onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
Documentation
/*******************************************************************************
* Copyright 2021 Intel Corporation
* Copyright 2022 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_UTILS_JIT_IO_HELPER_HPP
#define CPU_AARCH64_UTILS_JIT_IO_HELPER_HPP

#include <map>
#include <memory>
#include <unordered_set>

#include "common/optional.hpp"

#include "cpu/aarch64/cpu_isa_traits.hpp"
#include "cpu/aarch64/jit_generator.hpp"

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

namespace io {

class io_conf_t {
public:
    io_conf_t() = default;
    io_conf_t(const bool nt_stores_enabled);
    io_conf_t(const io_conf_t &other) = default;

    io_conf_t &operator=(const io_conf_t &other) = default;

    bool nt_stores_enabled_ = false;
};

class io_tail_conf_t {
public:
    io_tail_conf_t(const std::size_t simd_w, const std::size_t tail_size,
            const Xbyak_aarch64::PReg &tail_opmask, const int tail_vmm_mask_idx,
            const Xbyak_aarch64::XReg &reg_tmp,
            const Xbyak_aarch64::XReg &reg_tmp1);
    io_tail_conf_t(const io_tail_conf_t &other) = default;

    io_tail_conf_t &operator=(const io_tail_conf_t &other) = default;

    std::size_t simd_w_ = 0;
    std::size_t tail_size_ = 0;
    Xbyak_aarch64::PReg tail_opmask_ = Xbyak_aarch64::PReg(1);
    int tail_vmm_mask_idx_ = 0;
    Xbyak_aarch64::XReg reg_tmp_ = Xbyak_aarch64::XReg(26);
    Xbyak_aarch64::XReg reg_tmp1_ = Xbyak_aarch64::XReg(25);
};

class io_saturation_conf_t {
public:
    io_saturation_conf_t(const int vreg_zero_saturation_idx,
            const int vreg_saturation_ubound_idx,
            const Xbyak_aarch64::XReg &reg_tmp);
    io_saturation_conf_t(const io_saturation_conf_t &other) = default;

    io_saturation_conf_t &operator=(const io_saturation_conf_t &other)
            = default;

    int vreg_zero_saturation_idx_ = 0;
    int vreg_saturation_ubound_idx_ = 0;
    Xbyak_aarch64::XReg reg_tmp_ = Xbyak_aarch64::XReg(26);
};

class io_gather_conf_t {
public:
    io_gather_conf_t(const std::size_t simd_w,
            const Xbyak_aarch64::PReg &full_opmask, const int full_vmm_mask_idx,
            const Xbyak_aarch64::XReg &reg_tmp,
            const Xbyak_aarch64::XReg &reg_tmp1,
            const utils::optional_t<int> &vmm_tmp_idx = utils::nullopt);
    io_gather_conf_t(const io_gather_conf_t &other) = default;

    io_gather_conf_t &operator=(const io_gather_conf_t &other) = default;

    std::size_t simd_w_ = 0;
    Xbyak_aarch64::PReg full_opmask_ = Xbyak_aarch64::PReg(0);
    int full_vmm_mask_idx_ = 0;
    Xbyak_aarch64::XReg reg_tmp_ = Xbyak_aarch64::XReg(26);
    Xbyak_aarch64::XReg reg_tmp1_ = Xbyak_aarch64::XReg(25);
    // It is needed, when io_helper use emulation for gather
    // and it is not needed for sse.
    utils::optional_t<int> vmm_tmp_idx_ = utils::nullopt;
};

template <typename Vmm>
class jit_io_multi_dt_helper_t;

template <typename Vmm>
class jit_io_helper_t {
public:
    friend class jit_io_multi_dt_helper_t<Vmm>;

    jit_io_helper_t(jit_generator_t *host, const cpu_isa_t &isa,
            const data_type_t &data_type, const io_conf_t &io_conf,
            const utils::optional_t<io_tail_conf_t> &tail_conf = utils::nullopt,
            const utils::optional_t<io_saturation_conf_t> &saturation_conf
            = utils::nullopt,
            const utils::optional_t<io_gather_conf_t> &gather_conf
            = utils::nullopt);
    jit_io_helper_t(jit_io_helper_t &&) = default;
    jit_io_helper_t &operator=(jit_io_helper_t &&) = default;

    ~jit_io_helper_t();
    void prepare_tail_mask();
    void prepare_full_mask();
    /*
     * Sometimes the values in the register can be nan at the
     * beginning of the kernel, then using vcmpps(vmm, vmm, vmm)
     * will not set all bits to 1, instead of that this instruction will
     * return zero. At the beginning, it is worth to zeroing
     * full mask vmm to be sure, that vcmpps work properly.
     */
    void init_full_mask();
    void init_saturate_f32() const;
    void gather(const Xbyak_aarch64::XReg &src_reg, const Vmm &indices_vmm,
            const Vmm &dst_vmm, const bool tail);
    void broadcast(const Xbyak_aarch64::XReg &src_addr, const int src_offt,
            const Vmm &dst_vmm);
    void load(const Xbyak_aarch64::XReg &src_addr, const int src_offt,
            const Vmm &dst_vmm, const bool tail);
    void store(const Vmm &src_vmm, const Xbyak_aarch64::XReg &dst_addr,
            const int offt, const bool tail);

private:
    void compute_cmp_mask(const Vmm &cmp_dst, const Vmm &cmp_src,
            const Vmm &cmp_src2, const unsigned int uimm);
    void prepare_opmask(const std::size_t how_many_bits_to_set,
            const Xbyak_aarch64::XReg &reg_tmp0,
            const Xbyak_aarch64::XReg &reg_tmp1,
            const Xbyak_aarch64::PReg &mask);

    void load_f32(const Xbyak_aarch64::XReg &src_addr, const int offt,
            const Vmm &dst_vmm, const bool tail,
            const Xbyak_aarch64::PReg &mask);
    void load_s32(const Xbyak_aarch64::XReg &src_addr, const int offt,
            const Vmm &dst_vmm, const bool tail,
            const Xbyak_aarch64::PReg &mask);
    void load_i8(const Xbyak_aarch64::XReg &src_addr, const int offt,
            const Vmm &dst_vmm, const Xbyak_aarch64::PReg &mask);
    void saturate(const Vmm &vmm);
    void store_f32(const Vmm &src_vmm, const Xbyak_aarch64::XReg &dst_addr,
            const int offt, const bool tail, const Xbyak_aarch64::PReg &mask);
    void store_i8(const Vmm &src_vmm, const Xbyak_aarch64::XReg &dst_addr,
            const int offt, const Xbyak_aarch64::PReg &mask);
    void convert_to_f32(const Vmm &dst_vmm, const Vmm &src_vmm,
            const data_type_t src_data_type);

    void store_i8_sdb(Xbyak_aarch64::XReg addr, const Vmm &src_vmm,
            const Xbyak_aarch64::PReg &mask);
    void store_i8_udb(Xbyak_aarch64::XReg addr, const Vmm &src_vmm,
            const Xbyak_aarch64::PReg &mask);

    jit_generator_t *host_;
    const cpu_isa_t isa_;
    const data_type_t data_type_;
    const io_conf_t io_conf_;
    const utils::optional_t<io_tail_conf_t> tail_conf_;
    const utils::optional_t<io_saturation_conf_t> saturation_conf_;
    const utils::optional_t<io_gather_conf_t> gather_conf_;
};

template <typename Vmm>
class jit_io_multi_dt_helper_t {
public:
    using data_types_t = std::unordered_set<data_type_t, std::hash<int>>;

    jit_io_multi_dt_helper_t(jit_generator_t *host, const cpu_isa_t &isa,
            const data_types_t &data_types, const io_conf_t &io_conf,
            const utils::optional_t<io_tail_conf_t> &tail_conf = utils::nullopt,
            const std::map<data_type_t, io_saturation_conf_t> &saturation_confs
            = std::map<data_type_t, io_saturation_conf_t> {},
            const utils::optional_t<io_gather_conf_t> &gather_conf
            = utils::nullopt);
    ~jit_io_multi_dt_helper_t();
    void prepare_tail_mask();
    void prepare_full_mask();
    void init_saturate_f32(const data_types_t &store_data_types);
    void init_full_mask();

    std::shared_ptr<jit_io_helper_t<Vmm>> at(const data_type_t dt) const;

private:
    std::unordered_map<data_type_t, std::shared_ptr<jit_io_helper_t<Vmm>>,
            std::hash<int>>
            storage_;
};

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

#endif