xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/a64_hgemm_8x24.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 #if defined(__aarch64__) && (defined(FP16_KERNELS) || defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC))
27 
28 #include "../performance_parameters.hpp"
29 #include "../std_transforms_fixed.hpp"
30 
31 namespace arm_gemm {
32 
33 // Actual kernel implementations
34 void a64_hgemm_asimd_8x24(const __fp16 *, const __fp16 *, __fp16 *, int, int, int);
35 void a64_hgemm_asimd_8x24_a55r1(const __fp16 *, const __fp16 *, __fp16 *, int, int, int);
36 void a64_hgemm_asimd_8x24_x1(const __fp16 *, const __fp16 *, __fp16 *, int, int, int);
37 
38 // 8x24 HGEMM "strategy" class.  Describes the kernel properties.
39 //
40 // The generic "gemm_opt" function will instantiate one of these (allowing
41 // the constructor to pick a kernel implementation).
42 class cls_a64_hgemm_8x24 {
43 public:
44     typedef __fp16 operand_type;
45     typedef __fp16 result_type;
46 
47     typedef void (*kern_type)(const __fp16 *, const __fp16 *, __fp16 *, int, int, int);
48 
49     /* Kernel blocking parameters */
out_width()50     static unsigned int out_width() {
51         return 24;
52     }
53 
out_height()54     static unsigned int out_height() {
55         return 8;
56     }
57 
k_unroll()58     static unsigned int k_unroll() {
59         return 1;
60     }
61 
62     // Use the standard fixed size transforms.
63     StdTransformsFixed<operand_type, result_type, 8, 24> transforms = {};
64 
65     template<typename T>
get_performance_parameters(const CPUInfo * ci)66     static PerformanceParameters get_performance_parameters(const CPUInfo *ci) {
67         switch (ci->get_cpu_model()) {
68             case CPUModel::A55r1:
69                 return {  7.16, 1.14, 0.67  };
70 
71             default:
72                 return { 12.67, 3.98, 1.16  };
73         }
74     }
75 
76     // Default to the generic kernel
77     kern_type kernel = a64_hgemm_asimd_8x24;
78 
cls_a64_hgemm_8x24(const CPUInfo * ci)79     cls_a64_hgemm_8x24(const CPUInfo *ci) {
80         auto model = ci->get_cpu_model();
81 
82         if (model == CPUModel::A55r1) {
83             kernel = a64_hgemm_asimd_8x24_a55r1;
84         } else if (model == CPUModel::X1) {
85             kernel = a64_hgemm_asimd_8x24_x1;
86         }
87     }
88 };
89 
90 } // namespace arm_gemm
91 
92 #endif // __aarch64__ && (FP16_KERNELS || __ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
93