1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
|
/*
* Copyright (c) 2019-2020 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#ifdef __aarch64__
#include "arm_gemm.hpp"
#include "kernels/a64_hybrid_u8u32_dot_16x4.hpp"
#include "kernels/a64_smallK_hybrid_u8u32_dot_4x6.hpp"
#include "kernels/a64_smallK_hybrid_u8u32_dot_4x8.hpp"
#include "kernels/sve_hybrid_u8u32_dot_4VLx4.hpp"
#include "kernels/sve_smallK_hybrid_u8u32_dot_1VLx8.hpp"
#include "gemm_hybrid_quantized.hpp"
#include "quantize_wrapper.hpp"
namespace arm_gemm {
static const GemmImplementation<uint8_t, uint8_t, Requantize32> gemm_quint8_methods[] =
{
#ifdef __ARM_FEATURE_SVE
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"smallK_hybrid_u8u32_dot_1VLx8",
[](const GemmArgs &args, const Requantize32 &) { return args._Ksize<=64; },
nullptr,
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_1VLx8, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"hybrid_u8u32_dot_4VLx4",
[](const GemmArgs &args, const Requantize32 &) { return args._Ksize>=16; },
[](const GemmArgs &args, const Requantize32 &) { return ((args._Ksize <= 128) && (args._Nsize <= 128)) || ((args._nmulti > 1) && ((args._Msize / args._maxthreads) < 8)); },
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<hybrid_u8u32_dot_4VLx4, uint8_t, uint8_t>(args, qp); }
},
#endif
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"smallK_hybrid_u8u32_dot_4x8",
[](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize<=32); },
nullptr,
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_4x8, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"smallK_hybrid_u8u32_dot_4x6",
[](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && (args._Nsize % 4 == 0) && (args._Ksize>32) && (args._Ksize<=64); },
nullptr,
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<smallK_hybrid_u8u32_dot_4x6, uint8_t, uint8_t>(args, qp); }
},
{
GemmMethod::GEMM_HYBRID_QUANTIZED,
"hybrid_u8u32_dot_16x4",
[](const GemmArgs &args, const Requantize32 &) { return args._ci->has_dotprod() && args._Ksize>=16; },
[](const GemmArgs &args, const Requantize32 &) { return ((args._Nsize<=256) && (args._Ksize>128)) || (args._maxthreads >= 8); },
[](const GemmArgs &args, const Requantize32 &qp) { return new GemmHybridQuantized<hybrid_u8u32_dot_16x4, uint8_t, uint8_t>(args, qp); }
},
/** QUANTIZE_WRAPPER_2D enables 2D parallelisation hint for IScheduler in NEGEMMAssemblyDispatch */
{
GemmMethod::QUANTIZE_WRAPPER_2D,
"quantized_wrapper_2d",
nullptr,
[](const GemmArgs &args, const Requantize32 &) { return (args._maxthreads >= 8) && (args._Msize >= 8) && (args._Nsize >= 8);},
[](const GemmArgs &args, const Requantize32 &qp) { return new QuantizeWrapper<uint8_t, uint8_t, uint32_t>(args, qp); }
},
{
GemmMethod::QUANTIZE_WRAPPER,
"quantized_wrapper",
nullptr,
nullptr,
[](const GemmArgs &args, const Requantize32 &qp) { return new QuantizeWrapper<uint8_t, uint8_t, uint32_t>(args, qp); }
},
{
GemmMethod::DEFAULT,
"",
nullptr,
nullptr,
nullptr
}
};
template<>
const GemmImplementation<uint8_t, uint8_t, Requantize32> *gemm_implementation_list<uint8_t, uint8_t, Requantize32>() {
return gemm_quint8_methods;
}
template UniqueGemmCommon<uint8_t, uint8_t> gemm<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
template KernelDescription get_gemm_method<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
template std::vector<KernelDescription> get_compatible_kernels<uint8_t, uint8_t, Requantize32>(const GemmArgs &args, const Requantize32 &os);
} // namespace arm_gemm
#endif // __aarch64__
|
