h"
27
28 #include "unsupported/Eigen/CXX11/Tensor"
29
30 namespace android {
31 namespace nn {
32
33 namespace {
34
35 template <typename T>
GetBuffer(RunTimeOperandInfo * operand)36 inline T* GetBuffer(RunTimeOperandInfo* operand) {
37 return reinterpret_cast<T*>(operand->buffer);
38 }
39
40 template <typename T>
GetBuffer(const RunTimeOperandInfo * operand)41 inline const T* GetBuffer(const RunTimeOperandInfo* operand) {
42 return reinterpret_cast<const T*>(operand->buffer);
43 }
44
45 } // namespace
46
Multinomial(const Operation & operation,std::vector<RunTimeOperandInfo> & operands)47 Multinomial::Multinomial(const Operation& operation, std::vector<RunTimeOperandInfo>& operands) {
48 NNTRACE_TRANS("Multinomial::Multinomial");
49 input_ = GetInput(operation, operands, kInputTensor);
50 sample_count_ = getScalarData<int>(*GetInput(operation, operands, kSampleCountParam));
51 random_seeds_ = GetInput(operation, operands, kRandomSeedsTensor);
52
53 output_ = GetOutput(operation, operands, kOutputTensor);
54 }
55
Prepare(const Operation & operation,std::vector<RunTimeOperandInfo> & operands,Shape * outputShape)56 bool Multinomial::Prepare(const Operation& operation, std::vector<RunTimeOperandInfo>& operands,
57 Shape* outputShape) {
58 NNTRACE_TRANS("Multinomial::Prepare");
59 NN_CHECK_EQ(NumInputsWithValues(operation, operands), 3);
60 NN_CHECK_EQ(NumOutputs(operation), 1);
61
62 const RunTimeOperandInfo* input = GetInput(operation, operands, Multinomial::kInputTensor);
63 const Shape& inputShape = input->shape();
64
65 const uint32_t batch_size = SizeOfDimension(input, 0);
66 const uint32_t sample_count =
67 getScalarData<int>(*GetInput(operation, operands, kSampleCountParam));
68
69 outputShape->type = OperandType::TENSOR_INT32;
70 outputShape->dimensions = {batch_size, sample_count};
71 outputShape->offset = inputShape.offset;
72 outputShape->scale = inputShape.scale;
73
74 return true;
75 }
76
Eval()77 bool Multinomial::Eval() {
78 NNTRACE_COMP("Multinomial::Eval");
79 switch (input_->type) {
80 case OperandType::TENSOR_FLOAT16: {
81 std::vector<float> inputDataFloat32(getNumberOfElements(input_->shape()));
82 convertFloat16ToFloat32(GetBuffer<_Float16>(input_), &inputDataFloat32);
83 EvalFloat32(inputDataFloat32.data());
84 break;
85 }
86 case OperandType::TENSOR_FLOAT32: {
87 EvalFloat32(GetBuffer<float>(input_));
88 break;
89 }
90 default: {
91 LOG(ERROR) << "Unsupported data type: " << static_cast<int>(input_->type);
92 return false;
93 }
94 }
95 return true;
96 }
97
EvalFloat32(const float * inputData)98 void Multinomial::EvalFloat32(const float* inputData) {
99 const int batch_size = SizeOfDimension(input_, 0);
100 const int class_size = SizeOfDimension(input_, 1);
101
102 tensorflow::GuardedPhiloxRandom random_generator;
103 int32_t* seeds = GetBuffer<int32_t>(random_seeds_);
104 random_generator.Init(seeds[0], seeds[1]);
105
106 // PhiloxRandom produces results as 4 32-bit integers.
107 int sample_count_aligned = (sample_count_ + 3) / 4 * 4;
108 // The CPU operation uses 64-bit double values, so two results per sample.
109 sample_count_aligned *= 2;
110 auto random_generator_reserved =
111 random_generator.ReserveRandomOutputs(batch_size * sample_count_aligned, 256);
112 tensorflow::random::SimplePhilox simple_philox(&random_generator_reserved);
113
114 for (uint64_t b = 0; b < batch_size; ++b) {
115 const float* input_ptr_batch = inputData + b * class_size;
116 float max = std::numeric_limits<float>::lowest();
117 for (uint64_t j = 0; j < class_size; ++j) {
118 if (Eigen::numext::isfinite(input_ptr_batch[j])) {
119 max = std::max(max, input_ptr_batch[j]);
120 }
121 }
122 const double batch_max = static_cast<double>(max);
123 double total = 0;
124 std::vector<double> cdf;
125 cdf.resize(class_size);
126 for (uint64_t j = 0; j < class_size; ++j) {
127 if (Eigen::numext::isfinite(static_cast<float>(input_ptr_batch[j]))) {
128 total += exp(static_cast<double>(input_ptr_batch[j]) - batch_max);
129 }
130 cdf[j] = total;
131 }
132
133 auto* output_ptr_batch = GetBuffer<int32_t>(output_) + b * sample_count_;
134 for (uint64_t j = 0; j < sample_count_; ++j) {
135 const double target = simple_philox.RandDouble() * total;
136 auto found_iter = std::upper_bound(cdf.begin(), cdf.end(), target);
137 output_ptr_batch[j] = std::distance(cdf.begin(), found_iter);
138 }
139 }
140 }
141
142 } // namespace nn
143 } // namespace android
144