xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_s8_4x4.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2022 Arm Limited.
3  *
4  * SPDX-License-Identifier: MIT
5  *
6  * Permission is hereby granted, free of charge, to any person obtaining a copy
7  * of this software and associated documentation files (the "Software"), to
8  * deal in the Software without restriction, including without limitation the
9  * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10  * sell copies of the Software, and to permit persons to whom the Software is
11  * furnished to do so, subject to the following conditions:
12  *
13  * The above copyright notice and this permission notice shall be included in all
14  * copies or substantial portions of the Software.
15  *
16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19  * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22  * SOFTWARE.
23  */
24 #pragma once
25 
26 #ifdef __aarch64__
27 
28 #include "../std_transforms_fixed.hpp"
29 #include "../performance_parameters.hpp"
30 
31 namespace arm_gemm {
32 
33 // Load the actual kernel
34 void a64_gemm_s8_4x4(const int8_t *, const int8_t *, int32_t *, int, int, int);
35 
36 #include "arm_gemm.hpp"
37 
38 class cls_a64_gemm_s8_4x4 {
39 public:
40     typedef int8_t operand_type;
41     typedef int32_t result_type;
42 
43     typedef void (*kern_type)(const int8_t *, const int8_t *, int32_t *, int, int, int);
44 
45     /* Kernel blocking parameters */
out_width()46     static unsigned int out_width() {
47         return 4;
48     }
49 
out_height()50     static unsigned int out_height() {
51         return 4;
52     }
53 
k_unroll()54     static unsigned int k_unroll() {
55         return 16;
56     }
57 
58     // Use the standard fixed size transforms.
59     StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
60     StdTransformsFixed<operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
61 
62     template<typename T>
get_performance_parameters(const CPUInfo * ci)63     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
64         if (std::is_same<T, int32_t>::value) {
65             switch (ci->get_cpu_model()) {
66                 case CPUModel::A55r0:
67                 case CPUModel::A55r1:
68                     return { 3.12, 2.93, 1.84 };
69                 case CPUModel::A510:
70                     return { 3.32, 2.56, 2.63 };
71                 default:
72                     return { 7.97, 3.72, 7.31 };
73             }
74         }
75 
76         if (std::is_same<T, int8_t>::value) {
77             switch(ci->get_cpu_model()) {
78                 case CPUModel::A55r0:
79                 case CPUModel::A55r1:
80                     return { 3.12, 2.18, 0.09 };
81                 case CPUModel::A510:
82                     return { 3.33, 2.89, 0.09 };
83                 default:
84                     return { 7.97, 3.74, 0.34 };
85             }
86         }
87     }
88 
89     kern_type kernel=a64_gemm_s8_4x4;
90 
cls_a64_gemm_s8_4x4(const CPUInfo *)91     cls_a64_gemm_s8_4x4(const CPUInfo *) { }
92 };
93 
94 } // namespace arm_gemm
95 
96 #endif // __aarch64__
97 
98