xref: /aosp_15_r20/external/ComputeLibrary/src/core/NEON/kernels/arm_gemm/kernels/sve_hybrid_fp16_mla_6x4VL.hpp (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1 /*
2  * Copyright (c) 2019-2021 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
14  * all 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
21  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
22  * IN THE SOFTWARE.
23  */
24 #pragma once
25 #ifdef ARM_COMPUTE_ENABLE_SVE
26 
27 #include "../std_transforms_sve.hpp"
28 #include "../performance_parameters.hpp"
29 
30 #define ARGLIST  \
31     unsigned int, const unsigned int *, \
32     IndirectInputArg<__fp16>, \
33     size_t, size_t, \
34     const __fp16 *, \
35     IndirectOutputArg<__fp16>, \
36     const __fp16 *, Activation, bool
37 
38 namespace arm_gemm
39 {
40 // Actual kernel implementations
41 void sve_hybrid_fp16_mla_6x4VL( ARGLIST );
42 void sve_hybrid_fp16_mla_6x4VL_a64fx( ARGLIST );
43 
44 class cls_sve_hybrid_fp16_mla_6x4VL
45 {
46 public:
47     typedef __fp16 lhs_operand_type;
48     typedef __fp16 rhs_operand_type;
49     typedef __fp16 result_type;
50 
51     typedef void (*kern_type)( ARGLIST );
52 
53     /* Kernel blocking parameters */
out_height()54     static constexpr unsigned int out_height()
55     {
56         return 6;
57     }
58 
out_width()59     static unsigned int out_width()
60     {
61         return get_vector_length<__fp16>() * 4;
62     }
63 
k_unroll()64     static constexpr unsigned int k_unroll()
65     {
66         return 1;
67     }
68 
supports_accumulate()69     static constexpr bool supports_accumulate()
70     {
71         return true;
72     }
73 
74     StdTransformsSVE<rhs_operand_type, result_type, 6, 4, 1> transforms = {};
75     template<typename T>
get_performance_parameters(const CPUInfo * ci)76     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
77     {
78 
79         if (std::is_same<T, __fp16>::value) {
80             switch (ci->get_cpu_model()) {
81                 default:
82                     return { 15.90 };
83                 case CPUModel::A510:
84                     return { 12.44 };
85                 case CPUModel::V1:
86                     return { 31.51 };
87             }
88         }
89 
90         return { 1.0 };
91     }
92 
93     // Default to the generic kernel
94     kern_type kernel=sve_hybrid_fp16_mla_6x4VL;
cls_sve_hybrid_fp16_mla_6x4VL(const CPUInfo * ci)95     cls_sve_hybrid_fp16_mla_6x4VL(const CPUInfo *ci)
96     {
97         switch(ci->get_cpu_model()) {
98             default:
99                 break;
100             case CPUModel::A64FX:
101                 kernel=sve_hybrid_fp16_mla_6x4VL_a64fx;
102                 break;
103         }
104     }
105 };
106 
107 } // namespace arm_gemm
108 
109 #undef ARGLIST
110 
111 #endif // __aarch64__
112