onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
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

/*******************************************************************************
* Copyright 2023 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#ifndef GPU_INTEL_CONV_JIT_V2_BRIDGE_HPP
#define GPU_INTEL_CONV_JIT_V2_BRIDGE_HPP

#include "gpu/intel/conv/config.hpp"
#include "gpu/intel/conv/jit/v2/plan.hpp"
#include "gpu/intel/conv/jit/v2/problem.hpp"
#include "gpu/intel/conv/jit/v2/tensor_utils.hpp"
#include "gpu/intel/jit/ir/tensor.hpp"
#include "gpu/intel/jit/ir/tensor_config.hpp"
#include "gpu/intel/jit/ir/v2/tensor.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace conv {
namespace jit {
namespace v2 {

using namespace intel::jit::v2;

inline tile_t to_shape(const pd_t *pd) {
    tile_t shape;
    shape[pvars::mb] = pd->MB();
    shape[pvars::ic] = ir_utils::safe_div(pd->IC(), pd->G());
    shape[pvars::oc] = ir_utils::safe_div(pd->OC(), pd->G());
    shape[pvars::g] = pd->G();
    shape[pvars::id] = pd->ID();
    shape[pvars::ih] = pd->IH();
    shape[pvars::iw] = pd->IW();
    shape[pvars::od] = pd->OD();
    shape[pvars::oh] = pd->OH();
    shape[pvars::ow] = pd->OW();
    shape[pvars::kd] = pd->KD();
    shape[pvars::kh] = pd->KH();
    shape[pvars::kw] = pd->KW();
    shape[pvars::sd] = pd->KSD();
    shape[pvars::sh] = pd->KSH();
    shape[pvars::sw] = pd->KSW();
    shape[pvars::dd] = pd->KDD();
    shape[pvars::dh] = pd->KDH();
    shape[pvars::dw] = pd->KDW();
    shape[pvars::pd] = pd->padFront();
    shape[pvars::ph] = pd->padT();
    shape[pvars::pw] = pd->padL();
    memory_desc_wrapper mdw_src(pd->invariant_src_md());
    memory_desc_wrapper mdw_wei(pd->invariant_wei_md());
    memory_desc_wrapper mdw_dst(pd->invariant_dst_md());

    bool src_strided = (mdw_src.is_plain() && !mdw_src.is_dense());
    bool wei_strided = (mdw_wei.is_plain() && !mdw_wei.is_dense());
    bool dst_strided = (mdw_dst.is_plain() && !mdw_dst.is_dense());
    auto get_stride = [&](const memory_desc_t *md, int index,
                              int non_spatial_ndims = 2) -> dim_t {
        if (pd->ndims() > index) {
            index = index >= non_spatial_ndims
                    ? pd->ndims() - (index + 1 - non_spatial_ndims)
                    : index;
            if (md->dims[index] == 1) return 0;
            return md->format_desc.blocking.strides[index];
        }
        return 0;
    };
    if (src_strided) {
        shape[prb_stride(pvars::mb, tensor_kind_t::src)]
                = get_stride(pd->invariant_src_md(), 0);
        shape[prb_stride(pvars::ic, tensor_kind_t::src)]
                = get_stride(pd->invariant_src_md(), 1);
        shape[prb_stride(pvars::g, tensor_kind_t::src)]
                = shape[prb_stride(pvars::ic, tensor_kind_t::src)]
                * shape[pvars::ic];
        shape[prb_stride(pvars::id, tensor_kind_t::src)]
                = get_stride(pd->invariant_src_md(), 4);
        shape[prb_stride(pvars::ih, tensor_kind_t::src)]
                = get_stride(pd->invariant_src_md(), 3);
        shape[prb_stride(pvars::iw, tensor_kind_t::src)]
                = get_stride(pd->invariant_src_md(), 2);
    }
    if (wei_strided) {
        shape[prb_stride(pvars::g, tensor_kind_t::wei)] = pd->with_groups()
                ? get_stride(pd->invariant_wei_md(), 0, 3)
                : 0;
        shape[prb_stride(pvars::oc, tensor_kind_t::wei)]
                = get_stride(pd->invariant_wei_md(), 1, 3);
        shape[prb_stride(pvars::ic, tensor_kind_t::wei)]
                = get_stride(pd->invariant_wei_md(), 2, 3);
        shape[prb_stride(pvars::kd, tensor_kind_t::wei)]
                = get_stride(pd->invariant_wei_md(), 5, 3);
        shape[prb_stride(pvars::kh, tensor_kind_t::wei)]
                = get_stride(pd->invariant_wei_md(), 4, 3);
        shape[prb_stride(pvars::kw, tensor_kind_t::wei)]
                = get_stride(pd->invariant_wei_md(), 3, 3);
    }
    if (dst_strided) {
        shape[prb_stride(pvars::mb, tensor_kind_t::dst)]
                = get_stride(pd->invariant_dst_md(), 0);
        shape[prb_stride(pvars::oc, tensor_kind_t::dst)]
                = get_stride(pd->invariant_dst_md(), 1);
        shape[prb_stride(pvars::g, tensor_kind_t::dst)]
                = shape[prb_stride(pvars::oc, tensor_kind_t::dst)]
                * shape[pvars::oc];
        shape[prb_stride(pvars::od, tensor_kind_t::dst)]
                = get_stride(pd->invariant_dst_md(), 4);
        shape[prb_stride(pvars::oh, tensor_kind_t::dst)]
                = get_stride(pd->invariant_dst_md(), 3);
        shape[prb_stride(pvars::ow, tensor_kind_t::dst)]
                = get_stride(pd->invariant_dst_md(), 2);
    }
    return shape;
}

inline problem_t to_problem(const pd_t *pd, const impl::engine_t *engine) {
    auto prop = pd->desc()->prop_kind;
    auto src = make_layout_tag(
            tensor_kind_t::src, pd->ndims(), *pd->invariant_src_md());
    auto wei = make_layout_tag(
            tensor_kind_t::wei, pd->ndims(), *pd->invariant_wei_md());
    auto dst = make_layout_tag(
            tensor_kind_t::dst, pd->ndims(), *pd->invariant_dst_md());
    auto shape = to_shape(pd);

    problem_t prb;
    prb.set_hw(make_ir_hw(engine));
    prb.set_prop(prop);
    prb.set_with_groups(pd->with_groups());
    prb.set_bias_type(to_ir(pd->invariant_bia_md()->data_type));
    prb.set_src_tag(src);
    prb.set_wei_tag(wei);
    prb.set_dst_tag(dst);
    prb.set_shape(shape);
    if (pd->attr()->post_ops_.len() > 0) prb.set_with_post_ops(true);
    prb.set_deterministic(pd->attr()->deterministic_);
    prb.normalize();

    return prb;
}

inline jit::layout_t to_layout(const layout_tag_t &_tag,
        const memory_desc_t &md, bool remove_a_dim = false) {
    auto tag = _tag.raw_tag();
    dim_idx_t non_spatial_ndims = tag.ndims() - 3;
    if (remove_a_dim) {
        tag.remove_dim('a');
        non_spatial_ndims--;
    }
    while (tag.ndims() > into<dim_idx_t>(md.ndims)) {
        tag.remove_dim(dim_idx::as_tag(non_spatial_ndims));
    }
    return jit::make_layout(md, tag.str());
}

inline jit::layout_t to_layout(const layout_tag_t &_tag, const tile_t &shape) {
    int ndims = _tag.desc().ndims();
    const auto &tag = _tag.raw_tag();
    std::vector<dim_t> dims(ndims);
    for (int i = 0; i < ndims; i++) {
        auto d = _tag.desc().prb_dim(i);
        dims[i] = shape.at(d);
    }
    return jit::make_layout(_tag.type(), expr_t(0), tag.str(), dims);
}

inline jit::grid_info_t to_grid_info(const grid_t &grid, const tile_t &tile) {
    std::vector<dim_t> dims;
    std::vector<expr_t> idxs;
    for (int i = 0; i < grid_t::N; i++) {
        dim_t size = grid.size(i, tile);
        dims.push_back(size);
        idxs.push_back(size == 1 ? 0 : grid.index_var(i));
    }
    return jit::grid_info_t(dims, idxs);
}

} // namespace v2
} // namespace jit
} // namespace conv
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl

#endif