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

/**
 * \file dnn/src/cuda/fake_quant/opr_impl.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 "./opr_impl.h"
#include "./kern.cuh"
#include "src/common/utils.h"
namespace megdnn {
namespace cuda {

void FakeQuantForwardImpl::exec(
        _megdnn_tensor_in input, _megdnn_tensor_in scale, _megdnn_tensor_in zero_point,
        _megdnn_tensor_out output, _megdnn_workspace workspace) {
    check_exec(
            input.layout, scale.layout, zero_point.layout, output.layout,
            workspace.size);

    if (!input.layout.is_contiguous() || !output.layout.is_contiguous()) {
        return exec_noncontig(input, scale, zero_point, output);
    }
    ElemwiseOpParamN<2> ele_param;
    ele_param[0] = scale;
    ele_param[0].layout = ele_param[0].layout.broadcast(input.layout);
    ele_param[1] = zero_point;
    ele_param[1].layout = ele_param[1].layout.broadcast(input.layout);
    ele_param.init_from_given_tensor();
    auto stream = cuda_stream(handle());

#define cb(DType)                                             \
    if (input.layout.dtype == DType()) {                      \
        using T = typename DTypeTrait<DType>::ctype;          \
        run_elemwise<FakeQuantKernOp<T>, T, 2>(               \
                ele_param, stream, {input, output, m_param}); \
        return;                                               \
    }
    cb(megdnn::dtype::Float32)
#undef cb
}

void FakeQuantForwardImpl::exec_noncontig(
        _megdnn_tensor_in input, _megdnn_tensor_in scale, _megdnn_tensor_in zero_point,
        _megdnn_tensor_out output) {
    ElemwiseOpParamN<4> ele_param;
    ele_param[0] = output;
    ele_param[1] = input;
    ele_param[2] = scale;
    ele_param[2].layout = ele_param[2].layout.broadcast(input.layout);
    ele_param[3] = zero_point;
    ele_param[3].layout = ele_param[3].layout.broadcast(input.layout);
    ele_param.init_from_given_tensor();
    auto stream = cuda_stream(handle());

#define cb(DType)                                                                      \
    if (input.layout.dtype == DType()) {                                               \
        using T = typename DTypeTrait<DType>::ctype;                                   \
        run_elemwise<FakeQuantKernOpNonContig<T>, T, 4>(ele_param, stream, {m_param}); \
        return;                                                                        \
    }
    cb(megdnn::dtype::Float32)
#undef cb
}

void FakeQuantBackwardImpl::exec(
        _megdnn_tensor_in diff, _megdnn_tensor_in input, _megdnn_tensor_in scale,
        _megdnn_tensor_in zero_point, _megdnn_tensor_out grad,
        _megdnn_workspace workspace) {
    check_exec(
            diff.layout, input.layout, scale.layout, zero_point.layout, grad.layout,
            workspace.size);

    if (!input.layout.is_contiguous() || !diff.layout.is_contiguous() ||
        !grad.layout.is_contiguous()) {
        return exec_noncontig(diff, input, scale, zero_point, grad);
    }
    ElemwiseOpParamN<2> ele_param;
    ele_param[0] = scale;
    ele_param[0].layout = ele_param[0].layout.broadcast(input.layout);
    ele_param[1] = zero_point;
    ele_param[1].layout = ele_param[1].layout.broadcast(input.layout);
    ele_param.init_from_given_tensor();
    auto m_param = param();
    auto stream = cuda_stream(handle());
#define cb(DType)                                                 \
    if (grad.layout.dtype == DType()) {                           \
        using T = typename DTypeTrait<DType>::ctype;              \
        run_elemwise<FakeQuantBwdKernOp<T>, T, 2>(                \
                ele_param, stream, {diff, input, grad, m_param}); \
        return;                                                   \
    }
    cb(megdnn::dtype::Float32)
#undef cb
}

void FakeQuantBackwardImpl::exec_noncontig(
        _megdnn_tensor_in diff, _megdnn_tensor_in input, _megdnn_tensor_in scale,
        _megdnn_tensor_in zero_point, _megdnn_tensor_out grad) {
    ElemwiseOpParamN<5> ele_param;
    ele_param[0] = grad;
    ele_param[1] = diff;
    ele_param[2] = input;
    ele_param[3] = scale;
    ele_param[3].layout = ele_param[3].layout.broadcast(input.layout);
    ele_param[4] = zero_point;
    ele_param[4].layout = ele_param[4].layout.broadcast(input.layout);
    ele_param.init_from_given_tensor();
    auto m_param = param();
    auto stream = cuda_stream(handle());
#define cb(DType)                                           \
    if (grad.layout.dtype == DType()) {                     \
        using T = typename DTypeTrait<DType>::ctype;        \
        run_elemwise<FakeQuantBwdKernOpNonContig<T>, T, 5>( \
                ele_param, stream, {m_param});              \
        return;                                             \
    }
    cb(megdnn::dtype::Float32)
#undef cb
}

}  // namespace cuda
}  // namespace megdnn