wehub-resource-sync
99 lines
3.8 KiB
C++
99 lines
3.8 KiB
C++
/*!
|
|
* Copyright (c) 2021-2026 Microsoft Corporation. All rights reserved.
|
|
* Copyright (c) 2026-2026 The LightGBM developers. All rights reserved.
|
|
* Licensed under the MIT License. See LICENSE file in the project root for
|
|
* license information.
|
|
*/
|
|
|
|
#ifndef LIGHTGBM_INCLUDE_LIGHTGBM_CUDA_CUDA_OBJECTIVE_FUNCTION_HPP_
|
|
#define LIGHTGBM_INCLUDE_LIGHTGBM_CUDA_CUDA_OBJECTIVE_FUNCTION_HPP_
|
|
|
|
#ifdef USE_CUDA
|
|
|
|
#include <LightGBM/cuda/cuda_utils.hu>
|
|
#include <LightGBM/objective_function.h>
|
|
#include <LightGBM/meta.h>
|
|
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
namespace LightGBM {
|
|
|
|
template <typename HOST_OBJECTIVE>
|
|
class CUDAObjectiveInterface: public HOST_OBJECTIVE, public NCCLInfo {
|
|
public:
|
|
explicit CUDAObjectiveInterface(const Config& config): HOST_OBJECTIVE(config) {
|
|
if (config.num_gpu <= 1) {
|
|
const int gpu_device_id = config.gpu_device_id >= 0 ? config.gpu_device_id : 0;
|
|
SetCUDADevice(gpu_device_id, __FILE__, __LINE__);
|
|
}
|
|
}
|
|
|
|
explicit CUDAObjectiveInterface(const std::vector<std::string>& strs): HOST_OBJECTIVE(strs) {}
|
|
|
|
void Init(const Metadata& metadata, data_size_t num_data) {
|
|
HOST_OBJECTIVE::Init(metadata, num_data);
|
|
cuda_labels_ = metadata.cuda_metadata()->cuda_label();
|
|
cuda_weights_ = metadata.cuda_metadata()->cuda_weights();
|
|
}
|
|
|
|
void SetNCCLInfo(
|
|
ncclComm_t nccl_communicator,
|
|
int nccl_gpu_rank,
|
|
int local_gpu_rank,
|
|
int gpu_device_id,
|
|
data_size_t global_num_data) override {
|
|
NCCLInfo::SetNCCLInfo(nccl_communicator, nccl_gpu_rank, local_gpu_rank, gpu_device_id, global_num_data);
|
|
}
|
|
|
|
virtual const double* ConvertOutputCUDA(const data_size_t num_data, const double* input, double* output) const {
|
|
return LaunchConvertOutputCUDAKernel(num_data, input, output);
|
|
}
|
|
|
|
double BoostFromScore(int class_id) const override {
|
|
return LaunchCalcInitScoreKernel(class_id);
|
|
}
|
|
|
|
bool IsCUDAObjective() const override { return true; }
|
|
|
|
void GetGradients(const double* scores, score_t* gradients, score_t* hessians) const override {
|
|
LaunchGetGradientsKernel(scores, gradients, hessians);
|
|
SynchronizeCUDADevice(__FILE__, __LINE__);
|
|
}
|
|
|
|
void GetGradientsWithSampledQueries(const double* scores, const data_size_t /*num_sampled_queries*/, const data_size_t* /*sampled_query_indices*/, score_t* gradients, score_t* hessians) const override {
|
|
LaunchGetGradientsKernel(scores, gradients, hessians);
|
|
SynchronizeCUDADevice(__FILE__, __LINE__);
|
|
}
|
|
|
|
void RenewTreeOutputCUDA(const double* score, const data_size_t* data_indices_in_leaf, const data_size_t* num_data_in_leaf,
|
|
const data_size_t* data_start_in_leaf, const int num_leaves, double* leaf_value) const override {
|
|
global_timer.Start("CUDAObjectiveInterface::LaunchRenewTreeOutputCUDAKernel");
|
|
LaunchRenewTreeOutputCUDAKernel(score, data_indices_in_leaf, num_data_in_leaf, data_start_in_leaf, num_leaves, leaf_value);
|
|
SynchronizeCUDADevice(__FILE__, __LINE__);
|
|
global_timer.Stop("CUDAObjectiveInterface::LaunchRenewTreeOutputCUDAKernel");
|
|
}
|
|
|
|
protected:
|
|
virtual void LaunchGetGradientsKernel(const double* scores, score_t* gradients, score_t* hessians) const = 0;
|
|
|
|
virtual double LaunchCalcInitScoreKernel(const int class_id) const {
|
|
return HOST_OBJECTIVE::BoostFromScore(class_id);
|
|
}
|
|
|
|
virtual const double* LaunchConvertOutputCUDAKernel(const data_size_t /*num_data*/, const double* input, double* /*output*/) const { return input; }
|
|
|
|
virtual void LaunchRenewTreeOutputCUDAKernel(
|
|
const double* /*score*/, const data_size_t* /*data_indices_in_leaf*/, const data_size_t* /*num_data_in_leaf*/,
|
|
const data_size_t* /*data_start_in_leaf*/, const int /*num_leaves*/, double* /*leaf_value*/) const {}
|
|
|
|
const label_t* cuda_labels_;
|
|
const label_t* cuda_weights_;
|
|
};
|
|
|
|
} // namespace LightGBM
|
|
|
|
#endif // USE_CUDA
|
|
|
|
#endif // LIGHTGBM_INCLUDE_LIGHTGBM_CUDA_CUDA_OBJECTIVE_FUNCTION_HPP_
|