xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_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 "../performance_parameters.hpp"
29 #include "../std_transforms_fixed.hpp"
30 
31 namespace arm_gemm {
32 
33 // Kernel definition
34 void a64_gemm_u8_4x4(const uint8_t *Apanel, const uint8_t *Bpanel, uint32_t *Cpanel, int ablocks, int bblocks, int K);
35 
36 class cls_a64_gemm_u8_4x4 {
37 public:
38     typedef uint8_t operand_type;
39     typedef uint32_t result_type;
40 
41     typedef void (*kern_type)(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
42 
43     /* Describes the data layout for A input */
44     static const int A_interleave = 4;
45     static const int A_block = 16;
46     static const bool A_transpose = false;
47 
48     /* Same for B input */
49     static const int B_interleave = 4;
50     static const int B_block = 16;
51     static const bool B_transpose = true;
52 
53     /* Kernel blocking parameters */
out_width()54     static unsigned int out_width() {
55         return 4;
56     }
57 
out_height()58     static unsigned int out_height() {
59         return 4;
60     }
61 
k_unroll()62     static unsigned int k_unroll() {
63         return 16;
64     }
65 
66     // Use the standard fixed size transforms.
67     StdTransformsFixed<operand_type, result_type, 4, 4, 16> transforms = {};
68     StdTransformsFixed<operand_type, result_type, 4, 4, 16, true> transforms_quantized = {};
69 
70     template<typename T>
get_performance_parameters(const CPUInfo * ci)71     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
72         if (std::is_same<T, uint32_t>::value) {
73             switch (ci->get_cpu_model()) {
74                 case CPUModel::A55r0:
75                 case CPUModel::A55r1:
76                     return { 2.25, 2.92, 1.84 };
77                 case CPUModel::A510:
78                     return { 2.64, 2.72, 2.64 };
79                 default:
80                     return { 7.95, 3.76, 7.27 };
81             }
82         }
83 
84         if (std::is_same<T, uint8_t>::value) {
85             switch(ci->get_cpu_model()) {
86                 case CPUModel::A55r0:
87                 case CPUModel::A55r1:
88                     return { 2.25, 2.18, 0.09 };
89                 case CPUModel::A510:
90                     return { 2.64, 1.79, 0.10 };
91                 default:
92                     return { 7.95, 4.09, 0.33 };
93             }
94         }
95 
96         return { 0.0 };
97     }
98 
99     kern_type kernel = a64_gemm_u8_4x4;
100 
cls_a64_gemm_u8_4x4(const CPUInfo *)101     cls_a64_gemm_u8_4x4(const CPUInfo *) { }
102 };
103 
104 } // namespace arm_gemm
105 
106 #endif // __aarch64__
107