• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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