onednn-src 0.1.13

/*******************************************************************************
* Copyright 2017 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_SSE41_1X1_CONVOLUTION_HPP
#define CPU_X64_JIT_SSE41_1X1_CONVOLUTION_HPP

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

#include "cpu/cpu_convolution_pd.hpp"
#include "cpu/dw_convolution_utils.hpp"
#include "cpu/platform.hpp"

#include "cpu/x64/jit_sse41_1x1_conv_kernel_f32.hpp"
#include "cpu/x64/jit_uni_1x1_conv_utils.hpp"
#include "cpu/x64/jit_uni_dw_convolution.hpp"

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

struct jit_sse41_1x1_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;

        pd_t(const pd_t &other) : cpu_convolution_fwd_pd_t(other) {
            if (copy(other) != status::success) is_initialized_ = false;
        }

        DECLARE_COMMON_PD_T(JIT_IMPL_NAME_HELPER("jit_1x1:", sse41, ""),
                jit_sse41_1x1_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(f32, f32, f32, f32, 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, f32),
                    VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_CONV(set_default_formats(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_CONV(
                    attr_.set_default_formats(dst_md(0)) == status::success,
                    VERBOSE_UNSUPPORTED_POSTOP);

            // TODO: make `init_conf` assign initialized object to `jcp_`
            CHECK(jit_sse41_1x1_conv_kernel_f32_t::init_conf(jcp_, *desc(),
                    *src_md(), *weights_md(), *dst_md(), *attr(),
                    dnnl_get_max_threads()));
            if (jcp_.with_dw_conv) CHECK(depthwise_po_init(engine));

            return status::success;
        }

        const memory_desc_t *dst_1x1_md(int index = 0) const {
            return cpu_convolution_fwd_pd_t::dst_md(index);
        }

        // NOLINTBEGIN(google-default-arguments)
        const memory_desc_t *dst_md(
                int index = 0, bool user_input = false) const override {
            return dw_conv_pd_ && jcp_.with_dw_conv
                    ? dw_conv_pd_->dst_md(index, user_input)
                    : cpu_convolution_fwd_pd_t::dst_md(index, user_input);
        }

        const memory_desc_t *arg_md(
                int arg, bool user_input = false) const override {
            if (dw_conv_pd_ && jcp_.with_dw_conv) {
                switch (arg) {
                    case DNNL_ARG_ATTR_POST_OP_DW | DNNL_ARG_SRC:
                        return cpu_convolution_fwd_pd_t::dst_md(0, user_input);
                    case DNNL_ARG_ATTR_POST_OP_DW | DNNL_ARG_WEIGHTS:
                        return dw_conv_pd_->weights_md(0);
                    case DNNL_ARG_ATTR_POST_OP_DW | DNNL_ARG_BIAS:
                        return dw_conv_pd_->weights_md(1);
                    default: break;
                }
            }
            return convolution_fwd_pd_t::arg_md(arg, user_input);
        }
        // NOLINTEND(google-default-arguments)

        arg_usage_t arg_usage(int arg) const override {
            if (arg == (DNNL_ARG_ATTR_POST_OP_DW | DNNL_ARG_WEIGHTS))
                return arg_usage_t::input;

            if (arg == (DNNL_ARG_ATTR_POST_OP_DW | DNNL_ARG_BIAS))
                return attr_post_op_dw_inputs() > 1 ? arg_usage_t::input
                                                    : arg_usage_t::unused;

            return convolution_fwd_pd_t::arg_usage(arg);
        }

        jit_1x1_conv_conf_t jcp_ = utils::zero<decltype(jcp_)>();
        using dw_pd_t = jit_sse41_dw_convolution_fwd_t::pd_t;
        std::unique_ptr<dw_pd_t> dw_conv_pd_;

    protected:
        bool set_default_formats() {
            using namespace format_tag;

            const memory_desc_wrapper src_d(&src_md_);
            const memory_desc_wrapper dst_d(&dst_md_);

            const auto dat_tag_nxc = utils::pick(ndims() - 3, nwc, nhwc, ndhwc);
            const auto dat_tag_nCx8c
                    = utils::pick(ndims() - 3, nCw8c, nChw8c, nCdhw8c);
            const auto curr_src_tag
                    = src_d.matches_one_of_tag(dat_tag_nxc, dat_tag_nCx8c);
            const auto curr_dst_tag
                    = dst_d.matches_one_of_tag(dat_tag_nxc, dat_tag_nCx8c);
            const auto is_data_layout_nxc
                    = IMPLICATION(curr_src_tag != dat_tag_nxc,
                              src_d.format_kind() == format_kind::any)
                    && IMPLICATION(curr_dst_tag != dat_tag_nxc,
                            dst_d.format_kind() == format_kind::any)
                    && utils::one_of(dat_tag_nxc, curr_src_tag, curr_dst_tag);
            auto dat_tag = is_data_layout_nxc ? dat_tag_nxc : dat_tag_nCx8c;
            auto wei_tag = with_groups()
                    ? utils::pick(ndims() - 3, gOIw8i8o, gOIhw8i8o)
                    : utils::pick(ndims() - 3, OIw8i8o, OIhw8i8o);

            return set_default_formats_common(dat_tag, wei_tag, dat_tag);
        }

        status_t copy(const pd_t &other) {
            jcp_ = other.jcp_;
            if (other.dw_conv_pd_) {
                dw_conv_pd_.reset(other.dw_conv_pd_->clone());
                if (!dw_conv_pd_) return status::out_of_memory;
            }
            return status::success;
        }

        status_t depthwise_po_init(engine_t *engine) {

            using namespace memory_tracking;
            auto &jcp_1x1 = jcp_;
            primitive_attr_t attr_1x1(*attr());
            if (!attr_1x1.is_initialized()) return status::out_of_memory;
            const auto &src_md = dst_md_;
            const memory_desc_wrapper src_d(src_md);
            const auto nthr = dnnl_get_max_threads();
            auto l2_cache = platform::get_per_core_cache_size(2) * nthr;

            // Note: A robust fusion implementation would be to check if both
            // 1x1 conv and dw conv that are considered here for fusion are
            // optimal independently. This would require creating a new
            // primitive_desc through primitive_iterator & check if they match.
            // Due to concern that these creations and/or checks could be heavy,
            // for 1x1: Check that no better ISA is available.
            // for dw: Always fuse with same ISA.
            // Caveat: May be a better dw conv exists.
            VDISPATCH_CONV_IC(!mayiuse(avx), VERBOSE_1x1CONV_HEURISTIC_FAIL,
                    "higher isa is supported");
            VDISPATCH_CONV_IC(
                    attr_1x1.post_ops_.find(primitive_kind::sum) == -1,
                    VERBOSE_UNSUPPORTED_FEATURE, "unsupported sum post-op");
            // TODO: Below may be further tuned.
            VDISPATCH_CONV_IC(l2_cache * 2 < src_d.size(),
                    VERBOSE_1x1CONV_HEURISTIC_FAIL, "cache check size failed");
            // load_grp_count check can be redundant due to l2 check
            // above. Adding it explicitly as the current driver doesn't
            // work if this condition fails.
            VDISPATCH_CONV_IC(jcp_1x1.load_grp_count < 2,
                    VERBOSE_1x1CONV_HEURISTIC_FAIL, "load grp count > 1");

            int dw_po_index
                    = attr_1x1.post_ops_.find(primitive_kind::convolution);

            convolution_desc_t cd_dw;
            primitive_attr_t attr_dw;

            CHECK(get_depthwise_conv_desc(
                    cd_dw, src_md, attr_1x1, attr_dw, dw_po_index));

            // The code below doesn't work because currently it requires `jcp_`
            // member which is not available from the common interface. In turn,
            // this means the common pd creation interface through an iterator
            // can't be used and a specific convolution implementation's pd is
            // required here. It restricts the usage of inherited
            // `convolution_pd_t` constructor.
            // ANCHOR: USING_INHERITED_IS_IMPOSSIBLE.
            //
            // ```cpp
            // primitive_desc_iterator_t it(
            //         engine, (op_desc_t *)&cd_dw, &attr_dw, nullptr);
            // if (!it.is_initialized()) return status::out_of_memory;
            // while (++it != it.end()) {
            //     dw_conv_pd_ = *it;
            //     break;
            // }
            // VDISPATCH_CONV_IC(dw_conv_pd_, "dw_conv_pd hasn't been created");
            // ```
            //
            // ```compiler output
            // error: ‘using element_type = struct dnnl::impl::primitive_desc_t’
            // {aka ‘struct dnnl::impl::primitive_desc_t’} has no member named
            // ‘jcp_’
            // auto &jcp_dw = dw_conv_pd_->jcp_;
            //                             ^~~~
            // ```
            //
            // TODO: figure out the way to initialize fused conv through a
            // normal interface without hacks accessing specific members.
            CHECK(safe_ptr_assign(
                    dw_conv_pd_, new dw_pd_t(&cd_dw, &attr_dw, nullptr)));
            CHECK(dw_conv_pd_->init(engine));
            auto &jcp_dw = dw_conv_pd_->jcp_;

            VDISPATCH_CONV_IC(
                    dnnl_memory_desc_equal(&src_md, dw_conv_pd_->src_md(0)),
                    VERBOSE_INCONSISTENT_MDS, "src_md", "dw_conv_pd_->src_md");
            VDISPATCH_CONV_IC(
                    jcp_1x1.oc_without_padding % jcp_1x1.oc_block == 0,
                    VERBOSE_1x1CONV_HEURISTIC_FAIL,
                    "output-channel is not an exact multiple of oc_block");
            VDISPATCH_CONV_IC(
                    IMPLICATION(jcp_dw.ow_block, jcp_dw.ow_block == jcp_dw.ow),
                    VERBOSE_1x1CONV_HEURISTIC_FAIL,
                    "ow_block does not equal output-width");

            assert(dw_conv_pd_->dst_md(0)->format_kind != format_kind::any);
            assert(dw_conv_pd_->weights_md(0)->format_kind != format_kind::any);
            assert(IMPLICATION(
                    dw_conv_pd_->weights_md(1)->data_type != data_type::undef,
                    dw_conv_pd_->weights_md(1)->format_kind
                            != format_kind::any));

            jcp_dw.is_fused_conv = true;
            // TODO: Support/experiment arbitary oc_work in dw conv.
            // Until then we keep oc_work perfectly divisible.
            while (jcp_1x1.nb_load % jcp_1x1.nb_load_blocking != 0)
                --jcp_1x1.nb_load_blocking;
            jcp_1x1.nb_load_blocking_max = jcp_1x1.nb_load_blocking;

            while (jcp_1x1.nb_load_blocking % jcp_dw.nb_ch_blocking != 0)
                --jcp_dw.nb_ch_blocking;

            jcp_dw.dw_conv_buffer_oc
                    = jcp_1x1.nb_load_blocking * jcp_1x1.oc_block;

            const auto dat_tag_nxc = utils::pick(ndims() - 3, format_tag::nwc,
                    format_tag::nhwc, format_tag::ndhwc);
            const bool is_data_nxc = utils::everyone_is(
                    dat_tag_nxc, jcp_1x1.src_tag, jcp_1x1.dst_tag);
            if (!is_data_nxc)
                jcp_1x1.bcast_loop_output_step = jcp_1x1.ur * jcp_1x1.load_block
                        * jcp_1x1.typesize_out;

            registrar_t scratchpad(scratchpad_registry_);
            registrar_t dw_scratchpad(scratchpad, names::prefix_fusion);

            size_t dw_conv_buffer_size_ = (size_t)nthr * jcp_dw.kh * jcp_dw.iw
                    * jcp_dw.dw_conv_buffer_oc;
            assert(dw_conv_buffer_size_);
            dw_scratchpad.book(memory_tracking::names::key_fusion_inout_buffer,
                    dw_conv_buffer_size_,
                    types::data_type_size(dw_conv_pd_->src_md()->data_type));

            jit_uni_dw_conv_fwd_kernel_t<sse41,
                    data_type::f32>::init_scratchpad(dw_scratchpad, jcp_dw);

            return status::success;
        }
    };

    jit_sse41_1x1_convolution_fwd_t(const pd_t *apd) : primitive_t(apd) {}

    using data_t = typename prec_traits_t<data_type::f32>::type;

    status_t init(engine_t *engine) override {
        CHECK(safe_ptr_assign(kernel_,
                new jit_sse41_1x1_conv_kernel_f32_t(
                        pd()->jcp_, *pd()->attr(), *pd()->dst_1x1_md(0))));
        CHECK(kernel_->create_kernel());
        if (pd()->jcp_.with_dw_conv) {
            CHECK(safe_ptr_assign(kernel_dw_,
                    new dw_conv_kernel_t(
                            pd()->dw_conv_pd_->jcp_, *pd()->dst_md(0))));
            return kernel_dw_->create_kernel();
        }

        return status::success;
    }

    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;
    void execute_forward_thr(const int ithr, const int nthr, const data_t *src,
            const data_t *weights, const data_t *bias, const data_t *weights_dw,
            const data_t *bias_dw, data_t *dst,
            const memory_tracking::grantor_t &scratchpad,
            const void *post_ops_binary_rhs_arg_vec,
            const void *post_ops_binary_rhs_arg_vec_dw) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
    std::unique_ptr<jit_sse41_1x1_conv_kernel_f32_t> kernel_;
    using dw_conv_kernel_t = jit_uni_dw_conv_fwd_kernel_f32_t<sse41>;
    std::unique_ptr<dw_conv_kernel_t> kernel_dw_;
};

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

#endif