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 #define LOG_TAG "Operations"
18
19 #include "Multinomial.h"
20
21 #include "CpuExecutor.h"
22 #include "CpuOperationUtils.h"
23 #include "HalInterfaces.h"
24 #include "Tracing.h"
25
26 #include "guarded_philox_random.h"
27 #include "philox_random.h"
28 #include "simple_philox.h"
29
30 #include <algorithm>
31 #include <limits>
32 #include <unsupported/Eigen/CXX11/Tensor>
33 #include <vector>
34
35 namespace android {
36 namespace nn {
37
38 namespace {
39
40 using namespace hal;
41
42 template <typename T>
GetBuffer(RunTimeOperandInfo * operand)43 inline T* GetBuffer(RunTimeOperandInfo* operand) {
44 return reinterpret_cast<T*>(operand->buffer);
45 }
46
47 template <typename T>
GetBuffer(const RunTimeOperandInfo * operand)48 inline const T* GetBuffer(const RunTimeOperandInfo* operand) {
49 return reinterpret_cast<const T*>(operand->buffer);
50 }
51
52 } // namespace
53
Multinomial(const Operation & operation,RunTimeOperandInfo * operands)54 Multinomial::Multinomial(const Operation& operation, RunTimeOperandInfo* operands) {
55 NNTRACE_TRANS("Multinomial::Multinomial");
56 input_ = GetInput(operation, operands, kInputTensor);
57 sample_count_ = getScalarData<int>(*GetInput(operation, operands, kSampleCountParam));
58 random_seeds_ = GetInput(operation, operands, kRandomSeedsTensor);
59
60 output_ = GetOutput(operation, operands, kOutputTensor);
61 }
62
Prepare(const Operation & operation,RunTimeOperandInfo * operands,Shape * outputShape)63 bool Multinomial::Prepare(const Operation& operation, RunTimeOperandInfo* operands,
64 Shape* outputShape) {
65 NNTRACE_TRANS("Multinomial::Prepare");
66 NN_CHECK_EQ(NumInputsWithValues(operation, operands), 3);
67 NN_CHECK_EQ(NumOutputs(operation), 1);
68
69 const RunTimeOperandInfo* input = GetInput(operation, operands, Multinomial::kInputTensor);
70 const Shape& inputShape = input->shape();
71
72 const uint32_t batch_size = SizeOfDimension(input, 0);
73 const uint32_t sample_count =
74 getScalarData<int>(*GetInput(operation, operands, kSampleCountParam));
75
76 outputShape->type = OperandType::TENSOR_INT32;
77 outputShape->dimensions = {batch_size, sample_count};
78 outputShape->offset = inputShape.offset;
79 outputShape->scale = inputShape.scale;
80
81 return true;
82 }
83
Eval()84 bool Multinomial::Eval() {
85 NNTRACE_COMP("Multinomial::Eval");
86 switch (input_->type) {
87 case OperandType::TENSOR_FLOAT16: {
88 std::vector<float> inputDataFloat32(getNumberOfElements(input_->shape()));
89 convertFloat16ToFloat32(GetBuffer<_Float16>(input_), &inputDataFloat32);
90 EvalFloat32(inputDataFloat32.data());
91 break;
92 }
93 case OperandType::TENSOR_FLOAT32: {
94 EvalFloat32(GetBuffer<float>(input_));
95 break;
96 }
97 default: {
98 LOG(ERROR) << "Unsupported data type: " << static_cast<int>(input_->type);
99 return false;
100 }
101 }
102 return true;
103 }
104
EvalFloat32(const float * inputData)105 void Multinomial::EvalFloat32(const float* inputData) {
106 const int batch_size = SizeOfDimension(input_, 0);
107 const int class_size = SizeOfDimension(input_, 1);
108
109 tensorflow::GuardedPhiloxRandom random_generator;
110 int32_t* seeds = GetBuffer<int32_t>(random_seeds_);
111 random_generator.Init(seeds[0], seeds[1]);
112
113 // PhiloxRandom produces results as 4 32-bit integers.
114 int sample_count_aligned = (sample_count_ + 3) / 4 * 4;
115 // The CPU operation uses 64-bit double values, so two results per sample.
116 sample_count_aligned *= 2;
117 auto random_generator_reserved =
118 random_generator.ReserveRandomOutputs(batch_size * sample_count_aligned, 256);
119 tensorflow::random::SimplePhilox simple_philox(&random_generator_reserved);
120
121 for (uint64_t b = 0; b < batch_size; ++b) {
122 const float* input_ptr_batch = inputData + b * class_size;
123 float max = std::numeric_limits<float>::lowest();
124 for (uint64_t j = 0; j < class_size; ++j) {
125 if (Eigen::numext::isfinite(input_ptr_batch[j])) {
126 max = std::max(max, input_ptr_batch[j]);
127 }
128 }
129 const double batch_max = static_cast<double>(max);
130 double total = 0;
131 std::vector<double> cdf;
132 cdf.resize(class_size);
133 for (uint64_t j = 0; j < class_size; ++j) {
134 if (Eigen::numext::isfinite(static_cast<float>(input_ptr_batch[j]))) {
135 total += exp(static_cast<double>(input_ptr_batch[j]) - batch_max);
136 }
137 cdf[j] = total;
138 }
139
140 auto* output_ptr_batch = GetBuffer<int32_t>(output_) + b * sample_count_;
141 for (uint64_t j = 0; j < sample_count_; ++j) {
142 const double target = simple_philox.RandDouble() * total;
143 auto found_iter = std::upper_bound(cdf.begin(), cdf.end(), target);
144 output_ptr_batch[j] = std::distance(cdf.begin(), found_iter);
145 }
146 }
147 }
148
149 } // namespace nn
150 } // namespace android
151