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
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230

/*******************************************************************************
* Copyright 2020 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.
*******************************************************************************/
#include "oneapi/dnnl/dnnl_graph.h"

#include "common/engine.hpp"

#include "graph/interface/allocator.hpp"
#include "graph/interface/tensor.hpp"

#include "graph/utils/utils.hpp"

#if DNNL_CPU_RUNTIME != DNNL_RUNTIME_SYCL
static const size_t DNNL_CPU_MEMALIGNMENT = 64;
#endif

#ifdef DNNL_WITH_SYCL
#include "oneapi/dnnl/dnnl_sycl.hpp"
static const size_t DNNL_SYCL_MEMALIGNMENT = 64;
#endif

#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
#include "xpu/ocl/engine_factory.hpp"
static const size_t DNNL_OCL_MEMALIGNMENT = 0;

#endif

using namespace dnnl::impl::graph;

static void *tensor_malloc(
        size_t size, const engine_t *eng, allocator_t::mem_type_t type) {
    const auto *alc = static_cast<dnnl::impl::graph::allocator_t *>(
            eng->get_allocator());
#ifdef DNNL_WITH_SYCL
    void *dev_ptr {nullptr};
    dnnl_sycl_interop_engine_get_device(const_cast<engine_t *>(eng), &dev_ptr);
    auto *dev = static_cast<sycl::device *>(dev_ptr);
    void *ctx_ptr {nullptr};
    dnnl_sycl_interop_engine_get_context(const_cast<engine_t *>(eng), &ctx_ptr);
    auto *ctx = static_cast<sycl::context *>(ctx_ptr);
#endif
    if (eng->kind() == engine_kind::cpu) {
#if DNNL_CPU_RUNTIME == DNNL_RUNTIME_SYCL
        return alc->allocate(size, *dev, *ctx, {type, DNNL_SYCL_MEMALIGNMENT});
#else
        return alc->allocate(size, {type, DNNL_CPU_MEMALIGNMENT});
#endif
    } else if (eng->kind() == engine_kind::gpu) {
#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_SYCL
        return alc->allocate(size, *dev, *ctx, {type, DNNL_SYCL_MEMALIGNMENT});
#elif DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
        auto *ocl_engine = utils::downcast<
                const dnnl::impl::gpu::intel::ocl::engine_t *>(eng);
        const cl_device_id &ocl_dev = ocl_engine->device();
        const cl_context &ocl_ctx = ocl_engine->context();
        return alc->allocate(
                size, ocl_dev, ocl_ctx, {type, DNNL_OCL_MEMALIGNMENT});
#else
        assertm(false, "Unsupported gpu runtime");
        return nullptr;
#endif
    } else {
        assertm(false, "Unsupported engine kind");
        return nullptr;
    }
}

static void tensor_free(void *p, const engine_t *eng) {
    const auto *alc = static_cast<dnnl::impl::graph::allocator_t *>(
            eng->get_allocator());
#ifdef DNNL_WITH_SYCL
    void *dev_ptr {nullptr};
    dnnl_sycl_interop_engine_get_device(const_cast<engine_t *>(eng), &dev_ptr);
    auto *dev = static_cast<sycl::device *>(dev_ptr);
    void *ctx_ptr {nullptr};
    dnnl_sycl_interop_engine_get_context(const_cast<engine_t *>(eng), &ctx_ptr);
    auto *ctx = static_cast<sycl::context *>(ctx_ptr);
#endif
    if (eng->kind() == engine_kind::cpu) {
#if DNNL_CPU_RUNTIME == DNNL_RUNTIME_SYCL
        alc->deallocate(p, *dev, *ctx, {});
#else
        alc->deallocate(p);
#endif
    } else if (eng->kind() == engine_kind::gpu) {
#if DNNL_GPU_RUNTIME == DNNL_RUNTIME_SYCL
        alc->deallocate(p, *dev, *ctx, {});
#elif DNNL_GPU_RUNTIME == DNNL_RUNTIME_OCL
        auto *ocl_engine = utils::downcast<
                const dnnl::impl::gpu::intel::ocl::engine_t *>(eng);
        const cl_device_id &ocl_dev = ocl_engine->device();
        const cl_context &ocl_ctx = ocl_engine->context();
        return alc->deallocate(p, ocl_dev, ocl_ctx, {});
#else
        assertm(false, "Unsupported gpu runtime");
#endif
    } else {
        assertm(false, "Unsupported engine kind");
    }
}

dnnl_graph_tensor::dnnl_graph_tensor(
        const logical_tensor_t &lt, const engine_t *eng, void *handle)
    : lt_(lt), eng_(eng) {
    if (handle == DNNL_MEMORY_ALLOCATE) {
        size_t num_bytes = logical_tensor_wrapper_t(lt).size();

        void *data
                = tensor_malloc(num_bytes, eng, allocator_t::mem_type_t::temp);
        assertm(data, "Can't allocate memory for a tensor!");
        handle_.reset(data, [eng](void *p) { tensor_free(p, eng); });
    } else if (lt.property == property_type::host_scalar) {
        if (lt.data_type == data_type::s32) {
            scalar_.s32_value = *static_cast<int32_t *>(handle);
            handle_.reset(&scalar_.s32_value, dummy_destructor);
        } else if (lt.data_type == data_type::f32) {
            scalar_.f32_value = *static_cast<float *>(handle);
            handle_.reset(&scalar_.f32_value, dummy_destructor);
        } else if (lt.data_type == data_type::s64) {
            scalar_.s64_value = *static_cast<int64_t *>(handle);
            handle_.reset(&scalar_.s64_value, dummy_destructor);
        } else {
            assertm(false, "Unsupported data type for host scalar");
        }
    } else {
        handle_.reset(handle, dummy_destructor);
    }
}

status_t DNNL_API dnnl_graph_tensor_create(tensor_t **tensor,
        const logical_tensor_t *logical_tensor, engine_t *eng, void *handle) {
    if (utils::any_null(tensor, logical_tensor, eng))
        return status::invalid_arguments;

    const auto ltw = logical_tensor_wrapper_t(logical_tensor);
    if (ltw.is_host_scalar()) return status::invalid_arguments;

    *tensor = new tensor_t {*logical_tensor, eng, handle};
    if (*tensor == nullptr) return status::out_of_memory;
    if (handle == DNNL_MEMORY_ALLOCATE
            && (*tensor)->get_data_handle() == nullptr) {
        delete *tensor;
        *tensor = nullptr;
        return status::out_of_memory;
    }
    return status::success;
}

status_t DNNL_API dnnl_graph_tensor_create_scalar(tensor_t **tensor,
        const logical_tensor_t *logical_tensor, void *handle) {
    if (utils::any_null(tensor, logical_tensor))
        return status::invalid_arguments;

    const auto ltw = logical_tensor_wrapper_t(logical_tensor);
    if (!ltw.is_host_scalar()) return status::invalid_arguments;

    // TODO(xxx): extend for library allocated host scalar?
    if (nullptr == handle || DNNL_MEMORY_ALLOCATE == handle) {
        return status::invalid_arguments;
    }

    *tensor = new tensor_t {*logical_tensor, nullptr, handle};
    if (*tensor == nullptr) return status::out_of_memory;

    return status::success;
}

status_t DNNL_API dnnl_graph_tensor_destroy(tensor_t *tensor) {
    delete tensor;
    return status::success;
}

status_t DNNL_API dnnl_graph_tensor_get_data_handle(
        const tensor_t *tensor, void **handle) {
    if (utils::any_null(tensor, handle)) return status::invalid_arguments;

    const auto ltw = logical_tensor_wrapper_t(tensor->get_logical_tensor());
    if (ltw.is_host_scalar()) {
        *handle = nullptr;
    } else {
        *handle = tensor->get_data_handle();
    }
    return status::success;
}

status_t DNNL_API dnnl_graph_tensor_set_data_handle(
        tensor_t *tensor, void *handle) {
    if (tensor == nullptr) return status::invalid_arguments;

    const auto ltw = logical_tensor_wrapper_t(tensor->get_logical_tensor());
    if (ltw.is_host_scalar()) return status::invalid_arguments;

    auto ret = tensor->set_data_handle(handle);
    return ret;
}

status_t DNNL_API dnnl_graph_tensor_get_engine(
        const tensor_t *tensor, engine_t **engine) {
    if (utils::any_null(tensor, engine)) return status::invalid_arguments;

    const auto ltw = logical_tensor_wrapper_t(tensor->get_logical_tensor());
    if (ltw.is_host_scalar()) {
        *engine = nullptr;
    } else {
        *engine = const_cast<engine_t *>(tensor->get_engine());
    }

    return status::success;
}

dnnl_status_t DNNL_API dnnl_graph_tensor_get_logical_tensor(
        const tensor_t *tensor, logical_tensor_t *logical_tensor) {
    if (utils::any_null(tensor, logical_tensor))
        return status::invalid_arguments;

    *logical_tensor = tensor->get_logical_tensor();
    return status::success;
}