Перешёл к произвольным периодам агрегации
This commit is contained in:
@@ -1,4 +1,5 @@
|
||||
#include "gpu_loader.hpp"
|
||||
#include "utils.hpp"
|
||||
#include <dlfcn.h>
|
||||
#include <map>
|
||||
#include <algorithm>
|
||||
@@ -29,58 +30,54 @@ bool gpu_is_available() {
|
||||
return false;
|
||||
}
|
||||
|
||||
gpu_aggregate_days_fn load_gpu_aggregate_days() {
|
||||
gpu_aggregate_periods_fn load_gpu_aggregate_periods() {
|
||||
void* h = get_gpu_lib_handle();
|
||||
if (!h) return nullptr;
|
||||
|
||||
auto fn = (gpu_aggregate_days_fn)dlsym(h, "gpu_aggregate_days");
|
||||
auto fn = (gpu_aggregate_periods_fn)dlsym(h, "gpu_aggregate_periods");
|
||||
return fn;
|
||||
}
|
||||
|
||||
bool aggregate_days_gpu(
|
||||
bool aggregate_periods_gpu(
|
||||
const std::vector<Record>& records,
|
||||
std::vector<DayStats>& out_stats,
|
||||
gpu_aggregate_days_fn gpu_fn)
|
||||
std::vector<PeriodStats>& out_stats,
|
||||
gpu_aggregate_periods_fn gpu_fn)
|
||||
{
|
||||
if (!gpu_fn || records.empty()) {
|
||||
return false;
|
||||
}
|
||||
|
||||
// Общий таймер всей функции
|
||||
int64_t interval = get_aggregation_interval();
|
||||
|
||||
double t_total_start = omp_get_wtime();
|
||||
|
||||
// Таймер CPU preprocessing
|
||||
double t_preprocess_start = omp_get_wtime();
|
||||
|
||||
// Группируем записи по дням и подготавливаем данные для GPU
|
||||
std::map<DayIndex, std::vector<size_t>> day_record_indices;
|
||||
std::map<PeriodIndex, std::vector<size_t>> period_record_indices;
|
||||
|
||||
for (size_t i = 0; i < records.size(); i++) {
|
||||
DayIndex day = static_cast<DayIndex>(records[i].timestamp) / 86400;
|
||||
day_record_indices[day].push_back(i);
|
||||
PeriodIndex period = static_cast<PeriodIndex>(records[i].timestamp) / interval;
|
||||
period_record_indices[period].push_back(i);
|
||||
}
|
||||
|
||||
int num_days = static_cast<int>(day_record_indices.size());
|
||||
int num_periods = static_cast<int>(period_record_indices.size());
|
||||
|
||||
// Подготавливаем массивы для GPU
|
||||
std::vector<GpuRecord> gpu_records;
|
||||
std::vector<int> day_offsets;
|
||||
std::vector<int> day_counts;
|
||||
std::vector<long long> day_indices;
|
||||
std::vector<int> period_offsets;
|
||||
std::vector<int> period_counts;
|
||||
std::vector<long long> period_indices;
|
||||
|
||||
gpu_records.reserve(records.size());
|
||||
day_offsets.reserve(num_days);
|
||||
day_counts.reserve(num_days);
|
||||
day_indices.reserve(num_days);
|
||||
period_offsets.reserve(num_periods);
|
||||
period_counts.reserve(num_periods);
|
||||
period_indices.reserve(num_periods);
|
||||
|
||||
int current_offset = 0;
|
||||
|
||||
for (auto& [day, indices] : day_record_indices) {
|
||||
day_indices.push_back(day);
|
||||
day_offsets.push_back(current_offset);
|
||||
day_counts.push_back(static_cast<int>(indices.size()));
|
||||
for (auto& [period, indices] : period_record_indices) {
|
||||
period_indices.push_back(period);
|
||||
period_offsets.push_back(current_offset);
|
||||
period_counts.push_back(static_cast<int>(indices.size()));
|
||||
|
||||
// Добавляем записи этого дня
|
||||
for (size_t idx : indices) {
|
||||
const auto& r = records[idx];
|
||||
GpuRecord gr;
|
||||
@@ -96,22 +93,19 @@ bool aggregate_days_gpu(
|
||||
current_offset += static_cast<int>(indices.size());
|
||||
}
|
||||
|
||||
// Выделяем память для результата
|
||||
std::vector<GpuDayStats> gpu_stats(num_days);
|
||||
std::vector<GpuPeriodStats> gpu_stats(num_periods);
|
||||
|
||||
double t_preprocess_ms = (omp_get_wtime() - t_preprocess_start) * 1000.0;
|
||||
std::cout << " GPU CPU preprocessing: " << std::fixed << std::setprecision(3)
|
||||
<< std::setw(7) << t_preprocess_ms << " ms" << std::endl << std::flush;
|
||||
|
||||
// Вызываем GPU функцию (включает: malloc, memcpy H->D, kernel, memcpy D->H, free)
|
||||
// Детальные тайминги выводятся внутри GPU функции
|
||||
int result = gpu_fn(
|
||||
gpu_records.data(),
|
||||
static_cast<int>(gpu_records.size()),
|
||||
day_offsets.data(),
|
||||
day_counts.data(),
|
||||
day_indices.data(),
|
||||
num_days,
|
||||
period_offsets.data(),
|
||||
period_counts.data(),
|
||||
period_indices.data(),
|
||||
num_periods,
|
||||
gpu_stats.data()
|
||||
);
|
||||
|
||||
@@ -120,23 +114,21 @@ bool aggregate_days_gpu(
|
||||
return false;
|
||||
}
|
||||
|
||||
// Конвертируем результат в DayStats
|
||||
out_stats.clear();
|
||||
out_stats.reserve(num_days);
|
||||
out_stats.reserve(num_periods);
|
||||
|
||||
for (const auto& gs : gpu_stats) {
|
||||
DayStats ds;
|
||||
ds.day = gs.day;
|
||||
ds.avg = gs.avg;
|
||||
ds.open_min = gs.open_min;
|
||||
ds.open_max = gs.open_max;
|
||||
ds.close_min = gs.close_min;
|
||||
ds.close_max = gs.close_max;
|
||||
ds.count = gs.count;
|
||||
out_stats.push_back(ds);
|
||||
PeriodStats ps;
|
||||
ps.period = gs.period;
|
||||
ps.avg = gs.avg;
|
||||
ps.open_min = gs.open_min;
|
||||
ps.open_max = gs.open_max;
|
||||
ps.close_min = gs.close_min;
|
||||
ps.close_max = gs.close_max;
|
||||
ps.count = gs.count;
|
||||
out_stats.push_back(ps);
|
||||
}
|
||||
|
||||
// Общее время всей GPU функции (включая preprocessing)
|
||||
double t_total_ms = (omp_get_wtime() - t_total_start) * 1000.0;
|
||||
std::cout << " GPU TOTAL (with prep): " << std::fixed << std::setprecision(3)
|
||||
<< std::setw(7) << t_total_ms << " ms" << std::endl << std::flush;
|
||||
|
||||
Reference in New Issue
Block a user