1 /*
2 * Copyright (c) 2017-2020 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
25 // This can only be built if the target/compiler supports FP16 arguments.
26 #ifdef __ARM_FP16_ARGS
27
28 #include "arm_gemm.hpp"
29
30 #include "gemm_common.hpp"
31 #include "gemm_hybrid.hpp"
32 #include "gemm_hybrid_indirect.hpp"
33 #include "gemm_implementation.hpp"
34 #include "gemm_interleaved.hpp"
35 #include "gemm_interleaved_pretransposed_2d.hpp"
36
37 #include "kernels/a32_sgemm_8x6.hpp"
38 #include "kernels/a64_hgemm_8x24.hpp"
39 #include "kernels/a64_hybrid_fp16_mla_6x32.hpp"
40 #include "kernels/a64_sgemm_8x12.hpp"
41 #include "kernels/sve_hybrid_fp16_mla_6x4VL.hpp"
42 #include "kernels/sve_interleaved_fp16_mla_8x3VL.hpp"
43
44 namespace arm_gemm {
45
46 static const GemmImplementation<__fp16, __fp16> gemm_fp16_methods[] = {
47 #if defined(__ARM_FEATURE_SVE)
48 {
49 GemmMethod::GEMM_HYBRID,
50 "sve_hybrid_fp16_mla_6x4VL",
51 nullptr,
__anona6326bd60102() 52 [](const GemmArgs &args) { return ((args._Ksize <= 256) && (args._Nsize <= 256)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
__anona6326bd60202() 53 [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>(args); }
54 },
55 {
56 GemmMethod::GEMM_INTERLEAVED,
57 "sve_interleaved_fp16_mla_8x3VL",
__anona6326bd60302() 58 [](const GemmArgs &args) { return (args._Ksize > 4); },
59 nullptr,
__anona6326bd60402() 60 [](const GemmArgs &args) { return new GemmInterleaved<cls_sve_interleaved_fp16_mla_8x3VL, __fp16, __fp16>(args); }
61 },
62 #endif
63
64 #if defined(__aarch64__) && (defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) || defined(FP16_KERNELS))
65 GemmImplementation<__fp16, __fp16>::with_estimate(
66 GemmMethod::GEMM_HYBRID,
67 "a64_hybrid_fp16_mla_6x32",
68 #ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
__anona6326bd60502() 69 [](const GemmArgs &args) { return args._ci->has_fp16(); },
70 #else
71 nullptr,
72 #endif
__anona6326bd60602() 73 [](const GemmArgs &args) { return GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>::estimate_cycles(args, cls_a64_hybrid_fp16_mla_6x32::get_performance_parameters(args._ci)); },
__anona6326bd60702() 74 [](const GemmArgs &args) { return new GemmHybridIndirect<cls_a64_hybrid_fp16_mla_6x32, __fp16, __fp16>(args); }
75 ),
76 GemmImplementation<__fp16, __fp16>::with_estimate(
77 GemmMethod::GEMM_INTERLEAVED,
78 "a64_hgemm_8x24",
79 #ifndef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
__anona6326bd60802() 80 [](const GemmArgs &args) { return args._ci->has_fp16(); },
81 #else
82 nullptr,
83 #endif
__anona6326bd60902() 84 [](const GemmArgs &args) { return GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16>::estimate_cycles(args, cls_a64_hgemm_8x24::get_performance_parameters(args._ci)); },
__anona6326bd60a02() 85 [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_hgemm_8x24, __fp16, __fp16>(args); }
86 ),
87 #endif // aarch64 && FP16
88 #ifdef __aarch64__
89 {
90 GemmMethod::GEMM_INTERLEAVED,
91 "a64_sgemm_8x12",
92 nullptr,
__anona6326bd60b02() 93 [](const GemmArgs &args) { return !args._ci->has_fp16(); },
__anona6326bd60c02() 94 [](const GemmArgs &args) { return new GemmInterleaved<cls_a64_sgemm_8x12, __fp16, __fp16>(args); }
95 },
96 #elif defined(__arm__)
97 {
98 GemmMethod::GEMM_INTERLEAVED,
99 "sgemm_8x6",
100 nullptr,
101 nullptr,
102 [](const GemmArgs &args) { return new GemmInterleaved<sgemm_8x6, __fp16, __fp16>(args); }
103 },
104 #else // not AArch64 or AArch32
105 # error Unknown Architecture
106 #endif
107 {
108 GemmMethod::DEFAULT,
109 "",
110 nullptr,
111 nullptr,
112 nullptr,
113 }
114 };
115
116 template<>
gemm_implementation_list()117 const GemmImplementation<__fp16, __fp16> *gemm_implementation_list<__fp16, __fp16>() {
118 return gemm_fp16_methods;
119 }
120
121 /* Explicitly instantiate the external functions for these types. */
122 template UniqueGemmCommon<__fp16, __fp16> gemm<__fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
123 template KernelDescription get_gemm_method<__fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
124 template std::vector<KernelDescription> get_compatible_kernels<__fp16, __fp16, Nothing>(const GemmArgs &args, const Nothing &);
125
126 } // namespace arm_gemm
127
128 #endif // __ARM_FP16_ARGS
129