1 /*
2  * Copyright (c) 2022 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 "../bfloat.hpp"
29 #include "../kernel_weight_format.hpp"
30 #include "../performance_parameters.hpp"
31 
32 #define ARGLIST  \
33     const bfloat16 *, const bfloat16 *, size_t, \
34     float *, int, size_t, int
35 
36 namespace arm_gemm
37 {
38 // Actual kernel implementations
39 void sve_ffinterleaved_bf16fp32_mmla_8x3VL( ARGLIST );
40 
41 class cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL
42 {
43 public:
44     typedef bfloat16 operand_type;
45     typedef float result_type;
46 
47     typedef void (*kern_type)( ARGLIST );
48 
49     /* Kernel blocking parameters */
out_height()50     static constexpr unsigned int out_height()
51     {
52         return 8;
53     }
54 
out_width()55     static unsigned int out_width()
56     {
57         return get_vector_length<float>() * 3;
58     }
stripe_width()59     static unsigned int stripe_width()
60     {
61         return get_vector_length<float>();
62     }
63 
kernel_weight_format()64     static KernelWeightFormat kernel_weight_format()
65     {
66         return KernelWeightFormat::VL2VL_BL64;
67     }
68 
k_unroll()69     static constexpr unsigned int k_unroll()
70     {
71         return 4;
72     }
73 
74 
75     StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2> transforms = {};
76     StdTransformsSVE<operand_type, result_type, 8, 6, 4, 2, true> transforms_quantized = {};
77     template<typename T>
get_performance_parameters(const CPUInfo * ci)78     static inline PerformanceParameters get_performance_parameters(const CPUInfo *ci)
79     {
80 
81         if (std::is_same<T, bfloat16>::value) {
82             switch (ci->get_cpu_model()) {
83                 default:
84                     return { 39.90, 8.55, 4.42 };
85             }
86         }
87 
88 
89         if (std::is_same<T, float>::value) {
90             switch (ci->get_cpu_model()) {
91                 default:
92                     return { 39.66, 5.18, 4.37 };
93             }
94         }
95 
96         return { 1.0 };
97     }
98 
99     // Default to the generic kernel
100     kern_type kernel=sve_ffinterleaved_bf16fp32_mmla_8x3VL;
cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL(const CPUInfo *)101     cls_sve_ffinterleaved_bf16fp32_mmla_8x3VL(const CPUInfo *)
102     {
103     }
104 };
105 
106 } // namespace arm_gemm
107 
108 #undef ARGLIST
109 #endif // ARM_COMPUTE_ENABLE_SVE
110