xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/a32_sgemm_8x6.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2017-2018 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 __arm__
27 
28 #include "../std_transforms_fixed.hpp"
29 
30 namespace arm_gemm {
31 
32 // Actual kernel implementations
33 void a32_sgemm_8x6(const float *, const float *, float *, int, int, int);
34 void a32_sgemm_8x6_a53(const float *, const float *, float *, int, int, int);
35 void a32_sgemm_8x6_a55r1(const float *, const float *, float *, int, int, int);
36 
37 // 8x6 SGEMM "strategy" class.
38 //
39 // This describes the characteristics of a family of kernels, in terms of
40 // the required interleave properties and the output block size.
41 //
42 // All kernels in the family must share these characteristics.  The actual
43 // kernel to be used can be chosen at runtime, based on the CPU_type
44 // structure.
45 class sgemm_8x6 {
46 public:
47     typedef float operand_type;
48     typedef float result_type;
49 
50     typedef void (*kern_type)(const float *, const float *, float *, int, int, int);
51 
52     /* Kernel blocking parameters */
out_width()53     static unsigned int out_width() {
54         return 8;
55     }
56 
out_height()57     static unsigned int out_height() {
58         return 6;
59     }
60 
k_unroll()61     static unsigned int k_unroll() {
62         return 1;
63     }
64 
65     // Use the standard fixed size transforms.
66     StdTransformsFixed<operand_type, result_type, 6, 8> transforms = {};
67 
68     kern_type kernel = a32_sgemm_8x6;
69 
sgemm_8x6(const CPUInfo * ci)70     sgemm_8x6(const CPUInfo *ci) {
71         switch(ci->get_cpu_model()) {
72             case CPUModel::A53:
73                 kernel = a32_sgemm_8x6_a53;
74                 break;
75 
76             case CPUModel::A55r1:
77                 kernel = a32_sgemm_8x6_a55r1;
78                 break;
79 
80             default:
81                 /* Generic kernel is selected by default. */
82                 break;
83         }
84     }
85 };
86 
87 } // namespace arm_gemm
88 #endif // __arm__
89