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
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201

/**
 * \file dnn/src/x86/elemwise/fma_util/fma_util.cpp
 * 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.
 */
#include <immintrin.h>
#ifdef WIN32
#include <avxintrin.h>
#include <fmaintrin.h>
#endif
#include <algorithm>
#include <cmath>

#include "./fma_util.h"

namespace megdnn {
namespace x86 {
namespace detail {

/*
 *  Case 1. Unary Optrs
 */

/*
 * Set initial value of the result tensor in the calculation of convolution-bias.
 */
// FAST_TANH
// tanh = x * (27 + x^2) / (27 + 9 * x^2) where (-3 <= x <= 3)
//      = -1                              where (x < -3)
//      = 1                               where (x > 3)
void fma_element_fast_tanh(size_t tsize, const float* src_ptr, float* dst_ptr) {
    size_t cur_pos = 0;
    __m256 valx, valxp2, denominator;
    __m256 val_27 = _mm256_set1_ps(27.f);
    __m256 val_9 = _mm256_set1_ps(9.f);
    __m256 one_val = _mm256_set1_ps(1.f);
    __m256 mone_val = _mm256_set1_ps(-1.f);

    for (; cur_pos + 7 < tsize; cur_pos += 8) {
        valx = _mm256_loadu_ps(src_ptr + cur_pos);
        valxp2 = _mm256_mul_ps(valx, valx);
        denominator = _mm256_add_ps(valxp2, val_27);  // use denominator as a temp var
        valx = _mm256_mul_ps(valx, denominator);      // use valx as fractions.

        denominator = _mm256_fmadd_ps(valxp2, val_9, val_27);
        valx = _mm256_div_ps(valx, denominator);
        valx = _mm256_max_ps(valx, mone_val);
        valx = _mm256_min_ps(valx, one_val);
        _mm256_storeu_ps(dst_ptr + cur_pos, valx);
    }
    for (; cur_pos < tsize; ++cur_pos) {
        float x = src_ptr[cur_pos];
        if (x > 3.f) {
            dst_ptr[cur_pos] = 1.f;
        } else if (x < -3.f) {
            dst_ptr[cur_pos] = -1.f;
        } else {
            dst_ptr[cur_pos] = x * (27.f + x * x) / (27.f + 9.f * x * x);
        }
    }
}

void fma_element_fast_tanh(const TensorND& src_tensor, const TensorND& dst_tensor) {
    size_t tsize = src_tensor.layout.total_nr_elems();
    float* src_ptr = src_tensor.ptr<float>();
    float* dst_ptr = dst_tensor.ptr<float>();
    fma_element_fast_tanh(tsize, src_ptr, dst_ptr);
}

/*
 * Case 2. Binary Optrs
 * 2.1 src0 has the same size with src1.
 * 2.2 src1 is a scalar.
 * 2.3 shape of src1 is {1, C, 1, 1}.
 * 2.4 some other optrs only for dtype float32
 */

/*
 * Case 3. Ternary Optrs
 * 3.1 src0, src1 and src2 has the same size.
 * 3.2 src0, src1 has the same size, src2 is a scalar
 * 3.3 shape of src0 and src2 is (1,C,1,1).
 */

// Case 3.1 src0, src1 and src2 has the same size.
#define FMA_TERNARY_OPTR_TEMPLATE_HEAD(optr_type)                            \
    void fma_element_##optr_type(                                            \
            size_t tsize, float* src0_ptr, float* src1_ptr, float* src2_ptr, \
            float* dst_ptr) {                                                \
        size_t cur_pos = 0;                                                  \
        __m256 val0, val1, val2;

#define FMA_TERNARY_OPTR_LOOP1                                                      \
    for (; cur_pos + 7 < tsize;                                                     \
         cur_pos += 8, src0_ptr += 8, src1_ptr += 8, src2_ptr += 8, dst_ptr += 8) { \
        val0 = _mm256_loadu_ps(src0_ptr);                                           \
        val1 = _mm256_loadu_ps(src1_ptr);                                           \
        val2 = _mm256_loadu_ps(src2_ptr);

#define FMA_TERNARY_OPTR_LOOP2 \
    }                          \
    for (; cur_pos < tsize; ++cur_pos, ++src0_ptr, ++src1_ptr, ++src2_ptr, ++dst_ptr) {
#define FMA_TERNARY_OPTR_TAIL \
    }                         \
    }

#define FMA_TERNARY_OPTR_DEF(optr_type)                                        \
    void fma_element_##optr_type(                                              \
            const TensorND& src0_tensor, const TensorND& src1_tensor,          \
            const TensorND& src2_tensor, const TensorND& dst_tensor) {         \
        size_t tsize = dst_tensor.layout.total_nr_elems();                     \
        float* dst_ptr = dst_tensor.ptr<float>();                              \
        float* src0_ptr = src0_tensor.ptr<float>();                            \
        float* src1_ptr = src1_tensor.ptr<float>();                            \
        float* src2_ptr = src2_tensor.ptr<float>();                            \
        fma_element_##optr_type(tsize, src0_ptr, src1_ptr, src2_ptr, dst_ptr); \
    }

FMA_TERNARY_OPTR_TEMPLATE_HEAD(fma3)
FMA_TERNARY_OPTR_LOOP1
val0 = _mm256_fmadd_ps(val0, val1, val2);
_mm256_storeu_ps(dst_ptr, val0);
FMA_TERNARY_OPTR_LOOP2
*dst_ptr = (*src0_ptr) * (*src1_ptr) + *src2_ptr;
FMA_TERNARY_OPTR_TAIL
FMA_TERNARY_OPTR_DEF(fma3)

// Case 3.2 src0, src1 has the same size, src2 is a scalar
void fma_element_fma3_scalar(
        size_t tsize, float* src0_ptr, float* src1_ptr, float* src2_ptr,
        float* dst_ptr) {
    size_t i = 0;
    __m256 val0, val1, val2;
    val2 = _mm256_broadcast_ss(src2_ptr);

    for (; i + 7 < tsize; i += 8, src0_ptr += 8, src1_ptr += 8, dst_ptr += 8) {
        val0 = _mm256_loadu_ps(src0_ptr);
        val1 = _mm256_loadu_ps(src1_ptr);
        val0 = _mm256_fmadd_ps(val0, val1, val2);
        _mm256_storeu_ps(dst_ptr, val0);
    }

    for (; i < tsize; ++i, ++src0_ptr, ++src1_ptr, ++dst_ptr) {
        *dst_ptr = (*src0_ptr) * (*src1_ptr) + (*src2_ptr);
    }
}
FMA_TERNARY_OPTR_DEF(fma3_scalar)

// Case 3.3 shape of src0 and src2 is (1,C,1,1).
#define FMA_TERNARY_OPTR_DEF_1C11(optr_type)                                  \
    void fma_element_##optr_type(                                             \
            size_t batch_size, size_t channel_size, size_t channel_stride,    \
            const TensorND& src0_tensor, const TensorND& src1_tensor,         \
            const TensorND& src2_tensor, const TensorND& dst_tensor) {        \
        float* dst_ptr = dst_tensor.ptr<float>();                             \
        float* src0_ptr = src0_tensor.ptr<float>();                           \
        float* src1_ptr = src1_tensor.ptr<float>();                           \
        float* src2_ptr = src2_tensor.ptr<float>();                           \
        fma_element_##optr_type(                                              \
                batch_size, channel_size, channel_stride, src0_ptr, src1_ptr, \
                src2_ptr, dst_ptr);                                           \
    }

void fma_element_fma3_1c11(
        size_t batch_size, size_t channel_size, size_t channel_stride, float* src0_ptr,
        float* src1_ptr, float* src2_ptr, float* dst_ptr) {
    size_t cur_pos = 0, chanwise_pos = 0, channel_offset = 0;
    __m256 src0, src1, src2;

    for (size_t batch = 0; batch < batch_size; ++batch) {
        chanwise_pos = 0;
        for (size_t chan = 0; chan < channel_size; ++chan, ++chanwise_pos) {
            src0 = _mm256_broadcast_ss(src0_ptr + chanwise_pos);
            src2 = _mm256_broadcast_ss(src2_ptr + chanwise_pos);
            channel_offset += channel_stride;

            for (; cur_pos + 7 < channel_offset;
                 cur_pos += 8, src1_ptr += 8, dst_ptr += 8) {
                src1 = _mm256_loadu_ps(src1_ptr);
                src1 = _mm256_fmadd_ps(src0, src1, src2);
                _mm256_storeu_ps(dst_ptr, src1);
            }
            float src0_val = src0_ptr[chanwise_pos];
            float src2_val = src2_ptr[chanwise_pos];
            for (; cur_pos < channel_offset; ++cur_pos, ++dst_ptr, ++src1_ptr) {
                *dst_ptr = src0_val * (*src1_ptr) + src2_val;
            }
        }
    }
}

FMA_TERNARY_OPTR_DEF_1C11(fma3_1c11)

}  // namespace detail
}  // namespace x86
}  // namespace megdnn