benchmark_device_reduce.parallel.hpp

// MIT License
//
// Copyright (c) 2022-2024 Advanced Micro Devices, Inc. All rights reserved.
//
// Permission is hereby granted, free of charge, to any person obtaining a copy
// of this software and associated documentation files (the "Software"), to deal
// in the Software without restriction, including without limitation the rights
// to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
// copies of the Software, and to permit persons to whom the Software is
// furnished to do so, subject to the following conditions:
//
// The above copyright notice and this permission notice shall be included in all
// copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
// FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
// AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
// LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
// OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
// SOFTWARE.

#ifndef ROCPRIM_BENCHMARK_DEVICE_REDUCE_PARALLEL_HPP_
#define ROCPRIM_BENCHMARK_DEVICE_REDUCE_PARALLEL_HPP_

#include "benchmark_utils.hpp"

// Google Benchmark
#include <benchmark/benchmark.h>

// HIP API
#include <hip/hip_runtime.h>

// rocPRIM HIP API
#include <rocprim/device/detail/device_config_helper.hpp>
#include <rocprim/device/device_reduce.hpp>

#include <string>
#include <vector>

#include <cstddef>

constexpr const char* get_reduce_method_name(rocprim::block_reduce_algorithm alg)
{
    switch(alg)
    {
        case rocprim::block_reduce_algorithm::raking_reduce: return "raking_reduce";
        case rocprim::block_reduce_algorithm::raking_reduce_commutative_only:
            return "raking_reduce_commutative_only";
        case rocprim::block_reduce_algorithm::using_warp_reduce:
            return "using_warp_reduce";
            // Not using `default: ...` because it kills effectiveness of -Wswitch
    }
    return "unknown_algorithm";
}

template<typename Config>
std::string config_name()
{
    const rocprim::detail::reduce_config_params config = Config();
    return "{bs:" + std::to_string(config.reduce_config.block_size)
           + ",ipt:" + std::to_string(config.reduce_config.items_per_thread)
           + ",method:" + std::string(get_reduce_method_name(config.block_reduce_method)) + "}";
}

template<>
inline std::string config_name<rocprim::default_config>()
{
    return "default_config";
}

template<typename T              = int,
         typename BinaryFunction = rocprim::plus<T>,
         typename Config         = rocprim::default_config>
struct device_reduce_benchmark : public config_autotune_interface
{
    std::string name() const override
    {
        return bench_naming::format_name("{lvl:device,algo:reduce,key_type:"
                                         + std::string(Traits<T>::name())
                                         + ",cfg:" + config_name<Config>() + "}");
    }

    static constexpr unsigned int batch_size = 10;
    static constexpr unsigned int warmup_size = 5;

    void run(benchmark::State&   state,
             size_t              size,
             const managed_seed& seed,
             hipStream_t         stream) const override
    {
        BinaryFunction reduce_op{};
        const auto     random_range = limit_random_range<T>(0, 1000);
        std::vector<T> input
            = get_random_data<T>(size, random_range.first, random_range.second, seed.get_0());

        T * d_input;
        T * d_output;
        HIP_CHECK(hipMalloc(reinterpret_cast<void**>(&d_input), size * sizeof(T)));
        HIP_CHECK(hipMalloc(reinterpret_cast<void**>(&d_output), sizeof(T)));
        HIP_CHECK(
            hipMemcpy(
                d_input, input.data(),
                size * sizeof(T),
                hipMemcpyHostToDevice
                )
        );
        HIP_CHECK(hipDeviceSynchronize());

        // Allocate temporary storage memory
        size_t temp_storage_size_bytes;
        void * d_temp_storage = nullptr;
        // Get size of d_temp_storage
        HIP_CHECK(
            rocprim::reduce<Config>(
                d_temp_storage, temp_storage_size_bytes,
                d_input, d_output, T(), size,
                reduce_op, stream
                )
        );
        HIP_CHECK(hipMalloc(&d_temp_storage,temp_storage_size_bytes));
        HIP_CHECK(hipDeviceSynchronize());

        // Warm-up
        for(size_t i = 0; i < warmup_size; i++)
        {
            HIP_CHECK(
                rocprim::reduce<Config>(
                    d_temp_storage, temp_storage_size_bytes,
                    d_input, d_output, T(), size,
                    reduce_op, stream
                    )
            );
        }
        HIP_CHECK(hipDeviceSynchronize());

        // HIP events creation
        hipEvent_t start, stop;
        HIP_CHECK(hipEventCreate(&start));
        HIP_CHECK(hipEventCreate(&stop));

        for(auto _ : state)
        {
            // Record start event
            HIP_CHECK(hipEventRecord(start, stream));

            for(size_t i = 0; i < batch_size; i++)
            {
                HIP_CHECK(
                    rocprim::reduce<Config>(
                        d_temp_storage, temp_storage_size_bytes,
                        d_input, d_output, T(), size,
                        reduce_op, stream
                        )
                );
            }
            HIP_CHECK(hipStreamSynchronize(stream));

            // Record stop event and wait until it completes
            HIP_CHECK(hipEventRecord(stop, stream));
            HIP_CHECK(hipEventSynchronize(stop));

            float elapsed_mseconds;
            HIP_CHECK(hipEventElapsedTime(&elapsed_mseconds, start, stop));
            state.SetIterationTime(elapsed_mseconds / 1000);
        }

        // Destroy HIP events
        HIP_CHECK(hipEventDestroy(start));
        HIP_CHECK(hipEventDestroy(stop));

        state.SetBytesProcessed(state.iterations() * batch_size * size * sizeof(T));
        state.SetItemsProcessed(state.iterations() * batch_size * size);

        HIP_CHECK(hipFree(d_input));
        HIP_CHECK(hipFree(d_output));
        HIP_CHECK(hipFree(d_temp_storage));
    }
};

#endif