onednn-src 0.1.13

/*******************************************************************************
* Copyright 2019 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 CPU_X64_JIT_UNI_DW_CONVOLUTION_HPP
#define CPU_X64_JIT_UNI_DW_CONVOLUTION_HPP

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

#include "cpu/cpu_convolution_pd.hpp"
#include "cpu/x64/cpu_barrier.hpp"
#include "cpu/x64/cpu_reducer.hpp"

#include "cpu/x64/jit_uni_dw_conv_kernel_utils.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace x64 {

template <cpu_isa_t isa, data_type_t src_type, data_type_t dst_type = src_type>
struct jit_uni_dw_convolution_fwd_t : public primitive_t {
    using jit_uni_dw_conv_fwd_kernel_inst_t
            = jit_uni_dw_conv_fwd_kernel_t<isa, src_type>;

    struct pd_t : public cpu_convolution_fwd_pd_t {
        // Note: check `USING_INHERITED_IS_IMPOSSIBLE` comment in other files
        // for details why this ctor can't be removed.
        pd_t(const op_desc_t *adesc, const primitive_attr_t *attr,
                const typename pd_t::base_class *hint_fwd_pd)
            : cpu_convolution_fwd_pd_t(adesc, attr, hint_fwd_pd) {}

        DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_dw:", jcp_.isa, ""),
                jit_uni_dw_convolution_fwd_t);

        status_t init(engine_t *engine) {
            using namespace data_type;
            VDISPATCH_CONV(is_fwd(), VERBOSE_BAD_PROPKIND);
            VDISPATCH_CONV(expect_data_types(src_type, src_type,
                                   data_type::undef, dst_type, f32),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(set_default_alg_kind(alg_kind::convolution_direct),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_CONV(!has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
            VDISPATCH_CONV(
                    attr()->has_default_values(
                            primitive_attr_t::skip_mask_t::post_ops, dst_type),
                    VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_CONV(
                    IMPLICATION(this->with_bias(),
                            utils::one_of(this->desc()->bias_desc.data_type,
                                    f32, bf16, f16)),
                    VERBOSE_UNSUPPORTED_BIAS_CFG);

            // TODO: make `init_conf` assign initialized object to `jcp_`
            CHECK(jit_uni_dw_conv_fwd_kernel_inst_t::init_conf(jcp_, *desc(),
                    src_md_, weights_md_, bias_md_, dst_md_, attr_));

            auto scratchpad = scratchpad_registry().registrar();
            jit_uni_dw_conv_fwd_kernel_inst_t::init_scratchpad(
                    scratchpad, jcp_);

            return status::success;
        }

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

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

    using f32_data_t = typename prec_traits_t<data_type::f32>::type;
    using bf16_data_t = typename prec_traits_t<data_type::bf16>::type;
    using f16_data_t = typename prec_traits_t<data_type::f16>::type;
    using data_t = typename prec_traits_t<src_type>::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_uni_dw_conv_fwd_kernel_inst_t(
                        pd()->jcp_, *pd()->dst_md(0))));
        return kernel_->create_kernel();
    }

    status_t execute(const exec_ctx_t &ctx) const override {
        execute_forward(ctx);
        return status::success;
    }

private:
    void execute_forward(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }

    std::unique_ptr<jit_uni_dw_conv_fwd_kernel_inst_t> kernel_;
};

using jit_avx512_common_dw_convolution_fwd_t
        = jit_uni_dw_convolution_fwd_t<avx512_core, data_type::f32>;
using jit_avx2_dw_convolution_fwd_t
        = jit_uni_dw_convolution_fwd_t<avx2, data_type::f32>;
using jit_sse41_dw_convolution_fwd_t
        = jit_uni_dw_convolution_fwd_t<sse41, data_type::f32>;

template <cpu_isa_t isa, data_type_t diff_dst_type,
        data_type_t diff_src_type = diff_dst_type>
struct jit_uni_dw_convolution_bwd_data_t : public primitive_t {
    struct pd_t : public cpu_convolution_bwd_data_pd_t {
        using cpu_convolution_bwd_data_pd_t::cpu_convolution_bwd_data_pd_t;

        DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_dw:", jcp_.isa, ""),
                jit_uni_dw_convolution_bwd_data_t);

        status_t init(engine_t *engine) {
            using namespace data_type;
            VDISPATCH_CONV(desc()->prop_kind == prop_kind::backward_data,
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_CONV(expect_data_types(diff_src_type, diff_dst_type,
                                   data_type::undef, diff_dst_type, f32),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(set_default_alg_kind(alg_kind::convolution_direct),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_CONV(!has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
            VDISPATCH_CONV(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);

            // TODO: make `init_conf` assign initialized object to `jcp_`
            using jit_uni_dw_conv_bwd_data_kernel_inst
                    = jit_uni_dw_conv_bwd_data_kernel_t<isa, diff_dst_type>;
            CHECK(jit_uni_dw_conv_bwd_data_kernel_inst::init_conf(
                    jcp_, *desc(), diff_src_md_, weights_md_, diff_dst_md_));

            auto scratchpad = scratchpad_registry().registrar();
            jit_uni_dw_conv_bwd_data_kernel_t<isa,
                    diff_dst_type>::init_scratchpad(scratchpad, jcp_);

            return status::success;
        }

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

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

    using diff_src_data_t = typename prec_traits_t<diff_src_type>::type;
    using diff_dst_data_t = typename prec_traits_t<diff_dst_type>::type;
    using wei_data_t = typename prec_traits_t<diff_dst_type>::type;

    status_t init(engine_t *engine) override {
        CHECK(safe_ptr_assign(kernel_,
                new jit_uni_dw_conv_bwd_data_kernel_t<isa, diff_dst_type>(
                        pd()->jcp_)));
        return kernel_->create_kernel();
    }

    status_t execute(const exec_ctx_t &ctx) const override {
        execute_backward_data(ctx);
        return status::success;
    }

private:
    void execute_backward_data(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }

    std::unique_ptr<jit_uni_dw_conv_bwd_data_kernel_t<isa, diff_dst_type>>
            kernel_;
};

using jit_avx512_common_dw_convolution_bwd_data_t
        = jit_uni_dw_convolution_bwd_data_t<avx512_core, data_type::f32>;
using jit_avx2_dw_convolution_bwd_data_t
        = jit_uni_dw_convolution_bwd_data_t<avx2, data_type::f32>;
using jit_sse41_dw_convolution_bwd_data_t
        = jit_uni_dw_convolution_bwd_data_t<sse41, data_type::f32>;

template <cpu_isa_t isa, data_type_t src_type,
        data_type_t diff_weights_type = src_type>
struct jit_uni_dw_convolution_bwd_weights_t : public primitive_t {
    struct pd_t : public cpu_convolution_bwd_weights_pd_t {
        using cpu_convolution_bwd_weights_pd_t::
                cpu_convolution_bwd_weights_pd_t;

        using jit_uni_dw_convolution_bwd_weights
                = jit_uni_dw_convolution_bwd_weights_t<isa, src_type,
                        diff_weights_type>;
        DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_dw:", jcp_.isa, ""),
                jit_uni_dw_convolution_bwd_weights);

        status_t init(engine_t *engine) {
            using namespace data_type;
            VDISPATCH_CONV(desc()->prop_kind == prop_kind::backward_weights,
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_CONV(expect_data_types(src_type, diff_weights_type,
                                   data_type::undef, src_type, f32),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(set_default_alg_kind(alg_kind::convolution_direct),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_CONV(!has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
            VDISPATCH_CONV(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_CONV(
                    IMPLICATION(this->with_bias(),
                            utils::one_of(
                                    this->desc()->diff_bias_desc.data_type, f32,
                                    bf16)),
                    VERBOSE_UNSUPPORTED_BIAS_CFG);

            const int max_threads
                    = dnnl_in_parallel() ? 1 : dnnl_get_max_threads();

            // TODO: make `init_conf` assign initialized object to `jcp_`
            using jit_uni_dw_conv_bwd_weights_kernel_inst
                    = jit_uni_dw_conv_bwd_weights_kernel_t<isa, src_type>;
            CHECK(jit_uni_dw_conv_bwd_weights_kernel_inst::init_conf(jcp_,
                    *desc(), src_md_, diff_weights_md_, diff_bias_md_,
                    diff_dst_md_, max_threads));

            auto scratchpad = scratchpad_registry().registrar();
            jit_uni_dw_conv_bwd_weights_kernel_t<isa,
                    src_type>::init_scratchpad(scratchpad, jcp_);

            return status::success;
        }

        jit_conv_conf_t jcp_ = utils::zero<decltype(jcp_)>();
    };
    jit_uni_dw_convolution_bwd_weights_t(const pd_t *apd);

    using f32_data_t = typename prec_traits_t<data_type::f32>::type;
    using bf16_data_t = typename prec_traits_t<data_type::bf16>::type;
    using src_data_t = typename prec_traits_t<src_type>::type;
    using diff_dst_data_t = typename prec_traits_t<src_type>::type;
    using diff_weights_data_t = typename prec_traits_t<diff_weights_type>::type;

    status_t init(engine_t *engine) override {
        CHECK(safe_ptr_assign(kernel_,
                new jit_uni_dw_conv_bwd_weights_kernel_t<isa, src_type>(
                        pd()->jcp_)));
        CHECK(kernel_->create_kernel());

        const auto jcp = &pd()->jcp_;
        const int reduction = jcp->nthr_mb * jcp->nthr_oh;
        if (reduction > 1 && isa != sse41) {
            CHECK(safe_ptr_assign(
                    acc_ker_, new cpu_accumulator_1d_t<data_type::f32>()));
            CHECK(acc_ker_->create_kernel());
        }
        return status::success;
    }

    status_t execute(const exec_ctx_t &ctx) const override;

private:
    void execute_backward_weights(const exec_ctx_t &ctx) const;
    void execute_reduction(const exec_ctx_t &ctx) const;
    void execute_backward_weights_nxc(const exec_ctx_t &ctx) const;
    void execute_reduction_nxc(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }

    std::unique_ptr<cpu_accumulator_1d_t<data_type::f32>> acc_ker_;
    std::unique_ptr<jit_uni_dw_conv_bwd_weights_kernel_t<isa, src_type>>
            kernel_;
};

using jit_avx512_common_dw_convolution_bwd_weights_t
        = jit_uni_dw_convolution_bwd_weights_t<avx512_core, data_type::f32>;
using jit_avx2_dw_convolution_bwd_weights_t
        = jit_uni_dw_convolution_bwd_weights_t<avx2, data_type::f32>;
using jit_sse41_dw_convolution_bwd_weights_t
        = jit_uni_dw_convolution_bwd_weights_t<sse41, data_type::f32>;

} // namespace x64
} // namespace cpu
} // namespace impl
} // namespace dnnl

#endif