xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/a64_gemm_u8_8x12.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2018,2021 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 // Load the actual kernel
34 void a64_gemm_u8_8x12(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
35 void a64_gemm_u8_8x12_a55r1(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
36 void a64_gemm_u8_8x12_x1(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
37 
38 class cls_a64_gemm_u8_8x12 {
39 public:
40     typedef uint8_t operand_type;
41     typedef uint32_t result_type;
42 
43     typedef void (*kern_type)(const uint8_t *, const uint8_t *, uint32_t *, int, int, int);
44 
45     /* Describes the data layout for A input */
46     static const int A_interleave = 8;
47     static const int A_block = 4;
48     static const bool A_transpose = false;
49 
50     /* Same for B input */
51     static const int B_interleave = 12;
52     static const int B_block = 4;
53     static const bool B_transpose = true;
54 
55     /* Kernel blocking parameters */
out_width()56     static unsigned int out_width() {
57         return 12;
58     }
59 
out_height()60     static unsigned int out_height() {
61         return 8;
62     }
63 
k_unroll()64     static unsigned int k_unroll() {
65         return 4;
66     }
67 
68     // Use the standard fixed sized transforms.
69     StdTransformsFixed<operand_type, result_type, 8, 12, 4> transforms = {};
70     StdTransformsFixed<operand_type, result_type, 8, 12, 4, true> transforms_quantized = {};
71 
72     template<typename T>
get_performance_parameters(const CPUInfo * ci)73     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
74         if (std::is_same<T, uint8_t>::value) {
75             switch (ci->get_cpu_model()) {
76                 case CPUModel::A510:
77                     return { 19.73, 3.38, 0.27 };
78 
79                 case CPUModel::A55r1:
80                     return { 15.361, 0.9341, 0.1636 };
81 
82                 case CPUModel::V1:
83                     return { 51.14, 7.38, 0.65 };
84 
85                 default:
86                     return { 29.0698, 3.9793, 0.4003 };
87             }
88         }
89 
90         if (std::is_same<T, uint32_t>::value) {
91             switch (ci->get_cpu_model()) {
92                 case CPUModel::A510:
93                     return { 19.73, 3.38, 3.70 };
94 
95                 case CPUModel::A55r1:
96                     return { 14.286, 1.171, 1.209 };
97 
98                 case CPUModel::V1:
99                     return { 61.58, 4.78, 10.83 };
100 
101                 default:
102                     return { 31.82, 3.51, 8.03 };
103             }
104         }
105 
106         return { 0.0 };
107     }
108 
109     kern_type kernel = a64_gemm_u8_8x12;
110 
cls_a64_gemm_u8_8x12(const CPUInfo * ci)111     cls_a64_gemm_u8_8x12(const CPUInfo *ci) {
112         auto mod = ci->get_cpu_model();
113 
114         if (mod == CPUModel::A55r1) {
115             kernel = a64_gemm_u8_8x12_a55r1;
116         } else if (mod == CPUModel::X1) {
117             kernel = a64_gemm_u8_8x12_x1;
118         }
119     }
120 };
121 
122 } // namespace arm_gemm
123 
124 #endif // __aarch64__
125