onednn-src 0.1.13

/*******************************************************************************
* Copyright 2021 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 GPU_INTEL_POOL_XE_GLOBAL_HPP
#define GPU_INTEL_POOL_XE_GLOBAL_HPP

#include "common/c_types_map.hpp"
#include "common/primitive.hpp"
#include "common/reduction_pd.hpp"
#include "gpu/intel/pool/config.hpp"
#include "gpu/intel/primitive.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace pool {

struct xe_global_fwd_t : public primitive_t {
    using primitive_t::primitive_t;
    struct pd_t : public fwd_pd_t {
        using fwd_pd_t::fwd_pd_t;

        DECLARE_COMMON_PD_T("ocl:xe_global:any", xe_global_fwd_t);

        status_t init(impl::engine_t *engine) {
            using namespace data_type;
            using namespace prop_kind;
            using namespace alg_kind;
            auto src_data_t = src_md()->data_type;
            auto dst_data_t = dst_md()->data_type;

            VDISPATCH_POOLING_SC(set_default_params(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_POOLING(utils::one_of(desc()->prop_kind, forward_training,
                                      forward_inference),
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_POOLING(utils::one_of(desc()->alg_kind, pooling_max,
                                      pooling_avg_include_padding,
                                      pooling_avg_exclude_padding),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_POOLING(!utils::one_of(f64, src_data_t, dst_data_t),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_POOLING(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);
            CHECK(init_conf(engine));

            bool is_training = desc_.prop_kind == forward_training;
            if (desc()->alg_kind == pooling_max && is_training) {
                // Required for storing spatial offsets into workspace for
                // pooling_max training due to use of int type.
                VDISPATCH_POOLING(conf.id * conf.ih * conf.iw <= INT_MAX,
                        VERBOSE_OFFSET_DT_MISMATCH, "kernel spatial", "int");
                init_default_ws(s32);
            }

            VDISPATCH_POOLING_SC(init_reduction(engine),
                    VERBOSE_PRIMITIVE_CREATION_FAIL, "reduction");
            init_scratchpad();
            return status::success;
        }

        status_t init_conf(impl::engine_t *engine);
        status_t init_kernel_ctx(compute::kernel_ctx_t &kernel_ctx) const;

        void init_scratchpad();

        status_t init_reduction(impl::engine_t *engine) {
            using namespace alg_kind;

            reduction_desc_t rdesc;
            memory_desc_t red_src_mem_desc(*src_md(0));
            red_src_mem_desc.data_type = src_md()->data_type;
            CHECK(reduction_desc_init(&rdesc,
                    desc()->alg_kind == pooling_max
                            ? dnnl_alg_kind_t::dnnl_reduction_max
                            : dnnl_alg_kind_t::dnnl_reduction_mean,
                    &red_src_mem_desc, dst_md(0), 0, 0));
            primitive_attr_t reduction_attr(*attr());
            if (!reduction_attr.is_initialized()) return status::out_of_memory;
            primitive_desc_iterator_t it(
                    engine, (op_desc_t *)&rdesc, &reduction_attr, nullptr);
            if (!it.is_initialized()) return status::invalid_arguments;
            reduction_pd_ = *(++it);
            if (reduction_pd_)
                return status::success;
            else { return status::invalid_arguments; }
        }

        std::shared_ptr<primitive_desc_t> reduction_pd_;
        conf_t conf;
        offsets_t off;
    };

    status_t init(impl::engine_t *engine) override {
        compute::kernel_ctx_t kernel_ctx;
        status_t status = pd()->init_kernel_ctx(kernel_ctx);
        CHECK(status);

        using namespace alg_kind;
        // TODO: max-pooling requires workspace to track indices for training config
        if (pd()->desc()->alg_kind != pooling_max
                && !utils::one_of(data_type::f64, pd()->dst_md()->data_type,
                        pd()->src_md()->data_type)) {
            if (create_nested_primitive(
                        reduction_p_, pd()->reduction_pd_, engine)
                    == status::success) {
                return status::success;
            }
        }
        // fallback
        CHECK(create_kernel(
                engine, &kernel_, "xe_global_pooling_fwd", kernel_ctx));
        if (!kernel_) return status::runtime_error;
        return status::success;
    }

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

private:
    status_t execute_forward(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
    compute::kernel_t kernel_;
    std::shared_ptr<impl::primitive_t> reduction_p_;
};

struct xe_global_bwd_t : public primitive_t {
    using primitive_t::primitive_t;
    struct pd_t : public bwd_pd_t {
        using bwd_pd_t::bwd_pd_t;

        DECLARE_COMMON_PD_T("ocl:xe_global:any", xe_global_bwd_t);

        status_t init(impl::engine_t *engine) {
            using namespace prop_kind;
            using namespace alg_kind;
            auto *intel_engine = utils::downcast<intel::engine_t *>(engine);

            auto diff_dst_dt = diff_dst_md()->data_type;
            auto diff_src_dt = diff_src_md()->data_type;

            VDISPATCH_POOLING_SC(set_default_params(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_POOLING(utils::one_of(desc()->prop_kind, backward_data),
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_POOLING(utils::one_of(desc()->alg_kind, pooling_max,
                                      pooling_avg_include_padding,
                                      pooling_avg_exclude_padding),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_POOLING(
                    (utils::everyone_is(data_type::f32,
                             diff_dst_md()->data_type, diff_src_md()->data_type)
                            || utils::everyone_is(data_type::f16,
                                    diff_dst_md()->data_type,
                                    diff_src_md()->data_type)
                            || utils::everyone_is(data_type::bf16,
                                    diff_dst_md()->data_type,
                                    diff_src_md()->data_type)),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_POOLING(
                    !utils::one_of(data_type::f64, diff_src_dt, diff_dst_dt),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_POOLING(
                    IMPLICATION(diff_src_md()->data_type == data_type::f16,
                            intel_engine->mayiuse(
                                    compute::device_ext_t::khr_fp16)
                                    && intel_engine->mayiuse(
                                            compute::device_ext_t::
                                                    intel_subgroups_short)),
                    VERBOSE_UNSUPPORTED_DT_CFG);
            VDISPATCH_POOLING(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);
            CHECK(init_conf(engine));

            if (desc()->alg_kind == pooling_max) {
                // Required for storing spatial offsets into workspace for
                // pooling_max training due to use of int type.
                VDISPATCH_POOLING(conf.id * conf.ih * conf.iw <= INT_MAX,
                        VERBOSE_OFFSET_DT_MISMATCH, "kernel spatial", "int");
                init_default_ws(data_type::s32);
                VDISPATCH_POOLING(
                        compare_ws(hint_fwd_pd_), VERBOSE_WS_MISMATCH);
            }
            return status::success;
        }

        status_t init_conf(impl::engine_t *engine);
        status_t init_kernel_ctx(compute::kernel_ctx_t &kernel_ctx) const;

        conf_t conf;
        offsets_t off;
    };

    status_t init(impl::engine_t *engine) override {
        compute::kernel_ctx_t kernel_ctx;
        status_t status = pd()->init_kernel_ctx(kernel_ctx);
        CHECK(status);

        CHECK(create_kernel(
                engine, &kernel_, "xe_global_pooling_bwd", kernel_ctx));
        if (!kernel_) return status::runtime_error;

        return status::success;
    }

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

private:
    status_t execute_backward(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
    compute::kernel_t kernel_;
};

} // namespace pool
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl

#endif