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_GEMM_BF16_INNER_PRODUCT_HPP
#define CPU_X64_GEMM_BF16_INNER_PRODUCT_HPP

#include <assert.h>

#include <memory>

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

#include "cpu/cpu_engine.hpp"
#include "cpu/gemm/gemm.hpp"
#include "cpu/gemm_inner_product_utils.hpp"

#include "cpu/x64/jit_uni_convert_xf16.hpp"

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

template <data_type_t dst_data_type>
struct gemm_bf16_inner_product_fwd_t : public primitive_t {
    struct pd_t : public cpu_inner_product_fwd_pd_t {
        using cpu_inner_product_fwd_pd_t::cpu_inner_product_fwd_pd_t;

        DECLARE_COMMON_PD_T(GEMM_IMPL_STR, gemm_bf16_inner_product_fwd_t);

        status_t init(engine_t *engine) {
            using namespace utils;
            using namespace data_type;

            VDISPATCH_INNER_PRODUCT(
                    DNNL_CPU_THREADING_RUNTIME != DNNL_RUNTIME_THREADPOOL,
                    VERBOSE_UNSUPPORTED_THREADPOOL_RUNTIME);
            // disabling verbose dispatch messages for unsupported isa for
            // better readability
            if (!mayiuse(avx512_core)) return status::unimplemented;
            VDISPATCH_INNER_PRODUCT(is_fwd(), VERBOSE_BAD_PROPKIND);
            VDISPATCH_INNER_PRODUCT(
                    !has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
            VDISPATCH_INNER_PRODUCT(everyone_is(bf16, src_md()->data_type,
                                            weights_md()->data_type),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(dst_data_type == dst_md()->data_type,
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(
                    IMPLICATION(with_bias(),
                            one_of(weights_md(1)->data_type, f32, bf16)),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(set_default_params() == status::success,
                    VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_INNER_PRODUCT(
                    attr()->has_default_values(
                            primitive_attr_t::skip_mask_t::post_ops,
                            dst_md()->data_type),
                    VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_INNER_PRODUCT(
                    attr()->post_ops_.check_sum_consistency(
                            dst_md()->data_type, /* is_int8 */ false),
                    VERBOSE_UNSUPPORTED_POSTOP);
            VDISPATCH_INNER_PRODUCT(inner_product_utils::post_ops_ok(
                                            attr()->post_ops_, &dst_md_),
                    VERBOSE_UNSUPPORTED_POSTOP);
            VDISPATCH_INNER_PRODUCT(dense_gemm_consitency_check(
                                            src_md(), weights_md(), dst_md()),
                    VERBOSE_INCOMPATIBLE_GEMM_FMT);
            VDISPATCH_INNER_PRODUCT(
                    attr_.set_default_formats(dst_md(0)) == status::success,
                    VERBOSE_UNSUPPORTED_POSTOP);

            dst_is_acc_ = dst_data_type == f32;

            init_scratchpad();

            return status::success;
        }

        bool dst_is_acc_ = false;

    protected:
        void init_scratchpad() {
            if (!dst_is_acc_) {
                auto scratchpad = scratchpad_registry().registrar();
                scratchpad.template book<acc_data_t>(
                        memory_tracking::names::key_iprod_int_dat_in_acc_dt,
                        MB() * OC());
            }
        }
    };

    gemm_bf16_inner_product_fwd_t(const pd_t *apd) : primitive_t(apd) {}

    using dst_data_t = typename prec_traits_t<dst_data_type>::type;
    using acc_data_t = typename prec_traits_t<data_type::f32>::type;
    using src_data_t = typename prec_traits_t<data_type::bf16>::type;
    using wei_data_t = typename prec_traits_t<data_type::bf16>::type;

    status_t init(engine_t *engine) override {
        const bool has_bias = pd()->with_bias();
        const bool has_eltwise
                = pd()->attr()->post_ops_.find(primitive_kind::eltwise) >= 0;
        const bool has_binary
                = pd()->attr()->post_ops_.find(primitive_kind::binary) >= 0;
        const bool has_sum_as_postops = !pd()->dst_is_acc_;
        postops_in_ip_ = false
                || !pd()->dst_is_acc_ /* includes has_sum_as_postops */
                || has_bias || has_eltwise || has_binary;
        if (postops_in_ip_)
            CHECK(safe_ptr_assign(pp_kernel_,
                    inner_product_utils::pp_kernel_t::create(
                            pd(), !has_sum_as_postops)));

        auto sum_idx = pd()->attr()->post_ops_.find(primitive_kind::sum);
        beta_ = sum_idx >= 0 && !has_sum_as_postops
                ? pd()->attr()->post_ops_.entry_[sum_idx].sum.scale
                : 0.0;

        return (pp_kernel_) ? pp_kernel_->create_kernel() : status::success;
    }

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

private:
    std::unique_ptr<inner_product_utils::pp_kernel_t> pp_kernel_;
    bool postops_in_ip_;
    float beta_;

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

template <data_type_t diff_src_data_type>
struct gemm_bf16_inner_product_bwd_data_t : public primitive_t {
    struct pd_t : public cpu_inner_product_bwd_data_pd_t {
        using cpu_inner_product_bwd_data_pd_t::cpu_inner_product_bwd_data_pd_t;

        DECLARE_COMMON_PD_T(GEMM_IMPL_STR, gemm_bf16_inner_product_bwd_data_t);

        status_t init(engine_t *engine) {
            using namespace data_type;
            VDISPATCH_INNER_PRODUCT(
                    DNNL_CPU_THREADING_RUNTIME != DNNL_RUNTIME_THREADPOOL,
                    VERBOSE_UNSUPPORTED_THREADPOOL_RUNTIME);
            // disabling verbose dispatch messages for unsupported isa for
            // better readability
            if (!mayiuse(avx512_core)) return status::unimplemented;
            VDISPATCH_INNER_PRODUCT(
                    desc()->prop_kind == prop_kind::backward_data,
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_INNER_PRODUCT(
                    !has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
            VDISPATCH_INNER_PRODUCT(
                    utils::everyone_is(bf16, weights_md()->data_type,
                            diff_dst_md()->data_type),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(
                    diff_src_data_type == diff_src_md()->data_type,
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_INNER_PRODUCT(
                    this->set_default_params() == status::success,
                    VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_INNER_PRODUCT(dense_gemm_consitency_check(diff_src_md(),
                                            weights_md(), diff_dst_md()),
                    VERBOSE_INCOMPATIBLE_GEMM_FMT);

            diff_src_is_acc_ = diff_src_data_type == data_type::f32;

            init_scratchpad();

            return status::success;
        }

        bool diff_src_is_acc_ = false;

    private:
        void init_scratchpad() {
            if (!diff_src_is_acc_) {
                auto scratchpad = scratchpad_registry().registrar();
                scratchpad.template book<acc_data_t>(
                        memory_tracking::names::key_iprod_int_dat_in_acc_dt,
                        MB() * IC_total_padded());
            }
        }
    };

    gemm_bf16_inner_product_bwd_data_t(const pd_t *apd) : primitive_t(apd) {}

    using diff_dst_data_t = typename prec_traits_t<data_type::bf16>::type;
    using acc_data_t = typename prec_traits_t<data_type::f32>::type;
    using diff_src_data_t = typename prec_traits_t<diff_src_data_type>::type;
    using wei_data_t = typename prec_traits_t<data_type::bf16>::type;

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

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

template <data_type_t diff_wei_data_type>
struct gemm_bf16_inner_product_bwd_weights_t : public primitive_t {
    struct pd_t : public cpu_inner_product_bwd_weights_pd_t {
        using cpu_inner_product_bwd_weights_pd_t::
                cpu_inner_product_bwd_weights_pd_t;

        DECLARE_COMMON_PD_T(
                GEMM_IMPL_STR, gemm_bf16_inner_product_bwd_weights_t);

        status_t init(engine_t *engine) {
            using namespace utils;
            using namespace data_type;
            VDISPATCH_INNER_PRODUCT(
                    DNNL_CPU_THREADING_RUNTIME != DNNL_RUNTIME_THREADPOOL,
                    VERBOSE_UNSUPPORTED_THREADPOOL_RUNTIME);
            // disabling verbose dispatch messages for unsupported isa for
            // better readability
            if (!mayiuse(avx512_core)) return status::unimplemented;
            VDISPATCH_INNER_PRODUCT(
                    desc()->prop_kind == prop_kind::backward_weights,
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_INNER_PRODUCT(
                    !has_zero_dim_memory(), VERBOSE_EMPTY_TENSOR, "");
            VDISPATCH_INNER_PRODUCT(everyone_is(bf16, src_md()->data_type,
                                            diff_dst_md()->data_type),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(
                    diff_wei_data_type == diff_weights_md()->data_type,
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(
                    IMPLICATION(with_bias(),
                            one_of(diff_weights_md(1)->data_type, f32, bf16)),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_INNER_PRODUCT(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_INNER_PRODUCT(set_default_params() == status::success,
                    VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_INNER_PRODUCT(dense_gemm_consitency_check(src_md(),
                                            diff_weights_md(), diff_dst_md()),
                    VERBOSE_INCOMPATIBLE_GEMM_FMT);

            diff_wei_is_acc_ = diff_wei_data_type == f32;
            bias_reduction_nthr_ = dnnl_get_max_threads();

            init_scratchpad();

            return status::success;
        }

        bool diff_wei_is_acc_ = false;
        int bias_reduction_nthr_ = 1;
        static const dim_t bias_blksize = 32;

        void get_bias_partitioning(
                dim_t &OC_per_thread, int &nthr_OCB, int &nthr_MB) const {
            dim_t OCB = utils::div_up(OC(), bias_blksize);
            dim_t OCB_per_thread = utils::div_up(OCB, bias_reduction_nthr_);

            OC_per_thread = OCB_per_thread * bias_blksize;
            nthr_OCB = utils::div_up(OCB, OCB_per_thread);
            nthr_MB = bias_reduction_nthr_ / nthr_OCB;

            assert(nthr_OCB * nthr_MB <= bias_reduction_nthr_);
        }

    private:
        void init_scratchpad() {
            using namespace memory_tracking;
            auto scratchpad = scratchpad_registry().registrar();

            if (!diff_wei_is_acc_)
                scratchpad.template book<acc_data_t>(
                        names::key_iprod_int_dat_in_acc_dt,
                        OC() * IC_total_padded());

            if (with_bias()) {
                dim_t OC_per_thread {0};
                int nthr_OCB {0}, nthr_MB {0};
                get_bias_partitioning(OC_per_thread, nthr_OCB, nthr_MB);

                const bool diff_bias_is_acc = nthr_MB == 1
                        && diff_weights_md(1)->data_type == data_type::f32;

                if (!diff_bias_is_acc)
                    scratchpad.template book<acc_data_t>(
                            names::key_iprod_bias_bf16_convert_wsp,
                            nthr_OCB * nthr_MB * OC_per_thread);
            }
        }
    };

    gemm_bf16_inner_product_bwd_weights_t(const pd_t *apd) : primitive_t(apd) {}

    status_t init(engine_t *engine) override {
        if (pd()->with_bias())
            CHECK(safe_ptr_assign(bias_reduction_,
                    new jit_cvt_xf16_to_ps_t(
                            data_type::bf16, true, (size_t)pd()->OC())));
        return status::success;
    }

    using diff_dst_data_t = typename prec_traits_t<data_type::bf16>::type;
    using acc_data_t = typename prec_traits_t<data_type::f32>::type;
    using src_data_t = typename prec_traits_t<data_type::bf16>::type;
    using diff_wei_data_t = typename prec_traits_t<diff_wei_data_type>::type;

    status_t execute(const exec_ctx_t &ctx) const override {
        return execute_backward_weights(ctx);
    }

private:
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }

    status_t execute_backward_weights(const exec_ctx_t &ctx) const;
    void execute_backward_bias(const exec_ctx_t &ctx) const;

    std::unique_ptr<jit_cvt_xf16_to_ps_t> bias_reduction_;
};

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

#endif

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