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
202
203
204

/**
 * \file dnn/src/naive/local/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/naive/local/opr_impl.h"

#include <cstring>
#include "src/common/utils.h"
#include "src/naive/handle.h"

using namespace megdnn;
using namespace naive;

LocalForwardImpl::float_noncontig_batch_kern LocalForwardImpl::
        dispatch_float_noncontig_batch(
                const TensorLayout& src, const TensorLayout& /*filter*/,
                const TensorLayout& /*dst*/) {
    if (src.dtype == dtype::Float32()) {
        if (param().mode == Mode::CROSS_CORRELATION) {
            return &naive_kern<true, float>;
        } else {
            return &naive_kern<false, float>;
        }
    } else if (DNN_FLOAT16_SELECT(src.dtype == dtype::Float16(), false)) {
        DNN_INC_FLOAT16(
                megdnn_assert(src.dtype == dtype::Float16());
                if (param().mode == Mode::CROSS_CORRELATION) {
                    return &naive_kern<true MEGDNN_COMMA dt_float16>;
                } else { return &naive_kern<false MEGDNN_COMMA dt_float16>; });
    } else {
        megdnn_assert_internal(false);
        return nullptr;
    }
}

template <bool is_xcorr, typename dtype>
void LocalForwardImpl::naive_kern(const FloatNoncontigBatchKernParam& param) {
    UNPACK_LOCAL_FLOAT_NONCONTIG_BATCH_KERN_PARAM(param, dtype);

    static_cast<void>(workspace);
    rep(n, N) rep(oc, OC) rep(oh, OH) rep(ow, OW) {
        auto& dval = dst[n * OUT_BS + oc * OH * OW + oh * OW + ow];
        dval = 0.0f;
        rep(fh, FH) rep(fw, FW) {
            size_t ih = SH * oh;
            size_t iw = SW * ow;
            if (is_xcorr) {
                ih += fh;
                iw += fw;
            } else {
                ih += FH - fh - 1;
                iw += FW - fw - 1;
            }
            ih -= PH;
            iw -= PW;
            if (ih < static_cast<size_t>(IH) && iw < static_cast<size_t>(IW)) {
                rep(ic, IC) {
                    auto sval = src[n * INP_BS + ic * IH * IW + ih * IW + iw];
                    auto fval =
                            filter[oh * OW * IC * FH * FW * OC +
                                   ow * IC * FH * FW * OC + ic * FH * FW * OC +
                                   fh * FW * OC + fw * OC + oc];
                    dval += sval * fval;
                }
            }
        }
    }
}

void LocalForwardImpl::exec(
        _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_out dst,
        _megdnn_workspace workspace) {
    exec_use_float_noncontig_batch(src, filter, dst, workspace);
}

LocalForwardImpl::FloatNoncontigBatchKernParam LocalForwardImpl::make_float_kern_param(
        _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_out dst,
        _megdnn_workspace workspace) const {
    return {src.get_ref_ptr(), filter.get_ref_ptr(), dst.get_ref_ptr(),
            // n
            src.layout.shape[0],
            // ic, ih, iw, oc, oh, ow, fh, fw
            src.layout.shape[1], src.layout.shape[2], src.layout.shape[3],
            dst.layout.shape[1], dst.layout.shape[2], dst.layout.shape[3],
            filter.layout.shape[3], filter.layout.shape[4],
            // ph, pw, sh, sw
            param().pad_h, param().pad_w, param().stride_h, param().stride_w,
            // inp_bs, out_bs
            src.layout.stride[0], dst.layout.stride[0], workspace.raw_ptr};
}

void LocalForwardImpl::exec_use_float_noncontig_batch(
        _megdnn_tensor_in src, _megdnn_tensor_in filter, _megdnn_tensor_out dst,
        _megdnn_workspace workspace) {
    check_exec(src.layout, filter.layout, dst.layout, workspace.size);
    auto fp = make_float_kern_param(src, filter, dst, workspace);
    auto kptr = dispatch_float_noncontig_batch(src.layout, filter.layout, dst.layout);
    auto kern = [fp, kptr]() { kptr(fp); };
    static_cast<naive::HandleImpl*>(handle())->dispatch_kern(kern);
}

void LocalBackwardDataImpl::exec(
        _megdnn_tensor_in filter, _megdnn_tensor_in diff, _megdnn_tensor_out grad,
        _megdnn_workspace workspace) {
    check_exec(filter.layout, diff.layout, grad.layout, workspace.size);
    size_t N = grad.layout.shape[0], IC = grad.layout.shape[1],
           IH = grad.layout.shape[2], IW = grad.layout.shape[3];
    size_t FH = filter.layout.shape[3], FW = filter.layout.shape[4];
    size_t OC = diff.layout.shape[1], OH = diff.layout.shape[2],
           OW = diff.layout.shape[3];
    size_t ph = param().pad_h, pw = param().pad_w;
    size_t sh = param().stride_h, sw = param().stride_w;
    auto mode = param().mode;
    auto kern = [=]() {
        auto gptr = grad.ptr<dt_float32>(), fptr = filter.ptr<dt_float32>(),
             hptr = diff.ptr<dt_float32>();
        memset(gptr, 0, sizeof(float_t) * N * IC * IH * IW);
        rep(n, N) rep(oc, OC) rep(oh, OH) rep(ow, OW) {
            // auto &hval = hptr[n*OC*OH*OW + oc*OH*OW + oh*OW + ow];
            auto& hval = hptr[((n * OC + oc) * OH + oh) * OW + ow];
            rep(ic, IC) rep(fh, FH) rep(fw, FW) {
                size_t ih = -ph + sh * oh;
                size_t iw = -pw + sw * ow;
                if (mode == Mode::CROSS_CORRELATION) {
                    ih += fh;
                    iw += fw;
                } else {
                    ih += FH - fh - 1;
                    iw += FW - fw - 1;
                }

                if (ih < IH && iw < IW) {
                    // auto &gval = gptr[n*IC*IH*IW + ic*IH*IW + ih*IW + iw];
                    // auto fval = fptr[oh*OW*IC*FH*FW*OC + ow*IC*FH*FW*OC +
                    //    ic*FH*FW*OC + fh*FW*OC + fw*OC + oc];

                    auto& gval = gptr[(((n * IC + ic) * IH) + ih) * IW + iw];
                    auto fval =
                            fptr[((((oh * OW + ow) * IC + ic) * FH + fh) * FW + fw) *
                                         OC +
                                 oc];
                    gval += fval * hval;
                }
            }
        }
    };
    MEGDNN_DISPATCH_CPU_KERN_OPR(kern());
}

void LocalBackwardFilterImpl::exec(
        _megdnn_tensor_in src, _megdnn_tensor_in diff, _megdnn_tensor_out grad,
        _megdnn_workspace workspace) {
    check_exec(src.layout, diff.layout, grad.layout, workspace.size);
    size_t N = src.layout.shape[0], IC = src.layout.shape[1], IH = src.layout.shape[2],
           IW = src.layout.shape[3];
    size_t FH = grad.layout.shape[3], FW = grad.layout.shape[4];
    size_t OC = diff.layout.shape[1], OH = diff.layout.shape[2],
           OW = diff.layout.shape[3];
    size_t ph = param().pad_h, pw = param().pad_w;
    size_t sh = param().stride_h, sw = param().stride_w;
    auto mode = param().mode;
    auto kern = [=]() {
        auto gptr = grad.ptr<dt_float32>(), sptr = src.ptr<dt_float32>(),
             hptr = diff.ptr<dt_float32>();
        memset(gptr, 0, sizeof(float_t) * OH * OW * IC * FH * FW * OC);
        rep(n, N) rep(oc, OC) rep(oh, OH) rep(ow, OW) {
            // auto &hval = hptr[n*OC*OH*OW + oc*OH*OW + oh*OW + ow];
            auto& hval = hptr[((n * OC + oc) * OH + oh) * OW + ow];
            rep(ic, IC) rep(fh, FH) rep(fw, FW) {
                size_t ih = -ph + sh * oh;
                size_t iw = -pw + sw * ow;
                if (mode == Mode::CROSS_CORRELATION) {
                    ih += fh;
                    iw += fw;
                } else {
                    ih += FH - fh - 1;
                    iw += FW - fw - 1;
                }

                if (ih < IH && iw < IW) {
                    // auto sval = sptr[n*IC*IH*IW + ic*IH*IW + ih*IW + iw];
                    // auto &gval = gptr[oh*OW*IC*FH*FW*OC + ow*IC*FH*FW*OC +
                    //    ic*FH*FW*OC + fh*FW*OC + fw*OC + oc];

                    auto sval = sptr[((n * IC + ic) * IH + ih) * IW + iw];
                    auto& gval =
                            gptr[((((oh * OW + ow) * IC + ic) * FH + fh) * FW + fw) *
                                         OC +
                                 oc];
                    gval += sval * hval;
                }
            }
        }
    };
    MEGDNN_DISPATCH_CPU_KERN_OPR(kern());
}

// vim: syntax=cpp.doxygen