Hendrik/xgboost
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Revert hist init optimization. (#4502)

Browse Source
This commit is contained in:
Jiaming Yuan 2019-05-26 08:57:41 +08:00 committed by GitHub
parent 8ddd2715ee
commit 55e645c5f5
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 26 additions and 77 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

103
src/common/hist_util.cc
View File

@ -50,7 +50,7 @@ void HistCutMatrix::Init(DMatrix* p_fmat, uint32_t max_num_bins) {
constexpr int kFactor = 8; constexpr int kFactor = 8;
std::vector<WXQSketch> sketchs; std::vector<WXQSketch> sketchs;
const size_t nthread = omp_get_max_threads(); const int nthread = omp_get_max_threads();
unsigned const nstep = unsigned const nstep =
static_cast<unsigned>((info.num_col_ + nthread - 1) / nthread); static_cast<unsigned>((info.num_col_ + nthread - 1) / nthread);
@ -68,85 +68,34 @@ void HistCutMatrix::Init(DMatrix* p_fmat, uint32_t max_num_bins) {
// Use group index for weights? // Use group index for weights?
bool const use_group_ind = num_groups != 0 && weights.size() != info.num_row_; bool const use_group_ind = num_groups != 0 && weights.size() != info.num_row_;
if (use_group_ind) { for (const auto &batch : p_fmat->GetRowBatches()) {
for (const auto &batch : p_fmat->GetRowBatches()) { size_t group_ind = 0;
size_t group_ind = this->SearchGroupIndFromBaseRow(group_ptr, batch.base_rowid); if (use_group_ind) {
#pragma omp parallel num_threads(nthread) firstprivate(group_ind, use_group_ind) group_ind = this->SearchGroupIndFromBaseRow(group_ptr, batch.base_rowid);
{
CHECK_EQ(nthread, omp_get_num_threads());
auto tid = static_cast<unsigned>(omp_get_thread_num());
unsigned begin = std::min(nstep * tid, ncol);
unsigned end = std::min(nstep * (tid + 1), ncol);
// do not iterate if no columns are assigned to the thread
if (begin < end && end <= ncol) {
for (size_t i = 0; i < batch.Size(); ++i) { // NOLINT(*)
size_t const ridx = batch.base_rowid + i;
SparsePage::Inst const inst = batch[i];
if (group_ptr[group_ind] == ridx &&
// maximum equals to weights.size() - 1
group_ind < num_groups - 1) {
// move to next group
group_ind++;
}
for (auto const& entry : inst) {
if (entry.index >= begin && entry.index < end) {
size_t w_idx = group_ind;
sketchs[entry.index].Push(entry.fvalue, info.GetWeight(w_idx));
}
}
}
}
}
} }
} else { #pragma omp parallel num_threads(nthread) firstprivate(group_ind, use_group_ind)
for (const auto &batch : p_fmat->GetRowBatches()) { {
const size_t size = batch.Size(); CHECK_EQ(nthread, omp_get_num_threads());
const size_t block_size = 512; auto tid = static_cast<unsigned>(omp_get_thread_num());
const size_t block_size_iter = block_size * nthread; unsigned begin = std::min(nstep * tid, ncol);
const size_t n_blocks = size / block_size_iter + !!(size % block_size_iter); unsigned end = std::min(nstep * (tid + 1), ncol);
std::vector<std::vector<std::pair<float, float>>> buff(nthread); // do not iterate if no columns are assigned to the thread
for (size_t tid = 0; tid < nthread; ++tid) { if (begin < end && end <= ncol) {
buff[tid].resize(block_size * ncol); for (size_t i = 0; i < batch.Size(); ++i) { // NOLINT(*)
} size_t const ridx = batch.base_rowid + i;
SparsePage::Inst const inst = batch[i];
std::vector<size_t> sizes(nthread * ncol, 0); if (use_group_ind &&
group_ptr[group_ind] == ridx &&
for (size_t iblock = 0; iblock < n_blocks; ++iblock) { // maximum equals to weights.size() - 1
#pragma omp parallel num_threads(nthread) group_ind < num_groups - 1) {
{ // move to next group
int tid = omp_get_thread_num(); group_ind++;
const size_t ibegin = iblock * block_size_iter + tid * block_size;
const size_t iend = std::min(ibegin + block_size, size);
auto* p_sizes = sizes.data() + ncol * tid;
auto* p_buff = buff[tid].data();
for (size_t i = ibegin; i < iend; ++i) {
size_t const ridx = batch.base_rowid + i;
bst_float w = info.GetWeight(ridx);
SparsePage::Inst const inst = batch[i];
for (auto const& entry : inst) {
const size_t idx = entry.index;
p_buff[idx * block_size + p_sizes[idx]] = { entry.fvalue, w };
p_sizes[idx]++;
}
} }
#pragma omp barrier for (auto const& entry : inst) {
#pragma omp for schedule(static) if (entry.index >= begin && entry.index < end) {
for (int32_t icol = 0; icol < static_cast<int32_t>(ncol); ++icol) { size_t w_idx = use_group_ind ? group_ind : ridx;
for (size_t tid = 0; tid < nthread; ++tid) { sketchs[entry.index].Push(entry.fvalue, info.GetWeight(w_idx));
auto* p_sizes = sizes.data() + ncol * tid;
auto* p_buff = buff[tid].data() + icol * block_size;
for (size_t i = 0; i < p_sizes[icol]; ++i) {
sketchs[icol].Push(p_buff[i].first, p_buff[i].second);
}
p_sizes[icol] = 0;
} }
} }
} }