xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/a64_sgemm_8x12.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2021 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 "../std_transforms_fixed.hpp"
29 #include "../performance_parameters.hpp"
30 
31 #include "../bfloat.hpp"
32 
33 namespace arm_gemm {
34 
35 // Actual kernel implementations
36 void a64_sgemm_asimd_8x12(const float *, const float *, float *, int, int, int);
37 void a64_sgemm_asimd_8x12_a53(const float *, const float *, float *, int, int, int);
38 void a64_sgemm_asimd_8x12_a55(const float *, const float *, float *, int, int, int);
39 void a64_sgemm_asimd_8x12_a55r1(const float *, const float *, float *, int, int, int);
40 void a64_sgemm_asimd_8x12_x1(const float *, const float *, float *, int, int, int);
41 
42 // 8x12 SGEMM "strategy" class.
43 //
44 // This describes the characteristics of a family of kernels, in terms of
45 // the required interleave properties and the output block size.
46 //
47 // All kernels in the family must share these characteristics.  The actual
48 // kernel to be used can be chosen at runtime, based on the CPU_type
49 // structure.
50 class cls_a64_sgemm_8x12 {
51 public:
52     typedef float operand_type;
53     typedef float result_type;
54 
55     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
56 
57     /* Kernel blocking parameters */
out_width()58     static unsigned int out_width() {
59         return 12;
60     }
61 
out_height()62     static unsigned int out_height() {
63         return 8;
64     }
65 
k_unroll()66     static unsigned int k_unroll() {
67         return 1;
68     }
69 
70     // Use the standard fixed size transforms.
71     StdTransformsFixed<operand_type, result_type, 8, 12> transforms = {};
72 
73     template<typename T>
get_performance_parameters(const CPUInfo * ci)74     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
75         if (std::is_same<T, float>::value) {
76             switch (ci->get_cpu_model()) {
77                 case CPUModel::A55r1:
78                     return { 3.954, 1.252, 1.141 };
79 
80                 case CPUModel::A53:
81                     return { 2.777, 0.987, 0.898 };
82 
83                 case CPUModel::A73:
84                     return { 2.885, 1.429, 1.163 };
85 
86                 case CPUModel::V1:
87                     return { 14.95, 9.95, 5.28 };
88 
89                 default:
90                     return { 7.2307, 3.876, 2.932 };
91             }
92         }
93 
94         if (std::is_same<T, bfloat16>::value) {
95             switch(ci->get_cpu_model()) {
96                 case CPUModel::A510:
97                     return { 4.98, 2.27, 3.05 };
98 
99                 default:
100                     return { 7.99, 5.06, 7.32 };
101             }
102         }
103     }
104 
105     kern_type kernel=a64_sgemm_asimd_8x12;
106 
cls_a64_sgemm_8x12(const CPUInfo * ci)107     cls_a64_sgemm_8x12(const CPUInfo *ci) {
108         // Select specific kernel if available
109         switch(ci->get_cpu_model()) {
110             case CPUModel::A53:
111                 kernel = a64_sgemm_asimd_8x12_a53;
112                 break;
113 
114             case CPUModel::A55r0:
115                 kernel = a64_sgemm_asimd_8x12_a55;
116                 break;
117 
118             case CPUModel::A55r1:
119                 kernel = a64_sgemm_asimd_8x12_a55r1;
120                 break;
121 
122             case CPUModel::X1:
123                 kernel = a64_sgemm_asimd_8x12_x1;
124                 break;
125 
126             default:
127                 /* Generic kernel is initialized by default. */
128                 break;
129         }
130     }
131 };
132 
133 } // namespace arm_gemm
134 
135 #endif // __aarch64__
136