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

/*******************************************************************************
* Copyright 2025 Intel Corporation
* Copyright 2025 Codeplay Software Limited
*
* 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 "gpu/generic/sycl/ref_group_normalization.hpp"

namespace dnnl::impl::gpu::generic::sycl {
status_t ref_group_normalization_fwd_t::pd_t::init(impl::engine_t *engine) {
    using namespace data_type;
    VDISPATCH_GNORM(set_default_formats_common(), VERBOSE_UNSUPPORTED_TAG);
    VDISPATCH_GNORM(is_fwd(), VERBOSE_BAD_PROPKIND);

    auto src_mdw = memory_desc_wrapper(arg_md(DNNL_ARG_SRC));
    auto dst_mdw = memory_desc_wrapper(arg_md(DNNL_ARG_DST));
    auto src_dt = src_mdw.data_type();
    auto dst_dt = dst_mdw.data_type();

    VDISPATCH_GNORM(utils::one_of(src_dt, f32, bf16, f16, s8, u8),
            VERBOSE_UNSUPPORTED_DT);
    VDISPATCH_GNORM(utils::one_of(dst_dt, f32, bf16, f16, s8, u8),
            VERBOSE_UNSUPPORTED_DT);

    VDISPATCH_GNORM(attr_.set_default_formats(dst_md()) == status::success,
            VERBOSE_UNSUPPORTED_TAG);
    VDISPATCH_GNORM(sycl_post_ops_t::post_ops_ok(attr()),
            "sycl post op initialization returns false");

    const primitive_attr_t::skip_mask_t attr_mask
            = primitive_attr_t::skip_mask_t::scales
            | primitive_attr_t::skip_mask_t::post_ops;
    VDISPATCH_GNORM(
            attr()->has_default_values(attr_mask), VERBOSE_UNSUPPORTED_ATTR);
    VDISPATCH_GNORM(attr_scales_ok(), VERBOSE_UNSUPPORTED_SCALES_CFG);
    VDISPATCH_GNORM(check_scale_shift_data_type(),
            "scale / shift data type must be data_type::f32");
    VDISPATCH_GNORM(
            sycl_post_ops_t::post_ops_ok(attr()), VERBOSE_UNSUPPORTED_POSTOP);
    VDISPATCH_GNORM(impl::is_dense_format_kind({src_md(), dst_md()}),
            VERBOSE_UNSUPPORTED_SPARSE_CFG);

    const auto &dims = src_mdw.dims();
    const auto num_groups = desc()->groups;
    VDISPATCH_GNORM(dims[1] % num_groups == 0,
            "number of groups must divide the channels evenly");

    auto batch_size = static_cast<std::size_t>(dims[0]);
    auto group_range = static_cast<std::size_t>(desc()->groups);

    auto device = utils::downcast<const impl::xpu::sycl::engine_impl_t *>(
            engine->impl())
                          ->device();

    // To avoid using excess registers error
    auto local_range = std::max(std::size_t(64),
            device.get_info<::sycl::info::device::max_work_group_size>() / 8);
    launch_range = ::sycl::nd_range<2>(
            {batch_size * local_range, group_range}, {local_range, 1});

    conf_ = sycl_group_norm_conf_t();
    conf_.src_desc = xpu::sycl::md_t(arg_md(DNNL_ARG_SRC));
    conf_.dst_desc = xpu::sycl::md_t(arg_md(DNNL_ARG_DST));
    conf_.use_global_stats = stats_is_src();
    conf_.num_groups = static_cast<int32_t>(group_range);
    conf_.num_channels_per_group = static_cast<int32_t>(dims[1] / group_range);
    conf_.use_scale = use_scale();
    conf_.use_shift = use_shift();
    conf_.src_scaling = !attr()->scales_.has_default_values(DNNL_ARG_SRC);
    conf_.dst_scaling = !attr()->scales_.has_default_values(DNNL_ARG_DST);
    conf_.eta = desc()->group_norm_epsilon;
    conf_.post_ops = {attr(), dst_mdw};
    return status::success;
}

status_t ref_group_normalization_fwd_t::init(impl::engine_t *engine) {
    auto kid = ::sycl::get_kernel_id<group_norm_fwd_t>();
    CHECK(create_kernel(engine, kid, &kernel_));
    return status::success;
}

status_t ref_group_normalization_fwd_t::execute(const exec_ctx_t &ctx) const {
    exec_args_t cloned_args(ctx.args());
    // Circumventing the const pointers when mean and value
    // use_global_stats is set, to avoid creating 2 kernels just
    // for this
    cloned_args[DNNL_ARG_MEAN]
            = memory_arg_t {cloned_args[DNNL_ARG_MEAN].mem(), false};
    cloned_args[DNNL_ARG_VARIANCE]
            = memory_arg_t {cloned_args[DNNL_ARG_VARIANCE].mem(), false};

    exec_ctx_t exec_ctx(ctx.stream(), std::move(cloned_args));
    auto &conf_ = pd()->conf_;
    auto launch_range = pd()->launch_range;

    parallel_for(exec_ctx, kernel_, [&](::sycl::handler &cgh) {
        ::sycl::local_accessor<float, 1> local_memory(
                launch_range.get_local_range()[0] + 2, cgh);
        cgh.parallel_for(launch_range,
                group_norm_fwd_t(conf_, local_memory, cgh, exec_ctx));
    });
    return status::success;
}

status_t ref_group_normalization_bwd_t::pd_t::init(impl::engine_t *engine) {
    using namespace data_type;
    // TODO: remove me
    VDISPATCH_GNORM(
            false, "The implementation doesn't return the correct result");

    VDISPATCH_GNORM(set_default_formats_common(), VERBOSE_UNSUPPORTED_TAG);
    VDISPATCH_GNORM(!is_fwd(), VERBOSE_BAD_PROPKIND);
    VDISPATCH_GNORM(attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);

    auto src_mdw = memory_desc_wrapper(arg_md(DNNL_ARG_SRC));
    auto diff_src_mdw = memory_desc_wrapper(arg_md(DNNL_ARG_DIFF_SRC));
    auto diff_dst_mdw = memory_desc_wrapper(arg_md(DNNL_ARG_DIFF_DST));
    VDISPATCH_GNORM(utils::one_of(src_mdw.data_type(), f32, bf16, f16),
            VERBOSE_UNSUPPORTED_DT);
    VDISPATCH_GNORM(utils::one_of(diff_src_mdw.data_type(), f32, bf16, f16),
            VERBOSE_UNSUPPORTED_DT);
    VDISPATCH_GNORM(utils::one_of(diff_dst_mdw.data_type(), f32, bf16, f16),
            VERBOSE_UNSUPPORTED_DT);
    VDISPATCH_GNORM(impl::is_dense_format_kind(
                            {src_md(), diff_src_md(), diff_dst_md()}),
            VERBOSE_UNSUPPORTED_SPARSE_CFG);

    auto device = utils::downcast<const impl::xpu::sycl::engine_impl_t *>(
            engine->impl())
                          ->device();

    dim_t num_channels = src_mdw.dims()[1];
    // To avoid using excess registers error
    auto local_range = std::max(std::size_t(64),
            device.get_info<::sycl::info::device::max_work_group_size>() / 8);
    launch_range = ::sycl::nd_range<1>(
            {static_cast<std::size_t>(num_channels) * local_range},
            {local_range});

    conf_ = sycl_gnorm_bwd_conf_t();
    conf_.src_desc = xpu::sycl::md_t(arg_md(DNNL_ARG_SRC));
    conf_.diff_src_desc = xpu::sycl::md_t(arg_md(DNNL_ARG_DIFF_SRC));
    conf_.diff_dst_desc = xpu::sycl::md_t(arg_md(DNNL_ARG_DIFF_DST));
    conf_.num_groups = desc()->groups;
    conf_.num_channels_per_group = src_mdw.dims()[1] / desc()->groups;
    conf_.scale_diff_required = use_scale();
    conf_.bias_diff_required = use_shift();
    conf_.used_global_stats = stats_is_src();
    conf_.eta = desc()->group_norm_epsilon;
    return status::success;
}

status_t ref_group_normalization_bwd_t::init(impl::engine_t *engine) {
    auto kid = ::sycl::get_kernel_id<group_norm_bwd_t>();
    CHECK(create_kernel(engine, kid, &kernel_));
    return status::success;
}

status_t ref_group_normalization_bwd_t::execute(const exec_ctx_t &ctx) const {
    auto &conf_ = pd()->conf_;
    auto launch_range = pd()->launch_range;

    parallel_for(ctx, kernel_, [&](::sycl::handler &cgh) {
        ::sycl::local_accessor<float, 1> local_memory(
                launch_range.get_local_range()[0], cgh);
        cgh.parallel_for(
                launch_range, group_norm_bwd_t(conf_, local_memory, cgh, ctx));
    });
    return status::success;
}

} // namespace dnnl::impl::gpu::generic::sycl