xref: /aosp_15_r20/external/XNNPACK/test/vlrelu-microkernel-tester.h (revision 4bdc94577ba0e567308109d787f7fec7b531ce36) 
1 // Copyright 2022 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 <cmath>
13 #include <cstddef>
14 #include <cstdlib>
15 #include <functional>
16 #include <limits>
17 #include <random>
18 #include <vector>
19 
20 #include <xnnpack.h>
21 #include <xnnpack/math.h>
22 #include <xnnpack/microfnptr.h>
23 #include <xnnpack/microparams-init.h>
24 
25 
26 class VLReLUMicrokernelTester {
27  public:
batch_size(size_t batch_size)28   inline VLReLUMicrokernelTester& batch_size(size_t batch_size) {
29     assert(batch_size != 0);
30     this->batch_size_ = batch_size;
31     return *this;
32   }
33 
batch_size()34   inline size_t batch_size() const {
35     return this->batch_size_;
36   }
37 
positive_scale(float positive_scale)38   inline VLReLUMicrokernelTester& positive_scale(float positive_scale) {
39     assert(positive_scale > 0.0f);
40     assert(std::isnormal(positive_scale));
41     this->positive_scale_ = positive_scale;
42     return *this;
43   }
44 
positive_scale()45   inline float positive_scale() const {
46     return this->positive_scale_;
47   }
48 
negative_scale(float negative_scale)49   inline VLReLUMicrokernelTester& negative_scale(float negative_scale) {
50     assert(std::isnormal(negative_scale));
51     this->negative_scale_ = negative_scale;
52     return *this;
53   }
54 
negative_scale()55   inline float negative_scale() const {
56     return this->negative_scale_;
57   }
58 
input_zero_point(int16_t input_zero_point)59   inline VLReLUMicrokernelTester& input_zero_point(int16_t input_zero_point) {
60     this->input_zero_point_ = input_zero_point;
61     return *this;
62   }
63 
input_zero_point()64   inline int16_t input_zero_point() const {
65     return this->input_zero_point_;
66   }
67 
output_zero_point(int16_t output_zero_point)68   inline VLReLUMicrokernelTester& output_zero_point(int16_t output_zero_point) {
69     this->output_zero_point_ = output_zero_point;
70     return *this;
71   }
72 
output_zero_point()73   inline int16_t output_zero_point() const {
74     return this->output_zero_point_;
75   }
76 
iterations(size_t iterations)77   inline VLReLUMicrokernelTester& iterations(size_t iterations) {
78     this->iterations_ = iterations;
79     return *this;
80   }
81 
iterations()82   inline size_t iterations() const {
83     return this->iterations_;
84   }
85 
Test(xnn_qs8_vlrelu_ukernel_function vlrelu,xnn_init_qs8_lrelu_params_fn init_params)86   void Test(xnn_qs8_vlrelu_ukernel_function vlrelu, xnn_init_qs8_lrelu_params_fn init_params) const {
87     ASSERT_GE(input_zero_point(), std::numeric_limits<int8_t>::min());
88     ASSERT_LE(input_zero_point(), std::numeric_limits<int8_t>::max());
89     ASSERT_GE(output_zero_point(), std::numeric_limits<int8_t>::min());
90     ASSERT_LE(output_zero_point(), std::numeric_limits<int8_t>::max());
91 
92     std::random_device random_device;
93     auto rng = std::mt19937(random_device());
94     std::uniform_int_distribution<int32_t> i8dist(
95       std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max());
96 
97     std::vector<int8_t> input(batch_size() + XNN_EXTRA_BYTES / sizeof(int8_t));
98     std::vector<int8_t> output(batch_size());
99     std::vector<int8_t> output_ref(batch_size());
100     for (size_t iteration = 0; iteration < iterations(); iteration++) {
101       std::generate(input.begin(), input.end(), [&]() { return i8dist(rng); });
102       std::fill(output.begin(), output.end(), INT8_C(0xA5));
103 
104       union xnn_qs8_lrelu_params params;
105       init_params(&params, positive_scale(), negative_scale(), input_zero_point(), output_zero_point());
106 
107       // Call optimized micro-kernel.
108       vlrelu(batch_size() * sizeof(int8_t), input.data(), output.data(), &params);
109 
110       // Compute reference results
111       const int32_t positive_multiplier = (int32_t) lrintf(-256.0f * positive_scale());
112       const int32_t negative_multiplier = (int32_t) lrintf(-256.0f * negative_scale());
113       for (size_t i = 0; i < batch_size(); i++) {
114         const int32_t input_value = (input_zero_point() - input[i]) << 7;
115         const int32_t multiplier = input_value <= 0 ? positive_multiplier : negative_multiplier;
116         int32_t output_value = math_asr_s32(input_value * multiplier + INT32_C(0x4000), 15) + output_zero_point();
117         output_value = std::min<int32_t>(output_value, std::numeric_limits<int8_t>::max());
118         output_value = std::max<int32_t>(output_value, std::numeric_limits<int8_t>::min());
119         output_ref[i] = static_cast<int8_t>(output_value);
120       }
121 
122       // Verify results.
123       for (size_t i = 0; i < batch_size(); i++) {
124         ASSERT_EQ(int32_t(output[i]), int32_t(output_ref[i]))
125           << "at " << i << " / " << batch_size()
126           << ", x[" << i << "] = " << int32_t(input[i]);
127       }
128     }
129   }
130 
Test(xnn_qu8_vlrelu_ukernel_function vlrelu,xnn_init_qu8_lrelu_params_fn init_params)131   void Test(xnn_qu8_vlrelu_ukernel_function vlrelu, xnn_init_qu8_lrelu_params_fn init_params) const {
132     ASSERT_GE(input_zero_point(), std::numeric_limits<uint8_t>::min());
133     ASSERT_LE(input_zero_point(), std::numeric_limits<uint8_t>::max());
134     ASSERT_GE(output_zero_point(), std::numeric_limits<uint8_t>::min());
135     ASSERT_LE(output_zero_point(), std::numeric_limits<uint8_t>::max());
136 
137     std::random_device random_device;
138     auto rng = std::mt19937(random_device());
139     std::uniform_int_distribution<int32_t> u8dist(
140       std::numeric_limits<uint8_t>::min(), std::numeric_limits<uint8_t>::max());
141 
142     std::vector<uint8_t> input(batch_size() + XNN_EXTRA_BYTES / sizeof(uint8_t));
143     std::vector<uint8_t> output(batch_size());
144     std::vector<uint8_t> output_ref(batch_size());
145     for (size_t iteration = 0; iteration < iterations(); iteration++) {
146       std::generate(input.begin(), input.end(), [&]() { return u8dist(rng); });
147       std::fill(output.begin(), output.end(), UINT8_C(0xA5));
148 
149       union xnn_qu8_lrelu_params params;
150       init_params(&params, positive_scale(), negative_scale(), input_zero_point(), output_zero_point());
151 
152       // Call optimized micro-kernel.
153       vlrelu(batch_size() * sizeof(uint8_t), input.data(), output.data(), &params);
154 
155       // Compute reference results
156       const int32_t positive_multiplier = (int32_t) lrintf(-256.0f * positive_scale());
157       const int32_t negative_multiplier = (int32_t) lrintf(-256.0f * negative_scale());
158       for (size_t i = 0; i < batch_size(); i++) {
159         const int32_t input_value = (input_zero_point() - input[i]) << 7;
160         const int32_t multiplier = input_value <= 0 ? positive_multiplier : negative_multiplier;
161         int32_t output_value = math_asr_s32(input_value * multiplier + INT32_C(0x4000), 15) + output_zero_point();
162         output_value = std::min<int32_t>(output_value, std::numeric_limits<uint8_t>::max());
163         output_value = std::max<int32_t>(output_value, std::numeric_limits<uint8_t>::min());
164         output_ref[i] = static_cast<uint8_t>(output_value);
165       }
166 
167       // Verify results.
168       for (size_t i = 0; i < batch_size(); i++) {
169         ASSERT_EQ(int32_t(output[i]), int32_t(output_ref[i]))
170           << "at " << i << " / " << batch_size()
171           << ", x[" << i << "] = " << int32_t(input[i]);
172       }
173     }
174   }
175 
176  private:
177   float positive_scale_ = 1.75f;
178   float negative_scale_ = 0.75f;
179   int16_t input_zero_point_ = 1;
180   int16_t output_zero_point_ = 5;
181   size_t batch_size_ = 1;
182   size_t iterations_ = 15;
183 };
184