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

/**
 * \file dnn/src/cuda/tensor_remap/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 "src/cuda/tensor_remap/opr_impl.h"

#include "src/cuda/tensor_remap/tensor_remap.cuh"
#include "src/cuda/utils.h"

namespace megdnn {
namespace cuda {

void IndexingRemapForwardImpl::exec(
        _megdnn_tensor_in src, _megdnn_tensor_in map, _megdnn_tensor_out dst,
        _megdnn_workspace workspace) {
    check_exec(src.layout, map.layout, dst.layout, workspace.size);
    constexpr auto NDIM = TensorShape::MAX_NDIM;
    array_wrapper<int, NDIM> sstride;
    array_wrapper<int, NDIM> dstride;
    array_wrapper<uint32_t, NDIM> dshape;
    // Initialize array_wrappers.
    for (size_t i = 0_z; i < src.layout.ndim; ++i) {
        sstride.data[i] = src.layout.stride[i];
    }
    for (size_t i = 0_z; i < dst.layout.ndim; ++i) {
        dstride.data[i] = dst.layout.stride[i];
    }
    for (size_t i = 0_z; i < dst.layout.ndim; ++i) {
        dshape.data[i] = dst.layout.shape[i];
    }
    // Invoke kernel
#define cb(dt)                                                              \
    if (src.layout.dtype.enumv() == DTypeTrait<dt>::enumv) {                \
        using ctype = DTypeTrait<dt>::ctype;                                \
        tensor_remap::forward<ctype>(                                       \
                src.ptr<ctype>(), map.ptr<dt_int32>(), dst.ptr<ctype>(),    \
                src.layout.ndim, dst.layout.ndim, sstride, dstride, dshape, \
                cuda_stream(handle()));                                     \
        return;                                                             \
    }
    cb(dtype::Float32) cb(dtype::Int32)
#undef cb
            megdnn_throw(ssprintf(
                    "cuda indexing remap forward only support "
                    "float32/int32 dtype, got %s",
                    src.layout.dtype.name()));
}

void IndexingRemapBackwardImpl::exec(
        _megdnn_tensor_in diff, _megdnn_tensor_in map, _megdnn_tensor_out grad,
        _megdnn_workspace workspace) {
    check_exec(diff.layout, map.layout, grad.layout, workspace.size);
    constexpr auto NDIM = TensorShape::MAX_NDIM;
    array_wrapper<int, NDIM> sstride;
    array_wrapper<int, NDIM> dstride;
    array_wrapper<uint32_t, NDIM> sshape;
    array_wrapper<uint32_t, NDIM> dshape;
    // Intialize array_wrappers.
    for (size_t i = 0_z; i < grad.layout.ndim; ++i) {
        sstride.data[i] = grad.layout.stride[i];
    }
    for (size_t i = 0_z; i < diff.layout.ndim; ++i) {
        dstride.data[i] = diff.layout.stride[i];
    }
    for (size_t i = 0_z; i < grad.layout.ndim; ++i) {
        sshape.data[i] = grad.layout.shape[i];
    }
    for (size_t i = 0_z; i < diff.layout.ndim; ++i) {
        dshape.data[i] = diff.layout.shape[i];
    }

    // Invoke kernel
#define cb(dt)                                                                        \
    if (diff.layout.dtype.enumv() == DTypeTrait<dt>::enumv) {                         \
        using ctype = DTypeTrait<dt>::ctype;                                          \
        tensor_remap::backward<ctype>(                                                \
                diff.ptr<ctype>(), map.ptr<dt_int32>(), grad.ptr<ctype>(),            \
                grad.layout.ndim, diff.layout.ndim, sstride, dstride, sshape, dshape, \
                param().is_non_overlapping, cuda_stream(handle()));                   \
        return;                                                                       \
    }
    cb(dtype::Float32) cb(dtype::Int32)

            megdnn_throw(ssprintf(
                    "cuda indexing remap forward only support "
                    "float32/int32 dtype, got %s",
                    diff.layout.dtype.name()));
}

}  // namespace cuda
}  // namespace megdnn

// vim: syntax=cpp.doxygen