• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright 2019 Google LLC
2 //
3 // This source code is licensed under the BSD-style license found in the
4 // LICENSE file in the root directory of this source tree.
5 
6 #pragma once
7 
8 #include <gtest/gtest.h>
9 
10 #include <algorithm>
11 #include <cassert>
12 #include <cstddef>
13 #include <cstdlib>
14 #include <functional>
15 #include <random>
16 #include <vector>
17 
18 #include <xnnpack.h>
19 #include <xnnpack/params-init.h>
20 #include <xnnpack/params.h>
21 
22 
23 class VUnOpMicrokernelTester {
24  public:
25   enum class OpType {
26     Sigmoid,
27   };
28 
29   enum class Variant {
30     Native,
31     Scalar,
32   };
33 
batch_size(size_t batch_size)34   inline VUnOpMicrokernelTester& batch_size(size_t batch_size) {
35     assert(batch_size != 0);
36     this->batch_size_ = batch_size;
37     return *this;
38   }
39 
batch_size()40   inline size_t batch_size() const {
41     return this->batch_size_;
42   }
43 
inplace(bool inplace)44   inline VUnOpMicrokernelTester& inplace(bool inplace) {
45     this->inplace_ = inplace;
46     return *this;
47   }
48 
inplace()49   inline bool inplace() const {
50     return this->inplace_;
51   }
52 
qmin(uint8_t qmin)53   inline VUnOpMicrokernelTester& qmin(uint8_t qmin) {
54     this->qmin_ = qmin;
55     return *this;
56   }
57 
qmin()58   inline uint8_t qmin() const {
59     return this->qmin_;
60   }
61 
qmax(uint8_t qmax)62   inline VUnOpMicrokernelTester& qmax(uint8_t qmax) {
63     this->qmax_ = qmax;
64     return *this;
65   }
66 
qmax()67   inline uint8_t qmax() const {
68     return this->qmax_;
69   }
70 
iterations(size_t iterations)71   inline VUnOpMicrokernelTester& iterations(size_t iterations) {
72     this->iterations_ = iterations;
73     return *this;
74   }
75 
iterations()76   inline size_t iterations() const {
77     return this->iterations_;
78   }
79 
80   void Test(xnn_f32_vunary_ukernel_function vunary, OpType op_type, Variant variant = Variant::Native) const {
81     std::random_device random_device;
82     auto rng = std::mt19937(random_device());
83     auto f32rng = std::bind(std::uniform_real_distribution<float>(-125.0f, 125.0f), rng);
84 
85     std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float));
86     std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0));
87     std::vector<double> y_ref(batch_size());
88     for (size_t iteration = 0; iteration < iterations(); iteration++) {
89       if (inplace()) {
90         std::generate(y.begin(), y.end(), std::ref(f32rng));
91       } else {
92         std::generate(x.begin(), x.end(), std::ref(f32rng));
93         std::fill(y.begin(), y.end(), nanf(""));
94       }
95       const float* x_data = inplace() ? y.data() : x.data();
96 
97       // Compute reference results.
98       for (size_t i = 0; i < batch_size(); i++) {
99         switch (op_type) {
100           case OpType::Sigmoid:
101           {
102             const double e = std::exp(double(x_data[i]));
103             y_ref[i] = e / (1.0 + e);
104             break;
105           }
106         }
107       }
108       const float accumulated_min = *std::min_element(y_ref.cbegin(), y_ref.cend());
109       const float accumulated_max = *std::max_element(y_ref.cbegin(), y_ref.cend());
110       const float accumulated_range = accumulated_max - accumulated_min;
111       const float y_max = accumulated_range > 0.0f ?
112         (accumulated_max - accumulated_range / 255.0f * float(255 - qmax())) :
113         +std::numeric_limits<float>::infinity();
114       const float y_min = accumulated_range > 0.0f ?
115         (accumulated_min + accumulated_range / 255.0f * float(qmin())) :
116         -std::numeric_limits<float>::infinity();
117       for (size_t i = 0; i < batch_size(); i++) {
118         y_ref[i] = std::max<float>(std::min<float>(y_ref[i], y_max), y_min);
119       }
120 
121       // Prepare output parameters.
122       xnn_f32_output_params output_params = { };
123       switch (variant) {
124         case Variant::Native:
125           output_params = xnn_init_f32_output_params(y_min, y_max);
126           break;
127         case Variant::Scalar:
128           output_params = xnn_init_scalar_f32_output_params(y_min, y_max);
129           break;
130       }
131 
132       // Call optimized micro-kernel.
133       vunary(batch_size() * sizeof(float), x_data, y.data(), &output_params);
134 
135       // Verify results.
136       for (size_t i = 0; i < batch_size(); i++) {
137         ASSERT_NEAR(y[i], y_ref[i], 5.0e-6)
138           << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i];
139       }
140     }
141   }
142 
143  private:
144   size_t batch_size_{1};
145   bool inplace_{false};
146   uint8_t qmin_{0};
147   uint8_t qmax_{255};
148   size_t iterations_{15};
149 };
150