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

Span class. (#3548)

Browse Source 
* Add basic Span class based on ISO++20.

* Use Span<Entry const> instead of Inst in SparsePage.

* Add DeviceSpan in HostDeviceVector, use it in regression obj.
This commit is contained in:
trivialfis
2018-08-14 13:58:11 +08:00
committed by Rory Mitchell
parent 2b7a1c5780
commit 2c502784ff
28 changed files with 1927 additions and 138 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/tree/updater_basemaker-inl.h
View File

@@ -49,9 +49,9 @@ class BaseMaker: public TreeUpdater {
auto &batch = iter->Value();
for (bst_uint fid = 0; fid < batch.Size(); ++fid) {
auto c = batch[fid];
if (c.length != 0) {
if (c.size() != 0) {
fminmax_[fid * 2 + 0] = std::max(-c[0].fvalue, fminmax_[fid * 2 + 0]);
fminmax_[fid * 2 + 1] = std::max(c[c.length - 1].fvalue, fminmax_[fid * 2 + 1]);
fminmax_[fid * 2 + 1] = std::max(c[c.size() - 1].fvalue, fminmax_[fid * 2 + 1]);
}
}
}
@@ -106,9 +106,9 @@ class BaseMaker: public TreeUpdater {
inline static int NextLevel(const SparsePage::Inst &inst, const RegTree &tree, int nid) {
const RegTree::Node &n = tree[nid];
bst_uint findex = n.SplitIndex();
for (unsigned i = 0; i < inst.length; ++i) {
if (findex == inst[i].index) {
if (inst[i].fvalue < n.SplitCond()) {
for (const auto& ins : inst) {
if (findex == ins.index) {
if (ins.fvalue < n.SplitCond()) {
return n.LeftChild();
} else {
return n.RightChild();
@@ -250,7 +250,7 @@ class BaseMaker: public TreeUpdater {
auto it = std::lower_bound(sorted_split_set.begin(), sorted_split_set.end(), fid);
if (it != sorted_split_set.end() && *it == fid) {
const auto ndata = static_cast<bst_omp_uint>(col.length);
const auto ndata = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_uint ridx = col[j].index;
@@ -308,7 +308,7 @@ class BaseMaker: public TreeUpdater {
auto &batch = iter->Value();
for (auto fid : fsplits) {
auto col = batch[fid];
const auto ndata = static_cast<bst_omp_uint>(col.length);
const auto ndata = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_uint ridx = col[j].index;

20
src/tree/updater_colmaker.cc
View File

@@ -269,7 +269,7 @@ class ColMaker: public TreeUpdater {
const std::vector<GradientPair> &gpair) {
// TODO(tqchen): double check stats order.
const MetaInfo& info = fmat.Info();
const bool ind = col.length != 0 && col.data[0].fvalue == col.data[col.length - 1].fvalue;
const bool ind = col.size() != 0 && col[0].fvalue == col[col.size() - 1].fvalue;
bool need_forward = param_.NeedForwardSearch(fmat.GetColDensity(fid), ind);
bool need_backward = param_.NeedBackwardSearch(fmat.GetColDensity(fid), ind);
const std::vector<int> &qexpand = qexpand_;
@@ -281,8 +281,8 @@ class ColMaker: public TreeUpdater {
for (int j : qexpand) {
temp[j].stats.Clear();
}
bst_uint step = (col.length + this->nthread_ - 1) / this->nthread_;
bst_uint end = std::min(col.length, step * (tid + 1));
bst_uint step = (col.size() + this->nthread_ - 1) / this->nthread_;
bst_uint end = std::min(static_cast<bst_uint>(col.size()), step * (tid + 1));
for (bst_uint i = tid * step; i < end; ++i) {
const bst_uint ridx = col[i].index;
const int nid = position_[ridx];
@@ -363,8 +363,8 @@ class ColMaker: public TreeUpdater {
GradStats c(param_), cright(param_);
const int tid = omp_get_thread_num();
std::vector<ThreadEntry> &temp = stemp_[tid];
bst_uint step = (col.length + this->nthread_ - 1) / this->nthread_;
bst_uint end = std::min(col.length, step * (tid + 1));
bst_uint step = (col.size() + this->nthread_ - 1) / this->nthread_;
bst_uint end = std::min(static_cast<bst_uint>(col.size()), step * (tid + 1));
for (bst_uint i = tid * step; i < end; ++i) {
const bst_uint ridx = col[i].index;
const int nid = position_[ridx];
@@ -620,13 +620,13 @@ class ColMaker: public TreeUpdater {
int fid = feat_set[i];
const int tid = omp_get_thread_num();
auto c = batch[fid];
const bool ind = c.length != 0 && c.data[0].fvalue == c.data[c.length - 1].fvalue;
const bool ind = c.size() != 0 && c[0].fvalue == c[c.size() - 1].fvalue;
if (param_.NeedForwardSearch(fmat.GetColDensity(fid), ind)) {
this->EnumerateSplit(c.data, c.data + c.length, +1,
this->EnumerateSplit(c.data(), c.data() + c.size(), +1,
fid, gpair, info, stemp_[tid]);
}
if (param_.NeedBackwardSearch(fmat.GetColDensity(fid), ind)) {
this->EnumerateSplit(c.data + c.length - 1, c.data - 1, -1,
this->EnumerateSplit(c.data() + c.size() - 1, c.data() - 1, -1,
fid, gpair, info, stemp_[tid]);
}
}
@@ -734,7 +734,7 @@ class ColMaker: public TreeUpdater {
auto &batch = iter->Value();
for (auto fid : fsplits) {
auto col = batch[fid];
const auto ndata = static_cast<bst_omp_uint>(col.length);
const auto ndata = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_uint ridx = col[j].index;
@@ -865,7 +865,7 @@ class DistColMaker : public ColMaker {
auto &batch = iter->Value();
for (auto fid : fsplits) {
auto col = batch[fid];
const auto ndata = static_cast<bst_omp_uint>(col.length);
const auto ndata = static_cast<bst_omp_uint>(col.size());
#pragma omp parallel for schedule(static)
for (bst_omp_uint j = 0; j < ndata; ++j) {
const bst_uint ridx = col[j].index;

2
src/tree/updater_gpu.cu
View File

@@ -669,7 +669,7 @@ class GPUMaker : public TreeUpdater {
auto &batch = iter->Value();
for (int i = 0; i < batch.Size(); i++) {
auto col = batch[i];
for (const Entry* it = col.data; it != col.data + col.length;
for (const Entry* it = col.data(); it != col.data() + col.size();
it++) {
int inst_id = static_cast<int>(it->index);
fval->push_back(it->fvalue);

64
src/tree/updater_histmaker.cc
View File

@@ -496,13 +496,13 @@ class CQHistMaker: public HistMaker<TStats> {
}
inline void UpdateHistCol(const std::vector<GradientPair> &gpair,
const SparsePage::Inst &c,
const SparsePage::Inst &col,
const MetaInfo &info,
const RegTree &tree,
const std::vector<bst_uint> &fset,
bst_uint fid_offset,
std::vector<HistEntry> *p_temp) {
if (c.length == 0) return;
if (col.size() == 0) return;
// initialize sbuilder for use
std::vector<HistEntry> &hbuilder = *p_temp;
hbuilder.resize(tree.param.num_nodes);
@@ -514,46 +514,46 @@ class CQHistMaker: public HistMaker<TStats> {
}
if (TStats::kSimpleStats != 0 && this->param_.cache_opt != 0) {
constexpr bst_uint kBuffer = 32;
bst_uint align_length = c.length / kBuffer * kBuffer;
bst_uint align_length = col.size() / kBuffer * kBuffer;
int buf_position[kBuffer];
GradientPair buf_gpair[kBuffer];
for (bst_uint j = 0; j < align_length; j += kBuffer) {
for (bst_uint i = 0; i < kBuffer; ++i) {
bst_uint ridx = c[j + i].index;
bst_uint ridx = col[j + i].index;
buf_position[i] = this->position_[ridx];
buf_gpair[i] = gpair[ridx];
}
for (bst_uint i = 0; i < kBuffer; ++i) {
const int nid = buf_position[i];
if (nid >= 0) {
hbuilder[nid].Add(c[j + i].fvalue, buf_gpair[i]);
hbuilder[nid].Add(col[j + i].fvalue, buf_gpair[i]);
}
}
}
for (bst_uint j = align_length; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (bst_uint j = align_length; j < col.size(); ++j) {
const bst_uint ridx = col[j].index;
const int nid = this->position_[ridx];
if (nid >= 0) {
hbuilder[nid].Add(c[j].fvalue, gpair[ridx]);
hbuilder[nid].Add(col[j].fvalue, gpair[ridx]);
}
}
} else {
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (const auto& c : col) {
const bst_uint ridx = c.index;
const int nid = this->position_[ridx];
if (nid >= 0) {
hbuilder[nid].Add(c[j].fvalue, gpair, info, ridx);
hbuilder[nid].Add(c.fvalue, gpair, info, ridx);
}
}
}
}
inline void UpdateSketchCol(const std::vector<GradientPair> &gpair,
const SparsePage::Inst &c,
const SparsePage::Inst &col,
const RegTree &tree,
size_t work_set_size,
bst_uint offset,
std::vector<BaseMaker::SketchEntry> *p_temp) {
if (c.length == 0) return;
if (col.size() == 0) return;
// initialize sbuilder for use
std::vector<BaseMaker::SketchEntry> &sbuilder = *p_temp;
sbuilder.resize(tree.param.num_nodes);
@@ -565,18 +565,18 @@ class CQHistMaker: public HistMaker<TStats> {
}
// first pass, get sum of weight, TODO, optimization to skip first pass
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (const auto& c : col) {
const bst_uint ridx = c.index;
const int nid = this->position_[ridx];
if (nid >= 0) {
sbuilder[nid].sum_total += gpair[ridx].GetHess();
sbuilder[nid].sum_total += gpair[ridx].GetHess();
}
}
// if only one value, no need to do second pass
if (c[0].fvalue == c[c.length-1].fvalue) {
if (col[0].fvalue == col[col.size()-1].fvalue) {
for (size_t i = 0; i < this->qexpand_.size(); ++i) {
const int nid = this->qexpand_[i];
sbuilder[nid].sketch->Push(c[0].fvalue, static_cast<bst_float>(sbuilder[nid].sum_total));
sbuilder[nid].sketch->Push(col[0].fvalue, static_cast<bst_float>(sbuilder[nid].sum_total));
}
return;
}
@@ -589,35 +589,35 @@ class CQHistMaker: public HistMaker<TStats> {
// second pass, build the sketch
if (TStats::kSimpleStats != 0 && this->param_.cache_opt != 0) {
constexpr bst_uint kBuffer = 32;
bst_uint align_length = c.length / kBuffer * kBuffer;
bst_uint align_length = col.size() / kBuffer * kBuffer;
int buf_position[kBuffer];
bst_float buf_hess[kBuffer];
for (bst_uint j = 0; j < align_length; j += kBuffer) {
for (bst_uint i = 0; i < kBuffer; ++i) {
bst_uint ridx = c[j + i].index;
bst_uint ridx = col[j + i].index;
buf_position[i] = this->position_[ridx];
buf_hess[i] = gpair[ridx].GetHess();
}
for (bst_uint i = 0; i < kBuffer; ++i) {
const int nid = buf_position[i];
if (nid >= 0) {
sbuilder[nid].Push(c[j + i].fvalue, buf_hess[i], max_size);
sbuilder[nid].Push(col[j + i].fvalue, buf_hess[i], max_size);
}
}
}
for (bst_uint j = align_length; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (bst_uint j = align_length; j < col.size(); ++j) {
const bst_uint ridx = col[j].index;
const int nid = this->position_[ridx];
if (nid >= 0) {
sbuilder[nid].Push(c[j].fvalue, gpair[ridx].GetHess(), max_size);
sbuilder[nid].Push(col[j].fvalue, gpair[ridx].GetHess(), max_size);
}
}
} else {
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (const auto& c : col) {
const bst_uint ridx = c.index;
const int nid = this->position_[ridx];
if (nid >= 0) {
sbuilder[nid].Push(c[j].fvalue, gpair[ridx].GetHess(), max_size);
sbuilder[nid].Push(c.fvalue, gpair[ridx].GetHess(), max_size);
}
}
}
@@ -794,8 +794,8 @@ class QuantileHistMaker: public HistMaker<TStats> {
if (this->node2workindex_[nid] < 0) {
this->position_[ridx] = ~nid;
} else {
for (bst_uint j = 0; j < inst.length; ++j) {
builder.AddBudget(inst[j].index, omp_get_thread_num());
for (auto& ins : inst) {
builder.AddBudget(ins.index, omp_get_thread_num());
}
}
}
@@ -807,9 +807,9 @@ class QuantileHistMaker: public HistMaker<TStats> {
const bst_uint ridx = static_cast<bst_uint>(batch.base_rowid + i);
const int nid = this->position_[ridx];
if (nid >= 0) {
for (bst_uint j = 0; j < inst.length; ++j) {
builder.Push(inst[j].index,
Entry(nid, inst[j].fvalue),
for (auto& ins : inst) {
builder.Push(ins.index,
Entry(nid, ins.fvalue),
omp_get_thread_num());
}
}

26
src/tree/updater_skmaker.cc
View File

@@ -155,7 +155,7 @@ class SketchMaker: public BaseMaker {
this->UpdateSketchCol(gpair, batch[fidx], tree,
node_stats_,
fidx,
batch[fidx].length == nrows,
batch[fidx].size() == nrows,
&thread_sketch_[omp_get_thread_num()]);
}
}
@@ -174,13 +174,13 @@ class SketchMaker: public BaseMaker {
}
// update sketch information in column fid
inline void UpdateSketchCol(const std::vector<GradientPair> &gpair,
const SparsePage::Inst &c,
const SparsePage::Inst &col,
const RegTree &tree,
const std::vector<SKStats> &nstats,
bst_uint fid,
bool col_full,
std::vector<SketchEntry> *p_temp) {
if (c.length == 0) return;
if (col.size() == 0) return;
// initialize sbuilder for use
std::vector<SketchEntry> &sbuilder = *p_temp;
sbuilder.resize(tree.param.num_nodes * 3);
@@ -192,10 +192,10 @@ class SketchMaker: public BaseMaker {
}
}
if (!col_full) {
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (const auto& c : col) {
const bst_uint ridx = c.index;
const int nid = this->position_[ridx];
if (nid >= 0) {
if (nid > 0) {
const GradientPair &e = gpair[ridx];
if (e.GetGrad() >= 0.0f) {
sbuilder[3 * nid + 0].sum_total += e.GetGrad();
@@ -213,10 +213,10 @@ class SketchMaker: public BaseMaker {
}
}
// if only one value, no need to do second pass
if (c[0].fvalue == c[c.length-1].fvalue) {
if (col[0].fvalue == col[col.size()-1].fvalue) {
for (int nid : this->qexpand_) {
for (int k = 0; k < 3; ++k) {
sbuilder[3 * nid + k].sketch->Push(c[0].fvalue,
sbuilder[3 * nid + k].sketch->Push(col[0].fvalue,
static_cast<bst_float>(
sbuilder[3 * nid + k].sum_total));
}
@@ -231,17 +231,17 @@ class SketchMaker: public BaseMaker {
}
}
// second pass, build the sketch
for (bst_uint j = 0; j < c.length; ++j) {
const bst_uint ridx = c[j].index;
for (const auto& c : col) {
const bst_uint ridx = c.index;
const int nid = this->position_[ridx];
if (nid >= 0) {
const GradientPair &e = gpair[ridx];
if (e.GetGrad() >= 0.0f) {
sbuilder[3 * nid + 0].Push(c[j].fvalue, e.GetGrad(), max_size);
sbuilder[3 * nid + 0].Push(c.fvalue, e.GetGrad(), max_size);
} else {
sbuilder[3 * nid + 1].Push(c[j].fvalue, -e.GetGrad(), max_size);
sbuilder[3 * nid + 1].Push(c.fvalue, -e.GetGrad(), max_size);
}
sbuilder[3 * nid + 2].Push(c[j].fvalue, e.GetHess(), max_size);
sbuilder[3 * nid + 2].Push(c.fvalue, e.GetHess(), max_size);
}
}
for (int nid : this->qexpand_) {