onednn-src 0.1.13

/*******************************************************************************
* Copyright 2023 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_BNORM_REUSABLE_HPP
#define GPU_INTEL_BNORM_REUSABLE_HPP

#include "common/serialization.hpp"
#include "gpu/intel/bnorm/config.hpp"
#include "gpu/intel/compute/dispatch_reusable.hpp"
#include "gpu/intel/compute/kernel.hpp"
#include "gpu/intel/primitive.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace bnorm {

struct reusable_params_t {
    status_t create_generator(const intel::engine_t &engine,
            compute::kernel_bundle_t &bundle) const {
        auto status = engine.create_kernel_bundle(
                bundle, get_kernel_names(), get_kernel_ctx());
        return status;
    }

    const std::vector<const char *> &get_kernel_names() const {
        static const std::vector<const char *> kernel_names
                = {"reusable_bnorm_fwd", "reusable_calculate_mean",
                        "reusable_calculate_variance", "reusable_reduce_mean",
                        "reusable_reduce_variance", "reusable_calculate_stats",
                        "reusable_reduce_stats", "reusable_bnorm_bwd"};
        return kernel_names;
    }

#if __cplusplus >= 202002L
    bool operator==(const reusable_params_t &) const = default;
#endif
    serialization_stream_t serialize() const {
        DNNL_ASSERT_TRIVIALLY_SERIALIZABLE(reusable_params_t);
        return serialization_stream_t(*this);
    }

    static reusable_params_t deserialize(const serialization_stream_t &s) {
        reusable_params_t t {};
        deserializer_t d(s);
        d.pop(t);
        return t;
    }

    compute::kernel_ctx_t get_kernel_ctx() const;

    // 4 supported data types (s8, f32, bf16, f16)
    data_type_t data_type;

    // 2 options each line
    bool use_scale;
    bool use_shift;
    bool is_training;
    bool fuse_norm_relu;
    bool fuse_norm_add_relu;
    bool with_relu;
    bool with_leaky_relu;
    bool calculate_stats;
    bool require_stateless_addressing;

    uint8_t padding[3] = {0};

    // Close to one set of configurations per block layout (9 common cases)
    compute::dispatch_compile_params_t calc_stat_params;
    compute::dispatch_compile_params_t reduce_stat_params;
    compute::dispatch_compile_params_t gws_params;
};

struct reusable_runtime_params_t {
    dim_t reduce_dim_stride;
    compute::dispatch_runtime_params_t calc_stat_params;
    compute::dispatch_runtime_params_t reduce_stat_params;
    compute::dispatch_runtime_params_t gws_params;

    float relu_negative_slope;
    float eps;

    dim_t stat_ic, reduction_nelems, div, ic;
};

struct reusable_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:reusable", reusable_fwd_t);

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

            const auto attr_skip_mask = primitive_attr_t::skip_mask_t::post_ops;

            VDISPATCH_BNORM(is_fwd(), VERBOSE_BAD_PROPKIND);
            VDISPATCH_BNORM(
                    utils::one_of(src_md()->data_type, f32, bf16, f16, s8),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_BNORM(IMPLICATION(f16 == src_md()->data_type,
                                    intel_engine->mayiuse(
                                            compute::device_ext_t::khr_fp16)),
                    VERBOSE_UNSUPPORTED_DT_CFG);
            VDISPATCH_BNORM(src_md()->data_type == dst_md()->data_type,
                    VERBOSE_INCONSISTENT_DT, "src", "dst");
            VDISPATCH_BNORM(IMPLICATION(src_md()->data_type == s8,
                                    !is_training() && stats_is_src()),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_BNORM(check_scale_shift_data_type(),
                    VERBOSE_UNSUPPORTED_FEATURE,
                    "unsupported scale, shift or datatype configuration");
            VDISPATCH_BNORM(attr()->has_default_values(attr_skip_mask),
                    VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_BNORM(
                    IMPLICATION(!attr()->has_default_values(),
                            attr()->post_ops_.len() == 1
                                    && with_relu_post_op(is_training())),
                    VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_BNORM(
                    set_default_formats_common(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_BNORM(memory_desc_wrapper(src_md())
                            == memory_desc_wrapper(dst_md()),
                    VERBOSE_INCONSISTENT_MDS, "src", "dst");
            VDISPATCH_BNORM(intel_engine->mayiuse(
                                    compute::device_ext_t::intel_subgroups),
                    VERBOSE_UNSUPPORTED_DEVICE_FEATURE, "subgroups");

            if (is_training() && (fuse_norm_relu() || fuse_norm_add_relu())) {
                VDISPATCH_BNORM_SC(init_default_ws(8), VERBOSE_WS_INIT);
            }
            CHECK(init_conf(engine));
            init_scratchpad();

            return status::success;
        }

        status_t init_conf(impl::engine_t *engine);
        void init_scratchpad();

        reusable_params_t conf;
        reusable_runtime_params_t rt_conf;
    };

    status_t init(impl::engine_t *engine) override {
        if (pd()->has_zero_dim_memory()) return status::success;
        std::vector<const char *> kernel_names
                = {"reusable_bnorm_fwd", nullptr, nullptr, nullptr, nullptr};
        if (pd()->conf.calculate_stats) {
            kernel_names[1] = "reusable_calculate_mean";
            kernel_names[2] = "reusable_calculate_variance";
            kernel_names[3] = "reusable_reduce_mean";
            kernel_names[4] = "reusable_reduce_variance";
        }

        std::vector<compute::kernel_t> kernels;
        CHECK(create_kernels(engine, kernels, kernel_names, pd()->conf));

        kernel_ = kernels[0];
        calculate_mean_kernel_ = kernels[1];
        calculate_variance_kernel_ = kernels[2];
        reduce_mean_kernel_ = kernels[3];
        reduce_variance_kernel_ = kernels[4];

        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_;
    compute::kernel_t calculate_mean_kernel_;
    compute::kernel_t reduce_mean_kernel_;
    compute::kernel_t calculate_variance_kernel_;
    compute::kernel_t reduce_variance_kernel_;
    compute::kernel_t calculate_mean_variance_kernel_;
};

struct reusable_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:reusable", reusable_bwd_t);

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

            VDISPATCH_BNORM(!is_fwd(), VERBOSE_BAD_PROPKIND);
            VDISPATCH_BNORM(utils::one_of(src_md()->data_type, f32, bf16, f16),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_BNORM(IMPLICATION(f16 == src_md()->data_type,
                                    intel_engine->mayiuse(
                                            compute::device_ext_t::khr_fp16)),
                    VERBOSE_UNSUPPORTED_DT_CFG);
            VDISPATCH_BNORM(src_md()->data_type == diff_src_md()->data_type,
                    VERBOSE_INCONSISTENT_DT, "src", "diff_src");
            VDISPATCH_BNORM(
                    diff_src_md()->data_type == diff_dst_md()->data_type,
                    VERBOSE_INCONSISTENT_DT, "diff_src", "diff_dst");
            VDISPATCH_BNORM(check_scale_shift_data_type(),
                    VERBOSE_UNSUPPORTED_FEATURE,
                    "unsupported scale, shift or datatype configuration");
            VDISPATCH_BNORM(
                    attr()->has_default_values(), VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_BNORM(
                    set_default_formats_common(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_BNORM(memory_desc_wrapper(diff_src_md())
                            == memory_desc_wrapper(diff_dst_md()),
                    VERBOSE_INCONSISTENT_MDS, "diff_src", "diff_dst");

            if (fuse_norm_relu() || fuse_norm_add_relu()) {
                VDISPATCH_BNORM_SC(init_default_ws(8), VERBOSE_WS_INIT);
                VDISPATCH_BNORM(compare_ws(hint_fwd_pd_), VERBOSE_WS_MISMATCH);
            }
            CHECK(init_conf(engine));
            init_scratchpad();

            return status::success;
        }

        status_t init_conf(impl::engine_t *engine);
        void init_scratchpad();

        reusable_params_t conf;
        reusable_runtime_params_t rt_conf;
    };

    status_t init(impl::engine_t *engine) override {
        if (pd()->has_zero_dim_memory()) return status::success;

        std::vector<const char *> kernel_names = {"reusable_bnorm_bwd",
                "reusable_calculate_stats", "reusable_reduce_stats"};

        std::vector<compute::kernel_t> kernels;
        CHECK(create_kernels(engine, kernels, kernel_names, pd()->conf));

        kernel_ = kernels[0];
        calculate_stats_kernel_ = kernels[1];
        reduce_stats_kernel_ = kernels[2];

        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_;
    compute::kernel_t calculate_stats_kernel_;
    compute::kernel_t reduce_stats_kernel_;
};

} // namespace bnorm
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl

#endif