megenginelite-sys 1.8.2

A safe megenginelite wrapper in Rust
Documentation 
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
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174

/**
 * \file dnn/src/arm_common/elemwise_helper/kimpl/hswish.h
 * MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
 *
 * Copyright (c) 2014-2021 Megvii Inc. All rights reserved.
 *
 * Unless required by applicable law or agreed to in writing,
 * software distributed under the License is distributed on an
 * "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or
 * implied.
 */
#pragma once

#include "src/arm_common/elemwise_helper/kimpl/kern_macro_prologue.h"
#include "src/arm_common/elemwise_helper/kimpl/op_base.h"

namespace megdnn {
namespace arm_common {

template <typename src_ctype, typename dst_ctype = src_ctype>
struct HSwishOpBase : UnaryOpBase<src_ctype, dst_ctype> {
    using UnaryOpBase<src_ctype, dst_ctype>::UnaryOpBase;
    void operator()(const src_ctype& src, dst_ctype* dst) const {
        *dst = operator()(src);
    }
    dst_ctype operator()(const src_ctype& src) const {
        float tmp = src;
        tmp = tmp * std::max(std::min(tmp + 3.f, 6.f), 0.f) / 6.f;
        return (tmp);
    }
};

//! h_swish(x) = x * clip(x + 3, 0, 6) / 6
template <typename src_ctype, typename dst_ctype = src_ctype>
struct HSwishOp;

#define OP(_ctype, _neon_type, _neon_type2, _func_suffix, _simd_width)      \
    template <>                                                             \
    struct HSwishOp<_ctype> : HSwishOpBase<_ctype> {                        \
        using HSwishOpBase::HSwishOpBase;                                   \
        using HSwishOpBase::operator();                                     \
        constexpr static size_t SIMD_WIDTH = _simd_width;                   \
        void operator()(const _neon_type2& src, _ctype* dst) const {        \
            auto vitem = operator()(src);                                   \
            vst1q_##_func_suffix(dst, vitem.val[0]);                        \
            vst1q_##_func_suffix(dst + SIMD_WIDTH, vitem.val[1]);           \
        }                                                                   \
        void operator()(const _neon_type& src, _ctype* dst) const {         \
            auto vitem = operator()(src);                                   \
            vst1q_##_func_suffix(dst, vitem);                               \
        }                                                                   \
        _neon_type2 operator()(const _neon_type2& src) const {              \
            auto val1 = src.val[0];                                         \
            auto val2 = src.val[1];                                         \
            H_SWISH_KERN(_func_suffix, val1, val2);                         \
            return {{val1, val2}};                                          \
        }                                                                   \
        _neon_type operator()(const _neon_type& src) const {                \
            auto val_zero = vdupq_n_##_func_suffix(0.f);                    \
            auto val_six = vdupq_n_##_func_suffix(6.f);                     \
            auto val_three = vdupq_n_##_func_suffix(3.f);                   \
            auto val_rec_six = vdupq_n_##_func_suffix(1.f / 6.f);           \
            auto clip1 = vmaxq_##_func_suffix(                              \
                    vminq_##_func_suffix(                                   \
                            vaddq_##_func_suffix(src, val_three), val_six), \
                    val_zero);                                              \
            return vmulq_##_func_suffix(                                    \
                    vmulq_##_func_suffix(src, clip1), val_rec_six);         \
        }                                                                   \
    };

OP(dt_float32, float32x4_t, float32x4x2_t, f32, 4)
#if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
OP(__fp16, float16x8_t, float16x8x2_t, f16, 8)
#endif
#undef OP

template <>
struct HSwishOpBase<dt_qint32, dt_qint8> : UnaryOpBase<dt_qint32, dt_qint8> {
    using UnaryOpBase::UnaryOpBase;
    void operator()(const dt_qint32& src, dt_qint8* dst) const {
        *dst = operator()(src);
    }

    dt_qint8 operator()(const dt_qint32& src) const {
        float tmp = src.as_int32() * this->scale_src;
        tmp = tmp * std::max(std::min(tmp + 3.f, 6.f), 0.f) / 6.f;
        tmp *= this->scale_dst;
        return QConverter::convert<dt_qint8, float>(tmp);
    }
};

template <>
struct HSwishOpBase<dt_qint32, dt_quint8> : UnaryOpBase<dt_qint32, dt_quint8> {
    using UnaryOpBase::UnaryOpBase;
    void operator()(const dt_qint32& src, dt_quint8* dst) const {
        *dst = operator()(src);
    }

    dt_quint8 operator()(const dt_qint32& src) const {
        float tmp = src.as_int32() * this->scale_src;
        tmp = tmp * std::max(std::min(tmp + 3.f, 6.f), 0.f) / 6.f;
        tmp *= this->scale_dst;
        return QConverter::convert<dt_quint8, float>(tmp, zp);
    }
};

template <>
struct HSwishOp<dt_qint32, dt_qint8> : HSwishOpBase<dt_qint32, dt_qint8> {
    using HSwishOpBase::HSwishOpBase;
    using HSwishOpBase::operator();
    constexpr static size_t SIMD_WIDTH = 4;

    void operator()(const int32x4x2_t& vsrc, dt_qint8* dst) const {
        vst1_s8(reinterpret_cast<int8_t*>(dst), operator()(vsrc));
    }
    void operator()(const int32x4_t& vsrc, dt_qint8* dst) const {
        vst1_lane_s32(
                reinterpret_cast<int32_t*>(dst), (int32x2_t)(operator()(vsrc)), 0);
    }

    int8x8_t operator()(const int32x4x2_t& vsrc) const {
        auto vitem0 = vmulq_f32(vcvtq_f32_s32(vsrc.val[0]), this->vscale_src);
        auto vitem1 = vmulq_f32(vcvtq_f32_s32(vsrc.val[1]), this->vscale_src);

        H_SWISH_KERN(f32, vitem0, vitem1);
        vitem0 = vmulq_f32(vitem0, this->vscale_dst);
        vitem1 = vmulq_f32(vitem1, this->vscale_dst);

        return QConverter::convert<int8x8_t, float32x4x2_t>({{vitem0, vitem1}});
    }
    int8x8_t operator()(const int32x4_t& src) const {
        auto vitem0 = vmulq_f32(vcvtq_f32_s32(src), this->vscale_src);

        H_SWISH_KERN_N1(f32, vitem0);
        vitem0 = vmulq_f32(vitem0, this->vscale_dst);

        return QConverter::convert<int8x8_t, float32x4_t>(vitem0);
    }
    int8x8_t operator()(const float32x4_t& src) const {
        auto vitem0 = vmulq_f32(src, this->vscale_src);
        H_SWISH_KERN_N1(f32, vitem0);
        vitem0 = vmulq_f32(vitem0, this->vscale_dst);
        return QConverter::convert<int8x8_t, float32x4_t>(vitem0);
    }
};

template <>
struct HSwishOp<dt_qint32, dt_quint8> : HSwishOpBase<dt_qint32, dt_quint8> {
    using HSwishOpBase::HSwishOpBase;
    using HSwishOpBase::operator();
    constexpr static size_t SIMD_WIDTH = 4;

    void operator()(const int32x4x2_t& vsrc, dt_quint8* dst) const {
        vst1_u8(reinterpret_cast<uint8_t*>(dst), operator()(vsrc));
    }

    uint8x8_t operator()(const int32x4x2_t& vsrc) const {
        auto vitem0 = vmulq_f32(vcvtq_f32_s32(vsrc.val[0]), this->vscale_src);
        auto vitem1 = vmulq_f32(vcvtq_f32_s32(vsrc.val[1]), this->vscale_src);
        H_SWISH_KERN(f32, vitem0, vitem1);
        vitem0 = vmulq_f32(vitem0, this->vscale_dst);
        vitem1 = vmulq_f32(vitem1, this->vscale_dst);

        return QConverter::convert<uint8x8_t, float32x4x2_t>(
                {{vitem0, vitem1}}, this->vzp);
    }
};

}  // namespace arm_common
}  // namespace megdnn

#include "src/arm_common/elemwise_helper/kimpl/kern_macro_epilogue.h"
// vim: syntax=cpp.doxygen