• 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.h>
20 #include <xnnpack/params-init.h>
21 
22 
23 class RAddStoreExpMinusMaxMicrokernelTester {
24  public:
elements(size_t elements)25   inline RAddStoreExpMinusMaxMicrokernelTester& elements(size_t elements) {
26     assert(elements != 0);
27     this->elements_ = elements;
28     return *this;
29   }
30 
elements()31   inline size_t elements() const {
32     return this->elements_;
33   }
34 
iterations(size_t iterations)35   inline RAddStoreExpMinusMaxMicrokernelTester& iterations(size_t iterations) {
36     this->iterations_ = iterations;
37     return *this;
38   }
39 
iterations()40   inline size_t iterations() const {
41     return this->iterations_;
42   }
43 
Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax,xnn_init_f32_expminus_params_fn init_params)44   void Test(xnn_f32_raddstoreexpminusmax_ukernel_function raddstoreexpminusmax, xnn_init_f32_expminus_params_fn init_params) const {
45     std::random_device random_device;
46     auto rng = std::mt19937(random_device());
47     // Choose such range that expf(x[i]) overflows, but expf(x[i] - x_max) doesn't.
48     // However, the range is still narrow enough that double-precision exp doesn't overflow.
49     auto f32rng = std::bind(std::uniform_real_distribution<float>(90.0f, 100.0f), rng);
50 
51     std::vector<float> x(elements() + XNN_EXTRA_BYTES / sizeof(float));
52     std::vector<float> y(elements());
53     std::vector<double> y_ref(elements());
54     for (size_t iteration = 0; iteration < iterations(); iteration++) {
55       std::generate(x.begin(), x.end(), std::ref(f32rng));
56       std::fill(y.begin(), y.end(), std::nanf(""));
57 
58       // Compute reference results.
59       double sum_ref = 0.0f;
60       const float x_max = *std::max_element(x.begin(), x.begin() + elements());
61       for (size_t i = 0; i < elements(); i++) {
62         const double y_ref_value = exp(double(x[i]) - double(x_max));
63         y_ref[i] = y_ref_value;
64         sum_ref += y_ref_value;
65       }
66 
67       // Call optimized micro-kernel.
68       float sum = std::nanf("");
69       xnn_f32_expminus_params params;
70       init_params(&params);
71       raddstoreexpminusmax(elements() * sizeof(float), x.data(), &x_max, y.data(), &sum, &params);
72 
73       // Verify results.
74       for (size_t i = 0; i < elements(); i++) {
75       ASSERT_NEAR(y_ref[i], double(y[i]), std::abs(y_ref[i]) * 1.0e-6)
76         << "i = " << i << ", elements = " << elements() << ", x_max = " << x_max;
77       }
78       ASSERT_NEAR(sum_ref, double(sum), std::abs(sum_ref) * 1.0e-6)
79         << "elements = " << elements() << ", x_max = " << x_max;
80     }
81   }
82 
83  private:
84   size_t elements_{1};
85   size_t iterations_{15};
86 };
87