Files
ceres-solver/internal/ceres/parallel_for.h
T
Dmitriy Korchemkin dc7a859752 Single-threaded operations on small vectors
As pointed out by several users, introduction of parallel operations on
vectors severely impacts solver performance on small problems, with time
consumption increasing with the number of threads.

In order to minimize overhead of trying to execute small tasks using a
large number of threads, task scheduling mechanism was changed to avoid
scheduling all tasks at once.

However, there is still a large difference in exectuion time because the
main thread always launches the next thread before starting doing the
work. This leads to several orders of magnitude slowdown when going from
a single-threaded execution (which follows a fast-forward path to a
single loop over all indices, without any synchronization involved)
to a two-thread execution:

/bin/parallel_vector_operations_benchmark
-------------------------------------------
Benchmark                              Time
-------------------------------------------
SetZero/128                         12.8 ns
SetZeroParallel/128/1               16.6 ns
SetZeroParallel/128/2               2211 ns

In order to eliminate this effect, we limit the block-size of parallel
execution of vector operations to 2^16 elements (thus, starting parallel
execution only for vectors of at least 2^17 elements).

Threshold of 2^16 elements was choosen by evaluating thresholds from
2^10 to 2^20 (only powers of 2), with 2^14..2^20 significantly reducing
worst-case runtime degradation.

Details can be found in discussion of the issue at
https://github.com/ceres-solver/ceres-solver/issues/1016

Change-Id: I555c882d63ee53323ceb426743b970f989b65503
2023-10-09 13:55:30 +00:00

187 lines
7.8 KiB
C++

// Ceres Solver - A fast non-linear least squares minimizer
// Copyright 2023 Google Inc. All rights reserved.
// http://ceres-solver.org/
//
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are met:
//
// * Redistributions of source code must retain the above copyright notice,
// this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above copyright notice,
// this list of conditions and the following disclaimer in the documentation
// and/or other materials provided with the distribution.
// * Neither the name of Google Inc. nor the names of its contributors may be
// used to endorse or promote products derived from this software without
// specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
// POSSIBILITY OF SUCH DAMAGE.
//
// Authors: vitus@google.com (Michael Vitus),
// dmitriy.korchemkin@gmail.com (Dmitriy Korchemkin)
#ifndef CERES_INTERNAL_PARALLEL_FOR_H_
#define CERES_INTERNAL_PARALLEL_FOR_H_
#include <mutex>
#include <vector>
#include "ceres/context_impl.h"
#include "ceres/internal/eigen.h"
#include "ceres/internal/export.h"
#include "ceres/parallel_invoke.h"
#include "ceres/partition_range_for_parallel_for.h"
#include "glog/logging.h"
namespace ceres::internal {
// Use a dummy mutex if num_threads = 1.
inline decltype(auto) MakeConditionalLock(const int num_threads,
std::mutex& m) {
return (num_threads == 1) ? std::unique_lock<std::mutex>{}
: std::unique_lock<std::mutex>{m};
}
// Execute the function for every element in the range [start, end) with at most
// num_threads. It will execute all the work on the calling thread if
// num_threads or (end - start) is equal to 1.
// Depending on function signature, it will be supplied with either loop index
// or a range of loop indicies; function can also be supplied with thread_id.
// The following function signatures are supported:
// - Functions accepting a single loop index:
// - [](int index) { ... }
// - [](int thread_id, int index) { ... }
// - Functions accepting a range of loop index:
// - [](std::tuple<int, int> index) { ... }
// - [](int thread_id, std::tuple<int, int> index) { ... }
//
// When distributing workload between threads, it is assumed that each loop
// iteration takes approximately equal time to complete.
template <typename F>
void ParallelFor(ContextImpl* context,
int start,
int end,
int num_threads,
F&& function,
int min_block_size = 1) {
CHECK_GT(num_threads, 0);
if (start >= end) {
return;
}
if (num_threads == 1 || end - start < min_block_size * 2) {
InvokeOnSegment(0, std::make_tuple(start, end), std::forward<F>(function));
return;
}
CHECK(context != nullptr);
ParallelInvoke(context,
start,
end,
num_threads,
std::forward<F>(function),
min_block_size);
}
// Execute function for every element in the range [start, end) with at most
// num_threads, using user-provided partitions array.
// When distributing workload between threads, it is assumed that each segment
// bounded by adjacent elements of partitions array takes approximately equal
// time to process.
template <typename F>
void ParallelFor(ContextImpl* context,
int start,
int end,
int num_threads,
F&& function,
const std::vector<int>& partitions) {
CHECK_GT(num_threads, 0);
if (start >= end) {
return;
}
CHECK_EQ(partitions.front(), start);
CHECK_EQ(partitions.back(), end);
if (num_threads == 1 || end - start <= num_threads) {
ParallelFor(context, start, end, num_threads, std::forward<F>(function));
return;
}
CHECK_GT(partitions.size(), 1);
const int num_partitions = partitions.size() - 1;
ParallelFor(context,
0,
num_partitions,
num_threads,
[&function, &partitions](int thread_id,
std::tuple<int, int> partition_ids) {
// partition_ids is a range of partition indices
const auto [partition_start, partition_end] = partition_ids;
// Execution over several adjacent segments is equivalent
// to execution over union of those segments (which is also a
// contiguous segment)
const int range_start = partitions[partition_start];
const int range_end = partitions[partition_end];
// Range of original loop indices
const auto range = std::make_tuple(range_start, range_end);
InvokeOnSegment(thread_id, range, function);
});
}
// Execute function for every element in the range [start, end) with at most
// num_threads, taking into account user-provided integer cumulative costs of
// iterations. Cumulative costs of iteration for indices in range [0, end) are
// stored in objects from cumulative_cost_data. User-provided
// cumulative_cost_fun returns non-decreasing integer values corresponding to
// inclusive cumulative cost of loop iterations, provided with a reference to
// user-defined object. Only indices from [start, end) will be referenced. This
// routine assumes that cumulative_cost_fun is non-decreasing (in other words,
// all costs are non-negative);
// When distributing workload between threads, input range of loop indices will
// be partitioned into disjoint contiguous intervals, with the maximal cost
// being minimized.
// For example, with iteration costs of [1, 1, 5, 3, 1, 4] cumulative_cost_fun
// should return [1, 2, 7, 10, 11, 15], and with num_threads = 4 this range
// will be split into segments [0, 2) [2, 3) [3, 5) [5, 6) with costs
// [2, 5, 4, 4].
template <typename F, typename CumulativeCostData, typename CumulativeCostFun>
void ParallelFor(ContextImpl* context,
int start,
int end,
int num_threads,
F&& function,
const CumulativeCostData* cumulative_cost_data,
CumulativeCostFun&& cumulative_cost_fun) {
CHECK_GT(num_threads, 0);
if (start >= end) {
return;
}
if (num_threads == 1 || end - start <= num_threads) {
ParallelFor(context, start, end, num_threads, std::forward<F>(function));
return;
}
// Creating several partitions allows us to tolerate imperfections of
// partitioning and user-supplied iteration costs up to a certain extent
constexpr int kNumPartitionsPerThread = 4;
const int kMaxPartitions = num_threads * kNumPartitionsPerThread;
const auto& partitions = PartitionRangeForParallelFor(
start,
end,
kMaxPartitions,
cumulative_cost_data,
std::forward<CumulativeCostFun>(cumulative_cost_fun));
CHECK_GT(partitions.size(), 1);
ParallelFor(
context, start, end, num_threads, std::forward<F>(function), partitions);
}
} // namespace ceres::internal
#endif // CERES_INTERNAL_PARALLEL_FOR_H_