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

/*******************************************************************************
* Copyright 2021 Intel Corporation
* Copyright 2021 FUJITSU LIMITED
* Copyright 2025 Arm Ltd. and affiliates
*
* 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 CPU_AARCH64_JIT_SVE_512_X8S8S32X_CONVOLUTION_HPP
#define CPU_AARCH64_JIT_SVE_512_X8S8S32X_CONVOLUTION_HPP

#include "common/c_types_map.hpp"
#include "common/dnnl_thread.hpp"
#include "common/memory_tracking.hpp"
#include "common/primitive.hpp"
#include "common/utils.hpp"

#include "cpu/cpu_convolution_pd.hpp"

#include "cpu/aarch64/jit_sve_512_x8s8s32x_conv_kernel.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace aarch64 {

template <impl::data_type_t src_type, impl::data_type_t dst_type>
struct jit_sve_512_x8s8s32x_convolution_fwd_t : public primitive_t {
    struct pd_t : public cpu_convolution_fwd_pd_t {
        using cpu_convolution_fwd_pd_t::cpu_convolution_fwd_pd_t;

        DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_int8:", sve_512, ""),
                jit_sve_512_x8s8s32x_convolution_fwd_t);

        status_t init(engine_t *engine) {
            using smask_t = primitive_attr_t::skip_mask_t;
            bool ok = true && is_fwd()
                    && set_default_alg_kind(alg_kind::convolution_direct)
                    && expect_data_types(src_type, data_type::s8,
                            data_type::undef, dst_type, data_type::s32)
                    && IMPLICATION(with_bias(),
                            utils::one_of(bias_md_.data_type, data_type::f32,
                                    data_type::s32, data_type::s8,
                                    data_type::u8))
                    && attr()->has_default_values(smask_t::post_ops, dst_type)
                    && !has_zero_dim_memory();
            if (!ok) return status::unimplemented;

            status_t status = jit_sve_512_x8s8s32x_fwd_kernel_t::init_conf(jcp_,
                    *desc(), src_md_, weights_md_, dst_md_, bias_md_, *attr(),
                    dnnl_get_max_threads());
            if (status != status::success) return status;

            auto scratchpad = scratchpad_registry().registrar();
            jit_sve_512_x8s8s32x_fwd_kernel_t::init_scratchpad(
                    scratchpad, jcp_, *attr());

            return status;
        }

        jit_conv_conf_t jcp_ = utils::zero<decltype(jcp_)>();
    };

    jit_sve_512_x8s8s32x_convolution_fwd_t(const pd_t *apd)
        : primitive_t(apd) {}

    using src_data_t = typename prec_traits_t<src_type>::type;
    using wei_data_t = typename prec_traits_t<data_type::s8>::type;
    using dst_data_t = typename prec_traits_t<dst_type>::type;

    status_t init(engine_t *engine) override {
        CHECK(safe_ptr_assign(kernel_,
                new jit_sve_512_x8s8s32x_fwd_kernel_t(
                        pd()->jcp_, *pd()->attr())));
        return kernel_->create_kernel();
    }

    status_t execute(const exec_ctx_t &ctx) const override {
        const auto &_pd = pd();
        if (_pd->ndims() == 3)
            return execute_forward_1d(ctx);
        else if (_pd->ndims() == 4)
            if (_pd->jcp_.is_depthwise)
                return execute_forward_2d_dw(ctx);
            else
                return execute_forward_2d(ctx);
        else if (_pd->ndims() == 5)
            return execute_forward_3d(ctx);
        return status::unimplemented;
    }

private:
    status_t execute_forward_1d(const exec_ctx_t &ctx) const;
    status_t execute_forward_2d(const exec_ctx_t &ctx) const;
    status_t execute_forward_2d_dw(const exec_ctx_t &ctx) const;
    status_t execute_forward_3d(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }

    std::unique_ptr<jit_sve_512_x8s8s32x_fwd_kernel_t> kernel_;
};

} // namespace aarch64
} // namespace cpu
} // namespace impl
} // namespace dnnl

#endif

// vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s