/*******************************************************************************
* 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_NHWC_POOLING_HPP
#define CPU_NHWC_POOLING_HPP

#include <assert.h>

#include "c_types_map.hpp"
#include "cpu_engine.hpp"
#include "cpu_pooling_pd.hpp"
#include "mkldnn_thread.hpp"
#include "type_helpers.hpp"
#include "utils.hpp"
#include "bfloat16_utils.hpp"

namespace mkldnn {
namespace impl {
namespace cpu {

template <data_type_t d_type>
struct nhwc_pooling_fwd_t: public cpu_primitive_t {
    struct pd_t: public cpu_pooling_fwd_pd_t {
        pd_t(engine_t *engine, const pooling_desc_t *adesc,
                const primitive_attr_t *attr,
                const pooling_fwd_pd_t *hint_fwd_pd)
            : cpu_pooling_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}

        DECLARE_COMMON_PD_T("nhwc_pooling:any", nhwc_pooling_fwd_t);

        virtual status_t init() override {
            using namespace prop_kind;
            using namespace alg_kind;
            using namespace memory_format;
            assert(engine()->kind() == engine_kind::cpu);
            auto src_format = src_pd()->desc()->format;
            bool ok = true
                && set_default_params() == status::success
                && utils::one_of(desc()->prop_kind, forward_training,
                        forward_inference)
                && utils::one_of(desc()->alg_kind, pooling_max,
                        pooling_avg_include_padding,
                        pooling_avg_exclude_padding)
                && utils::everyone_is(d_type,
                        src_pd()->desc()->data_type,
                        dst_pd()->desc()->data_type)
                && utils::one_of(src_format, nhwc, ndhwc)
                && (src_format == dst_pd()->desc()->format)
                && attr()->has_default_values();
            if (!ok) return status::unimplemented;

            bool is_training = desc_.prop_kind == forward_training;
            if (desc()->alg_kind == pooling_max && is_training) {
                // Allocate dense workspace buffer based on logical dimensions
                // of the output dst
                memory_desc_t indices_desc;
                if (is_3d()) {
                    dims_t ws_dims = { MB(), C(), OD(), OH(), OW() };
                    mkldnn_memory_desc_init(&indices_desc, 5, ws_dims,
                            pooling_index_data_type(desc()),
                            memory_format::ndhwc);
                } else {
                    dims_t ws_dims = { MB(), C(), OH(), OW() };
                    mkldnn_memory_desc_init(&indices_desc, 4, ws_dims,
                            pooling_index_data_type(desc()),
                            memory_format::nhwc);
                }
                ws_pd_ = cpu_memory_t::pd_t(engine_, &indices_desc);
            }

            init_scratchpad();

            return status::success;
        }

        private:
            void init_scratchpad() {
                using namespace memory_tracking::names;
                if (src_pd()->desc()->data_type == data_type::bf16) {
                    size_t bf16cvt_sz_ = C() * mkldnn_get_max_threads();
                    auto scratchpad = scratchpad_registry().registrar();
                    scratchpad.book(key_pool_src_bf16cvt,
                            sizeof(float) * bf16cvt_sz_);
                    scratchpad.book(key_pool_dst_bf16cvt,
                            sizeof(float) * bf16cvt_sz_);
                }
            }
    };

    nhwc_pooling_fwd_t(const pd_t *apd, const input_vector &inputs,
            const output_vector &outputs)
        : cpu_primitive_t(apd, inputs, outputs) {}

    ~nhwc_pooling_fwd_t() {}

    typedef typename prec_traits<d_type>::type data_t;
    typedef typename prec_traits<data_type::f32>::type ker_data_t;

    virtual void execute(event_t *e) const {
        execute_forward();
        e->set_state(event_t::ready);
    }

private:
    void execute_forward() const;

    void array_max(const int n, ker_data_t *dst, const ker_data_t *src,
            unsigned char *ws, const size_t ws_offset, const data_type_t ws_dt,
            const int index) const {
        assert(IMPLICATION(pd()->desc()->alg_kind == alg_kind::pooling_max &&
                    pd()->desc()->prop_kind == prop_kind::forward_training,
                    nullptr != ws));
        PRAGMA_OMP_SIMD()
        for (int oc = 0; oc < n; ++oc) {
            auto s = src[oc];
            ker_data_t mv = dst[oc];

            // update index of maximum
#if defined __INTEL_COMPILER
            if (ws && (s > mv)) {
                assert(ws_dt == data_type::u8 || ws_dt == data_type::s32);
                if (ws_dt == data_type::u8) {
                    assert(0 <= index && index <= 255);
                    ws[ws_offset + oc] = index;
                } else
                    reinterpret_cast<int *>(ws)[ws_offset + oc] = index;
            }
#else
            // Need to add explicit predicates for GCC to vectorize this.
            // And although the resulting code is ugly, it is still 4 times
            // faster than scalar
            if (ws) {
                assert(ws_dt == data_type::u8 || ws_dt == data_type::s32);

                if (ws_dt == data_type::u8) {
                    assert(0 <= index && index <= 255);
                    unsigned char predicate = (s > mv) ? 0xff : 0;
                    unsigned char current_value = ws[ws_offset + oc];
                    current_value = (predicate & (unsigned char)index)
                        | ((~predicate) & current_value);
                    ws[ws_offset + oc] = current_value;
                } else {
                    auto wint = reinterpret_cast<int *>(ws);
                    unsigned int predicate = (s > mv) ? 0xffffffff : 0;
                    unsigned int current_value = wint[ws_offset + oc];
                    current_value = (predicate & (unsigned int)index)
                        | ((~predicate) & current_value);
                    wint[ws_offset + oc] = current_value;
                }
            }
#endif
            dst[oc] = nstl::max(s, mv);
        }
    }

    void array_initialize(const int n, ker_data_t *dst, unsigned char *ws,
            const size_t ws_offset, const data_type_t ws_dt) const {
        assert(IMPLICATION(pd()->desc()->alg_kind == alg_kind::pooling_max &&
                    pd()->desc()->prop_kind == prop_kind::forward_training,
                    nullptr != ws));
        for (int oc = 0; oc < n; ++oc) {
            if (ws) {
                assert(ws_dt == data_type::u8 || ws_dt == data_type::s32);
                if (ws_dt == data_type::u8) {
                    ws[ws_offset + oc] = 0;
                } else
                    reinterpret_cast<int *>(ws)[ws_offset + oc] = 0;
            }
            if (d_type == data_type::bf16) {
                dst[oc] = bf16_cvt_utils::cvt_bfloat16_to_float(
                        bf16_cvt_utils::approx_bfloat16_lowest());
            } else
                dst[oc] = nstl::numeric_limits<ker_data_t>::lowest();
        }
    }

    void array_div_by_const(const int n, ker_data_t *dst,
            const size_t num, ker_data_t *src) const {
        for (int i = 0; i < n; ++i)
        {
            float ftmp = (float)src[i];
            ftmp = ftmp / num;
            dst[i] = math::out_round<ker_data_t>(ftmp);
        }
    }

    template<typename ker_data_t>
    void array_add(const int n, ker_data_t *dst, const ker_data_t *src) const {
        for (int i = 0;  i < n; ++i)
        {
            dst[i] += src[i];
        }
    }

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

template <impl::data_type_t d_type>
struct nhwc_pooling_bwd_t: public cpu_primitive_t {
    struct pd_t: public cpu_pooling_bwd_pd_t {
        pd_t(engine_t *engine, const pooling_desc_t *adesc,
                const primitive_attr_t *attr,
                const pooling_fwd_pd_t *hint_fwd_pd)
            : cpu_pooling_bwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}

        DECLARE_COMMON_PD_T("nhwc:any", nhwc_pooling_bwd_t);

        virtual status_t init() override {
            using namespace prop_kind;
            using namespace alg_kind;
            using namespace memory_format;
            assert(engine()->kind() == engine_kind::cpu);
            auto diff_dst_format = diff_dst_pd()->desc()->format;
            bool ok = true
                && set_default_params() == status::success
                && utils::one_of(desc()->prop_kind, backward_data)
                && utils::one_of(desc()->alg_kind, pooling_max,
                        pooling_avg_include_padding,
                        pooling_avg_exclude_padding)
                && utils::everyone_is(d_type,
                        diff_dst_pd()->desc()->data_type,
                        diff_src_pd()->desc()->data_type)
                && utils::one_of(diff_dst_format, nhwc, ndhwc)
                && (diff_dst_format == diff_src_pd()->desc()->format)
                && attr()->has_default_values();
            if (!ok)
                return status::unimplemented;

            if (desc()->alg_kind == pooling_max) {
                bool ws_ok = true
                    && hint_fwd_pd_
                    && hint_fwd_pd_->workspace_pd()
                    && utils::one_of(
                            hint_fwd_pd_->workspace_pd()->desc()->format,
                            nhwc, ndhwc)
                    && hint_fwd_pd_->workspace_pd()->engine()->kind()
                            == engine_kind::cpu;
                if (!ws_ok) return status::unimplemented;

                ws_pd_ = *(cpu_memory_t::pd_t *)hint_fwd_pd_->workspace_pd();
            }

            init_scratchpad();

            return status::success;
        }

        private:
            void init_scratchpad() {
                using namespace memory_tracking::names;
                if (diff_src_pd()->desc()->data_type == data_type::bf16) {
                    size_t bf16cvt_sz_ = C() * mkldnn_get_max_threads();
                    auto scratchpad = scratchpad_registry().registrar();
                    scratchpad.book(key_pool_src_bf16cvt,
                            sizeof(float) * bf16cvt_sz_);
                    scratchpad.book(key_pool_dst_bf16cvt,
                            sizeof(float) * bf16cvt_sz_);
                }
            }
    };

    nhwc_pooling_bwd_t(const pd_t *apd, const input_vector &inputs,
            const output_vector &outputs)
        : cpu_primitive_t(apd, inputs, outputs) {}

    ~nhwc_pooling_bwd_t() {}

    typedef typename prec_traits<d_type>::type data_t;

    virtual void execute(event_t *e) const {
        execute_backward();
        e->set_state(event_t::ready);
    }

private:
    void execute_backward() const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }

};

}// namespace cpu
}// namespace impl
}// namespace mkldnn

#endif

// vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s